Intelligent GPU scheduling in a virtualization environment

1. A computing device for scheduling workload submissions for a virtualized graphics processing unit (GPU), the computing device comprising:

a processor;

a virtualization service to establish a virtualization environment comprising a plurality of virtual machines, each of the virtual machines comprising a graphics driver to communicate with the GPU and a plurality of command buffers to store GPU commands; and

a memory having a plurality of instructions stored therein that, when executed by the processor, causes the computing device to:

evaluate the GPU commands of all of the command buffers of all of the virtual machines;

dynamically select, in response to output of the evaluating of the GPU commands, a scheduling policy from a plurality of different scheduling policies;

switch the scheduling policy from a per-ring scheduling policy to a gang scheduling policy in response to detecting a cross-buffer dependency between two GPU commands that are in different command buffers of the same virtual machine;

schedule at least one of the GPU commands according to the gang scheduling policy; and

execute, by the GPU, the at least one of the GPU commands according to the gang scheduling policy.

2. The computing device of claim 1 , wherein the plurality of instructions further causes the computing device to:

scan the GPU commands of all of the command buffers of a virtual machine;

generate data indicative of cross-buffer dependencies between GPU commands in different command buffers of the same virtual machine; and

dynamically select the scheduling policy based on the data indicative of cross-buffer dependencies.

3. The computing device of claim 1 , wherein the plurality of instructions further causes the computing device to:

scan the GPU commands of all of the command buffers of a virtual machine, determine a command type for each of the GPU commands; and

dynamically select the scheduling policy for a GPU command based on the command type of the GPU command.

4. The computing device of claim 1 , wherein the plurality of instructions further causes the computing device to:

schedule GPU commands of one of the virtual machines according to the dynamically-selected scheduling policy; and

schedule GPU commands of another one of the virtual machines according to a different scheduling policy.

5. The computing device of claim 1 , wherein the plurality of instructions further causes the computing device to:

evaluate a GPU command of a different virtual machine; and

schedule the GPU command of the different virtual machine according to a scheduling policy that is not a gang scheduling policy.

6. The computing device of claim 1 , wherein each of the command buffers is embodied as a ring buffer, and wherein the plurality of instructions further causes the computing device to scan the GPU commands of a virtual machine for cross-ring synchronization primitives and select the scheduling policy for the GPU commands of the virtual machine based on the presence or absence of cross-ring synchronization primitives.

7. The computing device of claim 1 , wherein the plurality of instructions further causes the computing device to dynamically select different scheduling policies for different virtual machines based on the output of the evaluating of the GPU commands.

8. The computing device of claim 1 , wherein the plurality of instructions further causes the computing device to create a command database comprising data indicating dependencies between the command buffers of the virtual machine.

9. The computing device of claim 8 , wherein the plurality of instructions further causes the computing device to evaluate the command databases of all of the virtual machines and select the scheduling policy for at least one of the virtual machines based on the evaluating of all of the command databases.

10. The computing device of claim 1 , wherein the plurality of instructions further causes the computing device to:

scan the GPU commands of all of the command buffers of a virtual machine;

detect cross-buffer dependencies between GPU commands in different command buffers of the same virtual machine;

determine a frequency of occurrence of the cross-buffer dependencies over a period of time; and

change the scheduling policy based on the frequency of occurrence of the cross-buffer dependencies.

11. The computing device of claim 1 , wherein the plurality of instructions further causes the computing device to:

monitor for occurrences of cross-buffer dependencies across all of the command buffers of all of the virtual machines; and

change the scheduling policy from a per-ring policy to a gang policy if no cross-buffer dependencies are detected in a selected number of the command buffers.

12. A method for scheduling workload submissions for a virtualized graphics processing unit (GPU), the method comprising:

establishing a virtualization environment comprising a plurality of virtual machines, each of the virtual machines comprising a graphics driver to communicate with the GPU and a plurality of command buffers to store GPU commands;

evaluating the GPU commands of all of the command buffers of all of the virtual machines;

in response to output of the evaluating of the GPU commands, dynamically selecting a scheduling policy from a plurality of different scheduling policies;

switching the scheduling policy from a per-ring scheduling policy to a gang scheduling policy in response to detecting a cross-buffer dependency between two GPU commands that are in different command buffers of the same virtual machine;

scheduling at least one of the GPU commands according to the gang scheduling policy; and

executing, by the GPU, the at least one of the GPU commands according to the gang scheduling policy.

13. The method of claim 12 , comprising scanning the GPU commands of all of the command buffers of a virtual machine, generating data indicative of cross-buffer dependencies between GPU commands in different command buffers of the same virtual machine, and dynamically selecting the scheduling policy based on the data indicative of cross-buffer dependencies.

14. The method of claim 12 , comprising scanning the GPU commands of all of the command buffers of a virtual machine, determining a command type for each of the GPU commands, and dynamically selecting the scheduling policy for a GPU command based on the command type of the GPU command.

15. The method of claim 12 , comprising scheduling GPU commands of one of the virtual machines according to the dynamically-selected scheduling policy and scheduling GPU commands of another one of the virtual machines according to a different scheduling policy.

16. The method of claim 12 , comprising evaluating a GPU command of a different virtual machine and scheduling the GPU command of the different virtual machine according to a different scheduling policy that is not a gang scheduling policy.

17. The method of claim 12 , wherein each of the command buffers is embodied as a ring buffer, and the method comprises scanning the GPU commands of a virtual machine for cross-ring synchronization primitives and selecting the scheduling policy for the GPU commands of the virtual machine based on the presence or absence of cross-ring synchronization primitives.

18. The method of claim 12 , comprising dynamically selecting different scheduling policies for different virtual machines based on the output of the evaluating of the GPU commands.

19. One or more non-transitory machine readable storage media comprising a plurality of instructions stored thereon that in response to being executed result in a computing device scheduling workload submissions for a virtualized graphics processing unit (GPU), by:

establishing a virtualization environment comprising a plurality of virtual machines, each of the virtual machines comprising a graphics driver to communicate with the GPU and a plurality of command buffers to store GPU commands;

evaluating the GPU commands of all of the command buffers of all of the virtual machines;

in response to output of the evaluating of the GPU commands, dynamically selecting a scheduling policy from a plurality of different scheduling policies;

switching the scheduling policy from a per-ring scheduling policy to a gang scheduling policy in response to detecting a cross-buffer dependency between two GPU commands that are in different command buffers of the same virtual machine;

scheduling at least one of the GPU commands according to the gang scheduling policy; and

executing, by the GPU, the at least one of the GPU commands according to the gang scheduling policy.

20. The one or more non-transitory machine readable storage media of claim 19 , wherein the instructions when executed result in the computing device, for each of the virtual machines, creating a command database comprising data indicating dependencies between the command buffers of the virtual machine.

21. The one or more non-transitory machine readable storage media of claim 20 , wherein the instructions when executed result in the computing device evaluating the command databases of all of the virtual machines and selecting the scheduling policy for at least one of the virtual machines based on the evaluating of all of the command databases.

22. The one or more non-transitory machine readable storage media of claim 19 , wherein the instructions when executed result in the computing device scanning the GPU commands of all of the command buffers of a virtual machine, detecting cross-buffer dependencies between GPU commands in different command buffers of the same virtual machine, determining a frequency of occurrence of the cross-buffer dependencies over a period of time, and changing the scheduling policy based on the frequency of occurrence of the cross-buffer dependencies.

23. The one or more non-transitory machine readable storage media of claim 19 , wherein the instructions when executed result in the computing device monitoring for occurrences of cross-buffer dependencies across all of the command buffers of all of the virtual machines, and changing the scheduling policy from a per-ring policy to a gang policy if no cross-buffer dependencies are detected in a selected number of the command buffers.