[
https://issues.apache.org/jira/browse/SPARK-59779?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
ASF GitHub Bot updated SPARK-59779:
-----------------------------------
Labels: pull-request-available (was: )
> Document spark.sql.codegen.hugeMethodLimit as a tuning knob, or make it public
> ------------------------------------------------------------------------------
>
> Key: SPARK-59779
> URL: https://issues.apache.org/jira/browse/SPARK-59779
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 5.0.0
> Reporter: Max Gekk
> Priority: Major
> Labels: pull-request-available
>
> spark.sql.codegen.hugeMethodLimit is the only setting that acts on the JIT
> cliff: HotSpot does not compile a method past 8000 bytes
> (-XX:+DontCompileHugeMethods, on by default), so a whole-stage codegen stage
> whose per-row method crosses it runs interpreted, typically several times
> slower. Setting the limit to 8000 makes such a stage fall back to the
> non-whole-stage path instead, and the config's own doc already suggests that
> value on HotSpot.
> The config is .internal(), yet SPARK-59774 names it in a user-facing WARN as
> the remedy. Internal configs already appear as tuning knobs in
> docs/sql-performance-tuning.md (spark.sql.files.openCostInBytes,
> spark.sql.sources.parallelPartitionDiscovery.parallelism,
> spark.sql.requireAllClusterKeysForCoPartition), so there is precedent either
> way.
> Proposal: either make spark.sql.codegen.hugeMethodLimit public, or keep it
> internal and list it in the configuration tables of sql-performance-tuning.md
> with a short explanation of the cliff and when to set it to 8000.
> Context: https://github.com/apache/spark/pull/59020
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
