Disclosed are various embodiments of real-time simulation of the performance of a compute accelerator workload for distributed resource scheduling. A compute kernel of a compute accelerator workload is augmented to include instructions that increment an execution counter at artificial halting points. Execution of the compute accelerator workload is suspended at an artificial halting point. The compute accelerator workload is executed on a plurality of candidate hosts and a performance counter is incremented during the execution of the compute accelerator workload on the various hosts. The compute accelerator workload is migrated to a destination host selected using an efficiency metric that is identified using the performance counter.