bgawrych commented on issue #18244: URL: https://github.com/apache/incubator-mxnet/issues/18244#issuecomment-632106623
@szha I've done some testing in docker for parallel tests and have some overview. For testing I've used about 35 long tests. pytest -n 4 spawns much more threads for MKL/MKLDNN built and all the time during test CPU usage is 100% (24 cores used out of 24). So it's probably starvation problem. Threads spawned: MKL/MKLDNN build: most of the time it was about 250, but pick was before end and it was 265 CPU build : 140 (pick 150) I've tested time for MKLDNN built with set OMP_NUM_THREADS=n/4 (n - number of cores) and it was significat faster (about 1.7x) I'm gonna do some testing with MKL and OMP flags and point the best configuration and also figure out 'serial' case cc: @TaoLv ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
