[
https://issues.apache.org/jira/browse/SPARK-43526?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
caican updated SPARK-43526:
---------------------------
Attachment: application_1684208757063_0028_90.html
> when shuffle hash join is enabled, q95 performance deteriorates
> ---------------------------------------------------------------
>
> Key: SPARK-43526
> URL: https://issues.apache.org/jira/browse/SPARK-43526
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 3.1.2, 3.2.0
> Reporter: caican
> Priority: Major
> Attachments: application_1684208757063_0028_90.html,
> image-2023-05-16-21-20-18-727.png, image-2023-05-16-21-21-35-493.png,
> image-2023-05-16-21-22-16-170.png, image-2023-05-16-21-22-44-532.png,
> image-2023-05-16-21-23-33-611.png, image-2023-05-16-21-23-35-237.png,
> image-2023-05-16-21-24-09-182.png, image-2023-05-16-21-28-11-514.png,
> image-2023-05-16-21-28-44-163.png
>
>
> Testing with 5TB dataset, the performance of q95 in tpcds deteriorates when
> shuffle hash join is enabled and the performance is better when sortMergeJoin
> is used.
>
> Performance difference: from 3.9min(sortMergeJoin) to
> 8.1min(shuffledHashJoin)
>
> 1. enable shuffledHashJoin, the execution plan is as follows:
> !image-2023-05-16-21-28-44-163.png|width=935,height=64!
> !image-2023-05-16-21-21-35-493.png|width=924,height=502!
> 2. disable shuffledHashJoin, the execution plan is as follows:
> !image-2023-05-16-21-28-11-514.png|width=922,height=67!
> !image-2023-05-16-21-22-16-170.png|width=934,height=477!
>
> And when shuffledHashJoin is enabled, gc is very serious,
> !image-2023-05-16-21-23-35-237.png|width=929,height=570!
>
> but sortMergeJoin executes without this problem.
> !image-2023-05-16-21-24-09-182.png|width=931,height=573!
>
> Any suggestions on how to solve it？Thanks!
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
