When I use MPS and trtexec to run 3 identical inference loads concurrently, 1 workload runs faster than the others obviously. When I run 5 identical inference loads concurrently, 2 workloads run faster than the others obviously. This situation does not happen in 2 and 4 processes.
What happened? How can I know which process will run faster before I run these workloads concurrently?
workload: resnet50
GPU: V100