Fix CUDA process enumeration race in checkpoint - #13933
Open
luiscape wants to merge 1 commit into
Open
Conversation
`preSaveCuda` enumerates CUDA processes once, then suspends them with `cuda-checkpoint --toggle`. The issue here is that processes may create new CUDA sessions after enumeration, leading to either checkpoint failues (because a new session now has been initialized) or restore failures (because a new in-flight session is initializing and can no longer complete). To fix this race, I am enumerating CUDA sessions repeatedly until no further sessions appear. This replicates what CRIU GPU does in ([`seize.c:collect_loop()`](https://github.com/checkpoint-restore/criu/blob/894a7c3f85a7c32cf6a88f956b6fb0381b5bfe06/criu/seize.c#L964-L995), which I think is a reasonable approach. I tested this on a 4 x A10G instance with multiple workloads and different combination of GPUs (1, 2, 4). I used PyTorch directly, but also vLLM and SGLang. Multi-GPU workloads are checkpointed and restored successfully after this change.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
preSaveCudaenumerates CUDA processes once, then suspends them withcuda-checkpoint --toggle. The issue here is that processes may create new CUDA sessions after enumeration, leading to either checkpoint failues (because a new session now has been initialized) or restore failures (because a new in-flight session is initializing and can no longer complete).To fix this race, I am enumerating CUDA sessions repeatedly until no further sessions appear. This replicates what CRIU GPU does in
seize.c:collect_loop(), which I think is a reasonable approach.I tested this on a 4 x A10G instance with multiple workloads and different combination of GPUs (1, 2, 4). I used PyTorch directly, but also vLLM and SGLang. Multi-GPU workloads are checkpointed and restored successfully after this change.