Power throttling of high performance computing (HPC) platform components
Embodiments of systems and methods for power throttling of High Performance Computing (HPC) components are described. In some embodiments, an HPC platform may include: a system Baseboard Management Controller (BMC), and an accelerator tray comprising a tray BMC coupled to a plurality of managed subsystems and to the system BMC, where the system BMC is configured to: in response to a power excursion event, instruct the tray BMC to throttle a first managed subsystem by a first amount and to throttle a second managed subsystem by a second amount.
1. A High Performance Computing (HPC) platform, comprising:
a system Baseboard Management Controller (BMC);
a host processor coupled to the system BMC and decoupled from a tray BMC; and
an accelerator tray comprising the tray BMC coupled to a plurality of managed subsystems and to the system BMC, wherein the system BMC is configured to:
in response to a power excursion event, instruct the tray BMC to throttle a first managed subsystem by a first amount and to throttle a second managed subsystem by a second amount; and
instruct the host processor to enter a throttling mode, at least in part, in response to the first and second amounts being insufficient to maintain a total power consumption of the HPC platform below a power cap.
2. The HPC platform of claim 1 , wherein the first and second managed subsystems comprise at least one of: a Graphics Processing Unit (GPU), a Field Programmable Gate Array (FPGA), an Intelligence Processing Unit (IPU), a Data Processing Unit (DPU), a Gaussian Neural Accelerator (GNA), an Audio and Contextual Engine (ACE), or a Vision Processing Unit (VPU).
3. The HPC platform of claim 1 , wherein the power excursion event is triggered by the detection of at least one of: a high output power supply current, a power supply input undervoltage, or an over temperature condition.
4. The HPC platform of claim 1 , wherein the system BMC is configured to identify an order in which each of the plurality of managed subsystems is to be throttled based, at least in part, upon a policy or a license.
5. The HPC platform of claim 1 , wherein the system BMC is configured to identify an order in which each of the plurality of managed subsystems is to be throttled based, at least in part, upon a priority of a workload.
6. The HPC platform of claim 1 , wherein the system BMC is configured to identify an order in which each of the plurality of managed subsystems is to be throttled based, at least in part, upon a priority of each managed subsystem.
7. The HPC platform of claim 1 , wherein the system BMC is configured to determine at least one of the first or second amounts based, at least in part, upon a policy or a license.
8. The HPC platform of claim 1 , wherein the system BMC is configured to determine the first amount based upon a first priority of a first workload assigned to the first managed subsystem and the second amount based upon a second priority of a second workload assigned to the second managed subsystem.
9. The HPC platform of claim 1 , wherein the system BMC is configured to determine the first amount based upon a first priority of the first managed subsystem and the second amount based upon a second priority of the second managed subsystem.
10. The HPC platform of claim 9 , wherein the system BMC is configured to communicate the first and second priorities to an Information Handling System (IHS).
11. The HPC platform of claim 10 , wherein the IHS is configured to assign a high-priority workload to a last-to-throttle managed subsystem.
12. The HPC platform of claim 1 , wherein the system BMC is configured to transmit, to the tray BMC, an indication of a programmable action to be performed with respect to the plurality of managed subsystems in response to a PWRBRK signal.
13. The HPC platform of claim 1 , wherein the system BMC is configured to determine a power consumption contribution of each of the plurality of managed subsystems based upon power consumption information received from at least one of: (a) the tray BMC via an Out-Of-Band (OOB) management link, or (b) directly from the plurality of managed subsystems.
14. The HPC platform of claim 13 , wherein the system BMC is configured to trigger migration of a workload executed by the first managed subsystem to the second managed subsystem based, at least in part, upon the power consumption information.
15. A system Baseboard Management Controller (BMC) deployable in a High Performance Computing (HPC) platform, the system BMC comprising:
a processing core; and
a memory coupled to the processing core, wherein the memory comprises program instructions, that, upon execution by the processing core, cause the system BMC to:
in response to a power excursion event, instruct a tray BMC of the HPC platform to throttle a first managed subsystem by a first amount and to throttle a second managed subsystem by a second amount; and
instruct a host processor coupled to the system BMC and decoupled from the tray BMC to enter a throttling mode, at least in part, in response to the first and second amounts being insufficient to maintain a total power consumption of the HPC platform below a power cap.
16. The system BMC of claim 15 , wherein the first and second managed subsystems comprise at least one of: a Graphics Processing Unit (GPU), a Field Programmable Gate Array (FPGA), an Intelligence Processing Unit (IPU), a Data Processing Unit (DPU), a Gaussian Neural Accelerator (GNA), an Audio and Contextual Engine (ACE), or a Vision Processing Unit (VPU).
17. The system BMC of claim 15 , wherein the program instructions, upon execution by the processing core, further cause the system BMC to identify an order in which each of the first and second managed subsystems is to be throttled based, at least in part, upon at least one of: a policy or a license, a priority of a workload, or a priority of each of the first and second managed subsystems.
18. A method, comprising:
in response to a power excursion event in a High Performance Computing (HPC) platform, a system Baseboard Management Controller (BMC) instructing a tray BMC of the HPC platform to throttle a first managed subsystem by a first amount and to throttle a second managed subsystem by a second amount; and
the system BMC instructing a host processor coupled to the system BMC and decoupled from the tray BMC to enter a throttling mode, at least in part, in response to the first and second amounts being insufficient to maintain a total power consumption of the HPC platform below a power cap.
19. The method of claim 18 , wherein the first and second managed subsystems comprise at least one of: a Graphics Processing Unit (GPU), a Field Programmable Gate Array (FPGA), an Intelligence Processing Unit (IPU), a Data Processing Unit (DPU), a Gaussian Neural Accelerator (GNA), an Audio and Contextual Engine (ACE), or a Vision Processing Unit (VPU).
20. The method of claim 18 , further comprising the system BMC identifying an order in which each of the first and second managed subsystems is to be throttled based, at least in part, upon at least one of: a policy or a license, a priority of a workload, or a priority of each of the first and second managed subsystems.