[jira] [Commented] (CASSANDRA-6934) Optimise Byte + CellName comparisons

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

    [ https://issues.apache.org/jira/browse/CASSANDRA-6934?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13960445#comment-13960445 ] 

Benedict commented on CASSANDRA-6934:
-------------------------------------

Initial patch available [here|https://github.com/belliottsmith/cassandra/tree/6934]

Optimises to some extent the various compare() implementations for AbstractCType, and at the same time slightly optimises compare in BTree to avoid unwrapping the special +/-Inf values except when absolutely necessary, and to perform (potentially) one fewer comparison per update() when not updating identical values, and to not perform a wasteful start-of-range compare() when _not_ inserting.

There are further performance improvements that can be made to AbstractCType.compare() and its inheritors, but they're a little more invasive, and since CASSANDRA-6694 will entail some optimisation work to make comparisons less expensive, I will wait until then to do anything more.

I need to make some tweaks to stress so I can properly test the impact of this patch on CQL, as there's no easy way to perform inserts of random columns. As shown with CASSANDRA-6553, there is a marked improvement for simple composites, and some quick and dirty benchmarking on my local box for thrift columns with only the general purpose improvements showed a lesser but still marked impact.



> Optimise Byte + CellName comparisons
> ------------------------------------
>
>                 Key: CASSANDRA-6934
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-6934
>             Project: Cassandra
>          Issue Type: Improvement
>            Reporter: Benedict
>            Assignee: Benedict
>              Labels: performance
>             Fix For: 2.1
>
>
> AbstractCompositeType is called a lot, so deserves some heavy optimisation. SimpleCellNameType can be optimised easily, but should explore other potential optimisations.



--
This message was sent by Atlassian JIRA
(v6.2#6252)