Patent Grant
11231966
This disclosure relates generally to the field of scheduling threads on a processor, and more specifically to a scheduler for an asymmetric multicore processor architecture.
A multi-core processor is a single computing component with two or more independent processing units called “cores.” Cores are units that read and execute program instructions. The single processor can run multiple instructions on separate cores at the same time, increasing overall speed for programs amenable to parallel computing. The multiple cores may be incorporated onto a single integrated circuit. There are at two major types of multicore processors: symmetric multicore processors (SMP) and asymmetric multicore processors (AMP). A symmetric multiprocessor (SMP) includes two or more identical processors controlled by a single operating system instance that treats all processors equally, reserving none for special purposes. Symmetric multicore processors have a plurality of cores of a same core type. Asymmetric multicore processors have a plurality of cores which may be of a different architecture from one another. Each core may or may not run an operating system
An AMP can have a first set of cores that is more efficient than a second set of cores. The second set of cores can be designed for performance. Efficiency is concerned with minimum energy consumed per instruction processed. Performance is concerned with maximum number of instructions processed per unit of time. In the prior art, to obtain efficiency, some cores would be set to idle or “dark” (shut down, not processing instructions). A substantial amount of time is required to bring a core back to a processing state, which can negatively affect performance.
In the prior art, when a system runs at high performance, and consumes a large amount of energy, a system may experience “battery droop” wherein the system may reset because there is not enough stored energy in, e.g., a storage cell or battery, to keep the system powered. In addition, in a portable device, which does have active heat mitigation, a system that is run in a performance mode for an extended period may become too hot to cool in a reasonable time, resulting in poor performance and overheating, and a poor user experience.
Objects of the systems and methods described herein include controlling system performance using measurements of performance metrics of groups of threads to make joint decisions on scheduling of threads and dynamic voltage and frequency scaling (DVFS) state for one or more clusters of cores in a multiprocessing system having a plurality of core types and one or more core of each core type. The performance metrics are fed to a closed loop control system that produces an output that is used to jointly decide how fast a core is to run and on which core type the threads of a thread group are to run. A thread group comprises one or more threads that are grouped together based on one or more characteristics that are used to determine a common goal or purpose of the threads in the thread group. Objects of the systems and methods described herein include minimizing thread scheduling latency for performance workloads, ensuring that performance workloads consistently find a performance core, maximizing throughput for performance workloads, and ensuring that efficiency workloads always find an efficient core. Objects can further include ensuring that cores are not powered down when threads are enqueued for processing. An additional object is to offload performance workloads when performance cores are oversubscribed. Threads are systematically guided to cores of the correct type for the workload.
Systems and methods are disclosed for scheduling threads on a processor that has at least two different core types each having one or more cores of that type, and the scheduler can schedule threads between at least two different core types. Each of the core types can have a different dynamic voltage and frequency scaling (DVFS) state. The processor can be a symmetric multiprocessing system (SMP) having a plurality of cores of the same design, but a first cluster of cores is run in a first configuration and a second cluster of cores is run in a second configuration. The processor can be an asymmetric multiprocessing system (AMP) that has at least two different core types, such as one or more performance cores and one or more efficiency cores. Each core type can run at a plurality of selectable voltage and frequency settings and both core types can run simultaneously.
Threads of processes can be grouped together into a thread group of related threads. Thread groups, generally, are intended to group work together to achieve a common goal, purpose, or type of work. Based upon samples of execution metrics of threads in a thread group, a closed loop performance controller (CLPC) can generate a control effort that represents an amount of performance, or efficiency, that the thread group should receive to meet performance targets for the thread group. CLPC can determine, for each of a plurality of thread groups that have been active on a core of the processor, a control effort for the thread group. The control effort for each thread group can be mapped to a performance map that recommends a core type and dynamic voltage and frequency scaling (DVFS) state for the thread group based on the control effort computed for the thread group. From the active thread groups, a maximum DVFS state can be determined for all threads that were recommended for a core of a first type (e.g., a performance core). A maximum DVFS state can be determined for all active thread groups that were recommended for a second core type (e.g. efficiency core type). A scheduler can coordinate setting the DVFS state for each core type with the scheduling of threads in the active thread groups.
In an embodiment, control effort limiter (CEL) in CLPC can generate a control effort limit that can limit the control effort output from the CLPC. A control effort limit can be based upon a fast die temperature, a peak power consumed over a short period of time, an average power consumed over any of a plurality of power zones, a skin temperature of an exterior of a device, a count of hardware throttling events, or other control metrics.
Threads can initially be grouped into default thread groups, such as grouping together threads of an application, called a “coalition.” A coalition is a grouping of closely related tasks/processes of an application which consist of threads, e.g. a web browser and application extensions associated with the web browser. Tasks can be combined into a single coalition by a launch services daemon, application programming interface (API), or framework. A thread group state is associated with each coalition. Thread grouping can be explicitly defined by a manual construct of a group of threads. Some types of workloads, e.g., an audio processing service, have complete knowledge of their organizations and can group threads accordingly. Threads can implicitly join thread groups by taking advantage of an existing kernel interaction. For example, when threads communicate via inter-process communication, they can optionally pass a voucher that references a thread group for the recipient to join. When a first thread wakes a second thread to do work on behalf of the first thread, the second thread can adopt the thread group of the first thread. When a first thread makes a second thread runnable, the second thread can adopt the thread group of the first thread. Threads can be grouped according to a type of work that the threads will perform as may be indicated by a daemon that launched the threads. An explicit thread grouping can be specified in source code, compiled code, or code packaged into a deliverable for installation. An explicit thread grouping for a work interval object can be specified in source code, compiled code, or a deliverable package for installation.
A work interval object (WIO) is an object that is used to represent periodic work where each period has a deadline. The WIO possesses a token and a specified time interval for one instance of the work. The WIO can be associated with a thread group. The thread group can either be created specifically for the WIO, or the WIO can be associated with an existing thread group. Threads that work to achieve a common purpose, intended to be performed within the specified time interval, can join the thread group of the WIO. A work interval object can be created when a user space process requests the work interval object via a specific system call. The kernel of an operating system can create the work interval object currently and automatically with a system call. By way of example, a work interval object may be, e.g., an object that represents one or more threads that composite multiple sources of audio and encode the audio for playback, all within a fixed period of time (work interval). A thread performing work having the purpose of the work interval object can opt into the work interval object thread group. Work performed by the opting-in threads is charged to the work interval object. A thread that receives a voucher containing a reference to a thread group may be grouped with the thread group referenced in the voucher. A voucher is a collection of attributes in a message sent via inter-process communication (IPC) from a first thread, T1, to a second thread, T2. One of the attributes that thread T1 puts in the voucher is the thread group to which T1 currently belongs. A second thread, T2, receives the IPC message from T1. As part of message reception, the voucher sent my T1 is “redeemed” (or adopted). Redeeming a voucher means processing all the attributes associated with the voucher and placing a pointer to the voucher on the thread. Thread T2 can now be considered to be running in the context of the voucher received from thread T1. As part of the action of “redeeming” a voucher, a voucher manager in the operating system kernel notices that thread T1 sent a thread group attribute in the voucher to thread T2, and the automatically joins thread T2 to the thread group specified in the voucher. After thread T2 has finished processing the message from thread T1, thread T1 can choose to discard the voucher. Discarding (or “dropping”) the voucher causes thread T2 to leave thread T1's thread group and causes thread T2 to re-join its “home” thread group (the one associated with the task in which T2 is running). Before thread T2 drops the voucher, however, thread T2 can also pass along the same voucher to another thread/process thread T3. Because the voucher contains a thread group attribute which points to thread T1's thread group, when thread T3 receives the message from thread T2 and adopts/redeems the voucher, thread T3 joins T1's thread group. Thread T3 will then be a part of T1's thread group for as long as that voucher is adopted by thread T3. Both threads T2 and T3 will leave thread T1's thread group as soon as threads T2 and T3 drop the voucher. This can be repeated indefinitely. Because the kernel voucher manager controls the contents of a voucher, and vouchers are automatically sent as part of a basic inter-process communication mechanism, when two processes communicate the thread which handles/receives the message will automatically join the thread group of the process who sent the message. In an embodiment, the thread may be backdated to be grouped with the thread group referenced in the voucher, prior to being made runnable, e.g. at a time when the thread receives the voucher.
A scheduler for the processor can schedule threads of the thread groups in accordance with a recommended core type and recommended DVFS state for the thread group. The recommended DVFS for the core type may be increased by the CLPC finding the maximum DVFS state for the core type, over all of the thread groups that are active on the system. The recommended DVFS state may be limited by the control effort limiter (CEL), limiting the DVFS state and/or number of cores of a core type that will execute threads.
The core types can have an associated interrupt and interrupt timer that can be used to delay a thread from being executed under some conditions. For example, if a thread is eligible for running on a performance core, but there are no idle performance cores, then the thread may be scheduled for an efficiency core. A timer can be programmed for the thread and the thread can be delayed from running. If the timer expires, the interrupt is triggered, and the performance-eligible thread is run on an efficiency core. But, if a performance core becomes available for thread execution before the interrupt time expires, then the timer can be canceled, and the thread can be executed on a performance core.
In a first embodiment, a method of running threads on a system having a processor that has a first cluster of one or more cores and a second cluster of one or more cores can include grouping threads of one or more processes into one or more thread groups. A first thread of a first thread group can be run on a core of the first cluster type. A plurality of thread execution metrics of the first thread can be incorporated with a plurality of thread execution metrics of the first thread group. A first control effort can be determined for the first thread group, based at least in part on samples of the plurality of thread group metrics. A first, or second, core type can be recommended for the first thread group type, and a DVFS state for the first thread group can be recommended. In an embodiment, the control effort can be mapped to a performance map of core clusters and DVFS states that are characterized by control effort. In an embodiment, the processor can comprise an asymmetric multiprocessing system (AMP) having a first core type of performance cores (P-cores) and a second core type of efficient cores (E-cores). A thread from a second group of threads can be executed, and performance metrics associated with execution of the second thread can be incorporated with execution metrics of the second thread group. A second control effort can be determined for the second thread group based at least in part on samples of the thread execution metrics of the second thread group. In response to determining that the first recommended core cluster for the first thread group and the second recommended core cluster for the second thread group are the same core cluster, a maximum DVFS state can be computed as the maximum of the first recommended DVFS state and the second recommended DVFS state, and the DVFS state for the core cluster recommended for the first and second thread groups can be set to the computed maximum DVFS state. In response to determining that the first recommended core cluster for the first thread group and the second recommended core cluster for the second thread group are different core clusters, the DVFS state for the first recommended core cluster can be set to the first recommended DVFS state and the DVFS state for the second recommended core cluster can be set to the second recommended DVFS state. Thread execution metrics can be determined periodically or in response to an asynchronous event. Samples of thread group execution metrics can be inputs to tunable controllers of a CLPC for the thread group. Control effort can map to a recommended core type and DVFS state. If the control effort maps to an overlapping part of the performance map, wherein either a P-core or E-core could be recommended, then a scheduler or CLPC could determine whether to recommend P-cores or E-cores based at least in part on a type of the work that the threads perform. For example, graphics rendering may be recommended to P-cores. I/O operations may be recommended to E-cores. In an embodiment, a time that a thread group has been resident on a core type and/or DVFS state can be considered in determining a core type and/or DVFS state to recommend for the thread group. In an embodiment, wherein cores of a first type are oversubscribed, and cores of a second type are idle, a scheduler may schedule threads of a thread group that are eligible for the first core type for execution on the second core type. A core of the first type can detect that a thread eligible for execution on the first type of core is actually running on the second type of core, and a core of the first type is going idle (e.g., in the process of transitioning into an idle state). The thread can be run on the now-idle core of the first type.
In a second embodiment a method of grouping threads into thread groups is performed on a system comprising an asymmetric multi-core processor having a plurality of cores of at least two different core types. The method includes launching an application program comprising a plurality of processes and assigning each of the plurality of processes to a first thread group. In response to determining that a process in the plurality of processes causes a work interval object to be generated, threads of the process can be assigned to a second thread group, different from the first thread group. At least one thread from the first thread group is executed. A plurality of thread execution metrics of threads in the first thread group is determined, and a processor core type for threads of the first thread group is recommended to a scheduler, based at least in part on the thread execution metrics of the first thread group. In an embodiment, in response to determining that a thread in the plurality of threads causes a work interval object to be generated, the thread can be assigned to a second thread group, different from the first thread group. At least one thread from the second thread group is executed and a plurality of thread execution metrics of threads in the second group is determined. A processor core type for threads of the second thread group is recommended to the scheduler, based at least in part upon samples of the thread execution metrics of the second thread group. In an embodiment, the method can further include receiving an indication that a thread from the first thread group has opted-in to the second thread group, executing the thread that opted-in to the second thread group, and attributing to the second group a plurality of thread execution metrics of the thread from the first group that opted-in to the second thread group. In an embodiment, the method can also include receiving an indication that the thread from the first thread group that opted-in to the second thread group has opted-out of the second thread group, assigning the opted-out thread to the first thread group, executing the opted-out thread, and attributing to the first thread group a plurality of thread execution metrics of the thread that opted-out of the second thread group.
A method of grouping threads for execution on an AMP can include determining that the application indicates that a process in the plurality of processes is to be assigned to a third thread group different from the first and second thread groups, executing at least one thread from the third thread group, determining a plurality of thread execution metrics of threads in the third group, and recommending to the scheduler a processor core type for threads of the third thread group based at least in part upon samples of thread execution metrics of the third thread group.
A method of grouping threads for execution on an AMP can include determining a type of work that a process in the plurality of processes performs and assigning a thread of the process to a fourth thread group, executing at least one thread in the fourth thread group, determining a plurality of thread execution metrics of threads in the fourth group, and recommending to the scheduler a processor core type for threads of the fourth thread group based at least in part upon samples of thread execution metrics of the fourth thread group. The fourth thread group can include threads from a daemon.
A method of grouping threads for execution on an AMP can include, in response to determining that a thread from the first thread group has caused a thread in a fifth thread group to be woken up and/or made runnable by the scheduler, assigning the woken up/and or made runnable thread from the fifth thread group to the first thread group. The method can further include executing at least one thread, determining a plurality of thread execution metrics of threads in the first thread group, and recommending to the scheduler a processor core type for threads of the first thread group based at least in part upon samples of the thread execution metrics of the first thread group.
A method of grouping threads for execution on an AMP can include, in response to determining that a thread from the first thread group has received a voucher to perform work on behalf of another thread, the voucher indicating a sixth thread group, assigning the thread to the sixth thread group. The method can further include executing at least one thread from the sixth thread group, determining a plurality of thread execution metrics of threads of the sixth group, and recommending to the scheduler a processor core type for threads in the sixth thread group base at least in part upon thread execution metrics of the sixth thread group. The method can also include determining that the thread which received the voucher and performed work on behalf of another thread has completed performing the work on behalf of the other thread and assigning the thread back to the first thread group.
A method of grouping threads for execution on an AMP can include launching an application program comprising a plurality of processes, assigning each of the plurality of processes to a first thread group, launching, by a daemon, at least one thread of a process in the first thread group, assigning, by the daemon, the at least one thread to a second thread group associated with a second daemon associated with the second thread group, executing at least one thread of the second thread group, determining a plurality of thread execution metrics of threads in the second thread group, and recommending to a scheduler a processor core type for threads of the second thread group based at least in part on samples of the thread execution metrics of the second thread group.
In a third embodiment, a method of processing work using a work interval object on an AMP can include creating a work interval object associated with a first thread group having a first thread group identifier. One or more threads associated with the work interval object can perform work toward a common goal of the work interval object within an identified interval of time. The method can also include receiving a request from a thread of a process associated with a second thread group having a second thread group identifier to adopt the first thread group identifier, attributing metrics of work performed by the thread to the first thread group identifier, in response to determining that the thread has adopted the first thread group identifier, and determining a processor core type to run one or more threads having the first thread group identifier based at least in part upon samples of metrics of work associated with the first thread group identifier. The work interval object can be instantiated in response to a call to a daemon or an operating system service. In an embodiment, the work interval object can be instantiated in response to a call by an application to an application programming interface (API) or framework. The method can further include receiving a request associated with the thread to opt out of the first thread group and associating the thread with the second thread group. In an embodiment, the method can also include attributing metrics of work performed by the thread to the second thread group and determining a processor type to run one more threads of the second thread group based at least in part upon samples of metrics of work associated with the second thread group identifier.
In a fourth embodiment, method of scheduling threads on a processor having a first cluster type having one or more cores and a second cluster type having one or more cores in a system that includes the processor, includes: determining a control effort limit for the processor and determining a DVFS state for at least one of the first cluster or second cluster of cores. The control effort limit can be used in conjunction with a power map that includes a plurality of DVFS states for each of the first and second cluster of core. The power map can be indexed by the control effort limit. A recommendation can be received that a thread scheduled for execution can be run on the at least one of the first or second cluster of cores at a DVFS state different than the DVFS state determined using the control effort limit. The DVFS state can be set for at least one of the first or second cluster of cores to the DVFS state determined using the control effort limit. In an embodiment, the method can further include determining that a total energy dissipated in the processor for a first predetermined period of time is greater than an instantaneous power target for the processor. The control effort limit for the processor is determined at least in part on the total energy dissipated in the processor for the first predetermined period. In an embodiment, the control effort limit can be reduced in response to determining one or more of an average power consumed by the processor for a second period of time is greater than an average power target for the processor, or an average energy dissipated per instruction is greater than the predetermined efficiency threshold for a predetermined period of time. In an embodiment, the system can monitor a plurality of power zones, each having a low pass filter that includes a power target, a time constant, a power delta, and a filter algorithm. At a first time, the filtered power can be less than a power target for a power zone, less the power delta for the power zone. At a second time, that is later than the first time, the filtered power can exceed the target power for the power zone. In an embodiment, the system can continue to monitor and filter power for the power zone that exceeded its power target, less the power delta for the power zone. The control effort limit can continue to be reduced until the filtered power for the power zone tracks the target power for the power zone. In an embodiment, the control effort limit can be reduced in response to a measured or estimated current for the processor exceeding a current limit for a predetermined period of time, or a number of instructions blocked from execution exceeds a peak throttle rate target for the predetermined period of time, wherein a throttling that blocks the instructions is done by hardware. In an embodiment, a maximum control effort is determined based at least in part on a maximum fast die temperature limit for at least one of the first or second cluster of cores. In an embodiment, a core can be masked off from executing. In an embodiment, the rate at which cores can be masked off can be limited. The processor can comprise an AMP processor having a first core cluster of P-cords and a second core cluster of E-cores.
A further embodiment provides for a processing system that includes a processor having at least two core types and at least one core of each core type. The system can also have a temperature sensor for each core type. The system can also have a scheduler that includes a scheduling queue for each core type. A CLPC can determine a control effort for threads of a thread group and recommend a core type and a DVFS state for threads of the thread group. A closed loop thermal management system can determine a control effort limit based at least in part on the temperature sensor for each core type. The system can comprise a system on a chip (SoC). One or both of the temperature sensors can be virtual sensors.
In a fifth embodiment, a method performed on a computing system having an asymmetric multiprocessing system (AMP) with a plurality of performance cores (P-cores) and a plurality of efficient cores (E-cores), can include determining that a thread, recommended for processing by a PC, is scheduled for processing by a PC of the AMP, and in response to determining that no P-cores are available for processing the thread and that an E-core is available for processing the thread, configuring a timer of an interrupt controller to generate an interrupt after expiration of a predetermined delay value. The method can further include, in response to determining, before the timer expires, that a PC is available to process the thread, canceling the interrupt, and processing the thread by the available PC. In an embodiment, the method can also include, in response to the timer expiring before a PC becomes available to process the thread, generating an interrupt to wake up an E-core and processing the thread by the E-core. The method can additionally include determining that a PC has become available, configuring a timer of the interrupt controller to generate an immediate interrupt to wake up the PC, generating the interrupt to wake up the PC, and processing the thread by the available PC.
A method of scheduling threads on an AMP processor can include determining that the thread has been recommended for processing by an E-core and that the thread is made runnable by a scheduler. In response to determining that an E-core is idle, configuring a timer of the interrupt controller to generate an immediate interrupt to wake up the E-core, otherwise configuring the timer interrupt controller to generate a delayed interrupt. The method can further include, in response to the interrupt controller generating an interrupt, waking up the E-core, and processing the thread by the E-core. The method can also include determining that the thread has been recommended for processing by a PC and that the thread has again been made runnable by the scheduler. The method can include configuring the timer of the interrupt controller to generate an immediate interrupt to wake up a PC. In response to the interrupt controller generating an interrupt, the method can include waking up the PC and processing the thread by the PC.
In an embodiment, a non-transitory computer readable can store executable instructions, that when executed by a processing system, can perform any of the functionality described above.
In yet another embodiment, a processing system coupled to a memory programmed with executable instructions can, when the instructions are executed by the processing system, perform any of the functionality described above.
A sixth embodiment provides for a data processing system comprising a processor complex including a plurality of core types, the plurality of core types to execute threads of a thread group. The data processing system additionally includes a closed loop performance controller (CLPC) to monitor and control performance of the threads of the thread group and a co-processor to execute a workload offloaded by a thread of the thread group. The CLPC can be configured to measure performance of the workload on the co-processor and associate the performance of the workload with the thread group.
An additional embodiment provides for a non-transitory machine-readable medium that stores instructions. Such instructions, when executed by one or more processors, cause the one or more processors to perform operations. The operations comprise receiving a message indicating that a thread of a thread group executing on a first processor of the one or more processors is to offload a workload to a second processor, the message including an identifier of the thread. The operations can additionally comprise inferring membership of the thread in the thread group based on the identifier of the thread and issuing a token in response to the message, the token to associate the workload with the thread group, the workload performed by one or more threads executed on the second processor. The operations additionally comprise associating performance metrics of the one or more threads executed on the second processor with the thread group.
A seventh embodiment provides for a non-transitory machine-readable medium storing instructions which, when executed by one or more processors, cause the one or more processors to perform operations comprising creating a work interval instance associated with a work interval object for a frame of content to be created, the work interval instance having an identifier and receiving notice of a submission of a workload for the frame of content on a co-processor, the notice including the identifier for the work interval instance. The operations additionally include associating the token to identify the workload with the identifier for the work interval instance and propagating the identifier for the work interval instance via the association with the token that identifies the workload and associating metrics for the workload on the co-processor with the frame of content based on the identifier propagated for the work interval instance. The seventh embodiment can additionally include instructions to perform operations comprising creating a work interval object associated with a thread group executing on the one or more processors, the thread group having a thread group identifier, where one or more threads associated with the work interval object perform work toward a common purpose within an identified interval of time.
An additional embodiment provides for a computer implemented method comprising creating a work interval instance associated with a work interval object for a frame of content to be created, the work interval instance having an identifier, receiving notice of a submission of a workload for the frame of content on a co-processor, the notice including the identifier for the work interval instance, associating the token to identify the workload with the identifier for the work interval instance, propagating the identifier for the work interval instance via the association with the token that identifies the workload, and associating metrics for the workload on the co-processor with the frame of content based on the identifier propagated for the work interval instance.
An additional embodiment provides for a data processing system comprising a processor complex including a plurality of core types, the plurality of core types to execute threads of a thread group, a co-processor to execute a workload offloaded by a thread of the thread group, and a closed loop performance controller (CLPC) to monitor and control performance of the threads of the thread group, to measure performance of the workload on the co-processor and associate the performance of the workload with the thread group. The CLPC can be configured to receive notice of a submission of a workload for a frame of content on a co-processor, the notice including the identifier for the work interval instance, associate the token to identify the workload with an identifier for a work interval instance associated with the frame of content on the co-processor, and associate metrics for the workload on the co-processor with the frame of content based on an identifier propagated for the work interval instance, the identifier propagated via an association with the token that identifies the workload. The CLPC can be additionally configured to create multiple work interval instances for multiple frames of scrolling content, generate the multiple frames of scrolling content on the co-processor, and track separate metrics for each of the multiple frames of scrolling content based on identifiers for the multiple work interval instances.
Some embodiments described herein can include one or more application programming interfaces (APIs) in an environment with calling program code interacting with other program code being called through the one or more interfaces. Various function calls, messages or other types of invocations, which further may include various kinds of parameters, can be transferred via the APIs between the calling program and the code being called. In addition, an API may provide the calling program code the ability to use data types or classes defined in the API and implemented in the called program code.
Other features and advantages will be apparent from the accompanying drawings and from the detailed description.
Embodiments of the invention are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings in which like reference numerals refer to similar elements.
In the following detailed description of embodiments, reference is made to the accompanying drawings in which like references indicate similar elements, and in which is shown by way of illustration manners in which specific embodiments may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention, and it is to be understood that other embodiments may be utilized, and that logical, mechanical, electrical, functional and other changes may be made without departing from the scope of the present disclosure. The following detailed description is, therefore, not to be taken in a limiting sense, and the scope of the present invention is defined only by the appended claims.
Reference in the specification to “one embodiment” or “an embodiment” or “some embodiments” means that a particular feature, structure, or characteristic described in conjunction with the embodiment can be included in at least one embodiment. The appearances of the phrase “embodiment” in various places in the specification do not necessarily all refer to the same embodiment. It should be noted that there could be variations to the flow diagrams or the operations described therein without departing from the embodiments described herein. For instance, operations can be performed in parallel, simultaneously, or in a different order that illustrated.
Hardware 110 can include a processor complex 111 with a plurality of core types or multiple processors of differing types. Processor complex 111 can comprise a multiprocessing system having a plurality of clusters of cores, each cluster having one or more cores of a core type, interconnected with one or more buses. Processor complex 111 can comprise a symmetric multiprocessing system (SMP) having a plurality of clusters of a same type of core wherein at least one cluster of cores is configured differently from at least one other cluster of cores. Cluster configurations can include, e.g., different configurations of DVFS states, different cache hierarchies, or differing amounts or speeds of cache. Processor complex 111 can additionally comprise an asymmetric multiprocessing system (AMP) having a plurality of clusters of cores wherein at least one cluster of cores has a different core type than at least one other cluster of cores. Each cluster can have one or more cores. Core types can include performance cores, efficiency cores, graphics cores, digital signal processing cores, and arithmetic processing cores. A performance core can have an architecture that is designed for very high throughput and may include specialized processing such as pipelined architecture, floating point arithmetic functionality, graphics processing, or digital signal processing. A performance core may consume more energy per instruction than an efficiency core. An efficient core may consume less energy per instruction than a performance core. In an embodiment, processor complex 111 can comprise a system on a chip (SoC) that may include one or more of the hardware elements in hardware 110.
Hardware 110 can further include an interrupt controller 112 having interrupt timers for each core type of processor complex 111. Interrupt controller 112 can be used, with interrupt timers, to implement deferred inter-processor interrupts (DIPI) as described more fully below with reference to
Hardware 110 can also include one or more thermal sensors 113. In an embodiment, wherein processor complex 111 comprises an SoC, one more thermal sensors 113 can be included in the SoC 111. In an embodiment, at least one thermal sensor 113 can be included on the SoC 111 for each core type of the processor complex 111. In an embodiment, a thermal sensor 113 can comprise a virtual thermal sensor 113. A virtual thermal sensor 113 can comprise a plurality of physical thermal sensors 113 and logic that estimates one or more temperature values at location(s) other than the location of the physical thermal sensors 113.
Hardware 110 can additionally include memory 114, storage 115, audio processing 116, one or more power sources 117, and one or more energy and/or power consumption sensors 118. Memory 114 can be any type of memory including dynamic random-access memory (DRAM), static RAM, read-only memory (ROM), flash memory, or other memory device. Storage can include hard drive(s), solid state disk(s), flash memory, USB drive(s), network attached storage, cloud storage, or other storage medium. Audio 116 can include an audio processor that may include a digital signal processor, memory, one or more analog to digital converters (ADCs), digital to analog converters (DACs), digital sampling hardware and software, one or more coder-decoder (codec) modules, and other components. Hardware can also include video processing hardware and software (not shown), such as one or more video encoders, camera, display, and the like. Power source 117 can include one or more storage cells or batteries, an AC/DC, power converter, or other power supply. Power source 117 may include one or more energy or power sensors 118. Power sensors 118 may also be included in specific locations, such as power consumed by the processor complex 111, power consumed by a particular subsystem, such as a display, storage device, network interfaces, and/or radio and cellular transceivers. Computing system 100 can include the above components, and/or components as described with reference to
Operating system 120 can include a kernel 121 and other operating system services 127. Kernel 121 can include a processor complex scheduler 705 for the processor complex 111. Processor complex scheduler 210 can include interfaces to processor complex 111 and interrupt controller 112 as described below with reference to
In an embodiment, IPC 125 can receive a message from a thread that references a voucher. A voucher is a collection of attributes in a message sent via inter-process communication (IPC) from a first thread, T1, to a second thread, T2. One of the attributes that thread T1 can put in the voucher is the thread group to which T1 currently belongs. IPC 125 can pass the voucher from a first thread to a second thread. The voucher can include a reference to a thread group that the second thread is to adopt before performing work on behalf of the first thread. Voucher management 126 can manage vouchers within operating system 120, user space 130, and system space 140. Operating system (OS) services 127 can include input/output (I/O) service for such devices as memory 114, storage 115, network interface(s) (not shown), and a display (not shown) or other I/O device. OS services 127 can further audio and video processing interfaces, data/time service, and other OS services.
User space 130 can include one or more application programs 131-133, closed loop thermal management (CLTM) 134, and one or more work interval object(s) 135. CLTM 134 is described more fully, below, with reference to
System space 140 can include a launch daemon 141 and other daemons, e.g. media service daemon 142 and animation daemon 143. In an embodiment, threads that are launched by a daemon that perform a particular type of work, e.g. daemons 142 and 143, can adopt the thread group of the daemon. Execution metrics of a thread that adopted the thread group of the daemon that launched the thread are attributable to the thread group of the daemon for purposes of CLPC 300 operation.
System 200 can include a kernel 121 that is part of an operating system, such as operating system 120 of
Processor complex 111 can comprise a plurality of processor core types of an asymmetric multiprocessing system (AMP) or a symmetric multiprocessing system (SMP). In an AMP, a plurality of core types can include performance cores (P-cores) and efficiency cores (E-cores). In an SMP, a plurality of cores types can include a plurality of cores configured in a plurality of different configurations. Processor complex 111 can further include a programmable interrupt controller (PIC) 112. In an embodiment, PIC 112 can have one or more programmable timers that can generate an interrupt to a core at a programmable delay time. In an embodiment, PIC 112 can have a programmable timer for the processor complex 111. In an embodiment, PIC 112 can have a programmable timer for each core type in the processor complex 111. For example, PIC 112 can have a programmable timer for all P-cores 222 and another programmable timer for all E-cores 222. In an embodiment, PIC 112 can have a programmable timer for each core of each core type.
Processor complex scheduler 210 can include a thread queue manager 211, thread group performance data manager 212, thread group recommendation manager 213, and a plurality of thread queues for each of a plurality of processor core types. In an example processor complex 111, processor complex scheduler 210 can have an E-core thread queue 215 and a P-core thread queue 220.
Processor complex scheduler 705 thread queue manager 211 can manage the scheduling of threads for each of the plurality of cores types of processor complex 111, e.g., E-core thread queue 215 and P-core thread queue 220. Thread queue manager 211 can further include logic to program interrupt controller 223 for immediate and/or deferred interrupts.
Thread group performance data manager 212 of the processor complex scheduler 210 can collect thread execution metrics for each of a plurality of thread groups executing on processor complex 111. A plurality of thread execution metrics 231 can be sampled from the collected thread execution metrics of thread group performance data manager 212 and provided to a plurality of tunable controllers 232 of CLPC 300 for each thread group. Tunable controllers 232 can be proportional-integral-derivate (PID) controllers. A PID controller has an output expressed as:
where KP is the proportional gain tuning parameter, Ki is the integral gain tuning parameter, Kd is the derivative gain tuning parameter, e(t) is the error between a set point and a process variable, t is the time or instantaneous time (the present), and τ is the variable of integration which takes on values from time 0 to the present time t.
Processor complex scheduler 210 thread group recommendation manager 213 can receive core type (cluster) recommendations from CLPC cluster recommendations 237 for each thread group that has been active on processor complex 111. Processor complex scheduler 210 thread queue manager 211 can utilize the cluster recommendations 237 for each thread group to program threads of each thread group onto an appropriate core type queue, e.g. 215 or 220.
CLPC 300 is a closed loop performance controller that determines, for each thread group active on a core, a control effort needed to ensure that threads of the thread group meet their performance goals. A performance goal can include ensuring a minimum scheduling latency, ensuring a block I/O completion rate, ensuring an instruction completion rate, maximizing processor complex utilization (minimizing core idles and restarts), and ensuring that threads associated with work interval objects complete their work in a predetermined period of time associated with the work interval object. Metrics can be periodically computed by CLPC 300 from inputs sampled by CLPC 300 either periodically or through asynchronous events from other parts of the system. In an embodiment, inputs can be sampled at an asynchronous event, such as the completion of a work interval object time period, or a storage event. A plurality of performance metrics 231 can be computed within CLPC 300 and each fed to a tunable controller 232. Tunable controllers 232 generate an output to a tunable thread group PID 233, which in turn outputs a control effort 234 needed for the thread group to meet its performance goals.
In an embodiment, a control effort 234 is a unitless value in the range 0 . . . 1 that can be mapped to a performance map and used to determine a recommended core cluster type 237 for the thread group. The cluster recommendations 237 are returned to thread group manager 213 in processor complex scheduler 705 for scheduling threads to core types. For each of thread groups 1 . . . n, a control effort CE 1 . . . n 234 collected by a cluster maximum control effort module 238. Cluster maximum control effort module 238 determines a maximum control effort value for all control efforts CE 1 . . . n 234 for each cluster type. Maximum control effort module 238 outputs maximum control effort for each cluster type to a respective cluster type mapping function, e.g. E-ce map 235 and P-ce map 236. E-ce map 235 determines a dynamic voltage and frequency scaling (DVFS) state for E-cores based upon the maximum E-cluster control effort output from maximum control effort module 238. Similarly, P-ce map 236 determines a DVFS state for P-cores based upon the maximum P-cluster control effort output from maximum control effort module 238. These respective maximum DVFS states may be limited by an output of control effort limiter (CEL) 400 of CLPC 300. Control effort limiter 400 is described further, below, with reference to
Control effort limiter 400 can receive input from a plurality of temperature control loops 261, a peak power manager 262, and a closed loop thermal manager (CLTM) 134. Control effort limiter 400 is described below with reference to
Work Interval Utilization 311 and 312
Many workloads are targeted towards a user-visible deadline, such as video/audio frame rate, for example. The processor complex 111 performance provided for such workloads needs to be sufficient to meet the target deadlines, without providing excess performance beyond meeting the respective deadlines, which is energy inefficient. Towards this end, for each video/audio frame (work interval), CLPC 300 receives timestamps from audio/rendering frameworks about when the processor complex 111 started working on the frame (start), when the processor complex 111 stopped working on the frame (finish) and what is the presentation deadline for the frame (deadline). CLPC 300 computes work interval utilization 311 or 312 for the frame as (finish-start)/(deadline-start). The work interval utilization is a measure of the proximity to the deadline. A value of 1.0 would indicate ‘just’ hitting the deadline. However, since the processor complex 111 is not the only agent in most workloads and dynamic voltage and frequency scaling (DVFS) operating points are discrete, and not continuous, a goal is to provide enough performance to meet the deadline with some headroom, but not so much headroom as to be energy inefficient.
Work interval-based control is reactive in nature. Hence, it is susceptible to offering a poor transient response when there is a sudden increase in the offered processor complex 111 load (for example, a frame that is inordinately more complex than the last ‘n’ frames). To achieve a degree of proactive response from the CLPC 300, video/audio APIs allow higher level frameworks to interact with CLPC 300 as soon as a new frame starts being processed and convey semantic information about the new frame, such as an estimate of the complexity of the new frame. Immediately after receiving complexity information, the CLPC 300 can adjust the workload's performance to meet the deadline, based on the observed running time of past work intervals with similar complexity.
Work interval utilization metric 311 is fed to tunable controller 321 having a target TPT. A difference between TPT and the work interval utilization metric 311 is determined and multiplied by a tuning constant, Ki, for the tunable controller 321. Similarly, work interval utilization metric 311 is fed to tunable controller 322 having a target TRT. A difference between TRT and the work interval utilization metric 312 is determined and multiplied by a tuning constant Ki, for the tunable controller 322.
I/O Transaction Rate 313
An input/output (I/O) bound workload, such as block storage I/O 115, interacts heavily with non-processor complex subsystems such as storage or a network. Such workloads typically exhibit low processor complex utilization and might appear uninteresting from a processor complex performance standpoint. However, the critical path of the workload includes some time spent on the processor complex 111 for managing meta-data or data going to or from the non-processor complex subsystem. This is typically time spent within kernel drivers such as a Block Storage Driver (for storage) and Networking Drivers (e.g. for Wi-Fi/mobile-data transfers). Hence processor complex 111 performance can become a bottleneck for the I/O. The I/O rate metric computes the number of I/O transactions measured over a sampling period and extrapolates it over a time period, e.g. one second. I/O rate metric 313 is fed to tunable controller 323 having a target TI/O. A difference between TI/O and the I/O rate metric 313 is determined and multiplied by a tuning constant, Ki, for the tunable controller 323.
Processor Complex Utilization 314
Processor complex scheduler 210 can accumulate statistics that measure processor complex utilization 304, scheduling latency 305, and cluster residency 306. Processor complex utilization 304 can measure an amount, such as a percentage, of utilization of the processor complex cores that are utilized over a window of time. The measured or computed value for processor complex utilization 304 can be sampled and be fed as a metric to processor complex utilization 314. A purpose of the processor complex utilization metric 314 is to characterize the ability of a workload to exhaust the serial cycle capacity of the system at a given performance level, where the serial cycle capacity examines the utilization of the processor complex as a whole. For each thread group, CLPC 300 can periodically compute the processor complex utilization metric 314 as (time spent on core by at least a single thread of the group)/(sampling period). The processor complex utilization metric 314 can be defined as a “running utilization”, i.e., it only captures the time spent on-core by threads. Processor complex utilization metric 314 can be sampled or computed from metrics provided by the processor complex scheduler 210. The processor complex scheduler 210 can determine a portion of time during a sample period that thread(s) from a thread group were using a core of the processor complex 111. Processor complex utilization metric 314 is fed to tunable controller 324 having a target TUTILIZATION. A difference between TUTILIZATION and the processor complex utilization metric 314 is determined and multiplied by a tuning constant, Ki, for the tunable controller 324.
In an embodiment, the “runnable utilization” of a thread group can be measured, which is computed through the time spent in a runnable state (running or waiting to run) by any thread of the group. This has the advantage of capturing thread contention for limited processor complex cores; a thread group that spends time waiting for processor complex 111 access will exhibit higher runnable utilization. Considering thread contention takes into account the period in which a thread is able to be run, relative to the amount of time in which the thread is running. When a large number of threads are contending for access to processor cores, threads will spend a larger amount of time in a runnable state before going on-core.
Performing closed loop control around the processor complex utilization metric 314 for a thread group will give higher execution throughput to this thread group once it eventually goes on-core, the idea being to try and pull in the completion time of the threads of the thread group to better approximate what they would have been in an un-contended system.
Thread Scheduling Latency 315
Scheduling latency 305 can measure an amount of latency that threads in a thread group experience between a time that a thread of a thread group is scheduled and the time that the thread is run on a core of the processor complex 111. Scheduling latency 305 can be sampled for a window of time for a thread group and provided to CLPC 300 as a scheduling latency metric 315. In one embodiment, thread scheduling latency metric 315 serves as a proxy for the runnable utilization of a thread group if runnable utilization cannot be directly determined from the processor complex 111. Scheduling latency metric 315 can be provided by the processor complex scheduler, e.g. processor complex scheduler 210 of
Each of the above metrics 311-315 can be fed to a tunable controller, e.g. 321-325 that outputs a contribution to a control effort for threads of the thread group. Each tunable controller, e.g. 321-325, can have a target value, e.g., TPT for working interval utilization 311, and a tuning constant Ki. An integrator 340 sums the contributions and generates a unitless control effort for the thread group in the range of 0 . . . 1 that is used as an index into a performance map 345.
Cluster Residency 316
Cluster residency 306 can measure an amount of time that threads of a thread group are resident on a cluster of cores, such as E-cores or P-cores. Cluster residency 306 can be sampled for a window of time for a thread group and provided as a metric to cluster residency metric 316. In an embodiment, cluster residency metric 316 can have sample metric for each of one or more cluster of core types, such as E-cores and P-cores. In an embodiment, cluster residency metric 316 comprises E-cluster residency metric 317 and P-cluster residency metric 318, and RS Occupancy Rate metric 309. E-cluster residency metric 317 is a measure of an amount of time that a thread group executes on a cluster of efficiency cores. P-cluster residency metric 318 is a measure of an amount of time that a thread group executes on a cluster of performance cores. RS Occupancy Rate metric 309 is a measure of reservation station occupancy, which is a measure of how long a workload waits in a ready state before being dispatched to a processor pipeline. Control effort for cluster residency for a thread group can be determined from cluster residency metric 316, including E-cluster residency metric 317 and P-cluster residency metric 318, and RS Occupancy rate 309, by feeding the metrics 316 through 318 to controllers 330 through 332, respectively.
The CLPC 300 output is a control effort, an abstract value on the unit interval that expresses the relative machine performance requirement for a workload. The control effort is used as an index into a performance map 345 to determine a recommended cluster type and dynamic voltage and frequency scaling (DVFS) state for the thread group. The recommended DVFS state for E-cores for each of a plurality of thread groups that have been active on a core, is input into a Max( ) function 367 to determine a recommended maximum DVFS state for E-cores. The recommended DVFS state for P-cores for each of a plurality of thread groups that have been on a core is input into a Max( ) function 366 to determine a recommended maximum DVFS for P-cores. The maximum DVFS state recommended for E-cores (output from Max( ) function 367) and the maximum DVFS state recommended for P-cores (output from Max( ) function 366) is sent to control effort limiter (CEL) 400 to determine whether the recommended DVFS states for P-cores and E-cores should be limited. Recommended DVFS states may be limited to reduce heat and/or to conserve power. CEL 400 outputs, to power manger 240, a DVFS state for each cluster of cores, e.g. E-cores DVFS states 371 and P-core DVFS states 372. In an embodiment, DVFS states 371 and 372 can include a bit map that can mask off one or more cores of a cluster, based on control effort limiting by CEL 400. CEL 400 is described below with reference to
Control effort limiter 400 can receive input from a fast-die thermal management control loop 401. Each cluster of cores of a type, e.g. E-cores and P-cores, can have a fast die temperature sensor that is fed into a closed loop thermal controller that is closed around a cluster of cores to regulate the temperature to a target. The fast-die thermal management control loop 401 outputs a maximum control effort for each core type that is input to CEL 400. In an embodiment, fast-die thermal management control loop 401 outputs a P-core maximum control effort 402 and an E-core maximum control effort 403.
Peak power manager (PPM) 262 can set a processor complex instantaneous peak power target 405 that is input to peak power controller 410 of CEL 400. PPM 262 can measure current consumed at the processor complex 111. PPM 262 can receive a signal from power manager (PMGR) 240 indicating that the power source, e.g. battery, indicating whether there is insufficient energy supply (“droop”) 373 to meet the demand for current at the processor complex 111. PPM 262 produces a processor complex instantaneous power target 405 to limit the amount of current consumed by the processor complex 111 in response to the signal of insufficient energy supply (droop 373) from PMGR 240. A state of a power source 117, e.g. a battery, can limit or constrain the output of the signal from PMGR 240, such as when the battery is too cold or has experienced a number of charging cycles beyond the rating of the battery, indicating that there may be insufficient energy supply to meet the demand for current at processor complex 111. Embodiments of a Peak Power Manager 262 are described in U.S. Provisional Patent Application 62/557,067, entitled, “COMPONENT POWER CONSUMPTION MANAGEMENT DETERMINING WHETHER THE POWER AVAILABILITY OF THE POWER SOURCE EXCEEDS THE EXPECTED POWER CONSUMPTION,” to Michael Eng, et al., filed Sep. 11, 2017, which is hereby incorporated by reference in its entirety to the extent that it is consistent with this disclosure.
Closed loop thermal manager (CLTM) 134 can feed measurements from one or more thermal sensors into a plurality of tunable closed loop controllers to determine a processor complex average power target 406 that is input to a controller 415 of CEL 400 to determine a control effort limit on the control effort output from CLPC 300. Controller 415 is described below with reference to
Processor complex power 404 can be measured using a power sensor, e.g. 118 of
Control Effort Limiter 400
Peak power controller 410 receives a processor complex instantaneous power target 405 from peak power manager (PPM) 262 and a processor complex power 404 as inputs. To avoid battery voltage droop 373 that can induce unexpected system shutdowns, peak power manager 262 limits processor complex power under certain physical conditions (e.g., low temperature, low battery state of charge, high cell impedance). It does this by setting a processor complex instantaneous power target 405, which is observed in one of several ways. In one embodiment, the power target 405 is used as the control target in a loop (e.g., proportional-integral control) closed around a measurement or estimate of total energy dissipated in the processor complex 111 (e.g., processor complex power 404). In another embodiment, the power target 405 is used to search a pre-characterized table of worst-case workload powers in each processor complex core type DVFS state. The latter method is conservative in delivered performance for workloads that are not worst-case. The output of either embodiment is a control effort limit that indexes into a power map 450 that is global for the entire processor complex 111 (described below). The processor complex instantaneous power target input 405 is asynchronous with respect to the periodic updates of the power limiting control system, and changes at a time scale of hundreds of milliseconds or more.
Average power controller 415 is described below with reference to
Efficiency controller 417. To allow the system to satisfy battery life goals, the amount of energy dissipated per retired instruction is limited to an efficiency target, TEFFICIENCY. TEFFICIENCY is the control target for a loop (e.g., proportional-integral control) closed around a measure of efficiency expressed in nanojoules (nJ) per instruction. The efficiency metric is calculated by first determining the number of instructions retired during a sample interval (as reported by processor complex performance counters) and dividing this number by the amount of energy dissipated in the processor complex 111 (using a measurement or estimate of processor complex power) over that interval. The output of the closed-loop efficiency controller 417 is a limit control effort that indexes into a power map 450 that is global for the entire processor complex (described below). The efficiency metric is sampled synchronously with respect to the periodic updates of the power limiting control system.
Throttle rate controller 420: To prevent voltage droops in the processor complex power delivery network (i.e., between the processor complex voltage regulator and the processor complex 111), an automatic hardware throttle mechanism blocks instruction or micro-operations issued (i.e., from reservation stations into execution pipelines) when measured or estimated processor complex current exceeds a limit. This peak power throttle reduces the rate at which the processor complex executes work, but it does not reduce processor complex frequency or voltage. As a result, it can be inefficient to complete a significant amount of work with the throttle engaged; efficiency is improved by lowering the processor complex core type DVFS state. The maximum rate of throttle assertion is limited to a throttle rate target, TTHROTTLE. This is the control target for a loop (e.g., proportional-integral control) closed around a measure of throttle assertions expressed in blocked issue slots per processor complex cycle. The throttle rate metric is calculated by first determining the number of issue slots that were blocked during a sample interval (as reported by processor complex performance counters) and dividing this number by the number of cycles clocked by the processor complex (again, as reported by processor complex performance counters) over that interval. The output of the closed-loop controller is a limit control effort that indexes into a power map that is global for the entire processor complex (described below). The throttle rate metric is sampled synchronously with respect to the periodic updates of the power limiting control system.
Average power controller 415 is described with reference to
Some power limiting structures update asynchronously at a time scale of ones of milliseconds. These structures have response latency requirements that are faster than the periodic updates of the power limiting control system.
Returning to
The P-core maximum control effort, received from fast-die temperature controller 401, can be used to index into power map 345 to determine a DVFS state for P-cores. The E-core maximum control effort, received from fast-die temperature controller 401, can be used to index into power map 345 to determine a DVFS state for E-cores. A minimum control effort, min( ), can be determined of the control efforts from each of peak power controller 410, average power controller 426, efficiency controller 417, and throttle rate controller 420. The minimum control effort can be mapped to power map 450 to determine a DVFS state for each core type. CLPC 300 outputs a maximum of recommended DVFS state for each core type, which is then limited 460 by the minimum of: the DVFS state for each core type recommended by the CLPC and the maximum DVFS state permitted by the control effort limit 400.
The limited 460 maximum control effort is also compared against thresholds at which cores within the processor complex subset are de-recommended to the processor complex scheduler 210. When the maximum control effort falls below a threshold Θ (e.g., Θ−1P, Θ−2P, Θ−1E, Θ−2E) one or more bits are cleared in the recommended cores bitfield 374 that is presented to the processor complex scheduler. The processor complex scheduler is described further below.
Power Map 500
The output of the closed-loop controllers of the CLPC 300 as limited by control effort limiter 400 is a scalar limit control effort that jointly restricts DVFS states and scheduling preferences. The energy dissipated in each tuple of active processor complex cores and processor complex core type DVFS states is characterized in a power map 500 that is indexed by control effort 520 determined by the control effort limiter 400. Each entry 501 . . . 510 in the power map 500 summarizes the calculated power using estimates of switched capacitance, part-specific binned voltages and processor complex core type frequencies of each (active cores, DVFS states) tuple for a maximum power workload. Since the control effort 520 is associated with a relative power limit 580, the power map 500 can be searched for a tuple of active cores 540 and DVFS states 560 satisfying the relative power limit 580 indicated by the control effort 520. In one embodiment, output of the power map 500 includes DVFS state limits and a recommended cores bitfield 374. In order to satisfy the relative power limit 580 expressed by the control effort 520, it may be necessary to limit the maximum DVFS state used by each subset of processor complex cores that share an instantaneous DVFS state. It may further be necessary to restrict the set of processor complex cores that execute the workload. When the active-cores subset indicated in the power map 500 by the control effort 520 excludes one or more processor complex cores, the processor complex cores are removed from the recommended cores bitfield 374 that is presented to the processor complex scheduler as described below. Although
Configuring DVFS and Recommended Cores
The DVFS state 560 configured for each subset of processor complex cores that share an instantaneous DVFS state is limited. The DVFS state for a processor complex cores subset is the minimum of the state requested by active thread groups for that subset, the limit state calculated by the periodic power limiting control system for that subset and the limit state calculated by the asynchronous power limiting structures for that subset. In one embodiment, thread groups cannot override DVFS state limits. The set of schedulable processor complex cores is limited. The set of processor complex cores that are recommended for use by the processor complex scheduler, expressed in a recommended cores bitfield, is the intersection of the cores recommended by the power map 500 and the cores recommended in each processor complex subset by the maximum control efforts given by fast die temperature control. Once a bit for a given processor complex core is cleared in the bitfield, the processor complex scheduler 210 allows an active thread executing on that core to run until the next processor complex scheduler event (e.g., quantum expiration or blocking), but no new threads are dispatched to the core. Some processor complex core recommendation changes require more processor complex scheduler work than others (e.g., relocating runnable threads from one run queue to another). A hysteresis parameter limits the frequency of core recommendation changes. In one embodiment, thread groups cannot override the recommended cores bitfield 374.
When an application is launched, threads of an application are initially grouped into thread groups. In an embodiment, threads can be initially group in accordance with grouping specified within program code by whether one the threads perform functionality, such as animation or audio/video functionality, in which the thread structure is known, or whether the threads are launched by a daemon that assigns a thread to a thread group when the daemon launches the thread. Threads can also be initially grouped with other threads launched by a same application. During execution of a thread, the thread can opt-in to join a different thread group. At a later time, the thread can opt-out of the different thread group and return to its initial thread group.
One example of a thread opting-in to a different thread group occurs when a thread performs work having a common purpose with a work interval object (WIO). Threads having a common purpose describe threads that perform similar operations on similar data or resources, where those operations are designed to be performed within the same time interval. A WIO can have its own thread group. The thread group can either be created specifically for the WIO or the WIO can be associated with an existing thread group. A thread performing work toward the common purpose of the WIO, intended to be performed within the specified time interval, can opt-in to the WIO thread group. Thread execution metrics attributable to work performed by the thread on behalf of the WIO are accumulated with the thread group of the WIO. A work interval object is used to represent periodic work where each period has a deadline. The WIO possesses a token and a specified interval of time for one instance of the work. For example, an audio compositing application can include an audio compositing WIO 136 for compositing audio from one or more other applications into a single audio frame that must be completed within a predetermined period of time. Work interval objects can be created in several ways. There can be a set of predefined work interval objects in an operating system, daemon, framework, or application. A kernel of an operating system can create a work interval object explicitly, such as on behalf of a driver. A kernel of an operating system can implicitly create a work interval object on behalf of an application, such as in response to an application call to a framework. Two examples of creating a work interval object are described below.
WIO example 101 includes a work interval object 135 that is created by a media server daemon 142. App. 1131 can be, e.g., a media streaming application. Media service daemon 142 can create work interval object 135, and threads associated with the WIO 135 perform streaming support for applications, such as App. 1131. WIO 135 can be, e.g., a WIO to support streaming of audio, video, or both. To ensure continuity of media playback, threads associated with a WIO 135 must process media in accordance with a predetermined time for the WIO 135. App. 1131, e.g. can generate a thread, e.g. T1, that performs work that is needed by WIO 135. T1 can have an initial thread group, e.g. TG1, that is associated with App. 1131. WIO 135 can have its own thread group, e.g. TG10, associated with WIO 135. WIO 135 can call thread T1 to perform work on behalf of WIO 135. A thread, e.g. T1, can request, via media service daemon 142 and inter-process communication (IPC) 125, that thread T1 join WIO 135's thread group (TG10). Thread T1 can opt-in to joining WIO 135 thread group TG10. When thread T1 to performs work on behalf of WIO 135, the thread execution metrics for thread T1 are attributed to WIO 135 thread group TG10.
In WIO example 102, an application, e.g. App. 2132 can call an application programming interface (API) or framework that creates a work interval object 136. Threads associated with the work interval object 136 perform the functionality requested in the API or framework call by App. 2132. App. 2132 can be, e.g., a recording studio application. The record studio application may record multiple tracks of audio and/or video and may further need one or more audio effects applied to a recording. App. 3133 can be, e.g., an audio effects program. App. 2132 recording studio would like to utilize App. 3 audio effects on a recording. Threads associated with work interval object 136 can perform audio compositing on recorded audio by App. 2132 using audio effects of App. 3133. App. 2132 may generate a thread, e.g. T2, having thread group TG2 associated with App. 2132, that records audio. Appl. 3133 may generate a thread, e.g. T3, having thread group TG3 associated with App. 133, that applies audio effects to a recording. WIO 136 can have its own thread group, e.g. TG11. Threads T2 and T3 can generate audio and effects data for compositing audio associated with WIO 136. One, or both, of T2 or T3 can opt-in to WIO 136 thread group T11. If both threads T2 and T3 opt-in to WIO 136 thread group T11, then thread execution metrics associated with T2, T3, and WIO 136 will all be attributed to WIO 136 thread group T11. Either, or both, of T2 or T3 can later opt-out of WIO 136 thread group TG11 and return to their respective initial thread groups. For example, thread T2 may opt-out of WIO 136 thread group TG11 in response to App. 132 saving or closing an audio file. Thread T3, for example, may opt-out of TG11 when App. 3133 is closed, perhaps due to effects processing no longer being needed by a user.
A DIPI system can be used to increase performance in the processing system by delaying, for a short time, execution of a thread that is waiting on a particular core type to become available. In an example, in an asymmetric multiprocessing (AMP) system comprising performance cores (P-core) and efficiency cores (E-cores), a thread group may be recommended for running a P-core. Threads of the thread groups recommended for running on a P-core are said to be “P-eligible.” If a P-eligible thread is runnable but waiting on a P-core to become available, and an E-core is available, one solution to ensuring that the thread makes progress is to “spillover” the P-eligible thread to run on an E-core. However, this solution requires a certain amount of overhead to run the P-eligible thread on the E-core. An alternative solution is to have the P-eligible thread wait for a small amount of time for a P-core to become available and run the P-eligible thread on the P-core. While it is counter-intuitive that waiting can increase performance, this solution using DIPI recognizes that P-cores generally retire instructions faster than E-cores, and a P-core may become available sooner than the time it takes wake up an idle E-core and switch the P-eligible thread to run on an E-core.
States of a core can include a state wherein the core is executing instructions of a thread and a state wherein the core is idle, i.e., not executing instructions, but the core is not in a sleep state or in a low power state. A core state can further include a state in which the core is in a sleep state or low power state and requires a re-initialization of the core to at least the idle state so that the core can execute instructions.
A DIPI system 700 can include a processor, such as processor complex 111, and a processor complex scheduler 210 that can schedule threads for a plurality of core types. Processor complex 111 can comprise one or more E-cores 221 and one or more P-cores 222. Processor complex 111 can further include an E-cache 735 that is dedicated to caching threads for running on E-cores 221, and a P-cache 745 that is dedicated to running threads on P-cores 222. In an embodiment, processor complex 111 can further comprise a shared cached 750. Processor complex 111 also can include a programmable interrupt controller 755. In an embodiment, programmable interrupt controller 755 can have interrupts for each core type and programmable interrupt timers for each core type, e.g., E-core interrupt timers 760 and P-core interrupt timers 765.
Processor complex scheduler 210 can include a thread queue for each core type of processor complex cluster of cores 221 and 222, e.g. efficiency thread queue 215 (E-queue) and performance thread queue 220 (P-queue). Each queue type can have a plurality of queue entries, e.g. E-queue 215 can have E-threads 215-1, 215-2, . . . , 215-n pending in the E-queue 215, and P-queue 220 can have P-threads 220-1, 220-2, . . . , 220-n pending in the P-queue 220. Processor complex scheduler 210, E-queue 215, and P-queue 220 can communicate with processor complex 111 the components of processor complex 111. Processor complex scheduler 210 can further contain scheduling logic 725 to manage queues 215, 220, and 726, and implement DIPI using interrupt controller 755 and timers 760 and 765. In an embodiment, processor complex scheduler 210 can also include a global run queue 726 that enqueues threads that are runnable. In an embodiment, one or more run queue(s) 726 can comprise an E-core run queue 215 and a performance core run queue 220.
Processor complex scheduler 210 can use inter-processor interrupts (IPI) to schedule threads on multiple core types, e.g., E-cores 221 and P-cores 222. Although four (4) E-cores and two (2) P-cores are shown, this is by way of example only. Any number of E-cores and P-cores can form a cluster of cores of a core type. When a target core is powered-off, processor complex scheduler 210 can program a deferred IPI (DIPI) that triggers after a programmable set time interval, using E-interrupt timers 760 and P-interrupt timers 765. During that time interval, if a target core type becomes available, processor complex scheduler 210 can cancel the deferred IPI, and schedule the thread on the target core. Schedule logic 725 can integrate AMP core type information, dynamic thread power/performance recommendations, and scheduling intent to selectively and dynamically use deferred IPIs for AMP core scheduling. This results in increased performance and power efficiency on AMP systems, particularly in performance-intensive multi-threaded workloads.
When the processor complex scheduler 210 sends an IPI to a processor complex core, e.g. 221 or 222, the processor complex scheduler logic 725 determines the type of IPI to use: deferred IPI or immediate IPI. A deferred IPI is an interrupt programmed with a delay value, to defer the interrupt. An immediate IPI is an interrupt programmed without a delay, to generate an immediate interrupt. For deferred IPIs, processor complex scheduler logic 725 can adjust the deferred time interval using workload heuristics. Processor complex scheduler logic 725 can identify with high accuracy the instances where using interrupt deferral, with the opportunity to cancel the interrupt during the deferred interval, is expected to increase performance and/or power efficiency relative to immediate interrupts. Processor complex scheduler logic 725 uses heuristics and inputs from the processor complex scheduler 111, hardware, and power/performance control systems closed loop performance controller (CLPC) 300 and control effort limiter (CEL) 400.
Processor complex scheduler logic 725 can determine a scheduling event that can determine whether to use deferred or immediate IPI for each scheduling event that results in an IPI. For each scheduling event, processor complex scheduler logic 725 evaluates: (1) the scheduling intent/event type, the current core type, the target core type, and the recommended core type classification obtained from CLPC 300. Scheduling events can include: spilling a thread from a P-core to an E-core, rebalancing a P-eligible thread from an E-core to a P-core, and making a thread runnable. The current and target core types for an AMP can be an efficiency core (E-core) and a performance core (P-core). The current core type is the core type on which a thread has run, and a target core type is the core type on which the thread is enqueued to run. A CLPC 300 thread recommendation can include “P-eligible,” wherein a thread is eligible to run on a P-core or an E-core, and “E-eligible,” wherein a thread is eligible to run on an E-core. An example table of processor complex scheduler logic 725 decisions is described in
Additionally, processor complex scheduler logic 725 can determine the programmable deferred IPI interval length control loop by evaluating: P-core and E-core thread migration rates, and E-queue 215 and P-queue 220 lengths. Processor complex scheduler logic. 725 can use migration rate and run-queue length thresholds to increase or decrease the deferred time interval of deferred IPIs. This allows the processor complex scheduler logic 725 to control how aggressive the core wake-up prevention and thread migration hysteresis are, in response to workload characteristics.
Advantages of using DIPI in a processor having multiple core types, such as AMP, include increased performance and efficiency for high-performance multi-threaded workloads on AMP systems, due to improved scheduling latency and thread migration hysteresis, and energy-efficiency for low-demand workloads, using deferred IPIs that avoid unnecessary core shutdown and subsequent power-up.
Periodically, or in response to an asynchronous event, a plurality of thread execution metrics for a thread group are fed into a plurality of tunable controllers of a closed loop performance controller (CLPC) 300 for the thread group. An output of the CLPC 300 is a control effort needed for the threads of the thread group to meet performance goals for the thread group. The control effort can be a unitless value from 0 . . . 1. The control effort can be used to index into a performance map 800 that indicates a plurality of core types and dynamic voltage and frequency scaling (DVFS) states for the plurality of core types. In the example performance map 800, the plurality of core types comprise performance cores (P-cores) and efficiency cores (E-cores) of an asymmetric multiprocessing system (AMP). A plurality of DVFS states are shown for each of the P-cores and the E-cores of the AMP. The core types can alternatively be different core configurations of a symmetric multiprocessing system (SMP). The DVFS states shown for the core types are by way of non-limiting example. Different DVFS states and more, and/or different, core types can be used in the performance map.
In an embodiment, when a thread group first begins execution on a processor, the control effort can be zero, by default. As shown in performance map 800, a control effort of zero corresponds to recommending that threads of the thread group be run on an E-core at a DVFS state of 400 MHz. As threads of the thread group are executed, a plurality of thread execution metrics are accumulated for threads of the group, indicating progress on performance targets for the threads of the group. If performance targets for the threads of the thread group are not being met, then the control effort produced by CLPC 300 for the thread group can rise, indicating that more processing performance is needed to meet performance targets for thread group. A higher control effort can indicate, e.g. an E-core running at a DVFS state of 1000 MHz. The performance map can have one or more overlapping entries for a control effort. In the example performance map 800, a control effort can map to recommending an E-core running at a DVFS state of 1800 MHz or recommending a P-core running at a DVFS state of 200 MHz. In an embodiment, the workload for the thread group can be analyzed to determine which core type and DVFS state is more appropriate for the workload. For example, if the workload comprises substantial arithmetic computations, graphics rendering, or other workload type that is appropriate for a P-core, then the P-core at DVFS state 200 MHz can be recommended for the thread group. If the workload comprises substantial block I/O operations, integer arithmetic, character I/O, or other general-purpose processing, then an E-core can be recommended at DVFS state 1800 MHz. If a workload comprises background tasks, then an E-core can be recommended at DVFS state 1800 MHz, based on the control effort value and analysis of the workload.
Similarly, as the control effort needed to meet performance goals for the thread group continues to rise, P-cores with increasing DVFS states can be recommended for the thread group. When performance targets are being met, and the control effort falls, then lower DVFS states for a P-core type, or an E-core type, are recommended for the thread group. Control effort needed to meet performance goals can fall when, e.g., the overall system workload is reduced, or the recommended core type and DVFS state meets or exceeds performance goals for the thread group.
In operation 910, one or more application programs, processes, tasks, or objects can be launched.
In operation 1000, default thread groups can be determined for threads of the launched application(s), daemons, or other default grouping. An initial control effort, core type, and DVFS state can also be set for one, or more, or all thread groups. See
In operation 920, processor complex scheduler 705 can enqueue one or more threads for execution. Processor complex scheduler 705 can have a thread queue for each core type, e.g. efficiency cores (E-cores) and performance cores (P-cores). By default, the processor complex scheduler can initially schedule threads of all thread groups for execution on an E-core queue for execution at a default DVFS state.
In operation 930, one or more threads can be executed, thereby generating thread execution metrics for each of a plurality of threads. Thread execution metrics, e.g. thread execution metrics 311-316 of
In operation 1100, dynamic thread group changes can be made, if any. For example, a second thread may receive a voucher indicating that the second thread is to perform work on behalf of a first thread. A voucher is a collection of attributes in a message sent via inter-process communication (IPC) from a first thread to a second thread. One of the attributes that the first thread puts in the voucher is the thread group to which the first thread currently belongs. The second thread can adopt the thread group in the voucher before the next execution of the second thread. In another example, a thread associated with a work interval object may perform work on behalf of the work interval object. A work interval object is used to represent periodic work where each period has a deadline. The object possesses a token and a specified time interval for one instance of the work. A work interval object can be used to designate threads that are working together to achieve a common purpose in a specified time interval. Before performing work for the work interval object, the thread can adopt the thread group of the work interval object. Dynamic thread grouping is described in detail, below, with reference to
In operation 1200, for each thread group that was active on a core of the processor complex 111 within a time interval, a CLPC 300 for the thread group can receive a sample of a sample of plurality of thread execution metrics for the thread group into a plurality of tunable controllers. CLPC can determine a control effort for the thread group from the outputs of each tunable controller of the thread group. The control effort for the thread group can be mapped to a performance map, e.g. performance map 800 of
In operation 940, as may be appropriate, one or more threads that dynamically changed thread group in operation 1100 may be restored to the thread group that the thread had before being dynamically changed.
In operation 1010, all threads of an application can be assigned to a thread group for the application (termed a “coalition”) by default. A coalition is a grouping of closely related tasks/processes of an application which consist of threads, e.g. a web browser and application extensions associated with the web browser. Threads of an application can also be initially grouped in accordance with an “application bundle.” An application bundle is package that can include source code, compiled code, script commands, installation directive, references to one or more APIs, frameworks or other services upon which the application in the bundle may depend, and other meta data describing the contents of the application bundle. A coalition can also include a group of related threads that work toward a common goal, such as work interval objects, or threads launched by a daemon.
In operation 1015, it can be determined whether an object type is specified in the source code, object code, script, or application bundle for the object type that may override a default thread grouping for an object. An overriding object type is an object of a type whose thread group will be adopted by another thread, such as by a voucher or a work interval object. Thus, the object thread group is said to “override” an initial thread group for a thread. An object type can be a process, task, or group of threads that perform work toward a common goal or purpose. An overriding object type can have its own thread group. An overriding object type can be a program object that calls one or more threads of one or more applications to perform work on behalf of the object having the object type. For example, a work interval object can be specified as having its own thread group. If it is determined that an overriding object type is specified, then method 1000 continues at operation 1020, otherwise method 1000 continues at operation 1025.
In operation 1020, application threads associated with the work interval object can be assigned the thread group of the work interval object. In addition, threads of other applications that opt-in to performing work on behalf of the work interval object can also be grouped with the work interval object. In an embodiment, when a second thread is called to perform work on behalf of first thread, the second thread may optionally adopt the thread group of the first thread. The second thread is said to “opt-in” to the thread group of the first thread. A thread that opts-in a thread group may later opt-out of the thread group and return to its previous thread group. Method 1000 continues at operation 1060.
In operation 1025, it can be determined whether a thread group has been specified in the application source code, compiled code, script, or application bundle for one or more threads of an application. Source code may be annotated with a thread group indication. For example, source code may be annotated with a compiler directive that a particular function, library, or block of source code is to have a thread group as indicated in the compiler directive. An example directive follows: #THREADGROUP_ID=12345678; function xyz(void) { . . . }; #ENDTHREADGROUP. One or more application programming interfaces (APIs) or frameworks may be specified or included in the application that determine a thread group when the API or framework is called. For example, the source code for the API can be annotated with a compiler directive that all, or a portion, of the API is to have a thread group as indicated in the compiler directive. If a thread group is specified in operation 1025, then method 1000 continues at operation 1030, otherwise method 1000 continues at operation 1035.
In operation 1030, threads having a specified thread group are grouped together into a thread group. Method 1000 continues at operation 1060.
In operation 1035, it can be determined whether one or more threads of the application are associated with a well-known functionality, such as video, audio, animation, block I/O, or the like. Functionality that is well-known can have a named thread group identifier that can be referenced within source code, scripts, application bundles, and the like. For example, “audio,” “video,” “game,” “network,” etc. Processor complex scheduler thread grouping logic can convert a named thread group to a thread group identifier. If so, then method 1000 continues at operation 1040, otherwise method 1000 continues at operation 1045.
In operation 1040, threads associated with a well-known functionality are grouped together into a thread group. Well-known functionality can include, e.g. a library, API, or framework that is directed to video processing, audio processing, rendering, input/output (I/O) for block storage devices. The library, API, or framework, can contain an explicit compiler directive indicating its functionality. In an embodiment, the source code or compiled code for the library, API, or framework, can analyzed to determine the daemon(s) called by the library, API, or framework. Initial thread grouping logic 251 can perform this analysis and assign thread groups in accordance with the analysis. Method 1000 continues at operation 1060.
In operation 1045, it can be determined whether a thread was launched by a daemon that is associated with a thread group. If so, then method 1000 continues at operation 1050, otherwise method 1000 continues at operation 1055.
In operation 1050, the called daemon dynamic thread grouping logic 252 can assign the thread group to the thread launched by the daemon. Method 1000 continues at operation 1060.
In operation 1055, threads not otherwise grouped, above, stay assigned to their default application thread group.
In operation 1060, by default, a minimum control effort is recommended for each thread group. During initial thread grouping, all thread groups can be recommended for minimum control effort by initial thread grouping logic 251. Minimum control effort corresponds to a control effort value of zero. A control effort of zero maps to a performance map entry for recommending an efficiency core at the lowest DVFS state for efficiency cores in the performance map.
In operation 1105, it can be determined whether a thread performing work having a common purpose with a work interval object (WIO) 135, the work to be completed within the specified time interval of the WIO 136, has opted-in to joining the WIO 135. In an embodiment, when a second thread is called to perform work on behalf of first thread, the second thread may optionally adopt the thread group of the first thread. The second thread is said to “opt-in” to the thread group of the first thread. A thread that opts-in a thread group may later opt-out of the thread group and return to its previous thread group. If the thread opts-in to the thread group of the work interval object, then in operation 1100, the thread adopts the thread group of WIO 135 and method 1100 ends. Otherwise method 1100 continues at operation 1115.
In operation 1115, it can optionally be determined whether the thread is called to perform work on behalf of a calling thread or is otherwise made runnable by a calling thread. A calling thread is a first thread that calls a second thread to perform work on behalf of the first thread. If so, then in operation 1120, the thread is grouped with the calling thread and method 1100 ends. Otherwise method 1100 continues at operation 1125.
In operation 1125, it can be determined whether the thread received a message to perform work on behalf of another thread. A message may be, e.g., passed from a first thread to this thread via inter-process communication, and the message contains or references a voucher that includes or references a thread group. A voucher is a collection of attributes in a message sent via inter-process communication (IPC) from a first thread to a second thread. One of the attributes that the first thread puts in the voucher is the thread group to which the first thread currently belongs. If so, then in operation 1130, the thread adopts the thread group referenced in the message (voucher) and method 1100 ends. Otherwise, method 1100 continues at operation 1135.
In operation 1135, it can be determined whether a thread that previously opted-in to joining a work interval object 135 has now opted-out of the work interval object 135. If so, then in operation 1140, the thread is restored to the thread group that it had before the thread opted-in to joining the thread group of the work interval object 135. Otherwise, method 1100 continues at operation 1145.
In operation 1145, it can be determined whether a thread has been made runnable by another thread. If so, then method 1100 continues at operation 1150. Otherwise, method 1100 ends.
In operation 1150, the thread that is made runnable by the other thread can join the thread group of the other thread and method 1100 ends.
In operation 1205, a CLPC receives samples of a plurality of performance metrics for the thread group. A non-limiting example list of performance metrics can include, e.g. a work interval utilization for one or more work interval objects, block storage I/O rate for the thread group, a processor complex utilization for the thread group, scheduling latency for the thread group, and a cluster residency for the thread group.
In operation 1210, samples of the performance metrics for the thread group can be fed into a plurality of tunable controllers. In an embodiment, a tunable controller can be a proportional-integral-derivative (PID) controller.
In operation 1215, for the thread group, CLPC outputs a control effort (CE) value. In an embodiment, the control effort value is a unitless value from 0 to 1.
In operation 1220, the CLPC can map the control effort to a performance map to determine a recommended core type and DVFS state for the thread group, based on the control effort for the thread group.
In operation 1225, it can be determined whether the control effort maps to an area of the performance map that has overlap between core types, such that the control effort maps to both an E-core and DVFS state and a P-core and DVFS state. If so, then method 1200 continues at operation 1230, otherwise method 1200 continues at operation 1235.
In operation 1230, the workload in the thread group can be analyzed to determine whether the work is more suitable for one core type or another. For example, the thread group may contain threads that render a graphic, or perform floating point arithmetic, and are more suitable for a P-core than an E-core. Similarly, the work in the thread groups may indicate that the work is not complex, such as block I/O, and may be more suitable for an E-core. In response to the analysis of the work of the thread group, a recommended core and DVFS state is determined and method 1200 continues at operation 1240.
In operation 1235, the control effort did not map to a portion of the performance map that contains overlapping entries of core types, and thus the recommended core type and DVFS state can be directly derived from the performance map, using the control effort for the thread group. Method 1200 continues at operation 1240.
In operation 1240, it can be determined whether there are more thread groups that were active on a processor core during the sample period of the CLPC. If so, then method 1200 continues at operation 1210 with a next thread group, otherwise method 1200 continues at operation 1245, described below with reference to
In
In operation 1700, a closed loop thermal and power controller can determine a limit on the DVFS state for each core type, and a number of active cores for each core type. Operation 1700 is described in detail, below, with reference to
In operation 1250, the maximum DVFS state for each core type, determined in operation 1245 above, can be limited using the DVFS state limit determined for each core type in operation 1700 above. The number of cores per core type may also be limited by operation 1700. Operation 1700 can mask off one or more cores of each type to limit power consumption and, thus, heat generation.
In operation 1310, it can be determined whether the processor complex scheduler detects that (1) a P-eligible thread is enqueued and runnable, (2) no P-cores are idle, and (3) at least one E-core is idle. If no P-eligible threads are enqueued on the P-queue, or there are idle P-cores, or there are no idle E-cores, then method 1300 ends, otherwise method 1300 continues at operation 1315.
In operation 1315, processor complex scheduler obtains a runnable P-eligible thread from the P-queue and enqueues the runnable P-eligible thread for running on an E-core.
In operation 1320, optionally, the DVFS state of the E-cores can be synchronously increased to run the P-eligible thread.
In operation 1410, it can be determined whether an E-core is going idle. If not, then method 1400 ends, otherwise, method 1400 continues at operation 1415.
In operation 1415, it can be determined whether there is a runnable P-eligible thread waiting on an available P-core. If not, then method 1400 ends. Otherwise, method 1400 continues at operation 1420.
In operation 1420, the E-core going idle can obtain the P-eligible thread and run the P-eligible thread on the E-core.
In operation 1425, optionally, the DVFS state of the E-core cluster can be synchronously increased to run the P-eligible thread on the E-core.
In operation 1510, it can be determined whether there is a P-core that is idle or in the process of going idle. If not, then method 1500 ends. Otherwise, method 1500 continues at operation 1515.
In operation 1515, it can be determined whether a P-eligible thread is scheduled on the E-core scheduling queue. If not, then method 1500 ends, otherwise method 1500 continues at operation 1520.
In operation 1520, the runnable P-eligible thread that is scheduled on the E-core scheduling queue is enqueued on the P-core scheduling queue, and the P-eligible thread is executed on a P-core.
In operation 1525, the DVFS state for the E-core cluster can optionally be reset to the DVFS state recommended for the E-cores, prior to the E-core DVFS state being synchronously increased to run the P-eligible thread on an E-core.
In
In operation 1610, a thread T1 of App. 1 calls an API or framework that invokes a daemon or operation system (OS) service 127 to perform work.
In operation 1615, the called daemon or OS service 127 creates the work interval object (WIO) 135 having a thread group e.g., 10 (TG10).
In operation 1620, thread T1, performing work associated with WIO 135, can request that the thread join WIO thread group TG10.
In operation 1625, thread T1 opts-in to WIO 135 thread group TG10. Thread T1 then performs work on behalf of WIO 135. Thread execution metrics generated by the execution of thread T1 are accumulated with thread group TG10 of the WIO 135.
In operation 1630, it can be determined whether thread T1 opts-out of WIO 135 thread group TG10. If so, then method 1600 continues at operation 1645, otherwise method 1600 continues at operation 1640.
In operation 1640, thread T1 remains in thread group TG10 of WIO 135. Method 1600 continues at operation 1650.
In operation 1645, thread T1 returns to application thread group TG1. Method 1600 continues at operation 1650.
In operation 1650, optionally thread T1 may perform additional work. The thread execution metrics associated with execution of thread T1 are charged to the thread group (application TG1, or WIO TG10) to which the thread T1 belongs.
In
In operation 1605, an application, e.g. App. 1, is launched. By default, the threads in the application are assigned to a single group, e.g., thread group 1 (TG1).
In operation 1612, an application, e.g. App 2, is launched, with threads belonging to, e.g. thread group 2 (TG2).
In operation 1617, App. 2 creates the work interval object (WIO) 135 having a thread group e.g., 10 (TG10).
In operation 1620, thread T1, performing work associated with WIO 135, can request that thread T1 join WIO thread group TG10.
In operation 1625, thread T1 opts-in to WIO 135 thread group TG10. T1 then performs work for WIO 135. Thread execution metrics generated by the execution of thread T1 are accumulated with thread group TG10 of the WIO 135.
In operation 1630, it can be determined whether thread T1 opts-out of WIO 135 thread group TG10. If so, then method 1600 continues at operation 1640, otherwise method 1600 continues at operation 1645.
In operation 1640, thread T1 remains in thread group TG10 of WIO 135. Method 1600 continues at operation 1650.
In operation 1645, thread T1 returns to application thread group TG1. Method 1600 continues at operation 1650.
In operation 1650, optionally thread T1 may perform additional work. The thread execution metrics associated with execution of thread T1 are charged to the thread group (application TG1, or WIO TG10) to which the thread T1 belongs.
In operation 1705, it can be determined whether a fast die temperature of a processor complex 111 is greater than a fast die temperature limit for a predetermined period of time. A processor complex 111 may be a system on a chip (SoC) comprising a plurality of temperature sensors. In an embodiment, each cluster of cores of a core type can have a fast die temperature sensor. In an embodiment, one or more fast die temperature sensors can be virtual temperature sensors. The fast die temperature predetermined time limit may be on the order of less than 10 milliseconds (10 ms). If a fast die temperature is greater than a threshold temperature for a predetermined threshold time, then method 1700 continues at operation 1735, otherwise method 1700 continues at operation 1710.
In operation 1710, it can be determined whether the raw power consumed by the processor complex 111 exceeds a raw power consumption limit for a predetermined period of time. Raw power is power consumed that is not filtered by a control element, such as a low-pass filter of a power zone. Power consumed can be measured by an energy sensor and summed over the predetermined time limit to obtain power consumed. In an embodiment, an average energy per instruction can be determined and multiplied by a number of instructions executed over the predetermined time to obtain power consumed. This operation detects bursts of power consumed that can raise the temperature of the processor complex 111 or the system very quickly. Thus, a predetermined time for this burst may be on the order of less than 10 ms. If raw power consumed during the predetermined time limit exceeds the predetermined power consumption threshold, then method 1700 continues at operation 1735, otherwise method 1700 continues at operation 1715.
In operation 1715, it can be determined whether the filtered power consumed in one of a plurality of power zones exceeds power target for the zone, less a power delta for the zone, for a predetermined period of time. If so, then method 1700 continues at operation 1735, otherwise method 1700 continues at operation 1720.
In operation 1720, can be determined whether the filtered power consumed by the processor complex 111 (system on a chip, SoC) is greater than a predetermined filtered power limit. If so, then method 1700 continues at operation 1735, otherwise method 1700 continues at operation 1725.
In operation 1725, it can be determined whether a count of automatic hardware throttles has exceeded a predetermined count for a predetermined period of time. Automatic hardware throttling can block an instruction from being executed by a processor core, but is expensive and inefficient. If it is determined that automatic hardware throttling has occurred more than a predetermined number of times during a predetermined period of time, then method 1700 continues at operation 1735, otherwise method 1700 continues at operation 1730.
In operation 1730, it can be determined whether a temperature sensor, other than a processor complex 111 temperature sensor, has exceeded a temperature limit for a predetermined period of time. If so, then method 1700 continues at operation 1735, otherwise method 1700 ends. A non-processor complex temperature sensor can include one or more temperature sensors on a device housing (“skin temperature”), a temperature sensor on a subsystem such as a display, battery, battery charger, or other heat-generating component. A predetermined period of time may be on the order of 100's of milliseconds.
In operation 1735, the DVFS state, and/or number of active cores, can be reduced for one or more clusters of processor complex 111. Reducing DVFS state and/or number of active cores can reduce power consumption and thereby reduce generated heat.
Aspects of
In
In operation 1815, an interrupt can be set for an E-core with a E-core delay time. P-cores can retire instructions very quickly, and a P-core may become available to accept a scheduled thread within the E-core delay time. Method 1800 continues at
In
In operation 1825, an E-core interrupt time has expired and no P-core has become available. An E-core can be woken up and the P-eligible thread can be scheduled to run on the E-core, and method 1800 ends.
In operation 1830, the interrupt timer has not yet expired, and it can be determined whether a P-core has become available for thread execution. If not, then method 1800 resumes at operation 1820 to check the interrupt timer, otherwise method 1800 continues at operation 1835.
In operation 1835, a P-core has become available before the E-core interrupt timer has expired. The E-core interrupt can be canceled, the P-eligible thread can be scheduled on a P-core, and method 1800 ends.
Returning to
In operation 1845, a P-core interrupt can be set with an interrupt delay of zero, to generate an immediate interrupt. Method 1800 resumes at operation 1870.
In operation 1850, the scheduling event is “make runnable.” An interrupt can be set for the recommended core type for this thread. In an embodiment, if the thread is a member of a thread group that is recommended for a P-core, then a P-core interrupt is set. In an embodiment, if the thread is a member of a thread group that is recommended for an E-core, then an E-core interrupt is set.
In operation 1852, it can be determined whether the recommended core type is a P-core. If so, then method 1800 continues at operation 1860, otherwise method 1800 continues at operation 1855.
In operation 1855, it can be determined whether the recommended E-core type is available (not idle) for the thread group of the thread. If so, then method 1800 continues at operation 1860, otherwise method 1800 continues at operation 1865.
In operation 1860, a core of the recommended core type is available, thus the interrupt timer is set to zero so that the interrupt for the recommended core type will trigger as soon as the interrupt is programmed (immediate interrupt). Method 1800 continues at operation 1870.
In operation 1865, the interrupt time for the interrupt timer for the recommended core type for the thread group of the thread is set to the interrupt timer delay for the recommended core type (deferred interrupt). Method 1800 continues at operation 1870.
In operation 1870, the interrupt controller can be programmed with the interrupt type determined in operation 1850 and the interrupt delay value determined in either operation 1860 or operation 1865, to wake up the thread at the expiration of the interrupt timer and schedule the thread on a core of the recommended core type. Method 1800 continues at operation 1875 of
On
In operation 1880, it can be determined whether a core of the recommended type for the thread group of the thread has become available. If so, then method 1800 continues at operation 1885, otherwise method 1800 continues at operation 1875.
In operation 1885, the interrupt programmed for the thread can be canceled.
In operation 1890, the thread is woken up and scheduled on a core of the recommended core type for the thread group of the thread.
A table of IPI actions can use the current core type 1910 upon which a thread last ran, a target core type 1920 that the thread is scheduled to run upon, a scheduling recommendation 1930 for a core type for the thread, a scheduling event 1940 that has occurred, a target core idle state 1950 (e.g., idle/not idle) of a target core type, and an interrupt configuration decision 1960 as to deferred inter-process interrupts for the thread.
As an example, in line 1901, a thread has last run on a P-core, thus the current core type 1910 is “P”. The thread will be scheduled to run on an E-core, thus the target core type 1920 is “E”. The scheduling recommendation 1930 for the thread, as recommended by the CLPC 300, is for the thread to run on a P-core. The scheduling event 1940 is “spill from P to E-core,” which indicates that the thread, though-P eligible, is scheduled to be transferred from a P-core to an E-core. Rather than expend the overhead to nm the P-eligible thread on the E-core, DIPI can be used to let the P-eligible thread wait for a short time to see if a P-core will become available to run the P-eligible thread. Thus, the interrupt decision 1960 is “deferred,” without regard to target core idle state 1950. The interrupt controller is programmed with the P-core delay value and if a P-core becomes available before the interrupt timer expires, then the interrupt is canceled and the P-eligible thread is run on a P-core.
In example 1902, the current core type 1910 for the thread is an E-core and the target core type 1920 is a P-core. The scheduling recommendation 1930 is to run the thread on a P-core. The scheduling event 1940 is “rebalance from E-core to P-core.” The interrupt decision 1960 is “immediate,” without regard to target core idle state 1950. The interrupt controller timer is programmed cause an interrupt to wake up a P-core, enabling the thread to be scheduled.
In example 1903, the current core type 1910 for the thread is a P-core, and the target core type 1920 is an E-core. The scheduling recommendation 1930 for the thread group of the thread is for an E-core. The scheduling event 1940 is “make runnable,” which indicates to make the thread runnable on a core. The conditions in line item 1903 can occur when CEL 400 reduces the recommended core type from P-core to E-core, such as when performance targets for the thread group are being met, and less control effort is needed to meet the thread group performance targets. Alternatively, the system may be heavily loaded, may have consumed too much power, and/or generated too heat much as determined by CEL 400, and the thread group recommended has been reduced by CEL 400 from P-core to E-core. If the target core idle state 1950 is “idle”, the interrupt decision 1960 is set to deferred. The interrupt controller can be programmed for the E-core delay value to wait on an E-core to become ready to run the thread. If the target core idle state 1950 is “not idle”, the interrupt decision is set of immediate. The interrupt controller timer can be set to zero and an interrupt will be generated, enabling the thread to run on an E-core.
In example 1904, the current core type 1910 is an E-core, the target core type 1920 is a P-core, with a scheduling recommendation 1930 or P-core. Where the scheduling event 1940 is to make the thread runnable. The interrupt time for P-cores can be set to zero (e.g., interrupt decision 1960 of immediate), an interrupt will be generated, and the thread will be run on a P-core. In example 1904, the interrupt decision 1960 can be made without regard to the target core idle state 1950.
In example 1905, the current core type 1910 for the thread is an E-core, the target core type 1920 for the thread is an E-core, the scheduling recommendation 1930 is for E-core, and the scheduling event 1940 is to make the thread runnable. The interrupt decision 1960 can be deferred or immediate based on the target core idle state 1950. If the target core idle state 1950 is idle, then an E-core delay can be programmed into the E-core interrupt to generate an interrupt. Otherwise, the E-core timer can be programmed to generate an immediate interrupt.
In example 1906, the current core type 1910 for the thread is a P-core, the target core type 1920 for the thread is a P-core, the scheduling recommendation 1930 is for a P-core, and the scheduling event 1940 is to make the thread runnable. The P-core interrupt time can be programmed with a value to generate an immediate interrupt (e.g., interrupt decision 1960 of immediate) and the thread will be run on a P-core. In example 1906, the interrupt decision 1960 can be made without regard to the target core idle state 1950.
Note that the Service 2 has two APIs, one of which (Service 2 API 1) receives calls from and returns values to Application 1 and the other (Service 2 API 2) receives calls from and returns values to Application 2, Service 1 (which can be, for example, a software library) makes calls to and receives returned values from OS API 1, and Service 2 (which can be, for example, a software library) makes calls to and receives returned values from both as API 1 and OS API 2, Application 2 makes calls to and receives returned values from as API 2.
The computing system illustrated in
Computing system 2100 includes bus 2105 or other communication device to communicate information, and processor complex 2110 coupled to bus 2105 that may process information.
While computing system 2100 is illustrated with a single processor, computing system 2100 may include multiple processors and/or co-processors in processor complex 2110. Processor complex 2110 can include a plurality of core types. Processor complex 2110 can comprise a symmetric multiprocessing complex (SMP) having a plurality of cores that are configured in a plurality of different configurations. Processor complex 2110 can comprise an asymmetric multiprocessing system having a plurality of different core types, each having one or more cores. Core types can include performance cores, efficiency cores, graphics cores, and arithmetic processing cores. A performance core can have an architecture that is designed for very high throughput and may include specialized processing such as pipelined architecture, floating point arithmetic functionality, graphics processing, or digital signal processing. A performance core may consume more energy per instruction than an efficiency core. An efficient processor may include a general-purpose processor that can process input/output (I/O) such as for block storage, data streams, interfacing to a display, processing integer arithmetic, and other general processing functionality. An efficient core may consume less energy per instruction than a performance core. Processor complex 2110 can comprise a system on a chip (SoC).
Computing system 2100 further may include random access memory (RAM) or other dynamic storage device (referred to as main memory 2120), coupled to bus 2105 and may store information and instructions that may be executed by processor complex 2110. Main memory 2120 may also be used to store temporary variables or other intermediate information during execution of instructions by processor complex 2110.
Computing system 2100 may also include read only memory (ROM) 2130 and/or other static data storage device coupled to bus 2105 that may store static information and instructions for processor complex 2110. A non-volatile data storage device 2140 may be coupled to bus 2105 to store information and instructions. The non-volatile data storage device 2140 can include storage media such as, but not limited to flash memory or a magnetic disk, optical disc, solid state disc, writeable or rewriteable compact disc.
Computing system 2100 may further include power and thermal sensors 2145 that include power or energy sensors and thermal sensors. In an embodiment where processor complex 2110 comprises a system on a chip (SoC), one more thermal sensors can be included in the SoC. In an embodiment, at least one thermal sensor can be included on the SoC for each core type of the processor complex 2110. In an embodiment, a thermal sensor can comprise a virtual thermal sensor. A virtual thermal sensor can comprise a plurality of physical thermal sensors and logic that estimates one or more temperature values at location(s) other than the physical thermal sensors. Power sensors may also be included in specific locations, such as power consumed by the processor or SoC 111, power consumed by a particular subsystem, such as a display, storage device, network interfaces, and/or radio and cellular transceivers. In an embodiment, a power sensor within the power and thermal sensors 2145 can comprise or include an energy sensor. In an embodiment, the energy sensor can accumulate an average energy consumed over time to produce a power consumed value.
Computing system 2100 may also be coupled via bus 2105 to display device 2150, such as a liquid crystal display (LCD), light emitting diode (LED) display, or touch screen, to display information to a user. Computing system 2100 can also include an alphanumeric input device 2160, including alphanumeric and other keys, which may be coupled to bus 2105 to communicate information and command selections to processor complex 2110. An alphanumeric keypad can be implemented as keypad images on a touch screen display. Another type of user input device is cursor control 2165, such as a touchpad, a mouse, a trackball, touch screen input or cursor direction keys to communicate direction information and command selections to processor complex 2110 and to control cursor movement on display device 2150. Computing system 2100 may also receive user input from a remote device that is communicatively coupled to computing system 2100 via one or more network interface(s) 2180.
Computing system 2100 can further include an audio, video, or audio/video processor 2170. An audio processor may include a digital signal processor, memory, one or more analog to digital converters (ADCs), digital to analog converters (DACs), digital sampling hardware and software, one or more coder-decoder (coded) modules, and other components. A video processor can include one or more video encoders, camera, display, and the like.
Computing system 2100 further may include one or more network interface(s) 2180 to provide access to a network, such as a local area network. Network interface(s) 2180 may include, for example, a wireless network interface having antenna 2185, which may represent one or more antenna(e). Computing system 2100 can include multiple wireless network interfaces such as a combination of Wi-Fi, Bluetooth® and cellular telephony interfaces. Network interface(s) 2180 may also include, for example, a wired network interface to communicate with remote devices via network cable 2187, which may be, for example, an Ethernet cable, a coaxial cable, a fiber optic cable, a serial cable, or a parallel cable.
In one embodiment, network interface(s) 2180 may provide access to a local area network, for example, by conforming to IEEE 802.11 b and/or IEEE 802.11 g standards, and/or the wireless network interface may provide access to a personal area network, for example, by conforming to Bluetooth standards. Other wireless network interfaces and/or protocols can also be supported. In addition to, or instead of, communication via wireless LAN standards, network interface(s) 2180 may provide wireless communications using, for example, Time Division, Multiple Access (TDMA) protocols, Global System for Mobile Communications (GSM) protocols, Code Division, Multiple Access (CDMA) protocols, and/or any other type of wireless communications protocol.
Multi-Agent Workload Metrics
The system 2200 additionally includes the CLPC 300, which acts as the performance and power manager for the system. In some embodiments the CLPC 300 is integrated into the operating system 120, as illustrated in
The signal can be performed by calling the CLPC 300 or a software module associated with the CLPC 300. For example, the co-processor driver can use the I/O service 2210 to call a WorkSubmit interface to indicate details about the submitted workload, as well as information on the submitting thread of the workload. Furthermore, the WorkSubmit, WorkBegin, WorkEnd calls also include an identifier to the identity of calling driver & hardware, for example, an identifier to the GPU driver and GPU hardware.
The information on the submitted workload can include a priority or quality of service classification for the submitted workload and/or a priority or quality of service classification associated with the submitting thread. The I/O Service 2210 can then generate a token in response to the message, where the token is an identifier that can be used to tie metrics associated with the offloaded workload to the submitting thread. For example, in one embodiment the token can be used to identify the thread group associated with the submitting thread of the workload, where the thread group is the repository of metrics associated with the group. In one embodiment, the token also keeps the thread group alive by taking a reference on the thread group object. Accordingly, even if all other references to the group are released while the workload has been offloaded to a co-processor, the thread group and associated metrics will remain allocated because of the reference associated with the token.
In one embodiment, upon beginning of the offloaded workload on the co-processor, the co-processor driver, or another thread associated with the co-processor driver, can call a WorkBegin interface with the token received from the I/O service 2210. The CLPC 300 can tie the WorkBegin call to the previously WorkSubmit call using the token, even if the calling thread group of WorkBegin differs from the calling thread group of WorkSubmit. Upon completion of the workload, the co-processor driver, or an associated thread, can inform the CLPC 300 via a WorkEnd call that also includes the token. In one embodiment, the WorkBegin and WorkEnd call can each be used to return a collection of metrics for the workload that were gathered by the co-processor driver. In one embodiment, the WorkBegin call can provide metrics such as hardware performance counter values, execution statistics for completed work (e.g., preemption times), and similar metrics.
In one embodiment, upon receipt of the WorkEnd call, the CLPC 300 can retrieve metrics for the workload from the co-processor driver. The CLPC 300 can then integrate the workload metrics into the performance and efficiency metrics for the submitting thread group. If the submitting thread is part of a thread group that is associated with a work interval object, adjustments can be made to the DVFS state for the processor complex 111 or co-processor based on processor performance relative to the WIO deadlines. Additionally, the reference on the thread-group object taken by the token is released during the WorkEnd call, allowing the thread-group object to be released if no other references are held.
It will be understood that the concepts described herein can be applied to a system including any number of GPUs, neural engines, or other co-processors, and are not limited to systems having a single instance of these co-processors. Furthermore, when threads are offloaded from a processor to a co-processor, threads can offload work to some subset of the available co-processor instances within the system, rather than using all available co-processor instances.
Some operations described below can be performed by the hardware of a co-processor, firmware modules, or software modules associated with the co-processor. The methods can be used to track metrics for a variety of co-processors, including but not limited to GPUs, neural engines, image processors, audio processors, and other processors that can co-operate with application processors within a computing system. Furthermore, in some embodiments the concepts described herein can be applied to any variant or type of accelerator devices, including scalar/rotator blocks or encoder/decoder blocks.
In one embodiment, the method 2300 includes operation 2302, which executes threads of a thread group on a processor of the processor complex. A scheduler for the processor can schedule threads of the thread group on the recommended core type at the recommended DVFS state for the thread group.
In operation 2304, one or more of the threads can determine to offload a workload to a co-processor. The workload to offload may be a workload most suited for processing on the co-processor. For example, a graphics processing workload can be offloaded to a GPU. A facial recognition or face detection workload can be offloaded to a general-purpose graphics processing unit (GPGPU) or another parallel compute engine, such as the GPU 2230 and neural engine 2235 of
In operation 2306, the offloading thread of the thread group can submit a command buffer to the co-processor. The offloading thread can submit the command buffer via a driver associated with the co-processor, such as a GPU driver 2220, a neural engine driver 2225, or another software driver associated with the co-processor.
In operation 2307, the thread group can determine if any additional work is pending. If additional work is pending for the thread group, in operation 2309 the thread group process the next workload. If no additional work is pending, in operation 2311 the thread group can go idle and yield the processor to other threads. The application processor can then process additional threads or go idle if no additional threads are enqueued for processing.
In operation 2308, the thread group can receive notice of completion of the workload on the co-processor. Operation 2308 can be performed after the thread or thread group is resumed from an idle or sleep state if no work was available for processing during the offload. As a result of the thread or thread group going idle, the processor on which the thread or thread group was executing may have been transitioned into a different DVFS state.
In operation 2310, the thread or thread group can receive results of the offloaded workload. As a result of performance and efficiency metrics gathered from the co-processor during the offload, the DVFS state of the processor executing the thread or thread group can be pre-adjusted to the appropriate DVFS state to enable the efficient processing of the received results.
Method 2400 includes operation 2402, in which the CLPC receives a message indicating that a thread of a thread group is to offload a workload to a co-processor. The message can be sent from the co-processor or co-processor driver in response to receipt of one or more command buffers of commands to be executed on behalf of the offloading thread. An identifier of the thread that is offloading the workload can be received with the message. In one embodiment, the message is, or is associated with, a WorkSubmit call into a software interface for the CLPC. For example, a co-processor driver can use an interface provided via the I/O service to call into a software interface of the CLPC.
In operation 2404 the CLPC can infer membership of the thread in the thread group based on an identifier of the thread using stored or accessible information regarding thread groups and thread-group membership. In one embodiment, priority information associated with the workload can also be determined from context information associated with the thread group.
In operation 2406 the I/O service can issue a token to the co-processor driver. The token can be used to associate the workload with the inferred thread group. Some time period may lapse between the submission of the workload to the co-processor and the beginning of workload execution on the processor. The token can be used to connect subsequent calls regarding the workload to the initial WorkSubmit call, allowing different threads or thread groups to issue WorkBegin and WorkEnd calls on behalf of a workload. Internally, the CLPC can create data structure entries to record metrics for the workload. The metrics can be indexed within the data structure according to the token issued to the co-processor or co-processor driver. For example, the CLPC can record a timestamp associated with a time of submission of the workload to the co-processor. The timestamp can be stored in a data structure at a location identified by the token or an identifier associated with the token. In one embodiment, the token can be an object associated with metrics for the workload.
In operation 2408, the CLPC can receive notice of the beginning of the workload on the co-processor, the notice including the issued token for the workload. This notice can be associated with a WorkBegin call made via the I/O service. In one embodiment the CLPC can record a timestamp associated with the beginning of the workload on the co-processor. The notice of the beginning of the workload on the co-processor can be provided by a different thread group than the thread group that caused the submission of the message in operation 2402.
In operation 2410, the CLPC can receive notice of completion of workload on the co-processor, the notice including the issued token for the workload. In one embodiment the CLPC can record a timestamp associated with the end of the workload on the co-processor.
In operation 2412, the CLPC can retrieve performance and/or efficiency metrics for the completed workload. In one embodiment the performance and/or efficiency metrics for the completed workload include timing metrics for the submission, beginning, and end of processing for the workload on the co-processor. In one embodiment, the metrics additionally include performance and/or efficiency metrics gathered by the co-processor or co-processor driver and submitted to the CLPC. In one embodiment, performance metrics for the workload are stored in memory accessible to the CLPC, which can retrieve the metrics for a workload. In such embodiment, the metrics for a workload can be made accessible to the CLPC after the workload completes execution, although at least some of the metrics may be available during workload execution. Further in such embodiment, the stored metrics may be indexed by, or otherwise made accessible by an identifier based on the token issued by the CLPC upon notice of submission of the workload.
At operation 2414, the CLPC can associate the performance and/or efficiency metrics for the completed workload with the thread group to generate updated metrics for the thread group. The performance and/or efficiency metrics for the completed workload can be integrated with the existing or historical metrics for the thread group.
At operation 2416, the CLPC can adjust a recommended core type and DVFS state based on the updated performance and/or efficiency metrics. Based on the performance metrics for the workload during execution on the co-processor, the thread group may be recommended for a different core type and/or the DVFS state for the core executing the associated thread group can be adjusted. For example, under circumstances where the performance or efficiency of workload execution on a co-processor can be improved by increasing the frequency of the application processor, the DVFS state for the application processor can be adjusted. In one embodiment, the CLPC can also adjust the DVFS state of the co-processor on which the workload is executed. Additionally, other techniques can be applied to adjust the performance and/or efficiency of the co-processor, including limiting the number of co-processor cores used to execute a workload, duty cycling the co-processor, or other techniques which can be used to manage the performance, processing efficiency, or power consumption of a co-processor.
In one embodiment, method 2500 includes operation 2502 to receive a command buffer for a workload offloaded from a thread executing on an application processor, such as an application processor executing within the processor complex. For example, a thread on an application processor can submit a buffer of commands to a GPU to render a window of a user interface. A thread on the application processor can also submit a buffer of commands to a neural engine to perform a facial recognition operation. The command buffer can be received by a co-processor driver, which in one embodiment is also executing on the application processor.
In operation 2504, a thread group associated with the co-processor driver can call a WorkSubmit interface of the CLPC via the I/O service. The call can include an identifier of a thread associated with the command buffer received in operation 2502. The WorkSubmit interface can be called via a software library or module that provides a software interface to the CLPC. In one embodiment the co-processor driver can access the software interface to the CLPC via an I/O service (e.g., I/O service 2210) provided by an operating system of a computing system described herein (e.g., operating system 120 of system 2200). In one embodiment, the WorkSubmit interface can be used to convey priority or quality of service information about the workload to be offloaded. In one embodiment, priority or quality of service information can be determined automatically from context information of the submitting thread. In one embodiment the WorkSubmit interface can be used to provide hints regarding a workload, such as an estimated complexity associated with the frame. The complexity information associated with the frame can enable the CLPC to proactively adjust processor and/or co-processor performance during execution of the workload to be offloaded.
In operation 2506, the co-processor driver can receive a token to identify the workload. The token can be used to tie the beginning and end of processing for the workload with the initial workload submit call. In one embodiment the token can be used to index or identify performance metrics generated for the workload. While in one embodiment the token is received from the I/O service, the token used to track workloads can be generated by other components within the system, such as but not limited to the CLPC.
In operation 2508, the co-processor driver can prepare co-processor execution logic to execute commands from the command buffer for the workload. The co-processor driver can configure a thread dispatcher or scheduler on the co-processor to schedule internal co-processor operations based on corm/lands specified in the command buffer. For example, the internal co-processor operations can be performed by hardware threads within execution units of the co-processor. The internal execution architecture of the co-processor can vary between co-processors.
In one embodiment, as shown in operation 2510, a thread group associated with the co-processor driver can call a WorkBegin interface of the CLPC when the workload is ready to execute. The call to the WorkBegin interface can include the token or a reference to the token for the workload. In one embodiment, the call to the WorkBegin interface can be performed by a thread group of the operating system kernel. The kernel thread group can call the WorkBegin interface in conjunction with submitting a command buffer to the co-processor on behalf of the co-processor driver. In one embodiment, the call to the WorkBegin interface can be used to convey metrics about the current performance state of the co-processor the CLPC. In one embodiment, metadata to enable estimation of amount of time it will take to perform a task can be passed during the WorkBegin call. In one embodiment, current co-processor load information can be conveyed during the WorkBegin call. In one embodiment, where multiple available co-processor or co-processor cores of a given type are available, the WorkBegin call can convey which of the available co-processor cores will be used to process the workload on the co-processor. The CLPC can use this submitted information to balance the overall power consumption of the system while offloaded work is being performed.
The co-processor driver, in one embodiment, can perform an optional operation 2512 to track performance metrics of co-processor execution logic associated with the work load. In addition to timestamp-based metrics gathered by the CLPC based on the WorkSubmit, WorkBegin, and WorkEnd calls, the co-processor may also record internal performance metrics that can be gathered, recorded, or monitored by the co-processor driver. These performance metrics can be reported to the CLPC or stored in memory that is accessible by the CLPC.
In operation 2514, the co-processor driver can call a WorkEnd interface of the CLPC with the token when the workload completes execution. Optionally, performance metrics captured by the co-processor driver, or a reference (e.g., pointer) to such metrics, can be provided with or in association with the call to the WorkEnd interface of the CLPC.
The method 2600 includes operation 2602 to create a work interval object associated with a first thread group. A work interval object can be created in several ways. There can be a set of predefined work interval objects in an operating system, daemon, framework, or application. A kernel of an operating system can create a work interval object explicitly, such as on behalf of a driver. A kernel of an operating system can implicitly create a work interval object on behalf of an application, such as in response to an application call to a framework.
In operation 2604, the CLPC can receive a message indicating that a thread of the thread group is to offload a workload to a co-processor. Operation 2604 can be performed in a similar manner as operation 2402 of
In operation 2606, the CLPC can infer membership of the thread in the thread group based on an identifier of the thread. In one embodiment, priority information associated with the workload can also be determined from context information associated with the thread group.
In operation 2608, in response to the message, the I/O service used to interface a co-processor driver with the CLPC can issue a token to associate the workload with the work interval object. In operation 2610, the CLPC can receive notice of the beginning of the workload on the co-processor, the notice including the issued token for the workload. In one embodiment the CLPC can record a timestamp associated with the beginning of the workload on the co-processor. In operation 2612, the CLPC can receive notice of completion of workload on the co-processor, the notice including the issued token. In one embodiment the CLPC can record a timestamp associated with the end of the workload on the co-processor.
In operation 2614, the CLPC can assess performance metrics for the completed workload. Assessing the performance metrics can include analyzing timestamps recorded by the CLPC, or software associated with the CLPC, in response to receipt of the WorkSubmit, WorkBegin, and WorkEnd. Those timestamps can be used to determine the time between the submission of the workload to the co-processor and the beginning of the workload on the co-processor, as well as the time required to complete the workload on the co-processor.
In operation 2616, the CLPC can associate the performance metrics for the completed workload with the thread group to generate updated metrics for the thread group. The updated metrics can be generated by integrating the new metrics with the existing or historical metrics for the thread group.
At operation 2618, the CLPC can adjust a recommended core type and DVFS state based on the updated performance metrics and target deadlines. Based on the performance metrics for the workload during execution on the co-processor and the performance of the thread group relative to the work interval object deadlines, the thread group may be recommended for a different core type and/or the DVFS state for the core executing the associated thread group can be adjusted. In one embodiment, the CLPC can also adjust the DVFS state of the co-processor on which the workload is executed.
In one embodiment, to before adjusting the DVFS state of the co-processor on which the workload is executed to increase the voltage or frequency of that co-processor, the CLPC can reduce the voltage or frequency of other co-processors within the system to keep an overall system power consumption below a threshold. For example, the voltage and frequency of processors or co-processors within a system that are not actively performing operations can be reduced, while increasing the voltage and frequency of other co-processors within the system.
Per-Instance Metrics for Offloaded Workloads
In one embodiment, WIO workloads offloaded to co-processors can be tracked on a per-instance basis. The token that is associated with a workload can be associated with a specific instance of the workload. For example, a WIO can be associated with a thread group that is to generate content for each frame to be displayed by a graphical system. Each frame to be generated based on a WIO be assigned a work interval instance identifier that uniquely identifies the workload instance associated with that frame. In some embodiments, metadata for each workload can be tracked on a per-instance basis. For example, in one embodiment a thread priority, frame complexity, or quality of service classification associated with the workload are be tracked on a per-frame basis.
In one embodiment, the instance identifier can be associated with the token that is provided by the I/O service in response to a call to the WorkSubmit interface. The instance identifier can be used to allow metrics for multiple instances of workloads associated with the same WIO to be tracked. Enabling the tracking of separate instances of a WIO allows the generation of performance and efficiency metrics for agent-spanning workloads at a per-frame granularity. This per-frame granularity of metrics allows for fine-grained DVFS scaling across the various processing agents within a computing system. Additionally, work interval instancing allows hints to be provided in advance that a particular frame may be more complex than previous frames, allowing for pre-emptive scaling of execution resources.
In one embodiment, work interval instancing allows the tracking of offloaded metrics for pipelined work operations associated with an application. For example, a thread group for an application can pipeline workloads for multiple frames of content. Work for each frame can be tracked as a separate work instance interval. Tracking work interval objects on a per-frame, per-instance basis allows the CLPC to determine which of the individual frames that each portion of an application's operations are associated. For example, where a scrolling operation is to be performed, workloads for multiple frames can be prepared in a pipelined manner. Metrics for each frame can be tracked, enable fine-grained performance adjustments during the scrolling operation. Furthermore, complexity hints for a frame can be used to enable the CLPC to proactively adjust frequency for processors before the generation of the multiple frames of content.
In one embodiment, the deadline associated with an instance of a WIO is a timestamp in the future by which work associated with the WIO instance should be completed. Each frame, a new work interval instance of the WIO can be created, with each work interval instance having a new timestamp and a new identifier. The identifier for the work interval instance can be propagated with a workload as the workload is offloaded to the various co-processors. As shown in
In various embodiments, the identifier for the work interval instance can be propagated across the processing agents using one or more techniques. In some embodiments, automatic work interval instance identifier propagation can be enabled, such that the identifier for the most recently created work interval instance can automatically be associated with a workload to be offloaded. Various techniques for automatic work interval instance identifier propagation are performed in various embodiments. In one embodiment, automatic work interval instance identifier propagation can be performed using the voucher by which threads are associated with a thread group. The work interval instance identifier can then be automatically propagated via an association between the I/O service token and the voucher.
Alternatively, instance identifiers can be explicitly propagated via the WorkBegin call. In one embodiment, a co-processor driver, or another user-space thread, can receive a work interval instance identifier in response to the API call to create the work interval instance. The work interval instance identifier can then be propagated through the program logic through to the program code responsible for the WorkBegin call. The program code can then explicitly propagate the work interval instance identifier as a parameter of the WorkBegin call.
In one embodiment, the method 2800 includes operation 2802 in which the operating system can create a work interval instance for a frame of content to be created. The work interval instance can be created in response to an API call. In one embodiment, the API call is performed by a user mode thread to the kernel of the operating system. An identifier can then be created to identify the newly created work interval instance. The identifier can be a unique, or at least quasi-unique, identifier. After operation 2802, the method 2800 can receive notice of a submission of a workload (WorkSubmit) for the frame of content on a co-processor, where the notice includes the identifier for the work interval instance, as further shown by method 3000 of
The method 2800 additionally includes operation 2804 to receive notice of the beginning of a workload for the frame on the co-processor. The received notice can be associated with a WorkBegin call to the CLPC via the I/O system and can include a token to identify the workload. The token that identifies the workload which can be a token issued in response to a WorkSubmit call for the workload. In one embodiment, operation 2804 is performed by the CLPC. The CLPC, or another operating system component, can associate the token with the identifier for the work interval instance for the frame during operation 2806. In one embodiment, the CLPC can be configured to associate a token for a workload with the identifier for the most recently created work interval instance. In one embodiment the association between the token for the workload and the identifier for the work interval instance can be stored in a data structure of the CLPC. In one embodiment, the token for the workload can include a field to which a work interval instance identifier can be stored.
The method 2800 additionally includes to propagate the identifier for the work interval instance via the association with the token that identifies the workload, as shown at block 2808. Based on the association between the workload token and the work interval instance identifier, metrics that are tracked using the workload token can be associated with the work interval instance for which the workload is performed.
Automatic work interval instance identifier propagation as enabled by method 2800 can be more effective for some types of workloads relative to other types of workloads, depending on the pattern of parallelism employed by the thread groups that process the workload. In one embodiment, method 2800 can be employed to determine and propagate a work interval instance identifier in the event that no other work instance information is otherwise available for a workload.
For example, in one embodiment the method 2900 includes an operation 2902 to store an identifier for a work interval instance in a voucher containing a reference to a thread group. The identifier for the work interval instance can be returned from a call to an API to create the work interval instance. The method 2900 additionally includes an operation 2904 to send the voucher to a thread that will perform work on behalf of the thread group. The method 2900 can include an operation 2906 to associate a token that identifies the workload with the voucher containing the reference to the thread group.
The method 2900 additionally includes to propagate the identifier for the work interval instance via the association with the voucher that identifies the thread group and the token that identifies the workload, as shown at block 2908. Based on the association between the workload token and the voucher, metrics that are tracked using the workload token can be associated with the work interval instance for which the workload is performed by retrieving the identifier for the work interval instance from the voucher.
In one embodiment the method 3000 includes to perform an operation 3002 to create a work interval instance for a frame of content to be created, where the work interval instance has a unique (or at least quasi-unique) identifier. In one embodiment, a thread of the co-processor driver, or another user-space thread associated with the co-processor driver, can receive a work interval instance identifier in response to the API call to create the work interval instance. The method 3000 additionally includes to perform an operation 3004 to propagate the work interval instance identifier through the program logic through to the program code responsible for the WorkSubmit call. The method 3000 additionally includes to perform an operation 3006 to call a WorkSubmit interface with the work interval instance. The CLPC can receive the WorkSubmit call and can embed the work interval instance identification with the token generated in response to the WorkSubmit call. The method 3000 additionally includes to perform an operation 3008 to receive a token for the workload in response to the WorkSubmit call. The token can include an embedded instance identifier for the work interval instance.
The memory interface 3102 can be coupled to memory 3150, which can include high-speed random-access memory such as static random-access memory (SRAM) or dynamic random-access memory (DRAM) and/or non-volatile memory, such as but not limited to flash memory (e.g., NAND flash, NOR flash, etc.).
Sensors, devices, and subsystems can be coupled to the peripherals interface 3106 to facilitate multiple functionalities. For example, a motion sensor 3110, a light sensor 3112, and a proximity sensor 3114 can be coupled to the peripherals interface 3106 to facilitate the mobile device functionality. Other sensors 3116 can also be connected to the peripherals interface 3106, such as a positioning system (e.g., GPS receiver), a temperature sensor, a biometric sensor, or other sensing device, to facilitate related functionalities. A camera subsystem 3120 and an optical sensor 3122, e.g., a charged coupled device (CCD) or a complementary metal-oxide semiconductor (CMOS) optical sensor, can be utilized to facilitate camera functions, such as recording photographs and video clips.
Communication functions can be facilitated through one or more wireless communication subsystems 3124, which can include radio frequency receivers and transmitters and/or optical (e.g., infrared) receivers and transmitters. The specific design and implementation of the wireless communication subsystems 3124 can depend on the communication network(s) over which a mobile device is intended to operate. For example, a mobile device including the illustrated mobile device architecture 3100 can include wireless communication subsystems 3124 designed to operate over a GSM network, a CDMA network, an LTE network, a Wi-Fi network, a Bluetooth network, or any other wireless network. In particular, the wireless communication subsystems 3124 can provide a communications mechanism over which a client browser application can retrieve resources from a remote web server.
An audio subsystem 3126 can be coupled to a speaker 3128 and a microphone 3130 to facilitate voice-enabled functions, such as voice recognition, voice replication, digital recording, and telephony functions.
The I/O subsystem 3140 can include a touch screen controller 3142 and/or other input controller(s) 3145. The touch screen controller 3142 can be coupled to a touch sensitive display system 3146 (e.g., touch screen). The touch sensitive display system 3146 and touch screen controller 3142 can, for example, detect contact and movement and/or pressure using any of a plurality of touch and pressure sensing technologies, including but not limited to capacitive, resistive, infrared, and surface acoustic wave technologies, as well as other proximity sensor arrays or other elements for determining one or more points of contact with a touch sensitive display system 3146. Display output for the touch sensitive display system 3146 can be generated by a display controller 3143. In one embodiment the display controller 3143 can provide frame data to the touch sensitive display system 3146 at a variable frame rate.
In one embodiment a sensor controller 3144 is included to monitor, control, and/or processes data received from one or more of the motion sensor 3110, light sensor 3112, proximity sensor 3114, or other sensors 3116. The sensor controller 3144 can include logic to interpret sensor data to determine the occurrence of one of more motion events or activities by analysis of the sensor data from the sensors.
In one embodiment the I/O subsystem 3140 includes other input controller(s) 3145 that can be coupled to other input/control devices 3148, such as one or more buttons, rocker switches, thumb-wheel, infrared port, USB port, and/or a pointer device such as a stylus, or control devices such as an up/down button for volume control of the speaker 3128 and/or the microphone 3130.
In one embodiment, the memory 3150 coupled to the memory interface 3102 can store instructions for an operating system 3152, including portable operating system interface (POSIX) compliant and non-compliant operating system or an embedded operating system. The operating system 3152 may include instructions for handling basic system services and for performing hardware dependent tasks. In some implementations, the operating system 3152 can be a kernel.
The memory 3150 can also store communication instructions 3154 to facilitate communicating with one or more additional devices, one or more computers and/or one or more servers, for example, to retrieve web resources from remote web servers. The memory 3150 can also include user interface instructions 3156, including graphical user interface instructions to facilitate graphic user interface processing.
Additionally, the memory 3150 can store sensor processing instructions 3158 to facilitate sensor-related processing and functions; telephony instructions 3160 to facilitate telephone-related processes and functions; messaging instructions 3162 to facilitate electronic-messaging related processes and functions; web browser instructions 3164 to facilitate web browsing-related processes and functions; media processing instructions 3166 to facilitate media processing-related processes and functions; location services instructions including GPS and/or navigation instructions 3168 and Wi-Fi based location instructions to facilitate location based functionality; camera instructions 3170 to facilitate camera-related processes and functions; and/or other software instructions 3172 to facilitate other processes and functions, e.g., security processes and functions, and processes and functions related to the systems. The memory 3150 may also store other software instructions such as web video instructions to facilitate web video-related processes and functions; and/or web shopping instructions to facilitate web shopping-related processes and functions. In some implementations, the media processing instructions 3166 are divided into audio processing instructions and video processing instructions to facilitate audio processing-related processes and functions and video processing-related processes and functions, respectively. A mobile equipment identifier, such as an International Mobile Equipment Identity (IMEI) 3174 or a similar hardware identifier can also be stored in memory 3150.
Each of the above identified instructions and applications can correspond to a set of instructions for performing one or more functions described above. These instructions need not be implemented as separate software programs, procedures, or modules. The memory 3150 can include additional instructions or fewer instructions. Furthermore, various functions may be implemented in hardware and/or in software, including in one or more signal processing and/or application specific integrated circuits.
The systems and methods described herein can control system performance using measurements of performance metrics of groups of threads to make joint decisions on scheduling of threads and dynamic voltage and frequency scaling (DVFS) state for one or more clusters of cores in a multiprocessing system having a plurality of core types and one or more core of each core type. The performance metrics are fed to a closed loop control system that produces an output that is used to jointly decide how fast a core is to run and on which core type the threads of a thread group are to run. A thread group comprises one or more threads that are grouped together based on one or more characteristics that are used to determine a common goal or purpose of the threads in the thread group. Objects of the systems and methods described herein include minimizing thread scheduling latency for performance workloads, ensuring that performance workloads consistently find a performance core, maximizing throughput for performance workloads, and ensuring that efficiency workloads always find an efficient core. Objects can further include ensuring that cores are not powered down when threads are enqueued for processing. An additional object is to offload performance workloads when performance cores are oversubscribed. Threads are systematically guided to cores of the correct type for the workload.
Systems and methods are disclosed for scheduling threads on a processor that has at least two different core types each having one or more cores of that type, and the scheduler can schedule threads between at least two different core types. Each of the core types can have a different dynamic voltage and frequency scaling (DVFS) state. The processor can be a symmetric multiprocessing system (SMP) having a plurality of cores of the same design, but a first cluster of cores is run in a first configuration and a second cluster of cores is run in a second configuration. The processor can be an asymmetric multiprocessing system (AMP) that has at least two different core types, such as one or more performance cores and one or more efficiency cores. Each core type can run at a plurality of selectable voltage and frequency settings and both core types can run simultaneously.
Threads of processes can be grouped together into a thread group of related threads. Thread groups, generally, are intended to group work together to achieve a common goal, purpose, or type of work. Based upon samples of execution metrics of threads in a thread group, a closed loop performance controller (CLPC) can generate a control effort that represents an amount of performance, or efficiency, that the thread group should receive to meet performance targets for the thread group. CLPC can determine, for each of a plurality of thread groups that have been active on a core of the processor, a control effort for the thread group. The control effort for each thread group can be mapped to a performance map that recommends a core type and dynamic voltage and frequency scaling (DVFS) state for the thread group based on the control effort computed for the thread group. From the active thread groups, a maximum DVFS state can be determined for all threads that were recommended for a core of a first type (e.g., a performance core). A maximum DVFS state can be determined for all active thread groups that were recommended for a second core type (e.g. efficiency core type). A scheduler can coordinate setting the DVFS state for each core type with the scheduling of threads in the active thread groups.
In an embodiment, control effort limiter (CEL) in CLPC can generate a control effort limit that can limit the control effort output from the CLPC. A control effort limit can be based upon a fast die temperature, a peak power consumed over a short period of time, an average power consumed over any of a plurality of power zones, a skin temperature of an exterior of a device, a count of hardware throttling events, or other control metrics.
Threads can initially be grouped into default thread groups, such as grouping together threads of an application, called a “coalition.” A coalition is a grouping of closely related tasks/processes of an application which consist of threads, e.g. a web browser and application extensions associated with the web browser. Tasks can be combined into a single coalition by a launch services daemon, application programming interface (API), or framework. A thread group state is associated with each coalition. Thread grouping can be explicitly defined by a manual construct of a group of threads. Some types of workloads, e.g., an audio processing service, have complete knowledge of their organizations and can group threads accordingly. Threads can implicitly join thread groups by taking advantage of an existing kernel interaction. For example, when threads communicate via inter-process communication, they can optionally pass a voucher that references a thread group for the recipient to join. When a first thread wakes a second thread to do work on behalf of the first thread, the second thread can adopt the thread group of the first thread. When a first thread makes a second thread runnable, the second thread can adopt the thread group of the first thread. Threads can be grouped according to a type of work that the threads will perform as may be indicated by a daemon that launched the threads. An explicit thread grouping can be specified in source code, compiled code, or code packaged into a deliverable for installation. An explicit thread grouping for a work interval object can be specified in source code, compiled code, or a deliverable package for installation.
A work interval object (WIO) is an object that is used to represent periodic work where each period has a deadline. The WIO possesses a token and a specified time interval for one instance of the work. The WIO can be associated with a thread group. The thread group can either be created specifically for the WIO, or the WIO can be associated with an existing thread group. Threads that work to achieve a common purpose, intended to be performed within the specified time interval, can join the thread group of the WIO. A work interval object can be created when a user space process requests the work interval object via a specific system call. The kernel of an operating system can create the work interval object currently and automatically with a system call. By way of example, a work interval object may be, e.g., an object that represents one or more threads that composite multiple sources of audio and encode the audio for playback, all within a fixed period of time (work interval). A thread performing work having the purpose of the work interval object can opt into the work interval object thread group. Work performed by the opting-in threads is charged to the work interval object. A thread that receives a voucher containing a reference to a thread group may be grouped with the thread group referenced in the voucher. A voucher is a collection of attributes in a message sent via inter-process communication (IPC) from a first thread, T1, to a second thread, T2. One of the attributes that thread T1 puts in the voucher is the thread group to which T1 currently belongs. A second thread, T2, receives the IPC message from T1. As part of message reception, the voucher sent my T1 is “redeemed” (or adopted). Redeeming a voucher means processing all the attributes associated with the voucher and placing a pointer to the voucher on the thread. Thread T2 can now be considered to be running in the context of the voucher received from thread T1. As part of the action of “redeeming” a voucher, a voucher manager in the operating system kernel notices that thread T1 sent a thread group attribute in the voucher to thread T2, and the automatically joins thread T2 to the thread group specified in the voucher. After thread T2 has finished processing the message from thread T1, thread T1 can choose to discard the voucher. Discarding (or “dropping”) the voucher causes thread T2 to leave thread T1's thread group and causes thread T2 to re-join its “home” thread group (the one associated with the task in which T2 is running). Before thread T2 drops the voucher, however, thread T2 can also pass along the same voucher to another thread/process thread T3. Because the voucher contains a thread group attribute which points to thread T1's thread group, when thread T3 receives the message from thread T2 and adopts/redeems the voucher, thread T3 joins T1's thread group. Thread T3 will then be a part of T1's thread group for as long as that voucher is adopted by thread T3. Both threads T2 and T3 will leave thread T1's thread group as soon as threads T2 and T3 drop the voucher. This can be repeated indefinitely. Because the kernel voucher manager controls the contents of a voucher, and vouchers are automatically sent as part of a basic inter-process communication mechanism, when two processes communicate the thread which handles/receives the message will automatically join the thread group of the process who sent the message. In an embodiment, the thread may be backdated to be grouped with the thread group referenced in the voucher, prior to being made runnable, e.g. at a time when the thread receives the voucher.
A scheduler for the processor can schedule threads of the thread groups in accordance with a recommended core type and recommended DVFS state for the thread group. The recommended DVFS for the core type may be increased by the CLPC finding the maximum DVFS state for the core type, over all of the thread groups that are active on the system. The recommended DVFS state may be limited by the control effort limiter (CEL), limiting the DVFS state and/or number of cores of a core type that will execute threads.
The core types can have an associated interrupt and interrupt timer that can be used to delay a thread from being executed under some conditions. For example, if a thread is eligible for running on a performance core, but there are no idle performance cores, then the thread may be scheduled for an efficiency core. A timer can be programmed for the thread and the thread can be delayed from running. If the timer expires, the interrupt is triggered, and the performance-eligible thread is run on an efficiency core. But, if a performance core becomes available for thread execution before the interrupt time expires, then the timer can be canceled, and the thread can be executed on a performance core.
In a first embodiment, a method of running threads on a system having a processor that has a first cluster of one or more cores and a second cluster of one or more cores can include grouping threads of one or more processes into one or more thread groups. A first thread of a first thread group can be run on a core of the first cluster type. A plurality of thread execution metrics of the first thread can be incorporated with a plurality of thread execution metrics of the first thread group. A first control effort can be determined for the first thread group, based at least in part on samples of the plurality of thread group metrics. A first, or second, core type can be recommended for the first thread group type, and a DVFS state for the first thread group can be recommended. In an embodiment, the control effort can be mapped to a performance map of core clusters and DVFS states that are characterized by control effort. In an embodiment, the processor can comprise an asymmetric multiprocessing system (AMP) having a first core type of performance cores (P-cores) and a second core type of efficient cores (E-cores). A thread from a second group of threads can be executed, and performance metrics associated with execution of the second thread can be incorporated with execution metrics of the second thread group. A second control effort can be determined for the second thread group based at least in part on samples of the thread execution metrics of the second thread group. In response to determining that the first recommended core cluster for the first thread group and the second recommended core cluster for the second thread group are the same core cluster, a maximum DVFS state can be computed as the maximum of the first recommended DVFS state and the second recommended DVFS state, and the DVFS state for the core cluster recommended for the first and second thread groups can be set to the computed maximum DVFS state. In response to determining that the first recommended core cluster for the first thread group and the second recommended core cluster for the second thread group are different core clusters, the DVFS state for the first recommended core cluster can be set to the first recommended DVFS state and the DVFS state for the second recommended core cluster can be set to the second recommended DVFS state. Thread execution metrics can be determined periodically or in response to an asynchronous event. Samples of thread group execution metrics can be inputs to tunable controllers of a CLPC for the thread group. Control effort can map to a recommended core type and DVFS state. If the control effort maps to an overlapping part of the performance map, wherein either a P-core or E-core could be recommended, then a scheduler or CLPC could determine whether to recommend P-cores or E-cores based at least in part on a type of the work that the threads perform. For example, graphics rendering may be recommended to P-cores. I/O operations may be recommended to E-cores. In an embodiment, a time that a thread group has been resident on a core type and/or DVFS state can be considered in determining a core type and/or DVFS state to recommend for the thread group. In an embodiment, wherein cores of a first type are oversubscribed, and cores of a second type are idle, a scheduler may schedule threads of a thread group that are eligible for the first core type for execution on the second core type. A core of the first type can detect that a thread eligible for execution on the first type of core is actually running on the second type of core, and a core of the first type is going idle (e.g., in the process of transitioning into an idle state). The thread can be run on the now-idle core of the first type.
In a second embodiment a method of grouping threads into thread groups is performed on a system comprising an asymmetric multi-core processor having a plurality of cores of at least two different core types. The method includes launching an application program comprising a plurality of processes and assigning each of the plurality of processes to a first thread group. In response to determining that a process in the plurality of processes causes a work interval object to be generated, threads of the process can be assigned to a second thread group, different from the first thread group. At least one thread from the first thread group is executed. A plurality of thread execution metrics of threads in the first thread group is determined, and a processor core type for threads of the first thread group is recommended to a scheduler, based at least in part on the thread execution metrics of the first thread group. In an embodiment, in response to determining that a thread in the plurality of threads causes a work interval object to be generated, the thread can be assigned to a second thread group, different from the first thread group. At least one thread from the second thread group is executed and a plurality of thread execution metrics of threads in the second group is determined. A processor core type for threads of the second thread group is recommended to the scheduler, based at least in part upon samples of the thread execution metrics of the second thread group. In an embodiment, the method can further include receiving an indication that a thread from the first thread group has opted-in to the second thread group, executing the thread that opted-in to the second thread group, and attributing to the second group a plurality of thread execution metrics of the thread from the first group that opted-in to the second thread group. In an embodiment, the method can also include receiving an indication that the thread from the first thread group that opted-in to the second thread group has opted-out of the second thread group, assigning the opted-out thread to the first thread group, executing the opted-out thread, and attributing to the first thread group a plurality of thread execution metrics of the thread that opted-out of the second thread group.
A method of grouping threads for execution on an AMP can include determining that the application indicates that a process in the plurality of processes is to be assigned to a third thread group different from the first and second thread groups, executing at least one thread from the third thread group, determining a plurality of thread execution metrics of threads in the third group, and recommending to the scheduler a processor core type for threads of the third thread group based at least in part upon samples of thread execution metrics of the third thread group.
A method of grouping threads for execution on an AMP can include determining a type of work that a process in the plurality of processes performs and assigning a thread of the process to a fourth thread group, executing at least one thread in the fourth thread group, determining a plurality of thread execution metrics of threads in the fourth group, and recommending to the scheduler a processor core type for threads of the fourth thread group based at least in part upon samples of thread execution metrics of the fourth thread group. The fourth thread group can include threads from a daemon.
A method of grouping threads for execution on an AMP can include, in response to determining that a thread from the first thread group has caused a thread in a fifth thread group to be woken up and/or made runnable by the scheduler, assigning the woken up/and or made runnable thread from the fifth thread group to the first thread group. The method can further include executing at least one thread, determining a plurality of thread execution metrics of threads in the first thread group, and recommending to the scheduler a processor core type for threads of the first thread group based at least in part upon samples of the thread execution metrics of the first thread group.
A method of grouping threads for execution on an AMP can include, in response to determining that a thread from the first thread group has received a voucher to perform work on behalf of another thread, the voucher indicating a sixth thread group, assigning the thread to the sixth thread group. The method can further include executing at least one thread from the sixth thread group, determining a plurality of thread execution metrics of threads of the sixth group, and recommending to the scheduler a processor core type for threads in the sixth thread group base at least in part upon thread execution metrics of the sixth thread group. The method can also include determining that the thread which received the voucher and performed work on behalf of another thread has completed performing the work on behalf of the other thread and assigning the thread back to the first thread group.
A method of grouping threads for execution on an AMP can include launching an application program comprising a plurality of processes, assigning each of the plurality of processes to a first thread group, launching, by a daemon, at least one thread of a process in the first thread group, assigning, by the daemon, the at least one thread to a second thread group associated with a second daemon associated with the second thread group, executing at least one thread of the second thread group, determining a plurality of thread execution metrics of threads in the second thread group, and recommending to a scheduler a processor core type for threads of the second thread group based at least in part on samples of the thread execution metrics of the second thread group.
In a third embodiment, a method of processing work using a work interval object on an AMP can include creating a work interval object associated with a first thread group having a first thread group identifier. One or more threads associated with the work interval object can perform work toward a common goal of the work interval object within an identified interval of time. The method can also include receiving a request from a thread of a process associated with a second thread group having a second thread group identifier to adopt the first thread group identifier, attributing metrics of work performed by the thread to the first thread group identifier, in response to determining that the thread has adopted the first thread group identifier, and determining a processor core type to run one or more threads having the first thread group identifier based at least in part upon samples of metrics of work associated with the first thread group identifier. The work interval object can be instantiated in response to a call to a daemon or an operating system service. In an embodiment, the work interval object can be instantiated in response to a call by an application to an application programming interface (API) or framework. The method can further include receiving a request associated with the thread to opt out of the first thread group and associating the thread with the second thread group. In an embodiment, the method can also include attributing metrics of work performed by the thread to the second thread group, and determining a processor type to run one more threads of the second thread group based at least in part upon samples of metrics of work associated with the second thread group identifier.
In a fourth embodiment, method of scheduling threads on a processor having a first cluster type having one or more cores and a second cluster type having one or more cores in a system that includes the processor, includes: determining a control effort limit for the processor and determining a DVFS state for at least one of the first cluster or second cluster of cores. The control effort limit can be used in conjunction with a power map that includes a plurality of DVFS states for each of the first and second cluster of core. The power map can be indexed by the control effort limit. A recommendation can be received that a thread scheduled for execution can be run on the at least one of the first or second cluster of cores at a DVFS state different than the DVFS state determined using the control effort limit. The DVFS state can be set for at least one of the first or second cluster of cores to the DVFS state determined using the control effort limit. In an embodiment, the method can further include determining that a total energy dissipated in the processor for a first predetermined period of time is greater than an instantaneous power target for the processor. The control effort limit for the processor is determined at least in part on the total energy dissipated in the processor for the first predetermined period. In an embodiment, the control effort limit can be reduced in response to determining one or more of: an average power consumed by the processor for a second period of time is greater than an average power target for the processor, or an average energy dissipated per instruction is greater than the predetermined efficiency threshold for a predetermined period of time. In an embodiment, the system can monitor a plurality of power zones, each having a low pass filter that includes a power target, a time constant, a power delta, and a filter algorithm. At a first time, the filtered power can be less than a power target for a power zone, less the power delta for the power zone. At a second time, that is later than the first time, the filtered power can exceed the target power for the power zone. In an embodiment, the system can continue to monitor and filter power for the power zone that exceeded its power target, less the power delta for the power zone. The control effort limit can continue to be reduced until the filtered power for the power zone tracks the target power for the power zone. In an embodiment, the control effort limit can be reduced in response to a measured or estimated current for the processor exceeding a current limit for a predetermined period of time, or a number of instructions blocked from execution exceeds a peak throttle rate target for the predetermined period of time, wherein a throttling that blocks the instructions is done by hardware. In an embodiment, a maximum control effort is determined based at least in part on a maximum fast die temperature limit for at least one of the first or second cluster of cores. In an embodiment, a core can be masked off from executing. In an embodiment, the rate at which cores can be masked off can be limited. The processor can comprise an AMP processor having a first core cluster of P-cords and a second core cluster of E-cores.
A processing system can include a processor having at least two core types and at least one core of each core type. The system can also have a temperature sensor for each core type. The system can also have a scheduler that includes a scheduling queue for each core type. A CLPC can determine a control effort for threads of a thread group and recommend a core type and a DVFS state for threads of the thread group. A closed loop thermal management system can determine a control effort limit based at least in part on the temperature sensor for each core type. The system can comprise a system on a chip (SoC). One or both of the temperature sensors can be virtual sensors.
In a fifth embodiment, a method performed on a computing system having an asymmetric multiprocessing system (AMP) with a plurality of performance cores (P-cores) and a plurality of efficient cores (E-cores), can include determining that a thread, recommended for processing by a PC, is scheduled for processing by a PC of the AMP, and in response to determining that no P-cores are available for processing the thread and that an E-core is available for processing the thread: configuring a timer of an interrupt controller to generate an interrupt after expiration of a predetermined delay value. The method can further include, in response to determining, before the timer expires, that a PC is available to process the thread, canceling the interrupt, and processing the thread by the available PC. In an embodiment, the method can also include, in response to the timer expiring before a PC becomes available to process the thread: generating an interrupt to wake up an E-core and processing the thread by the E-core. The method can additionally include determining that a PC has become available, configuring a timer of the interrupt controller to generate an immediate interrupt to wake up the PC, generating the interrupt to wake up the PC, and processing the thread by the available PC.
A method of scheduling threads on an AMP processor can include determining that the thread has been recommended for processing by an E-core and that the thread is made runnable by a scheduler. In response to determining that an E-core is idle, configuring a timer of the interrupt controller to generate an immediate interrupt to wake up the E-core, otherwise configuring the timer interrupt controller to generate a delayed interrupt. The method can further include, in response to the interrupt controller generating an interrupt, waking up the E-core, and processing the thread by the E-core. The method can also include determining that the thread has been recommended for processing by a PC and that the thread has again been made runnable by the scheduler. The method can include configuring the timer of the interrupt controller to generate an immediate interrupt to wake up a PC. In response to the interrupt controller generating an interrupt, the method can include waking up the PC and processing the thread by the PC.
In an embodiment, a non-transitory computer readable can store executable instructions, that when executed by a processing system, can perform any of the functionality described above.
In yet another embodiment, a processing system coupled to a memory programmed with executable instructions can, when the instructions are executed by the processing system, perform any of the functionality described above.
A sixth embodiment provides for a data processing system comprising a processor complex including a plurality of core types, the plurality of core types to execute threads of a thread group. The data processing system additionally includes a closed loop performance controller (CLPC) to monitor and control performance of the threads of the thread group and a co-processor to execute a workload offloaded by a thread of the thread group. The CLPC can be configured to measure performance of the workload on the co-processor and associate the performance of the workload with the thread group.
An additional embodiment provides for a non-transitory machine-readable medium that stores instructions. Such instructions, when executed by one or more processors, cause the one or more processors to perform operations. The operations comprise receiving a message indicating that a thread of a thread group executing on a first processor of the one or more processors is to offload a workload to a second processor, the message including an identifier of the thread. The operations can additionally comprise inferring membership of the thread in the thread group based on the identifier of the thread and issuing a token in response to the message, the token to associate the workload with the thread group, the workload performed by one or more threads executed on the second processor. The operations additionally comprise associating performance metrics of the one or more threads executed on the second processor with the thread group.
A seventh embodiment provides for a non-transitory machine-readable medium storing instructions which, when executed by one or more processors, cause the one or more processors to perform operations comprising creating a work interval instance associated with a work interval object for a frame of content to be created, the work interval instance having an identifier and receiving notice of a submission of a workload for the frame of content on a co-processor, the notice including the identifier of the work interval instance. The operations additionally include associating a token to identify the workload with the identifier for the work interval instance and propagating the identifier for the work interval instance via the association with the token that identifies the workload and associating metrics for the workload on the co-processor with the frame of content based on the identifier propagated for the work interval instance. The seventh embodiment can additionally include instructions to perform operations comprising creating a work interval object associated with a thread group executing on the one or more processors, the thread group having a thread group identifier, where one or more threads associated with the work interval object perform work toward a common purpose within an identified interval of time.
In a further embodiment, the operations additionally comprise performing operations to generate the frame of content via the one or more threads associated with the work interval object and receiving a complexity estimate for the frame of content from the one or more threads associated with the work interval object and adjusting a performance of the co-processor or the one or more processors based on the complexity for the frame of content. In one embodiment, the medium can include instructions to perform operations comprising creating multiple work interval instances for multiple frames of work interval content, generating the multiple frames of work interval content on the co-processor, tracking separate metrics for each of the multiple frames of work interval content based on identifiers for the multiple work interval instances, and adjusting a performance of the co-processor or the one or more processors based on the metrics for one or more of the multiple frames of work interval content. Adjusting the performance of the co-processor or the one or more processor can include adjusting a frequency, a count of active cores, a type of active cores, or a subset of the set of active cores, where adjusting the subset of the set of active cores includes switching a workload or thread group from a first subset of the set of active cores to a second subset of the set of active cores or deactivating a first subset of the set of active cores and activating a second subset of the set of active cores.
An additional embodiment provides for a computer implemented method comprising creating a work interval instance associated with a work interval object for a frame of content to be created, the work interval instance having an identifier, receiving notice of a submission of a workload for the frame of content on a co-processor, the notice including the identifier for the work interval instance, associating the token to identify the workload with the identifier for the work interval instance, propagating the identifier for the work interval instance via the association with the token that identifies the workload, and associating metrics for the workload on the co-processor with the frame of content based on the identifier propagated for the work interval instance.
An additional embodiment provides for a data processing system comprising a processor complex including a plurality of core types, the plurality of core types to execute threads of a thread group, a co-processor to execute a workload offloaded by a thread of the thread group, and a closed loop performance controller (CLPC) to monitor and control performance of the threads of the thread group, to measure performance of the workload on the co-processor and associate the performance of the workload with the thread group. The CLPC can be configured to receive notice of a beginning of a workload for a frame of content on a co-processor, the notice including the identifier for the work interval instance, associate the token to identify the workload with an identifier for a work interval instance associated with the frame of content on the co-processor, and associate metrics for the workload on the co-processor with the frame of content based on an identifier propagated for the work interval instance, the identifier propagated via an association with the token that identifies the workload. The CLPC can be additionally configured to create multiple work interval instances for multiple frames of scrolling content, generate the multiple frames of scrolling content on the co-processor, and track separate metrics for each of the multiple frames of scrolling content based on identifiers for the multiple work interval instances.
Some embodiments described herein can include one or more application programming interfaces (APIs) in an environment with calling program code interacting with other program code being called through the one or more interfaces. Various function calls, messages, or other types of invocations, which further may include various kinds of parameters, can be transferred via the APIs between the calling program and the code being called. In addition, an API may provide the calling program code the ability to use data types or classes defined in the API and implemented in the called program code.
In the foregoing specification, the invention has been described with reference to specific embodiments thereof. It will, however, be evident that various modifications and changes can be made thereto without departing from the broader spirit and scope of the invention. The specification and drawings are, accordingly, to be regarded in an illustrative rather than a restrictive sense. Other features of the present embodiments will be apparent from the accompanying drawings and from the detailed description above. Accordingly, the true scope of the embodiments will become apparent to the skilled practitioner upon a study of the drawings, specification, and following claims.
This application is a continuation of co-pending application Ser. No. 15/996,469, having the title, “SCHEDULER FOR AMP ARCHITECTURE USING A CLOSED LOOP PERFORMANCE AND THERMAL CONTROLLER,” to John G. Dorsey, filed Jun. 2, 2018, which claims priority of U.S. Provisional Patent Application No. 62/514,930 having the title “SCHEDULER FOR AMP ARCHITECTURE,” to Jeremy C. Andrus, filed Jun. 4, 2017 and is incorporated by reference in its entirety to the extent that it is consistent with this disclosure. The following U.S. patent applications are incorporated herein by reference, in their entirety, to the extent that they are consistent with this application: (1) U.S. patent application Ser. No. 15/275,213 having the title, “PROCESSOR UNIT EFFICIENCY CONTROL,” to John G. Dorsey, filed Sep. 23, 2016; (2) U.S. patent application Ser. No. 14/576,963 having the title, “METHOD AND APPARATUS FOR ACTIVITY BASED EXECUTION SCHEDULING,” to James Michael Magee, filed Dec. 19, 2014; and (3) U.S. patent application Ser. No. 14/732,266 having the title, “SCHEDULER AND CPU PERFORMANCE CONTROLLER COOPERATION,” to Russell A. Blaine filed Jun. 5, 2015.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5345567 | Hayden et al. | Sep 1994 | A |
| 5738860 | Schonfeldt | Apr 1998 | A |
| 5812844 | Jones et al. | Sep 1998 | A |
| 5974438 | Neufeld | Oct 1999 | A |
| 6021425 | Waldron, III et al. | Feb 2000 | A |
| 6329877 | Bowen et al. | Dec 2001 | B1 |
| 7770034 | Nanja | Aug 2010 | B2 |
| 7774626 | Fleming | Aug 2010 | B2 |
| 7921313 | Ghiasi | Apr 2011 | B2 |
| 8452999 | Barth | May 2013 | B2 |
| 8516461 | Bellows et al. | Aug 2013 | B2 |
| 8662943 | Conroy et al. | Mar 2014 | B2 |
| 8707314 | Gummaraju et al. | Apr 2014 | B2 |
| 8738860 | Griffin et al. | May 2014 | B1 |
| 8892931 | Kruglick | Nov 2014 | B2 |
| 9304573 | Lee | Apr 2016 | B2 |
| 9329877 | Donaldson et al. | May 2016 | B2 |
| 9552230 | Jeong | Jan 2017 | B2 |
| 9652027 | Sharda | May 2017 | B2 |
| 9665161 | Hsu et al. | May 2017 | B2 |
| 9916162 | Gupta | Mar 2018 | B2 |
| 9984430 | Surti et al. | May 2018 | B2 |
| 10187282 | Park et al. | Jan 2019 | B2 |
| 10339023 | Wang et al. | Jul 2019 | B2 |
| 10599481 | Pistol et al. | Mar 2020 | B2 |
| 20030088610 | Kohn et al. | May 2003 | A1 |
| 20050262270 | Latorre | Nov 2005 | A1 |
| 20070294694 | Jeter et al. | Dec 2007 | A1 |
| 20080098254 | Altevogt et al. | Apr 2008 | A1 |
| 20080163183 | Li | Jul 2008 | A1 |
| 20080263324 | Sutardja et al. | Oct 2008 | A1 |
| 20090049451 | Bates | Feb 2009 | A1 |
| 20090144744 | Gunnels | Jun 2009 | A1 |
| 20090271646 | Talwar et al. | Oct 2009 | A1 |
| 20100077185 | Gopalan et al. | Mar 2010 | A1 |
| 20100146316 | Carter | Jun 2010 | A1 |
| 20110022870 | McGrane et al. | Jan 2011 | A1 |
| 20110047401 | Werner | Feb 2011 | A1 |
| 20110088041 | Alameldeen et al. | Apr 2011 | A1 |
| 20110113269 | Park | May 2011 | A1 |
| 20110113270 | Carter | May 2011 | A1 |
| 20110246995 | Fedorova | Oct 2011 | A1 |
| 20110296407 | Bhandari et al. | Dec 2011 | A1 |
| 20120173895 | Kim et al. | Jul 2012 | A1 |
| 20120272086 | Anderson et al. | Oct 2012 | A1 |
| 20120284732 | Griglock et al. | Nov 2012 | A1 |
| 20120315966 | Spata | Dec 2012 | A1 |
| 20120317582 | Craik et al. | Dec 2012 | A1 |
| 20130042250 | Lim et al. | Feb 2013 | A1 |
| 20130060555 | Thomson et al. | Mar 2013 | A1 |
| 20130124885 | Davis et al. | May 2013 | A1 |
| 20130151879 | Thomson et al. | Jun 2013 | A1 |
| 20130268943 | Solihin | Oct 2013 | A1 |
| 20140013330 | Wang et al. | Jan 2014 | A1 |
| 20140068285 | Lee et al. | Mar 2014 | A1 |
| 20140091624 | Park et al. | Apr 2014 | A1 |
| 20140165074 | Creamer et al. | Jun 2014 | A1 |
| 20140189302 | Subbsreddy et al. | Jul 2014 | A1 |
| 20140201411 | Kumar | Jul 2014 | A1 |
| 20140237477 | Cadambi et al. | Aug 2014 | A1 |
| 20140277815 | Hanumaiah et al. | Sep 2014 | A1 |
| 20140282507 | Plondke et al. | Sep 2014 | A1 |
| 20140304538 | Chen | Oct 2014 | A1 |
| 20140310467 | Shalf et al. | Oct 2014 | A1 |
| 20140372782 | Breternitz et al. | Dec 2014 | A1 |
| 20150026495 | Jain | Jan 2015 | A1 |
| 20150058646 | Kim | Feb 2015 | A1 |
| 20150073920 | Pashkevich et al. | Mar 2015 | A1 |
| 20150095666 | Ananthakrishnan et al. | Apr 2015 | A1 |
| 20150113542 | Cadambi et al. | Apr 2015 | A1 |
| 20150121105 | Ahn et al. | Apr 2015 | A1 |
| 20150169035 | Allen-Ware et al. | Jun 2015 | A1 |
| 20150205644 | Ito | Jul 2015 | A1 |
| 20150347178 | Magee et al. | Dec 2015 | A1 |
| 20150347509 | Ahmed | Dec 2015 | A1 |
| 20150355700 | Pusukuri et al. | Dec 2015 | A1 |
| 20160041594 | Hanson et al. | Feb 2016 | A1 |
| 20160062447 | Hsu et al. | Mar 2016 | A1 |
| 20160085596 | Nakagawa | Mar 2016 | A1 |
| 20160139964 | Chen et al. | May 2016 | A1 |
| 20160147280 | Thomas et al. | May 2016 | A1 |
| 20160188380 | Eastep et al. | Jun 2016 | A1 |
| 20160225042 | Tran | Aug 2016 | A1 |
| 20160239074 | Lee et al. | Aug 2016 | A1 |
| 20160282930 | Ramachandran et al. | Sep 2016 | A1 |
| 20160327999 | Chang et al. | Nov 2016 | A1 |
| 20160342198 | Hsu et al. | Nov 2016 | A1 |
| 20160378471 | Lerzer et al. | Dec 2016 | A1 |
| 20170031429 | Miyazaki et al. | Feb 2017 | A1 |
| 20170031724 | Xue et al. | Feb 2017 | A1 |
| 20170097854 | Shah et al. | Apr 2017 | A1 |
| 20170131754 | Zobel et al. | May 2017 | A1 |
| 20170177413 | Wisniewski et al. | Jun 2017 | A1 |
| 20170262030 | Lee et al. | Sep 2017 | A1 |
| 20170269652 | Lee et al. | Sep 2017 | A1 |
| 20170357302 | Dorsey et al. | Dec 2017 | A1 |
| 20180074853 | Aoyama | Mar 2018 | A1 |
| 20180210531 | Shahneous Bari et al. | Jul 2018 | A1 |
| 20180293102 | Ray et al. | Oct 2018 | A1 |
| 20180349175 | Andrus et al. | Dec 2018 | A1 |
| 20180349176 | Andrus et al. | Dec 2018 | A1 |
| 20180349177 | Pistol et al. | Dec 2018 | A1 |
| 20180349182 | Andrus et al. | Dec 2018 | A1 |
| 20180349186 | Andrus et al. | Dec 2018 | A1 |
| 20180349191 | Dorsey et al. | Dec 2018 | A1 |
| 20190034238 | Dorsey et al. | Jan 2019 | A1 |
| 20190384348 | Srinivasan et al. | Dec 2019 | A1 |
| Number | Date | Country |
|---|---|---|
| 2990909 | Mar 2016 | EP |
| 2016115000 | Jul 2016 | WO |
| Entry |
|---|
| PCT/US2018/024807, “Notification of Transmittal of the International Preliminary Report on Patentability” dated Sep. 18, 2019, 21 pages. |
| International Search Report and Written Opinion for PCT/US2018/024807, dated Jun. 29, 2018, 21 pgs. |
| Herbert, S. et al., Analysis of Dynamic Voltage/Frequency Scaling in Chip-Multiprocessors, 2007, ACM, pp. 38-43. |
| Joao, Jose A., et al. “Utility-Based Acceleration of Multithreaded Applications on Asymmetric CMPs,” ISCA '13 Tel-Aviv, Israel, 2013, 12 pages. |
| PCT/US2018/024811 “International Search Report and Written Opinion” dated Jul. 11, 2018, 13 pgs. |
| PCT/US2018/024813 “International Search Report and Written Opinion” dated Jul. 11, 2018, 15 pgs. |
| PCT/US2018/024814 “International Search Report and Written Opinion” dated Jul. 11, 2018, 16 pgs. |
| Huang et al., “Grouping-Based Dynamic Power Management for Multi-Threaded Programs in Chip-Multiprocessors”, 2009, IEEE, pp. 56-63. (Year: 2009). |
| Number | Date | Country | |
|---|---|---|---|
| 20190034238 A1 | Jan 2019 | US |
| Number | Date | Country | |
|---|---|---|---|
| 62514930 | Jun 2017 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 15996469 | Jun 2018 | US |
| Child | 16147077 | US |
