[jira] [Commented] (SPARK-19371) Cannot spread cached partitions evenly across executors

["Sean Owen (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-19371?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15840113#comment-15840113 ] 

Sean Owen commented on SPARK-19371:
-----------------------------------

There's a tension between waiting for locality and taking a free slot that's not local. You can increase spark.locality.wait to prefer locality even at the expense of delays.

If your underlying HDFS data isn't spread out well, it won't help though. You should make sure you have some HDFS replication enabled.

You can try 2x caching replication to make more copies available, obviously at the expense of more memory.

Shuffling is actually a decent idea; one way or the other if the data isn't evenly spread, some if it has to be copied.

I don't think this is a bug per se.

> Cannot spread cached partitions evenly across executors
> -------------------------------------------------------
>
>                 Key: SPARK-19371
>                 URL: https://issues.apache.org/jira/browse/SPARK-19371
>             Project: Spark
>          Issue Type: Bug
>    Affects Versions: 1.6.1
>            Reporter: Thunder Stumpges
>
> Before running an intensive iterative job (in this case a distributed topic model training), we need to load a dataset and persist it across executors. 
> After loading from HDFS and persisting, the partitions are spread unevenly across executors (based on the initial scheduling of the reads which are not data locale sensitive). The partition sizes are even, just not their distribution over executors. We currently have no way to force the partitions to spread evenly, and as the iterative algorithm begins, tasks are distributed to executors based on this initial load, forcing some very unbalanced work.
> This has been mentioned a [number|http://apache-spark-developers-list.1001551.n3.nabble.com/RDD-Partitions-not-distributed-evenly-to-executors-tt16988.html#a17059] of [times|http://apache-spark-user-list.1001560.n3.nabble.com/Spark-work-distribution-among-execs-tt26502.html] in [various|http://apache-spark-user-list.1001560.n3.nabble.com/Partitions-are-get-placed-on-the-single-node-tt26597.html] user/dev group threads.
> None of the discussions I could find had solutions that worked for me. Here are examples of things I have tried. All resulted in partitions in memory that were NOT evenly distributed to executors, causing future tasks to be imbalanced across executors as well.
> *Reduce Locality*
> {code}spark.shuffle.reduceLocality.enabled=false/true{code}
> *"Legacy" memory mode*
> {code}spark.memory.useLegacyMode = true/false{code}
> *Basic load and repartition*
> {code}
> val numPartitions = 48*16
> val df = sqlContext.read.
>     parquet("/data/folder_to_load").
>     repartition(numPartitions).
>     persist
> df.count
> {code}
> *Load and repartition to 2x partitions, then shuffle repartition down to desired partitions*
> {code}
> val numPartitions = 48*16
> val df2 = sqlContext.read.
>     parquet("/data/folder_to_load").
>     repartition(numPartitions*2)
> val df = df2.repartition(numPartitions).
>     persist
> df.count
> {code}
> It would be great if when persisting an RDD/DataFrame, if we could request that those partitions be stored evenly across executors in preparation for future tasks. 
> I'm not sure if this is a more general issue (I.E. not just involving persisting RDDs), but for the persisted in-memory case, it can make a HUGE difference in the over-all running time of the remaining work.



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

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