Number of CPU threads for the python process

[mahmoodn]

For a test, I didn't use --cuda in order to run a cpu version. While the CPU has 8 physical cores (16 threads), I see 400% cpu utilization for the python process.

Is that normal? How can I control the number of threads?
I use time python main.py --epochs 1 in word_language_model

[ezyang]

In this case, you are seeing multithread usage from OpenMP parallelized CPU kernels, which run on multiple threads. You can control the number of threads with https://pytorch.org/docs/stable/torch.html#torch.set_num_threads although this was broken on some versions of PyTorch. You can also try setting OMP_NUM_THREADS=1 and MKL_NUM_THREADS=1 if that doesn't work for you.

CC @ssnl, in case you have anything to add.

[ssnl]

@ezyang ‘s explanation & suggestion above is 100% correct!

[CoinCheung]

@ezyang @ssnl I have tested both methods, and I found that setting OMP_NUM_THREADS=1 helps to solve problem, while torch.set_num_threads() and MKL_NUM_THREADS=1 does not work. Why does setting torch.set_num_threads(1) not work, but using environment variables of OMP_NUM_THREADS=1 would work ?

[ezyang]

I mentioned that we accidentally broke set_num_threads on an older version of PyTorch. What version are you running on? (MKL controls MKL only so it won't affect our OpenMP parallelized kernels).

[CoinCheung]

@ezyang
I used pytorch installed with ananconda conda install pytorch, that would be 1.0 for now. Maybe it is the version that has this option broken. I will use the environment variable method. Thanks for explaining !!

[gchanan]

closing as this doesn't seem to be a current bug.