[jira] [Commented] (FLINK-1597) VertexCentricIterations create inefficient execution plans

["Martin Kiefer (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-1597?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14333162#comment-14333162 ] 

Martin Kiefer commented on FLINK-1597:
--------------------------------------

It appears that the magic word in the execution plan is "Forwad (Cached)". The UDFs are FINISHED before the Iteration begins. I will close the Issue.

> VertexCentricIterations create inefficient execution plans
> ----------------------------------------------------------
>
>                 Key: FLINK-1597
>                 URL: https://issues.apache.org/jira/browse/FLINK-1597
>             Project: Flink
>          Issue Type: Bug
>          Components: Gelly
>    Affects Versions: master
>            Reporter: Martin Kiefer
>
> I did experiments with optimized versions of a graph algorithm that should utilize a secondary sort on the edges and a trade off between superstep numbers and I/O. To my surprise the optimizations did barely affect the execution times. I narrowed it down to inefficient execution plans.
> I assumed that edge sets would be partitioned once at the beginning of a VertexCentricIteration and never be touched again because they can not change during the iteration. I think this should be the desired behavior. What actually happens is that UDFs creating the edge set are pulled inside the iteration and are executed every superstep. This harms the performance of graph algorithms significantly. 
> As a simple example have a look at the execution plan generated for the PageRankExample:
> https://gist.github.com/martinkiefer/28a63f953477e3987b5d



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)