[
https://issues.apache.org/jira/browse/CASSANDRA-8887?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14390606#comment-14390606
]
Benedict commented on CASSANDRA-8887:
-------------------------------------
I'm confused by the description of this ticket: how does this affect GC load
(beyond any other long term objects)? At least since LZ4 was made the internode
compression algorithm, there is no extra allocation on the decompression path
(it looks like Snappy was wasteful, so in 2.0 this might be true, but I doubt
we're going to be backporting this to 2.0)
There is still a benefit to be had, but I think we should consider replacing
our compression scheme wholesale, so that the compression algorithm is more
easily swapped in, like we have for files.
> Direct (de)compression of internode communication
> -------------------------------------------------
>
> Key: CASSANDRA-8887
> URL: https://issues.apache.org/jira/browse/CASSANDRA-8887
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Reporter: Matt Stump
> Assignee: Ariel Weisberg
>
> Internode compression is on by default. Currently we allocate one set of
> buffers for the raw data, and then compress which results in another set of
> buffers. This greatly increases the GC load. We can decrease the GC load by
> doing direct compression/decompression of the communication buffers. This is
> the same work as done in CASSANDRA-8464 but applied to internode
> communication.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
