Patent Application
20050138422
The following commonly assigned co-pending applications have some subject matter in common with the current application:
Attorney Docket Number RA-5311 “Authorization Key System for Selectively Controlling the Performance of a Data Processing System”, Ser. No. 09/676,162 filed Sep. 29, 2000, and which is incorporated herein by reference.
Attorney docket number RA-5639 entitled “System and Method for Scaling Performance of a Data Processing System” filed on even date herewith.
Attorney docket number TN301/USYS-0141 entitled “Method and System for Economic Valuation in Partitioned Computer Systems”, filed on even date herewith.
The current invention relates generally to data processing systems, and more particularly to methods and apparatus for selectively controlling the performance of data processing systems.
Many growing businesses are challenged with ensuring that their data processing systems keep pace with expanding demands. This is particularly true for rapidly growing e-commerce companies, but also applies to other companies as well.
Another challenge facing many businesses is that of predicting and handling the peak loads that will be required to keep up with the day-to-day operations. For example, if there is a delay in gathering year-end data, there may be little time to process the data before the results must be published or otherwise released. The processing power required to handle such year-end data on such short notice may exceed the processing power of the available computer resources. In another example, e-commerce servers may experience severe peak loads during certain times of the year, such as the Christmas season. The extent of these peak loads is also often difficult to predict.
One way to increase processing power is to acquire additional processing systems. This can be expensive, and is not desirable if the additional systems are only required to address peak loads that exist during relatively short time periods. Another way to increase processing power is to modify existing systems. This may involve installing additional processors or memory, for example. However, system updates may necessitate the termination of normal processing activities so that the system can be powered down or otherwise placed in a state that accommodates maintenance. This can significantly disrupt the operations of the business. Moreover, updating a system to take into account peak demand is undesirable if this worst-case scenario rarely occurs.
One way to address the foregoing challenges involves allowing for the temporary increase of resources only when those resources are required to achieve a desired performance level. This is accomplished by including additional resources such as processors and memory in the data processing system when it is provided to the customer. However, only the resources that are required to achieve the performance purchased by the customer are enabled for use during normal operation. To temporarily or permanently increase the performance level of the data processing system, the customer may purchase an authorization key to enable the use of additional hardware resources. The authorization key may, for example, identify which additional processing resources are being authorized for use, the maximum time the additional resources are authorized for use, and an expiration date. This authorization key thereby allows selective increases in performance level to accommodate unplanned increases in performance requirements. When peak demand has ended, the customer may return to average processing levels without incurring the cost burden associated with permanently upgrading a system or obtaining additional systems.
Commonly-assigned U.S. patent application entitled “Authorization Key System for Selectively Controlling the Performance of a Data Processing System”, Ser. No. 09/676,162 filed Sep. 29, 2000, and which is incorporated herein by reference in its entirety, discloses an exemplary system of the type described in the foregoing paragraph. According to one embodiment of the disclosed system, the customer purchases a first authorization key that is delivered with the system. This key enables a first set of processing resources. If the customer later desires the option of enabling additional resources to increase system performance, a second authorization key may be purchased.
Prior art systems such as that described above generally select a performance level by identifying the system resources that will be enabled. For example, authorization keys provided with the system specifically identify the processors that are enabled for use. If one of these identified processors encounters some type of hardware problem, the customer is not allowed to instead employ one of the other available processors that is not specified by the key. Thus, encountered hardware problems may result in degraded throughput.
Another aspect of prior art systems involves the fact that authorization keys specify the number of processors that may be enabled within the system, not the processing power actually available from those processors. However, the processing power that will be obtained from a predetermined number of processors varies based on architectural characteristics of the data processing system. For example, four processors that are coupled to a shared cache may provide significantly more processing throughput than if two processors are operating from a first shared cache, while the remaining two processors utilize a different shared cache. Thus, the customer may not always be obtaining peak performance from the enabled resources.
An additional consideration associated with prior art systems relates to the use of multiple partitions within a data processing system. A partition is a grouping of resources that are allocated to execute in a cooperative manner to perform one or more assigned tasks. For example, a partition may be formed that includes one or more predetermined instruction processors and Input/Output Processors (IOPs), and a predetermined memory range within a shared main memory. A second partition may be created to include different processors, IOPs, and another memory range. Each of these partitions may operate independently from the other so that a number of tasks may be executed in parallel within the system. When system needs change, the partitions can be re-defined. For instance, if needed, all resources may be allocated to the same partition and assigned to execute a high-priority task.
Some prior art keys are “partitionable”, meaning these keys support the use of partitioning. Partitionable keys can be activated in a single partition, or in multiple partitions. For example, assume a partitionable key allows six identified processors to be enabled. These processors may be allocated to the same partition. Alternatively, two partitions may be created, each including three of the identified processors. When all six of the identified processors are in use, the operating system prevents the use of any more processors in any of the partitions.
Prior art partitionable keys do not account for system characteristics. For example, assume in the above example that three of the six identified processors share a first cache, and the remaining three processors share another cache. In this type of configuration, a single partition containing all processors will deliver less processing power than two partitions that each include a cache and the respective processors. This is true because of the loss of throughput that occurs when data must be shared between two caches of the same partition. Because the partitionable keys do not take into account such architectural considerations, the customer may not always be obtaining peak performance from the enabled resources. Additionally, since one partitioning configuration may provide more processing power than another configuration, the keys are difficult to price fairly.
What is needed, therefore, is an improved system and method for controlling and scaling the performance of a data processing system in a manner that addresses the foregoing issues.
The following invention provides an improved system and method for metering usage and scaling performance of a data processing system. In one embodiment, an authorization key is purchased that specifies both a baseline performance level and a ceiling performance level. These performance levels may be specified using a metric that describes processing throughput, such as Millions of Instructions Per Second (MIPS).
The cost of the baseline performance levels is included in the price of the key. After this key is installed on the system, the utilized performance of the data processing system is monitored and averaged over predetermined time periods. The customer is periodically issued an invoice that charges for any time period during which this averaged system utilized performance exceeds the baseline performance level. So long as the averaged system utilized performance remains below the pre-paid baseline performance level, the customer is not charged an additional amount. Performance of the data processing system is not allowed to exceed the ceiling level obtained with the authorization key.
In one embodiment, the data processing system can be configured in any selected one of multiple configurations. Each configuration is associated with a maximum performance level. The utilized performance of the data processing system is calculated based on the time spent performing work by each processor within the selected configuration. The utilized performance is also based on the maximum performance level for the selected configuration.
According to one aspect of the invention, the customer may programmably change the ceiling level, and is charged accordingly on their invoice. If desired, the baseline level may be set to zero by the system provider such that all performance consumption is purchased by the customer as it is used. In another embodiment, the ceiling level may be set to 100 percent so that system performance is not throttled.
According to one embodiment, baseline and ceiling performance levels are specified without any restrictions on the hardware that may be used to achieve these levels. The customer may therefore select which resources within the data processing system will be enabled, as well as how those resources will be configured.
The concept of using performance levels to scale and meter a data processing system can best be appreciated by example. Assume that a data processing system includes multiple IPs. A customer may choose to enable any or all of these IPs to achieve up to the purchased ceiling performance level. The performance of each of the IPs will automatically be scaled so that the overall performance of the data processing system does not exceed the purchased ceiling performance level.
In one embodiment, the customer may create one or more processing partitions to include the one or more enabled IPs. For instance, all enabled IPs may be included in the same partition, or may be divided between multiple partitions. Characteristics associated with the system architecture will be taken into account when scaling the performance of each IP in a partition so that the system as a whole will provide up to the purchased ceiling performance level.
According to the foregoing embodiment, performance of an IP may be scaled using a scaling factor that is derived using one or more lookup tables. These tables contain data indicating the peak performance level that will be provided by any allowable configuration of the data processing system. After the customer selects a configuration, the applicable scaling factor is calculated by dividing the purchased ceiling performance level by the peak performance level for the selected configuration. This scaling factor is then used to scale the processing power of each IP that is enabled within the configuration so that performance of the system does not exceed the ceiling level.
In a variation of the foregoing, a customer may select a configuration that includes multiple processing partitions. Each partition is allocated a portion of the total ceiling performance level. A ceiling scaling factor is created for each partition by dividing the portion of the allocated ceiling performance level by the peak performance level for that partition. The ceiling scaling factor is then used to scale the performance level of each IP within the partition.
In an embodiment wherein multiple processing partitions are utilized, the processing activities of each processor in each partition are monitored. The utilized performance of each of the partitions may then be calculated based on the portion of time spent by the processors in the partition performing work, as well as the maximum performance level that may be provided by the partition. The utilized performance of the system may then be derived by adding the utilized performance for each of the partitions. This system utilized performance is averaged over a predetermined averaging time period. The customer is billed based on this averaged system utilized performance. In one embodiment, the customer is billed based on an amount this averaged system utilized performance exceeds the baseline performance level. According to one aspect of the invention, the customer is billed for consumption, which is determined as a product of the averaged system utilized performance and the averaging period.
In one embodiment, the invention provides a method of metering performance of a data processing system. The method includes monitoring the performance of the data processing system, and charging a customer based on utilized performance that exceeds a baseline performance level.
According to another embodiment, a data processing system is provided that includes one or more processors, a memory coupled to the processors, and Software Controlled Performance Facility (SCPF) software stored within the memory to monitor performance of at least one of the processors, and to charge a customer based on the performance that is utilized.
In yet another embodiment, a system for charging for performance of a data processing system is disclosed. The data processing system includes one or more processors, means for recording performance of the one or more processors, and means for determining system utilized performance of the data processing system from the recorded performance of the one or more processors. The system further includes means for charging a customer based on the system utilized performance of the data processing system.
Other scopes and aspects of the invention will become apparent from the following description and the accompanying drawings.
The system further includes Processing mODules (PODs) 20A and 20B (shown dashed), which provides the processing capability for the system. A greater or lesser number of PODs may be included in the system than are shown in
Each of the PODs is coupled to each of the MSU devices via a dedicated, point-to-point connection referred to as an MSU Interface (MI), individually shown as MIs 30A through 30D. For example, MI 30A interfaces POD 20A to MSU device 10A, MI 30B interfaces POD 20A to MSU 10B device, and so on.
Each POD includes two Sub-Processing modules (Sub-PODs) and a crossbar module (XBAR). For example, POD 20A includes sub-PODs 50A and 50B and XBAR 60A, and so on. Each sub-POD is interconnected to the respective crossbar module (XBAR) through a dedicated point-to-point interface.
The system of
In the exemplary system of
The system of
Also shown residing with MSU 10 is at least one instance of a Software Controlled Performance Facility (SCPF) 90. The SCPF may be implemented in the kernel of OS 85 as shown in
The system of
System console provides all initialization, maintenance, and recovery operations for the system via the scan interface. In addition, system console may be employed by an operator to perform configuration activities in a manner to be discussed below.
Finally, the data processing system is coupled to a billing authority system 98 via a network 100 such as the Internet, or any other suitable type of network capable of supporting secure data transfers. This billing authority system is a data processing system that will generally be located at a remote location as compared to the data processing system, and will execute billing software 99. The billing software 99 utilizes data obtained from SCPF 90 to generate invoices charging the customer for utilization of the data processing system. This will be discussed below in reference to the remaining drawings.
It will be appreciated that the system of
The exemplary key of
Assume the authorization key illustrated in
In one embodiment, a corresponding configuration file (not shown) is provided to map the identifiers “IP0” and “IP1” specified in the authorization key to specific hardware within the system. For example, a configuration file may correlate the name “IP0” with IP 80A of sub-POD 50A by identifying a slot and chip location that is populated by IP 80A.
As discussed above, SCPF 90 is a software utility that is provided to control which IPs are enabled, as well as the peak utilization rate that is allowable for each of the enabled IPs. If the customer attempts to enable, or “up”, any of the processors other than IP0 and IP1, SCPF will issue a warning message and prevent the enabling of the identified IP.
The exemplary authorization key of
The forced idle state may be implemented using the multitasking capabilities of the system. As is known in the art, a multitasking environment allows an IP to execute multiple tasks, with each task being executed during a predetermined quantum of time. After the time for execution of a given task expires, the OS causes the IP to begin executing another task for the next quantum of time, and so on.
To facilitate multitasking, the OS must re-gain control of the IP at somewhat regular time intervals. This can be accomplished in a number of ways. Generally, a task that is executing on an IP periodically requests a service from the OS, thereby relinquishing control of the IP. This can be used as an opportunity to allow the OS to initiate execution of another task on the IP. Occasionally, however, a task may execute for long periods of time without relinquishing control to the OS. To prevent such execution from continuing for an extended period of time, the OS uses a quantum timer such as timer 81A to regain control of the IP. If a task continues execution beyond its allocated quantum of time, the quantum timer will expire to interrupt task execution. Control is returned to the OS so that another task can begin execution.
The foregoing environment may be utilized to scale performance of an IP as follows. When the OS gains control after task execution has been interrupted in any of the ways described above, SCPF 90 may, if necessary, force the IP to execute in a looping construct in which no useful work is done. The amount of time spent in the forced idle loop will be adjusted as necessary to cause the partition to run at a predetermined performance level specified by the system authorization key. SCPF monitors a system clock to cause the IP to execute within the idle loop until the predetermined scaled performance level is achieved. Preferably, the increments of time spent within a forced idle state are sufficiently small so as not to be discernable by a user. After the time required for execution within the forced idle loop has elapsed, the IP may be directed to resume execution of the next scheduled processing task. This will be discussed further below.
According to the current embodiment, SCPF 90 will prevent a customer from attempting to increase the utilization of the available processors beyond the authorized maximum utilization percentage, which is also referred to as “the ceiling”.
Next, assume the customer is experiencing a workload that cannot be adequately handled by the normal authorization key. To address this situation, the customer may purchase an optional authorization key.
An optional key is generally adopted for relatively short-term use as compared to a normal authorization key. In a manner similar to normal keys, this type of key may include an expiration date and/or a maximum usage time. For example, the key of
As can be appreciated by the foregoing, the use of an optional authorization key is particularly suited for a situation wherein a customer is experiencing a short-term workload increase. If it is anticipated that the increased workload will be sustained, the customer may purchase a normal authorization key that increases system performance for a longer time period.
The prior art system and method discussed above provides a flexible approach to increasing system performance. Performance can be increased without disrupting normal operations. Moreover, the performance increase may be tailored to a customer's specific needs. The customer is only required to purchase the amount of additional processing power for the limited time that processing power is needed. While this provides significant advantages, the flexibility of the prior art system may be improved. For example, prior art normal authorization keys specifically identify the IPs that are available for use. As a result, the customer does not have the discretion to disable one IP and instead employ a different IP, as may be desirable if a failure occurs within one of the executing IPs.
Another aspect of the prior art system involves the way in which the performance is specified. As previously discussed, a key describes the purchased processing power in terms of the number of processors that are available for use, and the percentage of utilization for each of the available processors. However, in some system configurations, these specifications do not necessarily accurately describe a predetermined performance level.
The foregoing observation can be appreciated by considering the optional authorization key of
Another concern associated with prior art systems involves the use of processing partitions. As discussed above, a partition is comprised of resources that are allocated to execute in a cooperative manner to perform one or more assigned tasks. For example, a partition may be created that includes one or more predetermined IPs and I/O modules, and a predetermined memory range within MSU 10. A second partition may be defined to include different IPs, I/O modules and another memory range. Each of these partitions may operate independently to execute respectively assigned tasks in parallel with those tasks being executed by other partitions. Partitions may be re-defined as system requirements change.
Some prior art keys are “partitionable”, meaning these keys support the use of partitioning. Partitionable keys can be activated in a single partition, or in multiple partitions. For example, assume a partitionable key authorizes the use of six identified IPs. All of these IPs may be allocated to the same partition. Alternatively, two partitions may be created, each including three of the identified processors.
Prior art partitionable keys do not take into account performance differences between various partitioning alternatives. For example, two partitions that each includes three IPs deliver considerably more processing power than a single partition that includes six IPs. Thus, it is difficult to price a partitionable key fairly.
Finally, prior art keys are rated in terms of a maximum performance level. A customer must pay for this maximum level during the entire time the key is used on the system, even if system usage only approaches the maximum level infrequently during that time.
The current invention provides an improved system and method for allowing the customer to pay for the processing power that is actually used, rather than requiring the customer to purchase an estimated maximum performance level ahead of time. In one embodiment, the inventive system measures processing power by specifying a performance level delivered by the system, rather than the number of processors that will deliver the processing power.
Because benchmark programs are generally developed with a particular system architecture and operating system in mind, a given suite of benchmarks do not necessarily provide data that can be used to conduct a meaningful comparison between different system architectures. However, a given suite of benchmarks can provide meaningful comparison data when considering the performance of systems that are included within the same or related product families.
The throughput of systems such as the ClearPath plus CS7802 system commercially available from Unisys Corporation is established using a suite of benchmark programs analyzed by International Data Corporation (IDC). These programs measure throughput in a unit of measure referred to as “IDC MIPS”, which hereinafter will just be referred to as “MIPS” for simplicity. Other types of MIPS may be used in the alternative.
Returning to
The current invention monitors the amount of processing power that is used by the customer in a manner to be discussed below. Processing power is considered to be “used” when it is being used to execute tasks, manage tasks, or schedule tasks for execution. Processing power is not being “used” when the processor is idle. The processor may be idle because there are no tasks in a state for execution (referred to herein as “natural idle”), or because performance of the processor is being scaled (referred to as “forced idle”). When the amount of utilized processing power exceeds the pre-paid baseline, the level of usage is recorded. The customer is periodically billed for this additional processing power. The customer's usage is not allowed to exceed the predetermined ceiling amount that is set based on pricing levels associated with the authorization key.
The foregoing can best be understood by returning to
As can be appreciated by
A review of the table of
A table such as that shown in
The table of
In the current example, assume that the customer chooses to deploy all eight IPs in the same partition. The customer may desire to employ this “eight-way” single partition configuration because it provides the best response times when multiple users are submitting requests within a transaction-processing environment. As illustrated by the fifth entry of
Assume next that the customer wants to change the system configuration. This may be desirable to transition from the transaction-processing environment discussed above to a batch mode environment wherein system-critical tasks are processed in a serial manner. By their nature, some of these system-critical tasks are single-threaded and must be executed consecutively. To better accommodate these types of tasks, five of the eight processors that were running in the partition are disabled, or “downed”. Only three IPs remain executing within the partition.
From the second entry of
It may be noted that in reality, processors are enabled and disabled individually, so the system's ceiling performance level changes incrementally as the transition from an 8-way to a 3-way configuration occurs, and vice versa.
The above example involves a ceiling level for a single partition. Similar considerations are employed when scaling the ceiling performance level in a scenario involving multiple partitions. For instance, assume that the customer of the current example wants to utilize the key having a 450 MIPS ceiling level for a system that is executing multiple partitions. Recall that the customer has a system similar to that of
Next, assume that the customer desires to run an important application in a first partition while the less critical applications execute in the other partition. To ensure that sufficient processing power is available for the critical task, the customer chooses to allocate 300 of the 450 MIPS to limit the ceiling of the first partition. In one embodiment, this type of allocation may be performed by an operator using a predetermined operations or administration screen available on system console 95. This type of allocation may be subject to limits imposed by the system administrator. Alternatively, the allocation may occur when the OS is booted and reads performance data from a predetermined location in main memory, as discussed below.
After a performance level has been allocated to a partition, scaling factors may be calculated. For example, assume that all four IPs in the first partition are enabled. The maximum rated performance of the first partition is therefore 630 MIPS, as shown in the third entry of
Next, the remaining 156 MIPS that have not been allocated to the first: partition are automatically allotted to the ceiling of the second partition. Assume that in this other partition, three IPs are enabled. This partition therefore has a maximum rated performance of 500 MIPS. SCPF 90 scales the performance level of the partition to 150/500, or 30 percent. Thus, the performance of each IP is scaled such that no IP achieves greater than 30 percent of its maximum processing potential.
As can be appreciated by the foregoing examples, the current system and method allows IP performance levels to vary between partitions. For instance, the IPs of the first partition may execute at up to 48 percent of their maximum performance level, whereas the IPs in the other partition may only execute up to 30 percent of their maximum capacity.
In one embodiment, in addition to varying the ceiling levels between partitions, it is also possible to varying the ceiling levels of IPs within the same partition. However, because the IPs of a given partition are operating on the same tasks and may be sharing and/or passing data through shared memory in MSU 10, processing power is generally most efficiently utilized by distributing it evenly between the IPs of the same partition. For this reason, in one embodiment, all IPs within the same partition are scaled by the same scaling factor.
According to one embodiment of the invention, warning messages may be provided if a partition cannot achieve a desired performance level. For example, if a user or an automated process attempts to set a ceiling at 600 MIPS for a partition having 3 IPs in one sub-POD, a warning message will be provided indicating the maximum processing power that may be achieved by this partition is 500 MIPS. In such situations, SCPF 90 will allow each IP in the partition to execute at 100 percent, and all remaining MIPS will be available for allocation to a ceiling of one or more other partitions.
According to another aspect of the system, a warning message will be issued if the performance level of an IP is scaled below a predetermined minimum value. This warning is provided because, in one embodiment, the forced idling mechanism does not operate in a predictable manner when an IP is scaled below a certain performance level. In addition to the warning, SCPF 90 will “down” one or more processors until the remaining processors are executing at, or above, the predetermined minimum processing level. For example, if the customer attempts to run eight IPs in a partition with a ceiling level of 20 MIPS, SCPF will continue downing IPs until the remaining IPs in the partition are running at a scaled performance level that is above the minimum level. This will allow the 20 MIPS to be predictably supported.
The authorization key is registered with the system using a software utility that tracks licensing data, shown as licensing registration software 410. This software verifies that the data stored within system id field 404 of the key matches identification data 412 provided with the system. Such system identification data may be stored within a read-only memory or some other storage device, may be hardwired on a back panel, manually or automatically selected using switches, or provided in any other suitable way.
Key information may also be copied to memory available to system control software 96, as illustrated by key data 414, such as memory within system console 95 of
Next, system control software 96 may be used to create one or more processing partitions. Specifically, an operator may employ maintenance screens provided by system control software to select the hardware that is to be added to a given partition. In response to this selection, system control software 96 employs scan interface 97 to enable and/or disable the appropriate hardware interfaces, including memory, cache, and processor interfaces. This electrically isolates the hardware of one partition from another, while allowing the various IPs, caches, and memories of the same partition to function as a unit. IPs that are included within a partition are enabled to communicate with their respective shared cache, whereas IPs that are not being used are electrically isolated from their respective shared cache and are not executing until such time as they are enabled. As discussed above, in one embodiment, the hardware of
After system control software 96 configures hardware and allocates one or more memory ranges within MSU 10 to a partition such as partition A 420, an instance of the OS 85 is booted within the allocated memory range(s). For example, an instance of the operating system, shown as OS A, 85A, is booted in partition A 420. In this embodiment, OS A includes as part of its kernel an instance of SCPF, shown as SCPF A, 90A. The partition also includes at least one IP 80A, which has a quantum timer 81A that is used to facilitate multitasking.
Sometime before or after the OS is booted, key information 418 including the maximum available performance level provided by the performance key is copied to partition A memory 416. Partition A memory is a range of memory within MSU that is directly accessible to partition A. OS A will read the key information from a known location within partition A memory to obtain the performance level provided by the registered key.
In one embodiment, the OS will, by default, attempt to obtain the entire performance level of the key. For example, if the key provides a maximum performance level of 450 MIPS, SCPF A included within OS A 85A will attempt to obtain the entire 450 MIPS. SCPF A then issues a message to system control software 96 indicating that the entire 450 MIPS has been obtained. If the entire 450 MIPS was available for allocation, system control software 96 updates the authorization key data 414 to record that partition A is executing at a performance level of 450 MIPS. OS A then notifies SCPF A that the performance level is allowable. SCPF A will thereafter scale performance of the IPs within the partition to achieve this performance level, as will be discussed further below.
In a manner similar to that described above, an operator may utilize system control software 96 to create an additional partition B 430. Memory within MSU 100 that is accessible to this partition is shown as partition B memory 421. Sometime before or after an instance of the OS is booted within partition B, key information 419 is copied to partition B memory. This key information includes the maximum available performance level 406 provided by the key.
An instance of the OS, shown as OS B, 85B, is booted in this additional partition B. OS B reads the key information 419 from partition B memory 421 and attempts to obtain all of the 450 MIPS provided by the key. SCPF B, 90B, issues a message to system control software 96 indicating that partition B is currently set to a performance level of 450 MIPS. System control software 96 utilizes key information 414 to determine that 450 MIPS have already been allocated to partition A 420. System control software returns an error message indicating that no MIPS are available for use, and partition B will be halted.
The foregoing discusses one embodiment wherein an OS always attempts to obtain all available MIPS provided by the authorization key upon completing the boot process. In this embodiment, some type of intervention is required to allow multiple partitions to be employed. For example, according to one aspect, OS A provides a display screen on system console 95 that is available to an operator for entering performance data. The operator may utilize this screen to send a message to SCPF A, 90A, indicating that partition A is to operate at something less than the entire 450 MIPS. Any time after OS A is booted, for instance, the operator may send a message to SCPF A indicating that partition A is to run at 200 MIPS. Upon receipt of this message, SCPF A stores this performance data within key information 418, and modifies performance of the partition accordingly. In addition, SCPF A sends a message to system control software 96 indicating the performance level of partition A has been modified, and system control software 96 updates the authorization key data 414 to reflect that partition A is now running at 200 MIPS.
Next, assume that partition B is created and OS B 85B attempts to obtain all 450 MIPS. OS B issues a message to system control software 96, which determines that only 250 MIPS of the 450 MIPS are available for use. System control software records that 250 MIPS are being allocated to partition B, and returns a message to OS B indicating that partition B must run at 250 MIPS. SCPF B records that partition B will execute at 250 MIPS, and thereafter scales performance of the IPs in the partition to achieve this performance level.
SCPF A has access to one or more performance data tables 440 stored within partition A memory 416. Similarly, SCPF B has access to one or more performance data tables 442 residing within partition B memory 421. These tables, which are similar to that shown in
System configurations and performance level allocations may be changed, as discussed above. For example, an operator may change the amount of processing power that is allocated to a given partition, so long as the total processing power used by all partitions does not exceed the maximum performance level specified by the registered key. Similarly, an operator may change the configuration of a partition by enabling or disabling IPs in a sub-POD included within the partition. When either of these events occurs, SCPF re-calculates the amount of time each IP must spend in a forced idle loop to achieve the allocated performance level for the partition.
The above examples describe one embodiment wherein the instance of SCPF included within an OS attempts to obtain all processing power provided by an authorization key when the OS is booted. In another embodiment, performance data may be stored with key information to cause SCPF to attempt to obtain a different performance level. For example, performance information may be stored within key information 418 of partition A memory 416 indicating that partition A should optimally obtain 200 MIPS. When OS A is booted, SCPF A will read this key information from memory 416, and will attempt to obtain the optimal performance level of 200 MIPS. A message will be issued to system control software 414, and system control software will determine whether this performance level is available for allocation in the manner discussed above. If this performance level is not available, system control software 96 will return a message to SCPF A that specifies the performance level that is available. SCPF A will set the performance of the partition to this available level.
In one embodiment, key information 418 will include the minimum performance level that is desired for optimal operation of the partition. This minimum performance level, which is configured by the customer, specifies the minimum level of performance that must be allocated to the partition to allow that partition to continue completing processing tasks in an optimal manner. For example, it may be beneficial to assign this type of minimum performance level to a partition that is performing high-priority work as a guarantee that enough processing power is available within the partition to allow the work to complete in a timely manner. If this type of minimum performance level has been assigned to a partition, and if system control software 96 returns a message to partition A indicating the performance level available to that partition is less than this assigned minimum performance level, the SCPF will issue periodic warning messages to the system operator. These messages will be provided to a display screen on system console 95 to warn the operator that the partition is not running at the optimal level.
The foregoing discussion describes situations wherein the authorization keys are registered before partitions are created and the OS instances are booted. In another embodiment, this is not a requirement. In a scenario wherein a key is registered after partition creation, the partition will stop executing if a key is not registered within a certain period of time thereafter. When the key is registered, key information is copied to memory accessible by each partition, such as key information 418 for partition A, and key information 419 for partition B. A message is issued to the SCPF of each partition, which will then set the performance level of its partition using the key information and information provided by system control software 96 in the manner discussed above.
The system of
After a partition is configured, SCPF 90C tracks processing time for each IP in a manner similar to that performed by SCPF 90A and 90B. In this embodiment, SCPF 90C controls the performance level of each partition by informing an SCPF agent included within the partition's OS to enforce the allocated performance level for the partition. This agent then scales the performance of each of the IPs in the partition appropriately. The manner in which IP performance is scaled is discussed further below.
The authorization key may be a normal key that is to be used for a long time period, or may be an optional key that is generally used for a short period of time. In one embodiment, the performance key may be delivered with the system. For example, the key may be registered on the system before the system is delivered. In another embodiment, the performance key may be provided to the customer after the system has been installed at the customer site. The performance key may be provided to the customer on a tape, disk, via an email transmission, or using any other suitable mechanism. The customer will register the key on the system and any system identification provided with the key will be verified during the registration process in the manner discussed above.
The customer may select a system configuration to use with the predetermined performance levels (504). This may be accomplished using system control software 96. In general, any one of multiple configurations will be available for use with the performance levels. The customer may select the desired configuration based on the type of processing tasks that will be executed on the data processing system, for example. At any time, the customer may re-select a new configuration based on changing conditions (506). These conditions may include the scheduling of system maintenance, the occurrence of unexpected failures, or a change in the type of processing tasks to be executed on the system. If desired, the configuration may be modified automatically. For example, this could be accomplished using functionality embodied within software executing on system console 95.
In one embodiment, a console program such as the Single Point of Operations console application commercially available from Unisys Corporation may be used to automatically select or modify the configuration. This type of automated configuration modification may occur at predetermined times of the day or based on monitored system conditions. For instance, one configuration may be selected for executing processing tasks in batch mode during the evening, whereas a second configuration may be selected to support a transaction—processing environment during the workday.
If an authorization key expires, or the time associated with the key is exhausted, the ceiling and baseline performance levels may transition to default values such as the performance levels that are specified by a different key (508). For example, if a long-term key has been registered with the system at the time a short-term key expires, the system may transition to performance levels specified by that long-term key. This transition may occur automatically under the control of SCPF 90 and system control software, or may be performed manually by the customer after a warning message has been issued regarding the termination of the optional key. In one embodiment, if another key has not been registered on the system when key expiration occurs, system execution will halt until the customer obtains another key.
The customer may also obtain a different key when performance requirements change (510). This key may be a long-term or short-term key, and may provide increased or decreased performance levels as compared to the previous key, depending on the changing requirements of the customer. Using this new key, the customer may select a system configuration to use with the predetermined performance level specified by the new key (504), and the process may be repeated.
According to the method, a ceiling, or “maximum available”, performance level is obtained by purchasing an authorization key (600). This performance level may be specified in MIPS or in some other unit of measure that is suitable for describing the performance of a data processing system. A partition is created having at least one enabled IP (600). Some or all of the ceiling performance level is allocated to the partition (602). The configuration of the partition, including the number and location of the enabled IPs, is then used to determine the maximum possible performance of the partition (604). This can be accomplished using one or more tables such as the table shown in
Ceiling scaling factor (allocated performance)/(maximum performance level of the partition).
The ceiling scaling factor is used to scale the maximum performance of the IPs within the partition, as discussed in regards to
Next, it may optionally be determined whether the scaling factor is below a predetermined minimum level (608). As discussed above, in some systems, accurate performance scaling cannot be performed if an IP is running below a predetermined minimum performance level. In one embodiment, the predetermined minimum level used during this verification step may be programmably selected.
If the scaling factor is below a predetermined minimum value, one or more IPs may be disabled within the partition (610). A new scaling factor is derived by again consulting the look-up table to determine the new maximum performance level of the partition, then calculating the new scaling value, as shown in steps 604 and 606. If the scaling factor is again below the predetermined minimum level (608), the process is repeated until the scaling factor exceeds the minimum value. The scaling factor may then be used to scale performance of each IP in the partition (614).
If any portion of the ceiling performance level specified by the authorization key is unused (616), the user may optionally create another partition having at least one IP (618). Some or all of the remaining ceiling performance level may be allocated to the additional partition, and the process may be repeated, as shown by arrow 622. In one embodiment, if the user creates a partition that is not assigned a specific ceiling performance level, SCPF 90 will automatically allocate all of the remaining ceiling performance level to that partition. The process of
The foregoing discusses the manner in which the ceiling performance level is allocated to partitions. The following describes how the allocated performance levels are utilized to throttle the processing of an IP, and to obtain a billing amount to provide to a customer.
First, two accumulators, Tf and Ti, are defined for the IP (702). Tf accumulates time that is spent in the forced idle state, and is used to keep the performance of the IP below the ceiling performance level. Ti accumulates all idle time for the IP, including forced idle and natural idle times, and is used in calculating the utilized performance of the partition, as will be described in
Next, a process referred to as a “dispatcher” is engaged. The dispatcher is a process that allocates the processing resources of an IP to tasks that are queued for execution, usually according to a priority mechanism. The details associated with task prioritization is beyond the scope of the invention. It is only important to appreciate that the dispatcher regards the processing of tasks as “work”, and the absence of eligible tasks as “natural idle”.
When the dispatcher is determining whether work is available to be processed, it must first check to make sure that the IP's performance level is being kept below the purchased ceiling performance level (706). To do this, the dispatcher determines whether the portion of time spent in a forced idle state thus far during the elapsed time period is less than that dictated by the ceiling scaling factor, as follows:
Tf/elasped time<1−ceiling scaling factor ?
If accumulator Tf indicates that not enough time has been spent in the forced idle state (705), the IP enters the forced idle state, repeatedly using the system clock to update Tf and Ti (708). The process periodically returns to step 706 to check Tf against the ceiling scaling factor, and the process is repeated until sufficient time has been spent in the forced idle state.
Once sufficient forced idle state Tf has been accumulated, the IP determines whether there is any work to be performed (710). If one or more tasks are awaiting execution, a task is selected and the task's execution environment is established. After a task is identified for execution, various IP registers must be initialized using environment data stored when the task was last executed by the IP. As is known in the art, this type of data may be stored within a stack entry or some other storage structure in memory. If this is not a trusted system task, the IP's quantum timer is initialized for use in interrupting the task, if necessary, to ensure that control will be returned to the OS after a maximum quantum of time allotted to the task has expired. The IP proceeds to execute the task's instructions. The task will be executed until it requires some service from the OS, the task completes, or an interrupt occurs. Such an interrupt may be received because of expiry of the quantum timer or the completion of an input/output (I/O) request (712). Any of these events may result in the same or other tasks being queued for execution. Eventually, the IP will return to the dispatcher, looking again for the highest priority work to process (706), and the sequence is repeated.
Returning to step 710, if there are no tasks awaiting execution, the IP enters the natural idle state (714). The time at which this state is entered is recorded for later use. While in this state, the IP executes an instruction sequence that has minimal effect on the efficiency of the rest of the system's components. The IP is able to detect hardware interrupts, such as the completion of an I/O request. The IP can also detect whether work is available for processing. Time spent in the idle state is not regarded as utilization of the IP.
If a hardware interrupt occurs, as indicated by arrow 716, the interrupt handler determines whether the IP had been in the natural idle state. If so, the time spent in this state is added to the accumulator Ti (718). The interrupt handler processes the interrupt, possibly queuing a task for execution.
Next, the IP returns to the dispatcher, as indicated by arrow 720. The IP will determine whether any forced idle is required to keep performance below the ceiling, before it selects any task for execution, and the process is repeated.
Returning to step 714, if the IP is in natural idle state and determines that work has become available for processing, as indicated by arrow 722, the time spent in the natural idle state is added to accumulator Ti (724). The IP returns to the dispatcher to determine whether any forced idle is required so that the performance level is maintained below the ceiling (706).
As mentioned above,
The foregoing process is described as being performed on a single IP. It will be understood that this process is likewise performed for each IP within a partition. That is, respective variables Ti, Tf, and elapsed_time are defined for each IP within the partition. Idle time is accumulated individually for an IP based on that IP's processing activities.
After defining the recording interval, Tr, the total recording time may be calculated as follows (742):
Total recording time for the partition=Tr×(number of processors in the partition).
These values are used to record the performance of a partition as follows.
Upon expiration of each Tr interval (744), the Ti accumulators for all IPs in the partition are added to get a total idle time for the partition (746). Next, accumulators for all of the IPs in the partition are cleared (748). The total utilized time for the partition is then calculated as follows (750):
utilized time for the partition=total recording time−total idle time.
The maximum performance of the partition, which is obtained from the configuration table in
Partition utilized performance=partition maximum performance×(utilized time for the partition/total recording time).
This provides the utilized performance for the partition (752). For example, if the configuration were rated at 300 MIPS and the partition was idle for 20% of the time, then utilized performance for the partition over the time Tr would be 300 MIPS×0.8, or 240 MIPS. The utilized performance for the partition is recorded and time-stamped for use, as discussed below in
In
Returning to the process of
Averaged system utilized performance=(Σ system utilized performance for each time period Tr in Ta)/(Number of time periods Tr in Ta)
The resulting averaged system utilized performance is expressed in MIPS or some other comparable unit of measure. This value may be used to calculate excess system utilized performance, which is defined as the amount the averaged system utilized performance exceeds the baseline performance level during Ta (806). If the averaged system utilized performance does not exceed the baseline performance level, the excess system utilized performance is recorded as being zero.
Next, metrics involving processing consumption are derived. Consumption is described in units of “Performance Level×Time”, such as “MIPS-Seconds”. As may be appreciated, consumption is derived by multiplying a utilized performance level by a period of time. To determine the system consumption during averaging time period Ta, the averaged system utilized performance is multiple by time Ta (808). Similarly, excess consumption during time Ta is derived by multiplying excess system utilized performance by time period Ta (810).
In the preferred embodiment, the customer is to be billed for excess consumption during any of the averaging periods. In other embodiments, the customer may be billed for average consumption rather than the peaks exceeding the baseline. In this case, it may be sufficient to monitor system consumption without regard to excess consumption.
One or both of the consumption calculations are recorded for later use (812). In one embodiment, these values are recorded in protected memory or some other storage device residing at the customer's location, the system provider's site, or any other suitable place that cannot be write-accessed by the customer. The manner in which these values are used is discussed below in reference to
The embodiment described in reference to
In the foregoing scenario wherein two keys are used on the same system, the metering process of
Next, the excess consumption values for all time periods Ta included within time period Th are added to obtain the total excess consumption during time period Tb (902). Optionally, system consumption for all time periods Ta may be added to obtain system performance for time period Tb. This excess consumption, and optionally, system consumption, are reported to the system provider's billing authority via some secure electronic or manual process (904). For example, it could be transferred over a network 100 (
Optionally, the customer may request a report on the excess consumption used thus far during a current billing period time. The customer will receive a report based on the sum of all excess consumption recordings since expiration of the last billing period (906). The customer can also review, but not alter, the detailed records that were used to generate the report.
The billing authority provides the customer with a bill for the excess consumption at some predetermined rate. This bill may be generated by billing software 99 (
The foregoing approach may be used to scale performance levels on a short-term, or a more long-term, basis. According to one embodiment, the customer is allowed to lower the ceiling performance level at any time during the billing period. This change may be programmably selected by the customer, or may be updated by the service provider upon customer request. The next billing period will include any necessary charges for the modified performance level. After activation, the new performance level is used to monitor system performance for billing purposes in the manner discussed above. After the ceiling performance level is lowered, the customer may choose to raise it to any level up to that which was originally specified by the key.
In a similar embodiment, the customer may also change baseline levels during the billing period. However, since baseline levels are pre-paid and recorded in the performance key, this would involve issuing the customer a new performance key and billing the customer at the time a baseline level is increased on a prorated basis that takes into account the time remaining in the billing cycle.
In one embodiment, the monitoring and averaging methods illustrated in
According to another embodiment, the monitoring method of
Although the preferred embodiment discussed above utilizes performance-based keys to implement the ceiling and baseline performance levels, this is not required. In another embodiment, the performance levels could be specified by providing information similar to that illustrated in the keys of
Having thus described the preferred embodiments of the present invention, those of skill in the art will readily appreciate that the teachings found herein may be applied to yet other embodiments. For example, the type of data processing system described and illustrated herein will be understood to be merely exemplary, and any other type of data processing system may be used in the alternative. Additionally, although the metering techniques described above are discussed in reference to instruction processors, they could be applied to I/O processors as well. In yet another embodiment, the model employed for charging customers could be modified. For example, in an alternative embodiment, customers may be charged for utilized performance at predetermined time increments, rather than being billed for consumption. Thus, the embodiments presented herein are to be considered exemplary only, and the scope of the invention is indicated only by the claims that follow rather than by the foregoing description.
