Allow batching the output of a join #2310
Merged
+1,193
−397
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: Robert (Bobby) Evans <[email protected]>
revans2
added
feature request
revans2
commented
Apr 29, 2021
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
|
I have been adding in spilling for the gather maps which let me push things a bit further and found a bug in the gather implementation. |
abellina
requested changes
Apr 30, 2021
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
jlowe
reviewed
Apr 30, 2021
...park300/src/main/scala/com/nvidia/spark/rapids/shims/spark300/GpuBroadcastHashJoinExec.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
abellina
requested changes
Apr 30, 2021
sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuBoundAttribute.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Outdated
Show resolved
Hide resolved
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Show resolved
Hide resolved
|
I think I have addressed all of the review comments. I put together my own quick hack for rapidsai/cudf#8121 and it is not enough to be able to run q72. Even at a batch size of 26m and 200 partitions it took over 6 mins to finish one of the join tasks our of 200 and failed on the next one. We are going to have to really think about what we want to try and do to support query 72. But all of the others run with reasonable configurations. |
jlowe
reviewed
May 1, 2021
...park300/src/main/scala/com/nvidia/spark/rapids/shims/spark300/GpuBroadcastHashJoinExec.scala
Show resolved
Hide resolved
jlowe
reviewed
May 3, 2021
sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuShuffledHashJoinBase.scala
Show resolved
Hide resolved
abellina
reviewed
May 3, 2021
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuHashJoin.scala
Show resolved
Hide resolved
Tests on not passing with struct joins need to do some more debugging
|
build |
|
I upmerged and had to update the code for the new struct join support. A good thing too because it exposed a bug in my filtering code. It would only have been a performance regression before the struct code, but afterwards it became an error. This should be all ready to go now. The dependency is merged. |
jlowe
approved these changes
May 3, 2021
abellina
approved these changes
May 3, 2021
This was referenced May 3, 2021
nartal1
pushed a commit
to nartal1/spark-rapids
that referenced
this pull request
Jun 9, 2021
Signed-off-by: Robert (Bobby) Evans <[email protected]>
nartal1
pushed a commit
to nartal1/spark-rapids
that referenced
this pull request
Jun 9, 2021
Signed-off-by: Robert (Bobby) Evans <[email protected]>
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.
This is the first step for out of core join. This at least partially addresses #20
This depends on rapidsai/cudf#8118 to go in first.
For most cases that I have tested this is strictly better than what was before. If the output of the join fits in the output batch size then the join will happen just like it does today. If the output is larger than that we now can output it in multiple batches. The problem that I have found is that the gather map is not spillable and after a single batch is output the GPU Semaphore is released. This means that for contrived joins that explode evenly, each active task will have a potentially large gather map in memory. I think I can make it spillable without a lot of work. If I can then I might just do it. But I also want to spend some time running benchmarks to see if this can help fix some of the exploding join issues have have seen there.