[jira] [Resolved] (SPARK-19098) Shuffled data leak/size doubling in ConnectedComponents/Pregel iterations

["Hyukjin Kwon (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/SPARK-19098?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Hyukjin Kwon resolved SPARK-19098.
----------------------------------
    Resolution: Incomplete

> Shuffled data leak/size doubling in ConnectedComponents/Pregel iterations
> -------------------------------------------------------------------------
>
>                 Key: SPARK-19098
>                 URL: https://issues.apache.org/jira/browse/SPARK-19098
>             Project: Spark
>          Issue Type: Bug
>          Components: GraphX
>    Affects Versions: 2.0.2, 2.1.0
>         Environment: Linux x64
> Cloudera CDH 5.8.0 hadoop (roughly hadoop 2.7.0)
> Spark on YARN, dynamic allocation with shuffle service
> Input/Output data on HDFS
> kryo serialization turned on
> checkpointing directory set on HDFS
>            Reporter: Steven Ruppert
>            Priority: Minor
>              Labels: bulk-closed
>         Attachments: Screen Shot 2017-01-30 at 18.36.43-fullpage.png, doubling-season.png
>
>
> I'm seeing a strange memory-leak-but-not-really problem in a pretty vanilla ConnectedComponents use, notably one that works fine with identical code on spark 2.0.1, but not on 2.1.0.
> I unfortunately haven't narrowed this down to a test case yet, nor do I have access to the original logs, so this initial report will be a little vague. However, this behavior as described might ring a bell to somebody.
> Roughly: 
> {noformat}
> val edges: RDD[Edge[Int]] = _ // from file
> val vertices: RDD[(VertexId, Int)] = _ // from file
> val graph = Graph(vertices, edges)
> val components: RDD[(VertexId, ComponentId)] = ConnectedComponents
>   .run(graph, 10)
>   .vertices
> {noformat}
> Running this against my input of ~5B edges and ~3B vertices leads to a strange doubling of shuffle traffic in each round of Pregel (inside ConnectedComponents), increasing from the actual data size of ~50 GB, to 100GB, to 200GB, all the way to around 40TB before I killed the job. The data being shuffled was apparently an RDD of ShippableVertexPartition .
> Oddly enough, only the kryo-serialized shuffled data doubled in size. The heap usage of the executors themselves remained stable, or at least did not account 1 to 1 for the 40TB of shuffled data, for I definitely do not have 40TB of RAM. Furthermore, I also have kryo reference tracking turned on still, so whatever is leaking somehow gets around that.
> I'll update this ticket once I have more details, unless somebody else with the same problem reports back first.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org