[
https://issues.apache.org/jira/browse/SOLR-15840?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17506623#comment-17506623
]
Chris M. Hostetter commented on SOLR-15840:
-------------------------------------------
I'm not sure why Ishan didn't mention it when creating this ticket -- but above
and beyond questions about the "performance" of using the HTTP2, the linked
jetty ticket indicates actual buggy behavior in the client code causing threads
to hang forever.
I've opened SOLR-16099 to track this particular problem as a "Bug" (with a
workaround of forcing http1)
> Performance degradation with Http2 client
> -----------------------------------------
>
> Key: SOLR-15840
> URL: https://issues.apache.org/jira/browse/SOLR-15840
> Project: Solr
> Issue Type: Task
> Affects Versions: 8.10
> Reporter: Ishan Chattopadhyaya
> Priority: Major
>
> The HTTP/2 client for internode communication has some serious bottleneck
> when the system is at a load. For the same scenario, the HTTP1 client
> performs much better.
> Here's how to reproduce the test:
> {code}
> git clone https://github.com/fullstorydev/solr-bench
> mvn clean compile assembly:single
> ./cleanup.sh && ./stress.sh http2-bug.json
> {code}
> ^ This will run a stress test scenario using the HTTP1 client. After running
> this, change this line from http1=true to http1=false and re-run the test:
> https://github.com/fullstorydev/solr-bench/blob/master/http2-bug.json#L92
> I ran this test on my AMD Ryzen 5700G, 64GB RAM machine. It has 8 cores/16
> threads. The test with http1=true passes in about 30-40 minutes. The test
> with http1=false hangs indefinitely, with queries getting timed out all the
> time and the system experiencing some hang/stuck. All queries are sent to the
> node at port 50000, where we're changing the client type.
> Test plan is here:
> https://github.com/fullstorydev/solr-bench/blob/master/http2-bug.json
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
