Performance issue when fork on linux

[diandianliu]

hi
I have created a libmxnet.so use USE_MKLDNN=1 in config.mk file.
the use time is 290ms per image is faster than before( 3800ms not use MKLDNN).

But I use fork to run the process the time is 1300ms per image the time is too long.

cpu useage:CPUS(32),not set OMP_NUM_THREADS use default
no fork :1600%CPU
fork:100%CPU ;

in fork case, when I set OMP_NUM_THREADS more large,the cpu usage is more large too ,but the speed is not improve.

why the Performance is so slow when use fork?

[emfomenk]

Hi @diandianliu,

Seems quite similar to this StackOverflow thread.

Could you please try resetting the mask in the child process?

cpu_set_t cpuset;
CPU_ZERO(&cpuset);
for (int i = 0; i < nthr; ++i) CPU_SET(i, &cpuset);
sched_setaffinity(0, sizeof(cpuset), &cpuset);

Alternatively, if you set KMP affinity like KMP_AFFINITY=granularity=fine,compact, try omitting it (just to check if this helps).

[TaoLv]

Hi @diandianliu Not sure what does "use fork to run the process" mean. But I guess it relates to the discussion here:
apache/mxnet#13606 (comment)
and
apache/mxnet#13593

[rsdubtso]

This does not seem to be directly related to MKL-DNN, but rather to the way OpenMP CPU affinity interacts with fork(), so I am inclined to close this issue. Keeping open for a couple of days...

[diandianliu]

Hi @diandianliu,

Seems quite similar to this StackOverflow thread.

Could you please try resetting the mask in the child process?

cpu_set_t cpuset;
CPU_ZERO(&cpuset);
for (int i = 0; i < nthr; ++i) CPU_SET(i, &cpuset);
sched_setaffinity(0, sizeof(cpuset), &cpuset);

Alternatively, if you set KMP affinity like KMP_AFFINITY=granularity=fine,compact, try omitting it (just to check if this helps

thanks ,i find the quest , when use fork , mkl can only use one thread