[ 
https://issues.apache.org/jira/browse/BEAM-13073?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17447830#comment-17447830
 ] 

Luis commented on BEAM-13073:
-----------------------------

Ow. That is unfortunate. As described, the context of this change is that Beam 
SDK (and Dataflow) with Java11 was using SerialGC, which caused performance 
issues when migrating from Java 8 (that uses ParallelGC).

I am not (yet) familiar with these benchmark jobs. I validated the changes in 
some ad-hoc benchmark jobs, but the results were not very reproducible.

I agree that ParallelGC sounds more appropriate here. I can looking into 
opening a PR for that.

> Unexpected GC when using Java 11
> --------------------------------
>
>                 Key: BEAM-13073
>                 URL: https://issues.apache.org/jira/browse/BEAM-13073
>             Project: Beam
>          Issue Type: Bug
>          Components: sdk-java-harness
>            Reporter: Luis
>            Assignee: Kenneth Knowles
>            Priority: P1
>              Labels: java11, java9, performance
>             Fix For: 2.35.0
>
>         Attachments: perf_regression_java_11.png
>
>          Time Spent: 0.5h
>  Remaining Estimate: 0h
>
> Beam SDK has been supporting Java 11 for a while (I guess the support was 
> introduced here https://issues.apache.org/jira/browse/BEAM-2530). 
> Unfortunately, in Spotify we are still experiencing performance issues when 
> using Beam SDK 2.32, Google Dataflow and Java 11.
> Thanks to [~emilyye] and [~iht], they confirmed JVM 11 is using SerialGC, 
> while Java 8 uses ParallelGC. It sounds like ParallelGC is a good option for 
> high throughput / low latency jobs. For Java11 we'd expect to use G1GC or 
> ParallelGC.
> This SO question [1] clarifies that JVM chooses SerialGC when it treats the 
> machine as a "client". It looks like the Java SDK container could benefit 
> from using `-XX:+AlwaysActAsServerClassMachine`. Is that correct?
> Let me know if the ticket needs further context or adjustment. (It is my 
> first time creating a ticket here).
>  [1] 
> [https://stackoverflow.com/questions/52474162/why-is-serialgc-chosen-over-g1gc]



--
This message was sent by Atlassian Jira
(v8.20.1#820001)

Reply via email to