Demystifying the Placement Policies of the NVIDIA GPU Thread Block Scheduler for Concurrent Kernels

In this work, we empirically derive the scheduler's behavior under concurrent workloads for NVIDIA's Pascal, Volta, and Turing microarchitectures. In contrast to past studies that suggest the scheduler uses a round-robin policy to assign thread blocks to streaming multiprocessors (SMs), we instead find that the scheduler chooses the next SM based on the SM's local resource availability. We show how this scheduling policy can lead to significant, and seemingly counter-intuitive, performance degradation; for example, a decrease of one thread per block resulted in a 3.58X increase in execution time for one kernel in our experiments. We hope that our work will be useful for improving the accuracy of GPU simulators and aid in the development of novel scheduling algorithms.