Remove GPUClock Cost Function #4802
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
ax3l
added
backend: cuda
ax3l
force-pushed
the
topic-rm-gpuclock
branch
7 times, most recently
from
1d6f23d
to
37db926
Compare
Removing the GPUClock cost function due to the following reasons. Incomplete Implementation: The implementation is only added to selected kernels. The implementation is not generalized to work with varying occupancy of different kernels, even if it were used in all kernels. The implementation is verbose. Unused: Our host-side timer implementation was in the last years extended to synchronize kernels at minimal overhead cost. This and heuristic is actually used. Research scope shifted: In the last years, we realized that we do not need more precise scalar cost functions, but instead vector cost functions to build better load balance performance models from. Costly when used: The implementation uses an atomic add of each kernel, instead of, e.g., just using one per warp. This adds severe memory bandwidth strain. Costly, even if not used: The implementation adds about 4 registers unnecessary to all instrumented GPU kernels once compiled in (by default).
RemiLehe
approved these changes
Mar 25, 2024
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Outdated suggestions cannot be applied.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.
Removing the GPUClock cost function due to the following reasons.
Incomplete Implementation:
The implementation is only added to selected kernels. The implementation is not generalized to work with varying occupancy of different kernels, even if it were used in all kernels. The implementation is verbose.
Unused:
Our host-side timer implementation was in the last years extended to synchronize kernels at minimal overhead cost. This and heuristic is actually used.
Research scope shifted:
In the last years, we realized that we do not need more precise scalar cost functions, but instead vector cost functions to build better load balance performance models from.
Costly when used:
The implementation uses an atomic add of each GPU thread, instead of, e.g., just using one per warp or a generalized time that scales with actual GPU time. This adds severe memory bandwidth strain.
Costly, even if not used:
The implementation adds about 4 registers unnecessary to all instrumented GPU kernels once compiled in (by default).
Thanks to @AlexanderSinn for bringing this up again.