GPU selection for jobs that use more than 1 gpu

[yassersouri]

For jobs that need more than 1 gpu, it is beneficial if the gpus selected are dependent on the topology of the gpu p2p connections.

This information can be seen for example from nvidia-smi topo -m.

Currently it seems that the gpus are selected randomly.

[yassersouri]

I think an intermediate solution would be if the server allocated gpus in order.
For example if gpus 0, 1, 4, 5, 6, 7 are available, it is better if the server allocates 4, 5, 6, 7.

This is because usually the p2p connections are between gpus in order. Like for example in cvg32 machine in our lab, gpus 0-3 and 4-7 are connected with good p2p connections.

[alexanderrichard]

This should be easy to do. I don't have time this weekend but might be able to look into it next weekend.

[denseishin]

I have addressed this issue in a fork.
https://github.com/denseishin/queueing-tool
It tries to allocate GPU's in order first.
I had a look at the NVIDIA Management Library to figure out something topology-based, but the API is very poorly documented.