[
https://issues.apache.org/jira/browse/HAMA-369?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13008376#comment-13008376
]
Miklos Erdelyi commented on HAMA-369:
-------------------------------------
According to measurements with RandBench (repeated 20 times, 2048 64k messages
sent by each peer in 1 superstep) this patch gains 3.17% in terms of job
performance which should encourage us to include it in 0.2. In other use cases
the gain might be much more, i.e., in case of graph algorithms.
Also, separating the local and remote message queues is a step towards
implementing asynchronous messaging during the computation phase, i.e.,
messages received in the latter phase can be accumulated in the local queue for
the next BSP iteration while computation is running using the local queue for
the current BSP iteration.
Are the above points enough reasons to include this patch in 0.2 or shall I
provide more? ;-)
> Do not buffer messages intended for the local groom the same way as messages
> intended for remote grooms. Instead, append these messages directly to the
> local message queue.
> ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: HAMA-369
> URL: https://issues.apache.org/jira/browse/HAMA-369
> Project: Hama
> Issue Type: Improvement
> Components: bsp
> Affects Versions: 0.2.0
> Reporter: Miklos Erdelyi
> Fix For: 0.2.0
>
> Attachments: localsend.patch
>
> Original Estimate: 1h
> Remaining Estimate: 1h
>
> Currently BSPPeer processes all sent messages the same way. Those which are
> sent to the local groom should be delivered directly to the respective
> message queue without calling RPC functions or buffering unnecessarily.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
