Missed ability to not use memory registering if provider does not request it about aws-ofi-nccl HOT 6 CLOSED

Thanks Dmitry for the patch again and we apologize for overriding the previous changes. I have reviewed the patch and have one critical feedback.

The major blocker I see in testing psm3 provider is availability/access to appropriate hardware. Does your team have any continuous integration system that our PRs can use? This will help us prevent such issues in future.

from aws-ofi-nccl.

Hello Rashika (@rashikakheria),

May I ask whether your test\CI system have RoCE or IB (InfiniBand) network card?
If so, PSM3 can be tested there out of the box, it does not require Intel NICs.
PSM3 is fully opensource and shipped with libfabric.

Accessing to our team's test system is a tough topic, I would like to discuss easier approaches first.

BRs,
Denis

from aws-ofi-nccl.

May I ask whether your test\CI system have RoCE or IB (InfiniBand) network card?

No, we don't have access to systems with IB or RoCE network card.

from aws-ofi-nccl.

PR merged.

from aws-ofi-nccl.

May I ask whether your test\CI system have RoCE or IB (InfiniBand) network card?

No, we don't have access to systems with IB or RoCE network card.

Hello Rashika (@rashikakheria),
Sorry for late reply. Unfortunately accessing to our team internal CI infrastructure from outside is a topic required thorough discussion (primarily with infosec representatives). We initiated discussion, but currently the answer is that we are not allowed to do this. I will update if something changes, but chances are low.

from aws-ofi-nccl.

Hello Rashika (@rashikakheria), Sorry for late reply. Unfortunately accessing to our team internal CI infrastructure from outside is a topic required thorough discussion (primarily with infosec representatives). We initiated discussion, but currently the answer is that we are not allowed to do this. I will update if something changes, but chances are low.

Thanks for getting back! One idea is that you could possibly pull in the changes from aws-ofi-nccl repository for every PR and test it on your end. You could comment on the PRs if your CI system break (until the time this repository has direct access to your testing).

from aws-ofi-nccl.