Fix TSAN violations for Distribution ops #1768
Merged
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
GleasonK
requested changes
Sep 2, 2023
GleasonK
requested changes
Sep 6, 2023
GleasonK
requested changes
Sep 7, 2023
GleasonK
approved these changes
Sep 12, 2023
ghpvnist
force-pushed
the
thread_fix
branch
from
42492b1
to
a480caa
Compare
penagos
pushed a commit
to penagos/stablehlo
that referenced
this pull request
Sep 21, 2023
The TSAN violation is due to an unprotected insert shown above and replacing `llvm::RefCountedBase` with its thread-safe alternative `llvm::ThreadSafeRefCountedBase`. In the original implementation of `rendezvous`, there were some nuances that were undesirable. For example, it's preferred to clear out the entry once all processes have read the data and clear out the map instead of letting the next distribution op to clear it out. The current implementation is also incorrect since there is no guarantee that processGroup size is the same as previous op calling `rendezvous`. ``` if (channels_[channelKey].size() == processGroup.size()) channels_[channelKey].clear(); channels_[channelKey].insert(processId, operand); ``` The proposed update enhances the `rendezvous` logic such that each process has a shared pointer to the result object once out of scope, it is automatically deleted. The PR also reenables tests for distribution ops. closes openxla#1755
penagos
pushed a commit
to penagos/stablehlo
that referenced
this pull request
Sep 21, 2023
The TSAN violation is due to an unprotected insert shown above and replacing `llvm::RefCountedBase` with its thread-safe alternative `llvm::ThreadSafeRefCountedBase`. In the original implementation of `rendezvous`, there were some nuances that were undesirable. For example, it's preferred to clear out the entry once all processes have read the data and clear out the map instead of letting the next distribution op to clear it out. The current implementation is also incorrect since there is no guarantee that processGroup size is the same as previous op calling `rendezvous`. ``` if (channels_[channelKey].size() == processGroup.size()) channels_[channelKey].clear(); channels_[channelKey].insert(processId, operand); ``` The proposed update enhances the `rendezvous` logic such that each process has a shared pointer to the result object once out of scope, it is automatically deleted. The PR also reenables tests for distribution ops. closes openxla#1755
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.
The TSAN violation is due to an unprotected insert shown above and replacing
llvm::RefCountedBasewith its thread-safe alternativellvm::ThreadSafeRefCountedBase.In the original implementation of
rendezvous, there were some nuances that were undesirable.For example, it's preferred to clear out the entry once all processes have read the data and clear out the map instead of letting the next distribution op to clear it out. The current implementation is also incorrect since there is no guarantee that processGroup size is the same as previous op calling
rendezvous.The proposed update enhances the
rendezvouslogic such that each process has a shared pointer to the result object once out of scope, it is automatically deleted.The PR also reenables tests for distribution ops.
closes #1755