Use hostdevice_vector in kernel_error to avoid the pageable copy
#15140
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
vuule
added
cuIO
@abellina suspects this might be the case. It's possible that the shared bounce buffer used for pageable copies creates a bottleneck only in multi-threaded use case. This could explain why the regression is specific to spark. I opened the PR to enable @abellina to test the hypothesis. |
|
@vuule @GregoryKimball we will review/test this, sorry for the delayed response. |
|
@vuule I have done some testing with this but I need a bit more time. So far I think with the error code checking plus the in This: Is 10 seconds faster than: Which is much better than what I had measured before (with pageable we had a 20 seconds or 5% regression). I also want to not do that last |
|
@vuule here are my updates so far:
|
Correct. The synchronization there is needed because the null count code needs the updated data from the page and nesting info bufs that gets computed in the kernel. The |
But the call to |
Sorry the funny business is in my end. I also didn't call |
What if we change |
|
Got an extra data point by using Comparing it with a baseline that is the pageable version of this (the code before this PR), it's essentially unchanged overall. Some queries are faster, and some are slower. Query 9 is 9% faster. Details: |
Re-reading the comment again, I think it makes sense now, so I am not sure we need to reword. Hope we have enough data for this PR. If you need other tests let me know. |
…perf-hd_v-kernel_error
|
No measurable performance impact on libcudf |
…into perf-hd_v-kernel_error
…perf-hd_v-kernel_error
|
I cooked the |
|
LGTM! Thanks! |
etseidl
reviewed
Mar 1, 2024
…perf-hd_v-kernel_error
| */ | ||
| [[nodiscard]] std::string str() const | ||
| [[nodiscard]] static std::string to_string(value_type value) | ||
| { |
There was a problem hiding this comment.
Instead of static, shouldn't this be a free function instead of a member function?
There was a problem hiding this comment.
It could be, but I like how it reads at the call site: kernel_error::to_string(error) and it does not really change much else about the function.
I'm not insisting on this option, so, open to suggestions :)
hyperbolic2346
approved these changes
Mar 5, 2024
There was a problem hiding this comment.
I typically don't like to see if (auto ret = foo(); ret == true) type stuff, but I'll give you pass on this usage, @vuule
Yeah, it would be pretty pointless if not for the repeated synchronization this avoids. |
pmattione-nvidia
approved these changes
Mar 5, 2024
|
/merge |
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.
Description
Issue #15122
The addition of kernel error checking introduced a 5% performance regression in Spark-RAPIDS. It was determined that the pageable copy of the error back to host caused this overhead, presumably because of the CUDA's bounce buffer bottleneck.
This PR aims to eliminate most of the error checking overhead by using
hostdevice_vectorin thekernel_errorclass. Thehostdevice_vectoruses pinned memory so the copy is no longer pageable. The PR also removes the redundant sync after we read the error.Checklist