gpu-policy: add orpheus

[thomas]

2 Cards 'NVIDIA TITAN X (Pascal)' 12GB each from 2016/17

[donald]

LGTM, go ahead.
You can test that by calling in with "init" manually as root. It should return the number of gpu.

[donald]

You might want to add cuda_12gb (or cuda_10gb) to the tags of oprheus. That way a user could submit with "--gpu --prere cuda_10gb" to avoid the job being started on a card with 5gb. See jabberwocky.

[thomas]

Nameing it cuda_10gb/cuda_12gb is a bit of the question. finer grain, or classes?

Right now an user who needs more than 5GB must use cuda_10gb and thus ends on the A100. IMHO that should suffice until we see what other problems arise when we use different GPU-capabilities (Pascal/Ampere or 2016/2021) under the same 'hood'.

[donald]

Okay. Anyway, I think we should plan a tag schema now, so that a user can formulate the requirement of his job. If we are doing it 'as needed' it will probably become a mess not much better than --whitelist.

[thomas]

Hostconfig for orpheus is now
orpheus tag amd mx64 server cuda confidential cuda_10g

Anything else to consider?

[donald]

No, looks good.