Real-time simulation of compute accelerator workloads for distributed resource scheduling

1. A non-transitory, computer-readable medium comprising machine-readable instructions that, when executed by at least one processor, cause at least one computing device to at least:

identify a compute accelerator workload and a working set for the compute accelerator workload;

augment a compute kernel of the compute accelerator workload to be an augmented compute kernel comprising instructions that increment an execution counter at a plurality of artificial halting points inserted into the compute kernel;

execute the compute accelerator workload comprising the augmented compute kernel on a plurality of candidate hosts to select a destination host for the compute accelerator workload;

select the destination host from the plurality of candidate hosts based at least in part on a plurality of efficiency metrics corresponding to execution of the compute accelerator workload on the plurality of candidate hosts; and

assign the compute accelerator workload to execute on the destination host.

2. The non-transitory, computer-readable medium of claim 1 , wherein the instructions, when executed by the at least one processor, cause the at least one computing device to at least:

determine the plurality of efficiency metrics corresponding to the plurality of candidate hosts based at least in part on a plurality of performance counters incremented during execution of the compute accelerator workload on a respective candidate host.

3. The non-transitory, computer-readable medium of claim 2 , wherein a respective efficiency metric comprises at least one of: a non-local page reference velocity, a non-local page dirty velocity, a device-local page reference velocity, a device-local page dirty velocity, and an execution velocity.

4. The non-transitory, computer-readable medium of claim 3 , wherein the instructions, when executed by the at least one processor, cause the at least one computing device to at least:

query the plurality of performance counters in real time to determine the respective efficiency metric.

5. The non-transitory, computer-readable medium of claim 2 , wherein at least one of the plurality of performance counters comprises a hardware performance counter.

6. The non-transitory, computer-readable medium of claim 2 , wherein the instructions, when executed by the at least one processor, cause the at least one computing device to at least:

insert, into a compute kernel of the compute accelerator workload, instructions to increment the performance counter at a memory load instruction, a memory store instruction, or a halting point of the compute kernel.

7. The non-transitory, computer-readable medium of claim 1 , wherein the working set comprises at least one of: initialization data provided to the compute accelerator workload to begin execution, and intermediate data identified at a halting point of the compute accelerator workload.

8. A method, comprising:

augmenting a compute kernel of a compute accelerator workload to be an augmented compute kernel comprising instructions that increment an execution counter at a plurality of artificial halting points of the augmented compute kernel;

executing the compute accelerator workload on a plurality of candidate hosts to select a destination host for the compute accelerator workload;

determining a plurality of efficiency metrics corresponding to the plurality of candidate hosts, a respective efficiency metric being determined based at least in part on at least one performance counter comprising the execution counter incremented by the augmented compute kernel during execution of the compute accelerator workload on a compute accelerator of a respective candidate host;

selecting the destination host from the plurality of candidate hosts based at least in part on an efficiency metric for the compute accelerator workload on the destination host; and

assigning the compute accelerator workload to execute on the destination host.

9. The method of claim 8 , wherein the compute accelerator workload is cloned to the plurality of candidate hosts along with a working set for the compute accelerator workload.

10. The method of claim 9 , wherein the working set comprises initialization data provided to the compute accelerator workload to begin execution.

11. The method of claim 9 , wherein the compute accelerator workload is analyzed for migration from a source host, and the working set comprises intermediate data identified on the source host at a halting point of the compute accelerator workload.

12. The method of claim 8 , wherein the augmented compute accelerator workload includes instructions to increment a page reference counter.

13. The method of claim 8 , wherein the augmented compute accelerator workload includes instructions to increment a page dirty counter.

14. The method of claim 8 , wherein the artificial halting points are inserted at a plurality of loop iterations of the compute accelerator workload.