[
https://issues.apache.org/jira/browse/CASSANDRA-6409?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13833868#comment-13833868
]
Quentin Conner commented on CASSANDRA-6409:
-------------------------------------------
+1
Moving getEndpointToTokenMapForReading() did the trick. I tested the patch
against cassandra-1.2 branch and the persistent flapping node and high cpu use
symptom disappeared in my 500 node cluster setup (num_tokens=512).
> gossip performance improvement at node startup
> ----------------------------------------------
>
> Key: CASSANDRA-6409
> URL: https://issues.apache.org/jira/browse/CASSANDRA-6409
> Project: Cassandra
> Issue Type: Bug
> Reporter: Quentin Conner
> Attachments: 2013-11-26_17-40-08.png, endpointToTokenMapCPU.txt
>
>
> With large clusters (> 500 nodes) and num_tokens > 255 we sometimes see a
> node have trouble starting up. CPU usage for one thread is pegged.
> We see this concurrent with Gossip flaps on the node trying to learn the ring
> topology. Other nodes on the ring, that are already at steady state do not
> seem to suffer. It is the node joining the large ring that has trouble.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
