[jira] [Commented] (IGNITE-104) Need to enable thread-per-partition mode for atomic caches

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

    [ https://issues.apache.org/jira/browse/IGNITE-104?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14681122#comment-14681122 ] 

Valentin Kulichenko commented on IGNITE-104:
--------------------------------------------

This ticket is implemented, but there is a performance drop around 10% for atomic puts. Need to investigate. What can be done:
* Remove circular buffer. Currently it's required only for rebalancing, therefore is not needed on stable topology.
* Try to remove locking in {{GridDhtAtomicCache}}.

Also need to take a look at backward compatibility - currently it's broken.

> Need to enable thread-per-partition mode for atomic caches
> ----------------------------------------------------------
>
>                 Key: IGNITE-104
>                 URL: https://issues.apache.org/jira/browse/IGNITE-104
>             Project: Ignite
>          Issue Type: Sub-task
>          Components: cache
>            Reporter: Dmitriy Setrakyan
>            Assignee: Valentin Kulichenko
>            Priority: Critical
>
> Currently messages are processed in any order for atomic caches. We can add ordered processing by assigning a thread per partition. This way we will be able to remove any locking as well.
> {{Cache.invoke()}} method would be able to invoke the EntryProcessor on primary and backup independently. Right now we only invoke the EntryProcessor only on the primary node and the send the computed value to the backup node, which may be expensive.



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