[VL] Rss use row-based sort

[marin-ma]

No description provided.

[github-actions]

Thanks for opening a pull request!

Could you open an issue for this pull request on Github Issues?

https://github.com/apache/incubator-gluten/issues

Then could you also rename commit message and pull request title in the following format?

[GLUTEN-${ISSUES_ID}][COMPONENT]feat/fix: ${detailed message}

See also:

• Other pull requests

[github-actions]

Run Gluten Clickhouse CI

[github-actions]

Run Gluten Clickhouse CI

[github-actions]

Run Gluten Clickhouse CI

[github-actions]

Run Gluten Clickhouse CI

[github-actions]

Run Gluten Clickhouse CI

[github-actions]

Run Gluten Clickhouse CI

[marin-ma]

@kerwin-zk Could you help to check whether this patch could work? Besides, I noticed it requires some extra configurations to be set for celeborn. Here's what I found and please help to check if there are anything else needed to fully enable sort-based shuffle with celeborn. Thanks!

--conf spark.shuffle.manager=celeborn \
--conf spark.celeborn.client.spark.shuffle.writer=sort \
--conf spark.celeborn.client.shuffle.compression.codec=zstd

cc: @FelixYBW

[kerwin-zk]

@kerwin-zk Could you help to check whether this patch could work? Besides, I noticed it requires some extra configurations to be set for celeborn. Here's what I found and please help to check if there are anything else needed to fully enable sort-based shuffle with celeborn. Thanks!

--conf spark.shuffle.manager=celeborn \
--conf spark.celeborn.client.spark.shuffle.writer=sort \
--conf spark.celeborn.client.shuffle.compression.codec=zstd

cc: @FelixYBW

@marin-ma I'll try to test it first.

[FelixYBW]

@marin-ma I'll try to test it first.

Is there any more config? I'm testing it in a customer case.

[kerwin-zk]

--conf spark.shuffle.manager=celeborn
--conf spark.celeborn.client.spark.shuffle.writer=sort

@FelixYBW @marin-ma I haven't tested row-based sort + Celeborn yet. For columnar-based sort + Celeborn, the following settings are needed:

spark.shuffle.manager: org.apache.spark.shuffle.gluten.celeborn.CelebornShuffleManager
spark.celeborn.master.endpoints: master-1-1:9097
spark.shuffle.service.enabled: false
spark.celeborn.client.spark.push.sort.memory.threshold: 128m
spark.celeborn.client.spark.shuffle.writer: sort
spark.celeborn.client.shuffle.compression.codec: none
spark.sql.adaptive.localShuffleReader.enabled: false

[FelixYBW]

@floesing_pins you may take a look of this. You may set spark.gluten.sql.columnar.shuffle.sort.partitions.threshold=4000

@jingyuanzhang_pins FYI.

[clay4megtr]

fiele isSort of VeloxCelebornColumnarShuffleWriter need update

[github-actions]

Run Gluten Clickhouse CI on x86

[github-actions]

This PR is stale because it has been open 45 days with no activity. Remove stale label or comment or this will be closed in 10 days.

[github-actions]

This PR was auto-closed because it has been stalled for 10 days with no activity. Please feel free to reopen if it is still valid. Thanks.

[github-actions]

Run Gluten Clickhouse CI on x86

[github-actions]

Run Gluten Clickhouse CI on x86

[github-actions]

Run Gluten Clickhouse CI on x86

[marin-ma]

cc @kerwin-zk

[kerwin-zk]

@marin-ma Since our internal version depends on GLUTEN_RSS_SORT_SHUFFLE_WRITER, I suggest adding a configuration to control whether to use GLUTEN_SORT_SHUFFLE_WRITER or GLUTEN_RSS_SORT_SHUFFLE_WRITER.

[FelixYBW]

@marin-ma Since our internal version depends on GLUTEN_RSS_SORT_SHUFFLE_WRITER, I suggest adding a configuration to control whether to use GLUTEN_SORT_SHUFFLE_WRITER or GLUTEN_RSS_SORT_SHUFFLE_WRITER.

Is your internal version the same as upstream? We picked this up because one customer met OOM issue using upstream. Did you try GLUTEN_SORT_SHUFFLE_WRITER? Is there performance gap?

[kerwin-zk]

@marin-ma Since our internal version depends on GLUTEN_RSS_SORT_SHUFFLE_WRITER, I suggest adding a configuration to control whether to use GLUTEN_SORT_SHUFFLE_WRITER or GLUTEN_RSS_SORT_SHUFFLE_WRITER.

Is your internal version the same as upstream? We picked this up because one customer met OOM issue using upstream. Did you try GLUTEN_SORT_SHUFFLE_WRITER? Is there performance gap?

@FelixYBW The internal version is somewhat different from the upstream version, and ShuffleRead may have memory issues(#9069). I previously tested GLUTEN_SORT_SHUFFLE_WRITER, and there were cases where it failed to process very large datasets. Therefore, I think it would be more appropriate to control whether to use GLUTEN_SORT_SHUFFLE_WRITER or GLUTEN_RSS_SORT_SHUFFLE_WRITER based on a parameter.

[github-actions]

This PR is stale because it has been open 45 days with no activity. Remove stale label or comment or this will be closed in 10 days.

[FelixYBW]

@marin-ma Let's add a config to do the switch.

[FelixYBW]

@kerwin-zk The OOM issue is fixed #9221

[github-actions]

This PR is stale because it has been open 45 days with no activity. Remove stale label or comment or this will be closed in 10 days.

[github-actions]

This PR was auto-closed because it has been stalled for 10 days with no activity. Please feel free to reopen if it is still valid. Thanks.