System and method for normalization of GPU workloads based on real-time GPU data

• US 10,373,283 B2
• Filed: 03/14/2016
• Issued: 08/06/2019
• Est. Priority Date: 03/14/2016
• Status: Active Grant

• Current Assignee: Dell Products LP (Dell Technologies Inc.)
• Original Assignee: Dell Products LP (Dell Technologies Inc.)
• Inventors: Ragupathi, Dinesh Kunnathur, Palmer, John R.
• Primary Examiner(s): Nguyen, Hau H
• Application Number:
  US15/069,516
• Publication Number:
  US 20170262953A1
• Time in Patent Office: 1,240 Days
• CPC Class Codes: G06F 2209/504 Resource capping
• CPC Class Codes: G06F 9/5083 Techniques for rebalancing the load in a distributed system
• CPC Class Codes: G06T 1/20 Processor architectures; Processor configuration, e.g. pipelining
• CPC Class Codes: Y02D 10/00 Energy efficient computing, e.g. low power processors, power management or thermal management

Abstract

An information handling system includes a host processing system and a management controller. The host processing system includes a main processor that instantiates a management controller agent, a graphics processing unit (GPU), and a GPU throttle module. The management controller accesses the management controller via a first interface to obtain a performance status from the GPU, determine that the performance status is outside of a status threshold, and direct, via a second interface of the information handling system, the GPU throttle module to throttle the GPU to bring the performance status to within the status threshold.

Claims

1. An information handling system, comprising:

a host processing system including;

a main processor that instantiates a management controller agent;

a first graphics processing unit (GPU);

a second GPU; and

a GPU throttle module; and

a baseboard management controller that;

accesses the management controller agent via a first interface to obtain a first performance status from the first GPU and a second performance status from the second GPU, wherein the first interface is a keyboard controller style interface;

compares the first performance status with the second performance status;

determines whether or not to direct the GPU throttle module to throttle the first GPU based upon the comparison; and

directs, via a second interface of the information handling system, the GPU throttle module to throttle the first GPU when it is determined to direct the GPU throttle module to throttle the first GPU, wherein the second interface operates out-of-band from the host processing system.

2. The information handling system of claim 1, wherein in throttling the GPU, the GPU throttle module asserts a control signal to the first GPU.

3. The information handling system of claim 2, wherein the control signal causes the first GPU to slow its operation.

4. The information handling system of claim 3, wherein the first GPU slows its operation by reducing an internal clock frequency of the first GPU.

5. The information handling system of claim 3, wherein the first GPU throttle module asserts the control signal in accordance with a pulse width modulation scheme.

6. The information handling system of claim 5, wherein the first GPU slows its operation in proportion with a pulse width of the control signal.

7. The information handling system of claim 1, wherein the first performance status comprises a temperature of the first GPU.

8. The information handling system of claim 1, wherein the first performance status comprises a power consumption of the first GPU.

9. The information handling system of claim 1, wherein the first performance status comprises a memory utilization of the first GPU.

10. The information handling system of claim 1, wherein the first performance status comprises an I/O bandwidth utilization of the first GPU.

11. A method, comprising:

accessing, by a baseboard management controller of an information handling system, a management controller agent instantiated by a main processor of the information handling system to obtain a first performance status from a first graphics processing unit (GPU) of the information handling system and to obtain a second performance status from a second GPU of the information handling system, wherein the management controller is accessed by the baseboard management controller via a first interface, and wherein the first interface is a Universal Serial Bus Network Interface Card (USB-NIC) interface;

comparing, by the baseboard management controller, the first performance with the second performance status;

determining, by the baseboard management controller, whether or not to direct the GPU throttle module to throttle the first GPU based upon the comparison; and

directing, via a second interface of the information handling system, a GPU throttle module of the information handling system to throttle the first GPU when it is determined to direct the GPU throttle module to throttle the first GPU, wherein the second interface operates out-of-band from the host processing system.

12. The method of claim 11, wherein in throttling the first GPU, the method further comprises:

asserting, by the GPU throttle module, a control signal to the first GPU.

13. The method of claim 12, further comprising:

slowing the operation of the first GPU in response to the asserted control signal.

14. The method of claim 13, in slowing the operation of the first GPU, the method further comprises:

reducing an internal clock frequency of the first GPU.

15. The method of claim 13, wherein the GPU throttle module asserts the control signal in accordance with a pulse width modulation scheme.

16. A non-transitory computer-readable medium including code for performing a method, the method comprising:

accessing, by a baseboard management controller of an information handling system, a management controller agent instantiated by a main processor of the information handling system to obtain a first performance status from a first graphics processing unit (GPU) of the information handling system and to obtain a second performance status from a second GPU of the information handling system, wherein the management controller is accessed by the baseboard management controller via a first interface, and wherein the first interface is a Universal Serial Bus Network Interface Card (USB-NIC) interface;

comparing, by the baseboard management controller, the first performance status with the second performance status;

determining, by the baseboard management controller, whether or not to direct the GPU throttle module to throttle the first GPU based upon the comparison; and

directing, via a second interface of the information handling system, a GPU throttle module of the information handling system to throttle the first GPU when it is determined to direct the GPU throttle module to throttle the first GPU, wherein the second interface is an interface that operates out-of-band from the host processing system.

17. The computer-readable medium of claim 16, wherein in throttling the first GPU, the method further comprises:

asserting, by the GPU throttle module, a control signal to the first GPU.

18. The computer-readable medium of claim 17, the method further comprising:

slowing the operation of the first GPU in response to the asserted control signal.

19. The computer-readable medium of claim 18, in slowing the operation of the first GPU, the method further comprises:

reducing an internal clock frequency of the first GPU.

20. The computer-readable medium of claim 18, wherein the GPU throttle module asserts the control signal in accordance with a pulse width modulation scheme.