Don't acquire the semaphore for empty input while scanning #4476
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
Signed-off-by: Alessandro Bellina <abellina@nvidia.com>
|
hmm looks like the build is failing on not being able to create the docker container. |
jlowe
reviewed
Jan 10, 2022
sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuMultiFileReader.scala
Show resolved
Hide resolved
sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuParquetScanBase.scala
Show resolved
Hide resolved
jlowe
previously approved these changes
Jan 10, 2022
|
build |
|
I ran another iteration of NDS, with similar results. Most queries are slightly faster, with some more significant gains in those single digit queries like q2 q53 q55 where we are gaining up to 3 seconds. Overall I see ~42 seconds over all of NDS at 2TB. |
revans2
reviewed
Jan 10, 2022
sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuMultiFileReader.scala
Outdated
Show resolved
Hide resolved
abellina
changed the title
jlowe
approved these changes
Jan 11, 2022
|
build |
jbrennan333
approved these changes
Jan 11, 2022
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Outdated suggestions cannot be applied.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.
Signed-off-by: Alessandro Bellina abellina@nvidia.com
This PR removes instances where we acquired the GpuSemaphore within the scan code, specifically those scans that are
PartitionReaders. The change is to stop acquiring the semaphore here, because these cases will not be consumed downstream, given how thePartitionReaderis used.PartitionReaderprovides anext()method that returns false when there isn't work left. ThePartitionReaderIteratorwraps these readers, as far as I can tell, and it exposes the iterator interface (hasNext()returns false if thePartitionReaderreturned a false innext()). This allows us to stop requiring us to acquire the semaphore in thenext()method, when there isn't any work left in the reader.The PR also adds an acquire in the aggregate where we are producing rows out of nothing. This should be a noop, and likely just a miss that it wasn't there before (protected by upstream nodes doing the acquire for the agg).
The change has modest gains in NDS so far. I want to run a few more times, but so far I see: 10 queries that are 1.20x+ faster, 39 queries that are 1.10x+, and most queries (92) are above 1. I don't see a lot of regression here, except for q94 with 2 seconds, but q94 is falling back to the CPU and can be unpredictable. In other words, it seems to be mostly good, but I want to run a few more tests.
@revans2 wanted comments explaining why we are not acquiring the semaphore. I wasn't sure where to put them, and they are the same each time. I could see us moving them to the
PartitionReaderIteratoror somewhere else more appropriate.