[
https://issues.apache.org/jira/browse/CASSANDRA-8714?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14347435#comment-14347435
]
Ariel Weisberg commented on CASSANDRA-8714:
-------------------------------------------
Looking at the micro benchmark it isn't measuring the right thing. The same
call site has to potentially contain direct and non-direct buffers. The JVM
also does JVM wide class hierarchy analysis, but I don't know what it does for
call site specific analysis. There are actually several potential direct and
non-direct byte buffer classes. Take a look at the type hierarchy for
ByteBuffer it's pretty long.
This would only be a factor for very small copies Say 4-16 bytes or similar.
Say when serializing variable length fields that can be very small. You would
end up calling this which is basically a memcpy equivalent, and that is where
the overhead of starting the memcpy matters.
> row-cache: use preloaded jemalloc w/ Unsafe
> -------------------------------------------
>
> Key: CASSANDRA-8714
> URL: https://issues.apache.org/jira/browse/CASSANDRA-8714
> Project: Cassandra
> Issue Type: Sub-task
> Reporter: Robert Stupp
> Assignee: Robert Stupp
> Fix For: 3.0
>
> Attachments: 8714-2.txt, 8714-3.txt, 8714-4.txt, 8714.txt
>
>
> Using jemalloc via Java's {{Unsafe}} is a better alternative on Linux than
> using jemalloc via JNA.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
