[jira] [Commented] (HAMA-843) Message communication overhead between master aggregation and vertex computation supersteps

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

    [ https://issues.apache.org/jira/browse/HAMA-843?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13869331#comment-13869331 ] 

Edward J. Yoon commented on HAMA-843:
-------------------------------------

Here's PageRank performance test on 8 thousand vertices graph using Single machine.

patch applied version: 24 secs.
TRUNK version: 32 secs.

> Message communication overhead between master aggregation and vertex computation supersteps
> -------------------------------------------------------------------------------------------
>
>                 Key: HAMA-843
>                 URL: https://issues.apache.org/jira/browse/HAMA-843
>             Project: Hama
>          Issue Type: Improvement
>          Components: graph
>    Affects Versions: 0.6.3
>            Reporter: Edward J. Yoon
>             Fix For: 0.7.0
>
>         Attachments: HAMA-843.patch
>
>
> Within doAggregationUpdates() method, we sends unconsumed messages to next superstep using send() method. This is huge overhead.
> {code}
>     // in case we need to sync, we need to replay the messages that already
>     // are added to the queue. This prevents loosing messages when using
>     // aggregators.
>     if (firstVertexMessage != null) {
>       peer.send(peer.getPeerName(), firstVertexMessage);
>     }
>     GraphJobMessage msg = null;
>     while ((msg = peer.getCurrentMessage()) != null) {
>       peer.send(peer.getPeerName(), msg);
>     }
> {code}
> Once HAMA-842 is done, we can get rid of this overhead.



--
This message was sent by Atlassian JIRA
(v6.1.5#6160)