[
https://issues.apache.org/jira/browse/SPARK-24498?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16520056#comment-16520056
]
Takeshi Yamamuro commented on SPARK-24498:
------------------------------------------
The results of my investigation (sf=1 performance values, compile time, max
method codegen size, and max class codegen size in TPCDS) are here:
[https://docs.google.com/spreadsheets/d/1Mgdd9dfFaACXOUHqKfaeKrj09hB3X1j9sKTJlJ6UM6w/edit?usp=sharing]
Their performance values are not so different between each other and Jdk
compile time is much larger than Janino time in q28-q72.
It seems the max method codegen size in Jdk is a little smaller than that in
Janino though, the class size in Jdk is larger than that in Janino.
> Add JDK compiler for runtime codegen
> ------------------------------------
>
> Key: SPARK-24498
> URL: https://issues.apache.org/jira/browse/SPARK-24498
> Project: Spark
> Issue Type: New Feature
> Components: SQL
> Affects Versions: 2.3.0
> Reporter: Xiao Li
> Priority: Major
>
> In some cases, JDK compiler can generate smaller bytecode and take less time
> in compilation compared to Janino. However, in some cases, Janino is better.
> We should support both for our runtime codegen. Janino will be still our
> default runtime codegen compiler.
> See the related JIRAs in DRILL:
> - https://issues.apache.org/jira/browse/DRILL-1155
> - https://issues.apache.org/jira/browse/DRILL-4778
> - https://issues.apache.org/jira/browse/DRILL-5696
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
