Multi GPU mode did not operate on Dokcer container

[chainer]

Multi GPU mode did not operate on Dokcer container

[keisukefukuda]

Hi @yoshihingis , thanks for your report.

First, I directly edit your comment so it would look a bit nicer.

For the cases that the executions failed, I guess the reason is python is mis-spelled such as pyhton or pthon, as the error messages implied.

Will you fix them and try again?

Thanks!
Keisuke

[yoshihingis]

Dear keisukefukuda

Thank you for your reply.
Sorry ,I did mistake , I copied and pasted unnecessary information at case(2)(Normal Ubuntu)
I re-write my question.

At Normal Ubuntu, I set n=2 (-n 2),hence the train_mnist.py indicated two GPU's information about both GPU0 & GPU1.
On the Docker Container , I set n=2(-n 2) ,but the train_mnist.py indicated only one GPU's information about GPU0.

But the elapse times are almost the same values at case(1) and case(2).

I think that ; train_mnist.py indicated only GPU0 information on the docker container(case(1), but it used two GPUs (GPU0 & GPU1) during training.

I want to know the reason why only one GPU information was indicated at the train_mnist.py on the docker container.

Could you give me any advice?

[keisukefukuda]

@yoshihingis ,

First, if you want to have more casual discussion, you can reach me on Twitter (keisukefukuda). (possibly in Japanese, if you feel more comfortable :) )

The output

GPU: 0
# unit: 1000
# Minibatch-size: 100
# epoch: 20

should be shown only once if the MPI is working correctly, as indicated in the code
https://github.com/chainer/chainermn/blob/master/examples/mnist/train_mnist.py#L65

Thus, (1) looks working fine and (2) not.

I'm not sure what's going on in (2). Is it raw output? non-edited?
I don't see a "Using hierarchical communicator" message in (2).