Open bjoo opened 6 years ago
This definitely should be tested before embracing. When using multiple threads and a heavy CPU workload, using explicit binding will always be optimal. For Balint's case where everything is on the GPU and only a single-threaded workload is used, then I can imagine it wouldn't matter that much.
Hi Evan, I have stopped using -physcpubind= and started using -N 0 and -N 8 to bind to numa-nodes 0 (socket 0) and 8 (socket 1), again because the number of threads may vary and occasionally getting physcpubind wrong, gave me errors with numactl.
Best, B