[
https://issues.apache.org/jira/browse/HAMA-556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13268974#comment-13268974
]
Thomas Jungblut commented on HAMA-556:
--------------------------------------
It was the wrong patch :/
Anyways, my "hack" seems to be not very accurate for those things, I think we
should add aggregators in this issue and let it be done via a master task that
does that aggregation.
So the graph job has a list of aggregators, and after compute the value of each
vertex will be "observed" or aggregated and after all vertices has been
iterated, we are messaging a master task with the aggregated values.
It applies the aggregation once again and then make it available to the
vertices in the next superstep again.
This way we could also handle the number of updated vertices, it will then just
be a special SumAggregator.
The big fail in this scenario is that we need an additional sync operation for
the master task. I will need some time, maybe we can get the same without it.
> Graph package to support stopping the interations when the node changes are
> within the tolerance value as in the case of page rank
> ----------------------------------------------------------------------------------------------------------------------------------
>
> Key: HAMA-556
> URL: https://issues.apache.org/jira/browse/HAMA-556
> Project: Hama
> Issue Type: New Feature
> Components: bsp core
> Affects Versions: 0.5.0
> Reporter: praveen sripati
> Assignee: Thomas Jungblut
> Priority: Minor
> Fix For: 0.6.0
>
>
> Currently in the graph package, the iteration continues till the nodes are no
> longer updated (absolutely) or the maximum number of iterations has reached.
> It doesn't support testing the node changes with some tolerance and then stop
> the iterations as in the case of page rank.
> The above scenario might be applicable besides page rank also.
> org.apache.hama.graph.GraphJobRunner#bsp()
> while (updated && iteration < maxIteration) {
> ......
> ......
> }
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
