Hi I am using Nsight System to profile my pytorch code. While visualizing .qdred file in nsight-sys, I found that when waitpid, fork and sem_timedwait running all other processes stops. Is there any way to remove these processes or decrease there time. I am using number of worker = 10, batch_size=101 in pytorch code. Find my nsight-sys graph screenshot at https://drive.google.com/open?id=1HVvmLuNUVGnaXBOpdTPmHRQuNWbilxUG
Those calls are all designed to make other processes wait until it is safe to continue. You would have to go into your code to determine if all of them are required to have a correct program.