[jira] [Commented] (CASSANDRA-12017) Allow configuration of inter DC compression

["Edward Capriolo (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CASSANDRA-12017?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15485712#comment-15485712 ] 

Edward Capriolo commented on CASSANDRA-12017:
---------------------------------------------

Looking at trunk I think this is the spot in the code.

OutboundTcpConnection.java
{noformat}
    private static void writeHeader(DataOutput out, int version, boolean compressionEnabled) throws IOException
    {
        // 2 bits: unused.  used to be "serializer type," which was always Binary
        // 1 bit: compression
        // 1 bit: streaming mode
        // 3 bits: unused
        // 8 bits: version
        // 15 bits: unused
        int header = 0;
        if (compressionEnabled)
            header |= 4;
        header |= (version << 8);
        out.writeInt(header);
    }
{noformat}

Should we use 3 bits? Create an enum that maps 3 bit numbers to available compression options?

> Allow configuration of inter DC compression 
> --------------------------------------------
>
>                 Key: CASSANDRA-12017
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-12017
>             Project: Cassandra
>          Issue Type: Improvement
>            Reporter: Thom Valley
>             Fix For: 3.x
>
>
> With larger and more extensively geographically distributed clusters, users are beginning to need the ability to reduce bandwidth consumption as much as possible.
> With larger workloads, the limits of even large intercontinental data links (55MBps is pretty typical) are beginning to be stretched.
> InterDC SSL is currently hard coded to use the fastest (not highest) compression settings.  LZ4 is a great option, but being able to raise the compression at the cost of some additional CPU may save as much as 10% (perhaps slightly more depending on the data).  10% of a 55MBps link, if running at or near capacity is substantial.
> This also has a large impact on the overhead and rate possible for instantiating new DCs as well as rebuilding a DC after a failure.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)