[
https://issues.apache.org/jira/browse/HAMA-380?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13021537#comment-13021537
]
Miklos Erdelyi commented on HAMA-380:
-------------------------------------
* You're right that caching constructors minimizes the time needed to
instantiate classes. Thanks for adding this!
* However, in the BSP model there is no guarantee to the order of received
messages so you cannot make assumptions like 'receive 1 Foo class then receive
3 Bar classes'. That is why I originally didn't deal with the situation where
mixed classes might be added to a BSPMessageBundle.
* From a design point of view, you can always embed different message classes
into one container message class so the simpler solution (i.e., to support
buffering instances of only one class in a bundle) still works.
> Send messages in batches to reduce RPC overhead.
> ------------------------------------------------
>
> Key: HAMA-380
> URL: https://issues.apache.org/jira/browse/HAMA-380
> Project: Hama
> Issue Type: Improvement
> Components: bsp
> Affects Versions: 0.3.0
> Reporter: Miklos Erdelyi
> Assignee: Miklos Erdelyi
> Fix For: 0.3.0
>
> Attachments: HAMA-380-v2.patch, bspmsgbundle.patch,
> bspmsgbundle.v2.patch
>
> Original Estimate: 4h
> Remaining Estimate: 4h
>
> When profiling graph computations running on HAMA it turned out that huge
> amount of time is spent in RPC calls.
> To reduce this overhead I propose extending the GroomServer's API with a
> put(BSPMessageBundle) method so that a group of messages can be transferred
> from the caller to the callee at once. Also, the sync() operation should be
> improved by sending messages in batches rather than one-by-one.
> Using BSPMessageBundle makes it easier to add support for compression of
> messages later.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
