[
https://issues.apache.org/jira/browse/SPARK-5920?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14331986#comment-14331986
]
Patrick Wendell commented on SPARK-5920:
----------------------------------------
We should definitely do this.
> Use a BufferedInputStream to read local shuffle data
> ----------------------------------------------------
>
> Key: SPARK-5920
> URL: https://issues.apache.org/jira/browse/SPARK-5920
> Project: Spark
> Issue Type: Improvement
> Components: Shuffle
> Affects Versions: 1.3.0, 1.2.1
> Reporter: Kay Ousterhout
> Assignee: Kay Ousterhout
> Priority: Blocker
>
> When reading local shuffle data, Spark doesn't currently buffer the local
> reads into larger chunks, which can lead to terrible disk performance if many
> tasks are concurrently reading local data from the same disk. We should use
> a BufferedInputStream to mitigate this problem; we can lazily create the
> input stream to avoid allocating a bunch of in-memory buffers at the same
> time for tasks that read shuffle data from a large number of local blocks.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
