[
https://issues.apache.org/jira/browse/CASSANDRA-15810?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17176876#comment-17176876
]
Robert Stupp edited comment on CASSANDRA-15810 at 8/13/20, 9:36 AM:
--------------------------------------------------------------------
Yea - I think, the biggest "consumer" is probably the list of tokens, which is
also quite unique per node. So no savings for that one.
Potential saving might happen for "release-version", "dc", "rack" and current
"schema-version" - not particularly long strings so I doubt that it's actually
worth the {{String.intern()}} as the total savings are rather low. The other
values of {{VersionedValue}} are unique per peer and put unnecessary pressure
on the string-intern-map.
Other usages of {{String.intern()}} in the whole production code (i.e.
including JDK and libraries) are (intentionally) not on a hot-path.
This change feels IMO safe for 4.0 and also safe to be backported to 3.11 + 3.0.
was (Author: snazy):
Yea - I think, the biggest "consumer" is probably the list of tokens, which is
also quite unique per node. So no savings for that one.
Potential saving might happen for "release-version", "dc", "rack" and current
"schema-version" - not particularly long strings so I doubt that it's actually
not worth the {{String.intern()}} as the total savings are rather low. The
other values of {{VersionedValue}} are unique per peer and put unnecessary
pressure on the string-intern-map.
Other usages of {{String.intern()}} in the whole production code (i.e.
including JDK and libraries) are (intentionally) not on a hot-path.
This change feels IMO safe for 4.0 and also safe to be backported to 3.11 + 3.0.
> Default StringTableSize parameter causes GC slowdown
> ----------------------------------------------------
>
> Key: CASSANDRA-15810
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15810
> Project: Cassandra
> Issue Type: Improvement
> Components: Local/Config
> Reporter: Tom van der Woerdt
> Priority: Normal
> Labels: gc, performance
>
> While looking at tail latency on a Cassandra cluster, it came up that the
> default StringTableSize in Cassandra is set to a million:
> {code:java}
> # Larger interned string table, for gossip's benefit (CASSANDRA-6410)
> -XX:StringTableSize=1000003{code}
> This was done for CASSANDRA-6410 by [~jbellis] in '13, to optimize heap usage
> on a test case, running with 500 nodes and num_tokens=512.
> Until Java 13, this string table is implemented as native code, and has to be
> traversed entirely during the GC initial marking phase, which is a STW event.
> Some testing on my end shows that the pause time of a GC cycle can be reduced
> by approximately 10 milliseconds if we lower the string table size back to
> the Java 8 default of 60013 entries.
> Thus, I would recommend this patch (3.11 branch, similar patch for 4.0):
> {code:java}
> diff --git a/conf/jvm.options b/conf/jvm.options
> index 01bb1685b3..c184d18c5d 100644
> --- a/conf/jvm.options
> +++ b/conf/jvm.options
> @@ -107,9 +107,6 @@
> # Per-thread stack size.
> -Xss256k
> -# Larger interned string table, for gossip's benefit (CASSANDRA-6410)
> --XX:StringTableSize=1000003
> -
> # Make sure all memory is faulted and zeroed on startup.
> # This helps prevent soft faults in containers and makes
> # transparent hugepage allocation more effective.
> {code}
> It does need some testing on more extreme clusters than I have access to, but
> I ran some Cassandra nodes with {{-XX:+PrintStringTableStatistics}} which
> suggested that the Java default will suffice here.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
