Add cudaStreamSynchronize when a new device buffer is added to the spill framework #5485
Conversation
This file contains 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
…n requirements Signed-off-by: Alessandro Bellina <[email protected]>
jlowe
previously approved these changes
May 13, 2022
|
The fix 5ad4d43 helps with some queries I saw an obvious regression (NDS q20 specifically). This makes it so that contig split is followed by a sync in the write path for the RapidsShuffleManager. I am still not done assessing the performance impacts of this, I should have numbers Monday. |
|
All tests I have executed with this PR show results in the noise, so I do not see blockers. |
|
build |
jlowe
previously approved these changes
May 16, 2022
|
ok I have related test failures, due to some mock object matchers I had used. I'll fix that soon |
|
build |
jlowe
approved these changes
May 17, 2022
anthony-chang
pushed a commit
to anthony-chang/spark-rapids
that referenced
this pull request
May 17, 2022
…ill framework (NVIDIA#5485) * Adds a stream synchronize in addBuffer to ensure we safely spill * Small cleanup in copyBuffer, add note about createBuffer synchronation requirements Signed-off-by: Alessandro Bellina <[email protected]> * Remove extra nvtx range * When adding contiguous_split buffers in RapidsShuffleManager, synchronize once * Fix RapidsShuffleTestHelper * Fix RapidsShuffleClientSuite
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
Closes #4818.
This is to fix the stream-ordered violation we could see currently in the spill framework: a buffer is allocated and acted on in stream A, and added to the store without any synchronization. Then stream B runs OOM and spills such buffer, not having waited on stream A.
The idea here is that if a device buffer is added to the store, we force in most cases a call to cudaStreamSynchronize, which should take care of the cases where we return from cuDF and have not stream synchronized there. There is a lot of stream synchronization in cuDF (via thrift usually) so it seems like this is an unlikely issue. That said, adding this synchronize, would be more robust as now the spill framework knows for a fact that the buffer can be freely copied (spilled).
Note that I added a flag to not synchronize in some cases, this is used for UCX since we synchronize on writes after compressing buffers, and also on reads after copying from the bounce buffer, so those two cases we can reliably say we have synchronized. I don't know of other cases where the synchronize in the spill framework can be skipped.
I have been measuring the impact of this in our benchmarks. I do not see any big differences so far in UCX or non-UCX runs. I am going to continue testing it today. Adding as WIP for comments.