number of CPU used when training with GPU

[Lewis-YL]

Hi,

In the deepmd v2.1.3,

Before dp train input.json, I used the commands:
export OMP_NUM_THREADS=1
export TF_INTRA_OP_PARALLELISM_THREADS=1
export TF_INTER_OP_PARALLELISM_THREADS=1

However, It still shows a log like this:

This did not happen in the previous deepmd version.
When I used the other GPU or other CPU, it slows my current training speed by 25% percent.

Do you know how to fix it?

[njzjz]

Your log only shows one thread:

OMP: Info #254: KMPAFFINITY: pid 212948 tid 213000 thread 0 bound to os proc set 0

It looks expected. Did you show the entire OMP log?

[Lewis-YL]

Yes, that is the entire OMP log.
I want to use 1 CPU core of my 24 cores.

You can see the training time changed from 60s to 73s because I was running another job using a different GPU card.
In principle, this should not affect my current training job.

Do you know what is the possible cause?

I have not seen this in previous deepmd version.

[njzjz]

The log of another job will be helpful. Do they bound to the same proc set?