You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@giraph.apache.org by "Alessandro Presta (JIRA)" <ji...@apache.org> on 2013/01/03 20:16:12 UTC
[jira] [Commented] (GIRAPH-459) Group Vertex Mutations by Partition
ID
[ https://issues.apache.org/jira/browse/GIRAPH-459?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13543204#comment-13543204 ]
Alessandro Presta commented on GIRAPH-459:
------------------------------------------
It would be way easier to review this on ReviewBoard, although I think we're still having issues after the switch to git.
A few comments:
- Please restore the original indentation of createMessageStoreFactory().
- Just curious: why did you change the ordering, meaning you now process mutation requests first, then messages?
- When checking for vertices that have received messages, you're loading each partition no matter what. I would first call getPartitionDestinationVertices(partitionId) and then, if it's not empty, load the partition and iterate.
- Nothing to do with your patch, but can you also change "prepareSuperstep" to "resolveMutations" in the log message?
- Can you run a quick benchmark (e.g., PageRankBenchmark with edge input) to make sure there are no performance regressions?
Overall looks good!
> Group Vertex Mutations by Partition ID
> --------------------------------------
>
> Key: GIRAPH-459
> URL: https://issues.apache.org/jira/browse/GIRAPH-459
> Project: Giraph
> Issue Type: Improvement
> Components: graph
> Reporter: Claudio Martella
> Attachments: GIRAPH-459.patch
>
>
> Currently, vertex mutations, and implicit creations of vertices based on messages to non-existing vertices, are executed randomly partition-wise. The iterated vertices can belong to different partitions. This is bad when we work out-of-core, as we need to load and unload the whole partition for each vertex. We should group these operations per-partition, and batch-execute them.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira