flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Greg Hogan <c...@greghogan.com>
Subject Re: Gelly - which partitioning
Date Wed, 29 Mar 2017 13:50:58 GMT
Hi Marc,

I’ll defer to Vasia’s comment below from FLINK-1536 as she has much more knowledge and
experience with graph partitioning. This is certainly an area of interest so please let us
know if you would like to contribute!

The referenced list of papers is at:
  http://www.citeulike.org/user/vasiakalavri/tag/graph-partitioning

"In a distributed graph processing system, you first have to partition the graph before you
perform any computation. The performance of graph algorithms greatly depends on the resulting
partitioning. A bad partitioning might assign disproportionally more vertices to one partition
thus hurting load balancing or it might partition the graph so that the communication required
is too high (or both). Currently, we only support hash partitioning; that is, vertices are
randomly assigned to workers using the hash of their id. This strategy has very low overhead
and results in good load balancing unless the graphs are skewed.”

https://issues.apache.org/jira/browse/FLINK-1536?focusedCommentId=15733284&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15733284

Greg


> On Mar 29, 2017, at 5:50 AM, Kaepke, Marc <marc.kaepke@haw-hamburg.de> wrote:
> 
> Hi guys,
> 
> I can’t found on web which graph partitioning are supported by Gelly.
> 
> During my search I found this link. But the ticket is still open.
> https://cwiki.apache.org/confluence/display/FLINK/Flink+Gelly <https://cwiki.apache.org/confluence/display/FLINK/Flink+Gelly>
> 
> 
> Thanks for help!
> 
> Best
> Marc


Mime
View raw message