Accelerator devices, such as field programmable gate arrays (FPGAs), may be configured (e.g., by a bit stream defining a configuration of gates of the FPGA) to perform a set of functions, referred to herein as a kernel. The kernel may be configured, through time consuming effort on the part of a kernel developer, to establish a connection with another accelerator device kernel (e.g., another kernel in the same FPGA, a kernel in another FPGA, or a kernel of another type of accelerator device, such as a graphics processing unit (GPU), etc.) and share data with that other kernel to facilitate the completion of a workload (e.g., a set of operations that are to be performed). Typically, details of the communication protocol must be set out in the kernel by the kernel developer and, as the primary purpose of the kernel is to perform a set of operations (e.g., the workload) as efficiently as possible, the communication aspect of the kernel may be limited to communicating with a single type of accelerator device kernel if that accelerator device kernel happens to be available on a certain type of physical communication path (e.g., a PCIe bus). While some data centers may utilize pools of disaggregated resources (e.g., accelerator devices) available through various types of physical communication paths, also referred to I/O channels or communication channels, a given accelerator device kernel may be unable to access (e.g., cooperatively execute a workload with) a large percentage of those other accelerator devices present in the data center due to the limited communication faculties of the accelerator device kernel.
The concepts described herein are illustrated by way of example and not by way of limitation in the accompanying figures. For simplicity and clarity of illustration, elements illustrated in the figures are not necessarily drawn to scale. Where considered appropriate, reference labels have been repeated among the figures to indicate corresponding or analogous elements.
While the concepts of the present disclosure are susceptible to various modifications and alternative forms, specific embodiments thereof have been shown by way of example in the drawings and will be described herein in detail. It should be understood, however, that there is no intent to limit the concepts of the present disclosure to the particular forms disclosed, but on the contrary, the intention is to cover all modifications, equivalents, and alternatives consistent with the present disclosure and the appended claims.
References in the specification to “one embodiment,” “an embodiment,” “an illustrative embodiment,” etc., indicate that the embodiment described may include a particular feature, structure, or characteristic, but every embodiment may or may not necessarily include that particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the knowledge of one skilled in the art to effect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described. Additionally, it should be appreciated that items included in a list in the form of “at least one A, B, and C” can mean (A); (B); (C); (A and B); (A and C); (B and C); or (A, B, and C). Similarly, items listed in the form of “at least one of A, B, or C” can mean (A); (B); (C); (A and B); (A and C); (B and C); or (A, B, and C).
The disclosed embodiments may be implemented, in some cases, in hardware, firmware, software, or any combination thereof. The disclosed embodiments may also be implemented as instructions carried by or stored on a transitory or non-transitory machine-readable (e.g., computer-readable) storage medium, which may be read and executed by one or more processors. A machine-readable storage medium may be embodied as any storage device, mechanism, or other physical structure for storing or transmitting information in a form readable by a machine (e.g., a volatile or non-volatile memory, a media disc, or other media device).
In the drawings, some structural or method features may be shown in specific arrangements and/or orderings. However, it should be appreciated that such specific arrangements and/or orderings may not be required. Rather, in some embodiments, such features may be arranged in a different manner and/or order than shown in the illustrative figures. Additionally, the inclusion of a structural or method feature in a particular figure is not meant to imply that such feature is required in all embodiments and, in some embodiments, may not be included or may be combined with other features.
Referring now to
Ethernet. As described in more detail herein, resources within sleds in the data center 100 may be allocated to a group (referred to herein as a “managed node”) containing resources from one or more sleds to be collectively utilized in the execution of a workload. The workload can execute as if the resources belonging to the managed node were located on the same sled. The resources in a managed node may belong to sleds belonging to different racks, and even to different pods 110, 120, 130, 140. As such, some resources of a single sled may be allocated to one managed node while other resources of the same sled are allocated to a different managed node (e.g., one processor assigned to one managed node and another processor of the same sled assigned to a different managed node).
A data center comprising disaggregated resources, such as data center 100, can be used in a wide variety of contexts, such as enterprise, government, cloud service provider, and communications service provider (e.g., Telco's), as well in a wide variety of sizes, from cloud service provider mega-data centers that consume over 100,000 sq. ft. to single- or multi-rack installations for use in base stations.
The disaggregation of resources to sleds comprised predominantly of a single type of resource (e.g., compute sleds comprising primarily compute resources, memory sleds containing primarily memory resources), and the selective allocation and deallocation of the disaggregated resources to form a managed node assigned to execute a workload improves the operation and resource usage of the data center 100 relative to typical data centers comprised of hyperconverged servers containing compute, memory, storage and perhaps additional resources in a single chassis. For example, because sleds predominantly contain resources of a particular type, resources of a given type can be upgraded independently of other resources. Additionally, because different resources types (processors, storage, accelerators, etc.) typically have different refresh rates, greater resource utilization and reduced total cost of ownership may be achieved. For example, a data center operator can upgrade the processors throughout their facility by only swapping out the compute sleds. In such a case, accelerator and storage resources may not be contemporaneously upgraded and, rather, may be allowed to continue operating until those resources are scheduled for their own refresh. Resource utilization may also increase. For example, if managed nodes are composed based on requirements of the workloads that will be running on them, resources within a node are more likely to be fully utilized. Such utilization may allow for more managed nodes to run in a data center with a given set of resources, or for a data center expected to run a given set of workloads, to be built using fewer resources.
Referring now to
It should be appreciated that each of the other pods 120, 130, 140 (as well as any additional pods of the data center 100) may be similarly structured as, and have components similar to, the pod 110 shown in and described in regard to
Referring now to
In the illustrative embodiments, each sled of the data center 100 is embodied as a chassis-less sled. That is, each sled has a chassis-less circuit board substrate on which physical resources (e.g., processors, memory, accelerators, storage, etc.) are mounted as discussed in more detail below. As such, the rack 240 is configured to receive the chassis-less sleds. For example, each pair 310 of elongated support arms 312 defines a sled slot 320 of the rack 240, which is configured to receive a corresponding chassis-less sled. To do so, each illustrative elongated support arm 312 includes a circuit board guide 330 configured to receive the chassis-less circuit board substrate of the sled. Each circuit board guide 330 is secured to, or otherwise mounted to, a top side 332 of the corresponding elongated support arm 312. For example, in the illustrative embodiment, each circuit board guide 330 is mounted at a distal end of the corresponding elongated support arm 312 relative to the corresponding elongated support post 302, 304. For clarity of the Figures, not every circuit board guide 330 may be referenced in each Figure.
Each circuit board guide 330 includes an inner wall that defines a circuit board slot 380 configured to receive the chassis-less circuit board substrate of a sled 400 when the sled 400 is received in the corresponding sled slot 320 of the rack 240. To do so, as shown in
It should be appreciated that each circuit board guide 330 is dual sided. That is, each circuit board guide 330 includes an inner wall that defines a circuit board slot 380 on each side of the circuit board guide 330. In this way, each circuit board guide 330 can support a chassis-less circuit board substrate on either side. As such, a single additional elongated support post may be added to the rack 240 to turn the rack 240 into a two-rack solution that can hold twice as many sled slots 320 as shown in
In some embodiments, various interconnects may be routed upwardly or downwardly through the elongated support posts 302, 304. To facilitate such routing, each elongated support post 302, 304 includes an inner wall that defines an inner chamber in which interconnects may be located. The interconnects routed through the elongated support posts 302, 304 may be embodied as any type of interconnects including, but not limited to, data or communication interconnects to provide communication connections to each sled slot 320, power interconnects to provide power to each sled slot 320, and/or other types of interconnects.
The rack 240, in the illustrative embodiment, includes a support platform on which a corresponding optical data connector (not shown) is mounted. Each optical data connector is associated with a corresponding sled slot 320 and is configured to mate with an optical data connector of a corresponding sled 400 when the sled 400 is received in the corresponding sled slot 320. In some embodiments, optical connections between components (e.g., sleds, racks, and switches) in the data center 100 are made with a blind mate optical connection. For example, a door on each cable may prevent dust from contaminating the fiber inside the cable. In the process of connecting to a blind mate optical connector mechanism, the door is pushed open when the end of the cable approaches or enters the connector mechanism. Subsequently, the optical fiber inside the cable may enter a gel within the connector mechanism and the optical fiber of one cable comes into contact with the optical fiber of another cable within the gel inside the connector mechanism.
The illustrative rack 240 also includes a fan array 370 coupled to the cross-support arms of the rack 240. The fan array 370 includes one or more rows of cooling fans 372, which are aligned in a horizontal line between the elongated support posts 302, 304. In the illustrative embodiment, the fan array 370 includes a row of cooling fans 372 for each sled slot 320 of the rack 240. As discussed above, each sled 400 does not include any on-board cooling system in the illustrative embodiment and, as such, the fan array 370 provides cooling for each sled 400 received in the rack 240. Each rack 240, in the illustrative embodiment, also includes a power supply associated with each sled slot 320. Each power supply is secured to one of the elongated support arms 312 of the pair 310 of elongated support arms 312 that define the corresponding sled slot 320. For example, the rack 240 may include a power supply coupled or secured to each elongated support arm 312 extending from the elongated support post 302. Each power supply includes a power connector configured to mate with a power connector of the sled 400 when the sled 400 is received in the corresponding sled slot 320. In the illustrative embodiment, the sled 400 does not include any on-board power supply and, as such, the power supplies provided in the rack 240 supply power to corresponding sleds 400 when mounted to the rack 240. Each power supply is configured to satisfy the power requirements for its associated sled, which can vary from sled to sled. Additionally, the power supplies provided in the rack 240 can operate independent of each other. That is, within a single rack, a first power supply providing power to a compute sled can provide power levels that are different than power levels supplied by a second power supply providing power to an accelerator sled. The power supplies may be controllable at the sled level or rack level, and may be controlled locally by components on the associated sled or remotely, such as by another sled or an orchestrator.
Referring now to
As discussed above, the illustrative sled 400 includes a chassis-less circuit board substrate 602, which supports various physical resources (e.g., electrical components) mounted thereon. It should be appreciated that the circuit board substrate 602 is “chassis-less” in that the sled 400 does not include a housing or enclosure. Rather, the chassis-less circuit board substrate 602 is open to the local environment. The chassis-less circuit board substrate 602 may be formed from any material capable of supporting the various electrical components mounted thereon. For example, in an illustrative embodiment, the chassis-less circuit board substrate 602 is formed from an FR-4 glass-reinforced epoxy laminate material. Of course, other materials may be used to form the chassis-less circuit board substrate 602 in other embodiments.
As discussed in more detail below, the chassis-less circuit board substrate 602 includes multiple features that improve the thermal cooling characteristics of the various electrical components mounted on the chassis-less circuit board substrate 602. As discussed, the chassis-less circuit board substrate 602 does not include a housing or enclosure, which may improve the airflow over the electrical components of the sled 400 by reducing those structures that may inhibit air flow. For example, because the chassis-less circuit board substrate 602 is not positioned in an individual housing or enclosure, there is no vertically-arranged backplane (e.g., a backplate of the chassis) attached to the chassis-less circuit board substrate 602, which could inhibit air flow across the electrical components. Additionally, the chassis-less circuit board substrate 602 has a geometric shape configured to reduce the length of the airflow path across the electrical components mounted to the chassis-less circuit board substrate 602. For example, the illustrative chassis-less circuit board substrate 602 has a width 604 that is greater than a depth 606 of the chassis-less circuit board substrate 602. In one particular embodiment, for example, the chassis-less circuit board substrate 602 has a width of about 21 inches and a depth of about 9 inches, compared to a typical server that has a width of about 17 inches and a depth of about 39 inches. As such, an airflow path 608 that extends from a front edge 610 of the chassis-less circuit board substrate 602 toward a rear edge 612 has a shorter distance relative to typical servers, which may improve the thermal cooling characteristics of the sled 400. Furthermore, although not illustrated in
As discussed above, the illustrative sled 400 includes one or more physical resources 620 mounted to a top side 650 of the chassis-less circuit board substrate 602. Although two physical resources 620 are shown in
The sled 400 also includes one or more additional physical resources 630 mounted to the top side 650 of the chassis-less circuit board substrate 602. In the illustrative embodiment, the additional physical resources include a network interface controller (NIC) as discussed in more detail below. Of course, depending on the type and functionality of the sled 400, the physical resources 630 may include additional or other electrical components, circuits, and/or devices in other embodiments.
The physical resources 620 are communicatively coupled to the physical resources 630 via an input/output (I/O) subsystem 622. The I/O subsystem 622 may be embodied as circuitry and/or components to facilitate input/output operations with the physical resources 620, the physical resources 630, and/or other components of the sled 400. For example, the I/O subsystem 622 may be embodied as, or otherwise include, memory controller hubs, input/output control hubs, integrated sensor hubs, firmware devices, communication links (e.g., point-to-point links, bus links, wires, cables, waveguides, light guides, printed circuit board traces, etc.), and/or other components and subsystems to facilitate the input/output operations. In the illustrative embodiment, the I/O subsystem 622 is embodied as, or otherwise includes, a double data rate 4 (DDR4) data bus or a DDRS data bus.
In some embodiments, the sled 400 may also include a resource-to-resource interconnect 624. The resource-to-resource interconnect 624 may be embodied as any type of communication interconnect capable of facilitating resource-to-resource communications. In the illustrative embodiment, the resource-to-resource interconnect 624 is embodied as a high-speed point-to-point interconnect (e.g., faster than the I/O subsystem 622). For example, the resource-to-resource interconnect 624 may be embodied as a QuickPath Interconnect (QPI), an UltraPath Interconnect (UPI), or other high-speed point-to-point interconnect dedicated to resource-to-resource communications.
The sled 400 also includes a power connector 640 configured to mate with a corresponding power connector of the rack 240 when the sled 400 is mounted in the corresponding rack 240. The sled 400 receives power from a power supply of the rack 240 via the power connector 640 to supply power to the various electrical components of the sled 400. That is, the sled 400 does not include any local power supply (i.e., an on-board power supply) to provide power to the electrical components of the sled 400. The exclusion of a local or on-board power supply facilitates the reduction in the overall footprint of the chassis-less circuit board substrate 602, which may increase the thermal cooling characteristics of the various electrical components mounted on the chassis-less circuit board substrate 602 as discussed above. In some embodiments, voltage regulators are placed on a bottom side 750 (see
In some embodiments, the sled 400 may also include mounting features 642 configured to mate with a mounting arm, or other structure, of a robot to facilitate the placement of the sled 600 in a rack 240 by the robot. The mounting features 642 may be embodied as any type of physical structures that allow the robot to grasp the sled 400 without damaging the chassis-less circuit board substrate 602 or the electrical components mounted thereto. For example, in some embodiments, the mounting features 642 may be embodied as non-conductive pads attached to the chassis-less circuit board substrate 602. In other embodiments, the mounting features may be embodied as brackets, braces, or other similar structures attached to the chassis-less circuit board substrate 602. The particular number, shape, size, and/or make-up of the mounting feature 642 may depend on the design of the robot configured to manage the sled 400.
Referring now to
The memory devices 720 may be embodied as any type of memory device capable of storing data for the physical resources 620 during operation of the sled 400, such as any type of volatile (e.g., dynamic random access memory (DRAM), etc.) or non-volatile memory. Volatile memory may be a storage medium that requires power to maintain the state of data stored by the medium. Non-limiting examples of volatile memory may include various types of random access memory (RAM), such as dynamic random access memory (DRAM) or static random access memory (SRAM). One particular type of DRAM that may be used in a memory module is synchronous dynamic random access memory (SDRAM). In particular embodiments, DRAM of a memory component may comply with a standard promulgated by JEDEC, such as JESD79F for DDR SDRAM, JESD79-2F for DDR2 SDRAM, JESD79-3F for DDR3 SDRAM, JESD79-4A for DDR4 SDRAM, JESD209 for Low Power DDR (LPDDR), JESD209-2 for LPDDR2, JESD209-3 for LPDDR3, and JESD209-4 for LPDDR4. Such standards (and similar standards) may be referred to as DDR-based standards and communication interfaces of the storage devices that implement such standards may be referred to as DDR-based interfaces.
In one embodiment, the memory device is a block addressable memory device, such as those based on NAND or NOR technologies. A memory device may also include next-generation nonvolatile devices, such as Intel 3D XPoint™ memory or other byte addressable write-in-place nonvolatile memory devices. In one embodiment, the memory device may be or may include memory devices that use chalcogenide glass, multi-threshold level NAND flash memory, NOR flash memory, single or multi-level Phase Change Memory (PCM), a resistive memory, nanowire memory, ferroelectric transistor random access memory (FeTRAM), anti-ferroelectric memory, magnetoresistive random access memory (MRAM) memory that incorporates memristor technology, resistive memory including the metal oxide base, the oxygen vacancy base and the conductive bridge Random Access Memory (CB-RAM), or spin transfer torque (STT)-MRAM, a spintronic magnetic junction memory based device, a magnetic tunneling junction (MTJ) based device, a DW (Domain Wall) and SOT (Spin Orbit Transfer) based device, a thyristor based memory device, or a combination of any of the above, or other memory. The memory device may refer to the die itself and/or to a packaged memory product. In some embodiments, the memory device may comprise a transistor-less stackable cross point architecture in which memory cells sit at the intersection of word lines and bit lines and are individually addressable and in which bit storage is based on a change in bulk resistance.
Referring now to
In the illustrative compute sled 800, the physical resources 620 are embodied as processors 820. Although only two processors 820 are shown in
In some embodiments, the compute sled 800 may also include a processor-to-processor interconnect 842. Similar to the resource-to-resource interconnect 624 of the sled 400 discussed above, the processor-to-processor interconnect 842 may be embodied as any type of communication interconnect capable of facilitating processor-to-processor interconnect 842 communications. In the illustrative embodiment, the processor-to-processor interconnect 842 is embodied as a high-speed point-to-point interconnect (e.g., faster than the I/O subsystem 622). For example, the processor-to-processor interconnect 842 may be embodied as a QuickPath Interconnect (QPI), an UltraPath Interconnect (UPI), or other high-speed point-to-point interconnect dedicated to processor-to-processor communications.
The compute sled 800 also includes a communication circuit 830. The illustrative communication circuit 830 includes a network interface controller (NIC) 832, which may also be referred to as a host fabric interface (HFI). The NIC 832 may be embodied as, or otherwise include, any type of integrated circuit, discrete circuits, controller chips, chipsets, add-in-boards, daughtercards, network interface cards, or other devices that may be used by the compute sled 800 to connect with another compute device (e.g., with other sleds 400). In some embodiments, the NIC 832 may be embodied as part of a system-on-a-chip (SoC) that includes one or more processors, or included on a multichip package that also contains one or more processors. In some embodiments, the NIC 832 may include a local processor (not shown) and/or a local memory (not shown) that are both local to the NIC 832. In such embodiments, the local processor of the NIC 832 may be capable of performing one or more of the functions of the processors 820. Additionally or alternatively, in such embodiments, the local memory of the NIC 832 may be integrated into one or more components of the compute sled at the board level, socket level, chip level, and/or other levels.
The communication circuit 830 is communicatively coupled to an optical data connector 834. The optical data connector 834 is configured to mate with a corresponding optical data connector of the rack 240 when the compute sled 800 is mounted in the rack 240. Illustratively, the optical data connector 834 includes a plurality of optical fibers which lead from a mating surface of the optical data connector 834 to an optical transceiver 836. The optical transceiver 836 is configured to convert incoming optical signals from the rack-side optical data connector to electrical signals and to convert electrical signals to outgoing optical signals to the rack-side optical data connector. Although shown as forming part of the optical data connector 834 in the illustrative embodiment, the optical transceiver 836 may form a portion of the communication circuit 830 in other embodiments.
In some embodiments, the compute sled 800 may also include an expansion connector 840. In such embodiments, the expansion connector 840 is configured to mate with a corresponding connector of an expansion chassis-less circuit board substrate to provide additional physical resources to the compute sled 800. The additional physical resources may be used, for example, by the processors 820 during operation of the compute sled 800. The expansion chassis-less circuit board substrate may be substantially similar to the chassis-less circuit board substrate 602 discussed above and may include various electrical components mounted thereto. The particular electrical components mounted to the expansion chassis-less circuit board substrate may depend on the intended functionality of the expansion chassis-less circuit board substrate. For example, the expansion chassis-less circuit board substrate may provide additional compute resources, memory resources, and/or storage resources. As such, the additional physical resources of the expansion chassis-less circuit board substrate may include, but is not limited to, processors, memory devices, storage devices, and/or accelerator circuits including, for example, field programmable gate arrays (FPGA), application-specific integrated circuits (ASICs), security co-processors, graphics processing units (GPUs), machine learning circuits, or other specialized processors, controllers, devices, and/or circuits.
Referring now to
As discussed above, the individual processors 820 and communication circuit 830 are mounted to the top side 650 of the chassis-less circuit board substrate 602 such that no two heat-producing, electrical components shadow each other. In the illustrative embodiment, the processors 820 and communication circuit 830 are mounted in corresponding locations on the top side 650 of the chassis-less circuit board substrate 602 such that no two of those physical resources are linearly in-line with others along the direction of the airflow path 608. It should be appreciated that, although the optical data connector 834 is in-line with the communication circuit 830, the optical data connector 834 produces no or nominal heat during operation.
The memory devices 720 of the compute sled 800 are mounted to the bottom side 750 of the of the chassis-less circuit board substrate 602 as discussed above in regard to the sled 400. Although mounted to the bottom side 750, the memory devices 720 are communicatively coupled to the processors 820 located on the top side 650 via the I/O subsystem 622. Because the chassis-less circuit board substrate 602 is embodied as a double-sided circuit board, the memory devices 720 and the processors 820 may be communicatively coupled by one or more vias, connectors, or other mechanisms extending through the chassis-less circuit board substrate 602. Of course, each processor 820 may be communicatively coupled to a different set of one or more memory devices 720 in some embodiments. Alternatively, in other embodiments, each processor 820 may be communicatively coupled to each memory device 720. In some embodiments, the memory devices 720 may be mounted to one or more memory mezzanines on the bottom side of the chassis-less circuit board substrate 602 and may interconnect with a corresponding processor 820 through a ball-grid array.
Each of the processors 820 includes a heatsink 850 secured thereto. Due to the mounting of the memory devices 720 to the bottom side 750 of the chassis-less circuit board substrate 602 (as well as the vertical spacing of the sleds 400 in the corresponding rack 240), the top side 650 of the chassis-less circuit board substrate 602 includes additional “free” area or space that facilitates the use of heatsinks 850 having a larger size relative to traditional heatsinks used in typical servers. Additionally, due to the improved thermal cooling characteristics of the chassis-less circuit board substrate 602, none of the processor heatsinks 850 include cooling fans attached thereto. That is, each of the heatsinks 850 is embodied as a fan-less heatsink. In some embodiments, the heat sinks 850 mounted atop the processors 820 may overlap with the heat sink attached to the communication circuit 830 in the direction of the airflow path 608 due to their increased size, as illustratively suggested by
Referring now to
In the illustrative accelerator sled 1000, the physical resources 620 are embodied as accelerator circuits 1020. Although only two accelerator circuits 1020 are shown in
In some embodiments, the accelerator sled 1000 may also include an accelerator-to-accelerator interconnect 1042. Similar to the resource-to-resource interconnect 624 of the sled 600 discussed above, the accelerator-to-accelerator interconnect 1042 may be embodied as any type of communication interconnect capable of facilitating accelerator-to-accelerator communications. In the illustrative embodiment, the accelerator-to-accelerator interconnect 1042 is embodied as a high-speed point-to-point interconnect (e.g., faster than the I/O subsystem 622). For example, the accelerator-to-accelerator interconnect 1042 may be embodied as a QuickPath Interconnect (QPI), an UltraPath Interconnect (UPI), or other high-speed point-to-point interconnect dedicated to processor-to-processor communications. In some embodiments, the accelerator circuits 1020 may be daisy-chained with a primary accelerator circuit 1020 connected to the NIC 832 and memory 720 through the I/O subsystem 622 and a secondary accelerator circuit 1020 connected to the NIC 832 and memory 720 through a primary accelerator circuit 1020.
Referring now to
Referring now to
In the illustrative storage sled 1200, the physical resources 620 are embodied as storage controllers 1220. Although only two storage controllers 1220 are shown in
In some embodiments, the storage sled 1200 may also include a controller-to-controller interconnect 1242. Similar to the resource-to-resource interconnect 624 of the sled 400 discussed above, the controller-to-controller interconnect 1242 may be embodied as any type of communication interconnect capable of facilitating controller-to-controller communications. In the illustrative embodiment, the controller-to-controller interconnect 1242 is embodied as a high-speed point-to-point interconnect (e.g., faster than the I/O subsystem 622). For example, the controller-to-controller interconnect 1242 may be embodied as a QuickPath Interconnect (QPI), an UltraPath Interconnect (UPI), or other high-speed point-to-point interconnect dedicated to processor-to-processor communications.
Referring now to
The storage cage 1252 illustratively includes sixteen mounting slots 1256 and is capable of mounting and storing sixteen solid state drives 1254. Of course, the storage cage 1252 may be configured to store additional or fewer solid state drives 1254 in other embodiments. Additionally, in the illustrative embodiment, the solid state drivers are mounted vertically in the storage cage 1252, but may be mounted in the storage cage 1252 in a different orientation in other embodiments. Each solid state drive 1254 may be embodied as any type of data storage device capable of storing long term data. To do so, the solid state drives 1254 may include volatile and non-volatile memory devices discussed above.
As shown in
As discussed above, the individual storage controllers 1220 and the communication circuit 830 are mounted to the top side 650 of the chassis-less circuit board substrate 602 such that no two heat-producing, electrical components shadow each other. For example, the storage controllers 1220 and the communication circuit 830 are mounted in corresponding locations on the top side 650 of the chassis-less circuit board substrate 602 such that no two of those electrical components are linearly in-line with each other along the direction of the airflow path 608.
The memory devices 720 of the storage sled 1200 are mounted to the bottom side 750 of the of the chassis-less circuit board substrate 602 as discussed above in regard to the sled 400. Although mounted to the bottom side 750, the memory devices 720 are communicatively coupled to the storage controllers 1220 located on the top side 650 via the I/O subsystem 622. Again, because the chassis-less circuit board substrate 602 is embodied as a double-sided circuit board, the memory devices 720 and the storage controllers 1220 may be communicatively coupled by one or more vias, connectors, or other mechanisms extending through the chassis-less circuit board substrate 602. Each of the storage controllers 1220 includes a heatsink 1270 secured thereto. As discussed above, due to the improved thermal cooling characteristics of the chassis-less circuit board substrate 602 of the storage sled 1200, none of the heatsinks 1270 include cooling fans attached thereto. That is, each of the heatsinks 1270 is embodied as a fan-less heatsink.
Referring now to
In the illustrative memory sled 1400, the physical resources 620 are embodied as memory controllers 1420. Although only two memory controllers 1420 are shown in
In some embodiments, the memory sled 1400 may also include a controller-to-controller interconnect 1442. Similar to the resource-to-resource interconnect 624 of the sled 400 discussed above, the controller-to-controller interconnect 1442 may be embodied as any type of communication interconnect capable of facilitating controller-to-controller communications. In the illustrative embodiment, the controller-to-controller interconnect 1442 is embodied as a high-speed point-to-point interconnect (e.g., faster than the I/O subsystem 622). For example, the controller-to-controller interconnect 1442 may be embodied as a QuickPath Interconnect (QPI), an UltraPath Interconnect (UPI), or other high-speed point-to-point interconnect dedicated to processor-to-processor communications. As such, in some embodiments, a memory controller 1420 may access, through the controller-to-controller interconnect 1442, memory that is within the memory set 1432 associated with another memory controller 1420. In some embodiments, a scalable memory controller is made of multiple smaller memory controllers, referred to herein as “chiplets”, on a memory sled (e.g., the memory sled 1400). The chiplets may be interconnected (e.g., using EMIB (Embedded Multi-Die Interconnect Bridge)). The combined chiplet memory controller may scale up to a relatively large number of memory controllers and I/O ports, (e.g., up to 16 memory channels). In some embodiments, the memory controllers 1420 may implement a memory interleave (e.g., one memory address is mapped to the memory set 1430, the next memory address is mapped to the memory set 1432, and the third address is mapped to the memory set 1430, etc.). The interleaving may be managed within the memory controllers 1420, or from CPU sockets (e.g., of the compute sled 800) across network links to the memory sets 1430, 1432, and may improve the latency associated with performing memory access operations as compared to accessing contiguous memory addresses from the same memory device.
Further, in some embodiments, the memory sled 1400 may be connected to one or more other sleds 400 (e.g., in the same rack 240 or an adjacent rack 240) through a waveguide, using the waveguide connector 1480. In the illustrative embodiment, the waveguides are 64 millimeter waveguides that provide 16 Rx (i.e., receive) lanes and 16 Tx (i.e., transmit) lanes. Each lane, in the illustrative embodiment, is either 16 GHz or 32 GHz. In other embodiments, the frequencies may be different. Using a waveguide may provide high throughput access to the memory pool (e.g., the memory sets 1430, 1432) to another sled (e.g., a sled 400 in the same rack 240 or an adjacent rack 240 as the memory sled 1400) without adding to the load on the optical data connector 834.
Referring now to
Additionally, in some embodiments, the orchestrator server 1520 may identify trends in the resource utilization of the workload (e.g., the application 1532), such as by identifying phases of execution (e.g., time periods in which different operations, each having different resource utilizations characteristics, are performed) of the workload (e.g., the application 1532) and pre-emptively identifying available resources in the data center 100 and allocating them to the managed node 1570 (e.g., within a predefined time period of the associated phase beginning). In some embodiments, the orchestrator server 1520 may model performance based on various latencies and a distribution scheme to place workloads among compute sleds and other resources (e.g., accelerator sleds, memory sleds, storage sleds) in the data center 100. For example, the orchestrator server 1520 may utilize a model that accounts for the performance of resources on the sleds 400 (e.g., FPGA performance, memory access latency, etc.) and the performance (e.g., congestion, latency, bandwidth) of the path through the network to the resource (e.g., FPGA). As such, the orchestrator server 1520 may determine which resource(s) should be used with which workloads based on the total latency associated with each potential resource available in the data center 100 (e.g., the latency associated with the performance of the resource itself in addition to the latency associated with the path through the network between the compute sled executing the workload and the sled 400 on which the resource is located).
In some embodiments, the orchestrator server 1520 may generate a map of heat generation in the data center 100 using telemetry data (e.g., temperatures, fan speeds, etc.) reported from the sleds 400 and allocate resources to managed nodes as a function of the map of heat generation and predicted heat generation associated with different workloads, to maintain a target temperature and heat distribution in the data center 100. Additionally or alternatively, in some embodiments, the orchestrator server 1520 may organize received telemetry data into a hierarchical model that is indicative of a relationship between the managed nodes (e.g., a spatial relationship such as the physical locations of the resources of the managed nodes within the data center 100 and/or a functional relationship, such as groupings of the managed nodes by the customers the managed nodes provide services for, the types of functions typically performed by the managed nodes, managed nodes that typically share or exchange workloads among each other, etc.). Based on differences in the physical locations and resources in the managed nodes, a given workload may exhibit different resource utilizations (e.g., cause a different internal temperature, use a different percentage of processor or memory capacity) across the resources of different managed nodes. The orchestrator server 1520 may determine the differences based on the telemetry data stored in the hierarchical model and factor the differences into a prediction of future resource utilization of a workload if the workload is reassigned from one managed node to another managed node, to accurately balance resource utilization in the data center 100.
To reduce the computational load on the orchestrator server 1520 and the data transfer load on the network, in some embodiments, the orchestrator server 1520 may send self-test information to the sleds 400 to enable each sled 400 to locally (e.g., on the sled 400) determine whether telemetry data generated by the sled 400 satisfies one or more conditions (e.g., an available capacity that satisfies a predefined threshold, a temperature that satisfies a predefined threshold, etc.). Each sled 400 may then report back a simplified result (e.g., yes or no) to the orchestrator server 1520, which the orchestrator server 1520 may utilize in determining the allocation of resources to managed nodes.
Referring now to
The accelerator device 1622 includes slots 1634, 1636, similar to the slots 1630, 1632 described above. Further, each slot 1634, 1636 includes a corresponding kernel 1664, 1666 and communication abstraction logic unit 1644, 1646, similar to the communication abstraction logic units 1640, 1642 described above. Additionally, the accelerator sled 1612 includes accelerators devices 1624 and 1626. The accelerator device 1624, in the illustrative embodiment, is a graphics processing unit (GPU), which may be embodied as any device or circuitry (e.g., a programmable logic chip, a processor, etc.) configured to perform graphics-related computations (e.g., matrix multiplication, vector operations, etc.), and the accelerator device 1626, in the illustrative embodiment, is a vision processing unit (VPU), which may be embodied as any device or circuitry (e.g., a programmable logic chip, a processor, etc.) configured to perform operations related to machine vision, machine learning, and artificial intelligence. Each accelerator device 1624, 1626, in the illustrative embodiment, includes a corresponding kernel 1668, 1670 and communication abstraction logic unit 1648, 1650, similar to the communication abstraction logic units 1640, 1642 described above. While, in the illustrative embodiment, each slot of the accelerator devices 1620, 1622 (FPGAs) includes a corresponding communication abstraction logic unit, in other embodiments, there may be one communication abstraction logic unit per accelerator device, or one abstraction logic unit per sled.
As further shown, each of the sleds 1610, 1612, and 1614 include a network interface controller (NIC) 1618, 1628, 1684, respectively. Each NIC 1618, 1628, 1684 may be similar to the NIC 832 described relative to
Further, as shown, the orchestrator server 1616 includes a kernel analysis and decision logic unit 1617. The kernel analysis and decision logic unit 1617 may be embodied as any device or circuitry to obtain telemetry data indicative of resource usage and power consumption of the accelerator sleds 1610, 1612 and the compute sled 1617. Further, the kernel analysis and decision logic unit 1617 may evaluate the telemetry data relative to one or more policies (e.g., a service level agreement (SLA) or one or more quality-of-service (QoS) requirements). The policies may include thresholds, which, when exceeded, trigger the kernel analysis and decision logic unit 1617 to orchestrate actions to perform in response.
For instance, as further described herein, the kernel analysis and decision logic unit 1617 may cause a given accelerator device to migrate an accelerator kernel to a target accelerator device (e.g., on another accelerator sled) to maintain QoS requirements. For example, if resource usage on a given accelerator device is relatively high (e.g., due to execution of multiple workloads by accelerator kernels), performance may possibly falter, potentially resulting in a QoS violation. To address this, the kernel analysis and decision logic unit 1617 may migrate one or more of the kernels to other accelerator devices in the system 1600 to reduce load on the currently executing accelerator devices, and in turn support compliance with specified policies.
In addition, as further described herein, the kernel analysis and decision logic unit 1617 may monitor power consumption of the accelerator sleds 1610, 1612 relative to power consumption thresholds specified by policy. To ensure that power consumption is maintained at stable levels, the kernel analysis and decision logic unit 1617 may orchestrate a “scale-out” operation of accelerator kernels from a given accelerator sled to other accelerator sleds. In an embodiment, a scale-out operation refers to re-provisioning a workload currently executing on a given accelerator sled to one or more additional accelerator sleds. During a scale-out operation, the orchestrator server 1616 may migrate one or more accelerator kernels to accelerator devices (or cloned instances of the currently executing accelerator devices) to another accelerator sled. By scaling-out accelerator kernels to another accelerator sled, power consumption may be further distributed across the accelerator sleds executing the workload.
In addition, embodiments presented herein also disclose scaling-out accelerator kernels to accelerator devices of differing types. For example, in an inter-kernel communication topology, a given accelerator kernel may execute on an FPGA (e.g., FPGA 1620), while another accelerator kernel executes on a GPU (e.g., GPU 1624). Doing so allows the orchestrator server 1616 to assign certain workload tasks (e.g., artificial intelligence functions, deep learning, cryptographic functions, etc.) that are more suited to one type of accelerator device over the other and thus optimize a workload executing within the system 1600.
Further, embodiments presented herein also disclose techniques for registering automatic heartbeat notifications from accelerator kernels associated with a scale-out operation. More particularly, based on a workload specification, the orchestrator server 1616 may direct accelerator kernels to transmit a “heartbeat” indicative of a kernel execution at a specified interval to the orchestrator server 1616. In the event that the orchestrator server 1616 does not receive a heartbeat notification at the specified interval, the orchestrator server 1616 may generate an alert that may be transmitted to a user or an application associated with the workload.
Further still, the orchestrator server 1616 may “scale-up” a workload within a given accelerator sled. In an embodiment, a scale-up operation refers to distributing, within the accelerator sled, workload tasks across additional accelerator devices (or accelerator device instances). Scaling-up operations allow a given accelerator device to execute multiple workloads. However, one concern relating to scaling-up workloads is that a fragmentation situation may occur, in which a workload may be allocated to non-adjacent accelerator devices within the sled, which may affect workload performance. To address this, embodiments presented herein also disclose “defragmenting” accelerator devices associated with a scale-up operation. In particular, assume that an accelerator device is executing a workload A but not executing a workload B, and that the accelerator device is adjacent to another accelerator device executing the workload B. On completion of the workload A, the accelerator device may advertise availability of resources. The orchestrator server 1616 may then migrate accelerator kernels associated with workload B executing on non-adjacent accelerator devices to the available accelerator device. Doing so may improve performance of the workload execution.
In an embodiment, inter-kernel communication enables kernels to multicast resource usage and availability to other kernels in the system 1600. A given kernel may, upon receiving a multicast indicative of an availability of accelerator resources, request the accelerator resources from the kernel. Further, in an embodiment, inter-kernel communication channels enable an accelerator device to transmit, via the inter-kernel communication channels, intermediate data targeted toward a given kernel. For example, the accelerator device may send kernel bit streams over the communication channels to provision a given accelerator function with the intermediate data.
Referring now to
In block 1706, the NIC identifies one or more outstanding kernel functions of the workload that is capable of being performed by an accelerator device (e.g., FPGA 1619, 1629, 1685) in the NIC. More particularly, in block 1708, the NIC determines, from the metadata provided with the kernel topology, whether any outstanding kernel functions in the topology do not have any associated dependencies. Such kernel functions may be performed by the accelerator device in the NIC.
In block 1710, the NIC determines whether any kernel functions were identified in the topology. If not, then in block 1712, the NIC forwards the request to the target accelerator device. Otherwise, the method 1700 proceeds to block 1714, in which the NIC determines whether the reduction in data payload resulting from execution the function by the FPGA exceeds a given threshold (e.g., specified in a policy). If not, then the method 1700 proceeds to block 1712. Otherwise, if so, then in block 1716, the NIC determines whether the accelerator device associated with the NIC is available to execute the kernel function. If not, then the method 1700 proceeds to block 1712. If the NIC accelerator device is available to execute the kernel function, then in block 1718, the accelerator device executes, via an inline processing function, the identified one or more accelerator functions. Thereafter, in block 1720, the NIC may forward the resulting data payload and the request to the target accelerator device.
Referring now to
As shown, the method 1800 begins in block 1802, in which the orchestrator server 1616 receives, in response to execution of a kernel scale-out operation by one or more accelerator sleds, a specification of a heartbeat interval by each accelerator kernel associated with the scale-out operation. For example, to satisfy QoS requirements, a workload may specify that particular accelerator kernels complete functions within a particular deadline. The workload may provide such specifications to be provided by a heartbeat notification to be sent at a given interval. In an embodiment, in block 1804, the orchestrator server 1616 receives, in response to the kernel scale-out operation, a kernel topology including metadata specifying a heartbeat interval associated with each kernel in the topology. In block 1806, the orchestrator server 1616 registers the specified heartbeat intervals and expected kernel activity for each kernel associated with the scale-out operation.
Turning now to
If the orchestrator server 1616 receives a heartbeat notification from the accelerator kernel (e.g., in block 1907), the orchestrator server 1616 may evaluate the heartbeat notification to determine activity, e.g., a progress of the accelerator kernel relative to an underlying workload, associated with an execution of the kernel. In block 1910, the orchestrator server 1616 determines whether the actual progress matches an expected progress. If so, then in block 1914, the orchestrator server 1616 sends the heartbeat to the application. The method 1800 then returns to block 1902, in which the orchestrator server 1616 continues to monitor for heartbeat notifications from the accelerated kernel (e.g., until completion of the workload). In block 1912, if the actual progress does not match the expected progress, then in block 1912, the orchestrator server 1616 generates an alert indicating that the actual progress does not match the expected progress for the kernel. The method 1800 then proceeds to block 1906, in which the orchestrator server 1616 sends the generated alert to the application. The method 1800 may then return to block 1902.
Referring now to
As shown, the method 2000 begins in block 2002, in which an accelerator kernel may receive, from another kernel, an advertisement of currently available resources (or resources that are predicted to be available). More particularly, in block 2004, the accelerator kernel may receive an advertisement that includes a group identifier (e.g., an identifier indicative of a workload that the kernel is associated with), and amount of accelerator resources available, a duration of time in which the resources are or will be available, and a kernel identifier.
In block 2006, the accelerator kernel determines, based on one or more policies and relative to a current usage thereof, whether to request resources from the kernel. For example, in block 2008, the accelerator kernel evaluates a load balancing policy relative to the accelerator kernel. Other examples of policies include SLAs and QoS requirements associated with the workload, and the like.
Further, in block 2010, the accelerator determines whether any conflicts for the resources are present. More particularly, to claim the accelerator resources advertised to be available, a given accelerator kernel may multicast a request for the resources using the advertised kernel identifier, creating a potential conflict for another accelerator kernel requesting the resources. If no conflicts are present, then in block 2012, the accelerator kernel may request the advertised resources from the multicasting kernel. For instance, in block 2014, the accelerator kernel may multicast the request to kernels in the associated kernel topology. The request includes the identifier of the advertising kernel. If a conflict exists, the method 2000 may end if another accelerator device claims the resources. However, each accelerator kernel may also include logic to resolve a conflict situation, e.g., a priority may be defined for a given kernel, and kernels having higher priority may receive the advertised resources in event of a conflict.
Referring now to
In block 2106, the orchestrator server 1616 determines whether a specified threshold is exceeded. If not, then the method 2100 returns to block 2102. Otherwise, if a threshold is exceeded, then in block 2108, the orchestrator server 1616 identifies a target accelerator device (e.g., from a registry of accelerator devices in the system 1600) to which to migrate the kernel. For example, in block 2109, the orchestrator server 1616 transmits to a target accelerator device, a request to migrate the kernel. In such a case, the orchestrator server 1616 may identify the target accelerator device as a function of the request. Further, in block 2110, the orchestrator server 1616 may notify accelerator devices associated with the underlying inter-kernel topology that the kernel is to be migrated.
Turning now to
In block 2208, the orchestrator server 1616 performs a migration process of the kernel from the source accelerator device to the target accelerator device. For instance, to do so, in block 2210 the orchestrator server 1616 causes the source accelerator device to carry out the migration process. More particularly, in block 2212, the orchestrator server 1616 may cause the source accelerator to suspend operation of the accelerator kernel. Doing so results in requests targeted to the accelerator kernel to remain in egress queues of requesting devices. In block 2214, the orchestrator server 1616 causes the source accelerator device to serialize data associated with the operation of the accelerator kernel. Thereafter, in block 2216, the orchestrator server 1616 causes the source accelerator device to migrate the kernel to the target accelerator device. To do so, the orchestrator server 1616 may transmit the kernel to the target accelerator device, which in turn generates bit stream data from the kernel and configures a slot thereon. In block 2218, the orchestrator server 1616 causes the source accelerator device to direct standing requests to the target accelerator device. For example, the source accelerator device may broadcast, over the inter-kernel communication channels, that the kernel is migrated to the target accelerator device. In block 2220, the orchestrator server 1616 causes the target accelerator device to resume operation of the kernel. Further, as part of resuming operation of the kernel, in block 2222, the orchestrator server 1616 causes the target accelerator device to deserialize data associated with the kernel operation, including data associated with memory on the target accelerator device, data structures in the underlying bit stream data, and storage inside the target accelerator device.
Referring now to
However, if accelerator resources are available, then in block 2310, the orchestrator server 1616 scales-out the workload to one or more accelerator devices on a target accelerator sled. For instance, to do so, in block 2312, the orchestrator server 1616 may migrate the workload from one or more of the accelerator devices of the source accelerator sled to one or more accelerator devices of the target accelerator sled. As another example, in block 2314, the orchestrator server 1616 may migrate the workload from one or more of the accelerator devices of the source accelerator sled to one or more cloned instances of the accelerator device on the target accelerator sled. To do so, in block 2316, the orchestrator server 1616 causes the target accelerator device to initialize a specified amount of cloned instances of the underlying accelerator device. Because the scale-out operation may affect the composition of a node in the system 1600, in block 2318, the orchestrator server 1616 may update a managed node registry reflecting the change made by the scale-out operation.
Referring now to
As shown, the method 2400 begins in block 2402, in which the orchestrator server 1616 receives a notification of available accelerator devices of an accelerator sled. More particularly, in block 2404, the orchestrator server 1616 may receive a notification indicative of a completion of a workload by the accelerator sled. The notification may specify the accelerator devices associated with the workload.
In block 2406, the orchestrator server 1616 evaluates usage of accelerator devices to determine whether a fragmenting situation is present in the accelerator sled. For instance, in block 2408, the orchestrator server 1616 determines whether an accelerator device on the sled that is assigned to a workload is executing an additional workload. To do so, the orchestrator server 1616 may evaluate an inter-kernel topology of workloads in the system 1600 to identify collocated workloads on accelerator devices on the accelerator sled. The orchestrator server 1616 may also determine whether migration of one of the collocated workloads to any of the available accelerator devices would result in an adjacent positioning of accelerator devices executing the workload. Further, in block 2410, the orchestrator server 1616 determines one or more improved QoS characteristics resulting from migration of a portion of the workload to one or more of the available accelerator devices.
In block 2412, the orchestrator server 1616 determines, based on the evaluation, whether a fragmenting situation is present. If not, then the method 2400 ends. Otherwise, in block 2414, the orchestrator server 1616 migrates a portion of the workload to one or more of the available accelerator devices. More particularly, in block 2416, the orchestrator server 1616 migrates the portion of the workload to an accelerator device at a specified location in the accelerator sled. The specified location may correspond to a target available accelerator device that is adjacent to another accelerator device currently executing the workload on the accelerator sled. The source accelerator device (from which the workload is being migrated) may be an accelerator device that is not adjacent to any accelerator devices executing the workload.
Referring now to
As shown, the method 2500 begins in block 2502, in which the accelerator device receives intermediate data targeted to an accelerated function associated with a workload. The intermediate data may include bit stream data indicative of a result of processing of a workload task by another kernel. Further, the intermediate data may originate from another accelerator device. In block 2504, the accelerator device identifies a communication path to the kernel associated with the accelerated function. For example, to do so, the accelerator device may evaluate an inter-kernel topology and determine the path at which to transmit the intermediate data. Once identified, in block 2506, the accelerator device transmits, via the identified communication path to the kernel, the intermediate data to the accelerator function.
Referring now to
As shown, the method 2600 begins in block 2602, in which the orchestrator server 1616 detects a trigger to initiate a scale-out operation of one or more kernels associated with a workload. An example trigger includes a determination that resource usage in a given accelerator device or sled is relatively high. Another example trigger includes a determination that power consumption of a given accelerator sled is relatively high.
In block 2604, the orchestrator server 1616 determines, as a function of a policy (e.g., a SLA, QoS requirements, a load balancing policy, user-defined specifications in an application associated with the workload, etc.), one or more accelerator devices. Each accelerator device may correspond to one of multiple types (e.g., an FPGA, VPU, ASIC, GPU, etc.). Further, in block 2606, the orchestrator server 1616 may determine a configuration of accelerator devices to satisfy one or more QoS requirements, in which the configuration specifies accelerator devices of differing types.
In block 2608, the orchestrator server 1616 migrates one or more of the accelerator kernels associated with the workload to the accelerator devices identified in the configuration. For instance, to do so, in block 2610, the orchestrator server 1616 generates, from each kernel to be scaled-out to a given accelerator device, a corresponding kernel bit stream that is compatible with the accelerator device. For example, a kernel may be created using the OpenCL framework, which allows the kernel to be compiled into a bit stream compatible with a desired device. In block 2612, the orchestrator server 1616 configures the kernel bit streams on the accelerator devices. For example, to do so, the orchestrator server 1616 may program a slot in which the kernel, via the bit stream, performs the workload tasks. In block 2614, the orchestrator server 1616 initializes inter-kernel communication channels between the kernels associated with the workload. Doing so allows the kernels associated with the workload to send data to one another. Further, because the scale-out operation may affect the composition of a node in the system 1600, the orchestrator server 1616 may, in block 2616, update a managed node registry reflecting the change made by the scale-out operation.
Illustrative examples of the technologies disclosed herein are provided below. An embodiment of the technologies may include any one or more, and any combination of, the examples described below.
Example 1 includes an orchestrator server comprising circuitry to monitor resource usage of an accelerator kernel configured on a source accelerator device; determine whether the resource usage exceeds a threshold specified in one or more policies; upon a determination that the resource usage exceeds the threshold, identify a target accelerator device to which to migrate the accelerator kernel; and migrate the accelerator kernel from the source accelerator device to the target accelerator device.
Example 2 includes the subject matter of Example 1, and wherein to migrate the accelerator kernel from the source accelerator device to the target accelerator device comprises to cause the source accelerator device to suspend operation of the accelerator kernel on the source accelerator device; serialize data associated with the operation of the accelerator kernel; migrate, by the source accelerator device to the target accelerator device, the accelerator kernel; and deserialize the data associated with the operation of the accelerator kernel on the target accelerator device.
Example 3 includes the subject matter of any of Examples 1 and 2, and wherein the circuitry is further to monitor a power consumption of a source accelerator sled having one or more accelerator devices executing a workload relative to a power threshold specified in the one or more policies; and upon a determination that the power consumption threshold is exceeded, scale-out the workload to one or more accelerator devices on a target accelerator sled.
Example 4 includes the subject matter of any of Examples 1-3, and wherein to scale-out the workload to the one or more accelerator devices on the target accelerator sled comprises to migrate the workload from one or more of the accelerator devices of the source accelerator sled to the one or more accelerator devices of the target accelerator sled.
Example 5 includes the subject matter of any of Examples 1-4, and wherein to scale-out the workload to the one or more accelerator devices on the target accelerator sled further comprises to update a registry of managed nodes of a system including the orchestrator server.
Example 6 includes the subject matter of any of Examples 1-5, and wherein to scale-out the workload to the one or more accelerator devices on the target accelerator sled comprises to migrate the workload from one or more accelerator sleds to one or more instances of an accelerator device on the target accelerator sled.
Example 7 includes the subject matter of any of Examples 1-6, and wherein the circuitry is further to receive a notification of available accelerator devices of an accelerator sled.
Example 8 includes the subject matter of any of Examples 1-7, and wherein to receive the notification of available accelerator devices of an accelerator sled comprises to receive a notification of a completion of a workload by the accelerator sled, the notification specifying one or more accelerator devices performing the workload on the accelerator sled.
Example 9 includes the subject matter of any of Examples 1-8, and wherein the circuitry is further to determine, as a function of an evaluation of resource usage of second accelerator devices currently executing a second workload, whether a fragmenting situation is present in the accelerator devices.
Example 10 includes the subject matter of any of Examples 1-9, and wherein the circuitry is further to, upon a determination that the fragmentation situation is present, migrate a portion of the workload to one or more of the available accelerator devices.
Example 11 includes the subject matter of any of Examples 1-10, and wherein the circuitry is further to detect a trigger to initiate a scale-out operation of one or more accelerator kernels associated with a workload; determine, as a function of the one or more policies, one or more types of accelerator devices to which to scale-out the one or more accelerator kernels; and migrate the accelerator kernels to accelerator devices of the one or more types.
Example 12 includes the subject matter of any of Examples 1-11, and wherein to the migrate the accelerator kernels to the accelerator devices of the one or more types comprises to generate, from each accelerator kernel to be scaled-out to an accelerator device of a given type, a bit stream compatible for a corresponding type of accelerator device; configure, for each accelerator kernel, the bit stream on the accelerator device of the corresponding type; initializing communication channels between each of the accelerator kernels associated with the workload; and update a registry of managed nodes in a system including the orchestrator of the migration.
Example 13 includes the subject matter of any of Examples 1-12, and wherein the circuitry is further to register, for each migrated accelerator kernel, a specified interval for a heartbeat notification to be sent to the orchestrator server by the migrated accelerator kernel.
Example 14 includes the subject matter of any of Examples 1-13, and wherein the circuitry is further to, upon a determination that the heartbeat notification is not sent by one of the migrated accelerator kernels, generate an alert indicating that the heartbeat notification was not received at the specified interval.
Example 15 includes one or more machine-readable storage media comprising a plurality of instructions, which, when executed, causes an orchestrator server to monitor resource usage of an accelerator kernel configured on a source accelerator device; determine whether the resource usage exceeds a threshold specified in one or more policies; upon a determination that the resource usage exceeds the threshold, identify a target accelerator device to which to migrate the accelerator kernel; and migrate the accelerator kernel from the source accelerator device to the target accelerator device.
Example 16 includes the subject matter of Example 15, and wherein to migrate the accelerator kernel from the source accelerator device to the target accelerator device comprises to cause the source accelerator device to suspend operation of the accelerator kernel on the source accelerator device; serialize data associated with the operation of the accelerator kernel; migrate, by the source accelerator device to the target accelerator device, the accelerator kernel; and deserialize the data associated with the operation of the accelerator kernel on the target accelerator device.
Example 17 includes the subject matter of any of Examples 15 and 16, and wherein the circuitry is further to monitor a power consumption of a source accelerator sled having one or more accelerator devices executing a workload relative to a power threshold specified in the one or more policies; and upon a determination that the power consumption threshold is exceeded, scale-out the workload to one or more accelerator devices on a target accelerator sled.
Example 18 includes the subject matter of any of Examples 15-17, and wherein to scale-out the workload to the one or more accelerator devices on the target accelerator sled comprises to migrate the workload from one or more of the accelerator devices of the source accelerator sled to the one or more accelerator devices of the target accelerator sled.
Example 19 includes an orchestrator server comprising circuitry for monitoring resource usage of an accelerator kernel configured on a source accelerator device; means for determining whether the resource usage exceeds a threshold specified in one or more policies; means for identifying, upon a determination that the resource usage exceeds the threshold, a target accelerator device to which to migrate the accelerator kernel; and means for migrating the accelerator kernel from the source accelerator device to the target accelerator device.
Example 20 includes the subject matter of Example 19, and wherein the means for migrating the accelerator kernel from the source accelerator device to the target accelerator device comprises means for causing the source accelerator device to (i) suspend operation of the accelerator kernel on the source accelerator device, (ii) serialize data associated with the operation of the accelerator kernel, and (iii) migrate, by the source accelerator device to the target accelerator device, the accelerator kernel.