[
https://issues.apache.org/jira/browse/HADOOP-15548?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16528093#comment-16528093
]
Hudson commented on HADOOP-15548:
---------------------------------
SUCCESS: Integrated in Jenkins build Hadoop-trunk-Commit #14504 (See
[https://builds.apache.org/job/Hadoop-trunk-Commit/14504/])
HADOOP-15548: Randomize local dirs. Contributed by Jim Brennan. (ericp: rev
d36f6b9e93e4c30d24d0e837cb00bd24ffa8f274)
* (edit)
hadoop-common-project/hadoop-common/src/test/java/org/apache/hadoop/fs/TestLocalDirAllocator.java
* (edit)
hadoop-common-project/hadoop-common/src/main/java/org/apache/hadoop/fs/LocalDirAllocator.java
> Randomize local dirs
> --------------------
>
> Key: HADOOP-15548
> URL: https://issues.apache.org/jira/browse/HADOOP-15548
> Project: Hadoop Common
> Issue Type: Bug
> Reporter: Jim Brennan
> Assignee: Jim Brennan
> Priority: Minor
> Attachments: HADOOP-15548.001.patch, HADOOP-15548.002.patch
>
>
> shuffle LOCAL_DIRS, LOG_DIRS and LOCAL_USER_DIRS when launching container.
> Some applications will process these in exactly the same way in every
> container (e.g. roundrobin) which can cause disks to get unnecessarily
> overloaded (e.g. one output file written to first entry specified in the
> environment variable).
> There are two paths for local dir allocation, depending on whether the size
> is unknown or known. The unknown path already uses a random algorithm. The
> known path initializes with a random starting point, and then goes
> round-robin after that. When selecting a dir, it increments the last used by
> one and then checks sequentially until it finds a dir that satisfies the
> request. Proposal is to increment by a random value of between 1 and
> num_dirs - 1, and then check sequentially from there. This should result in
> a more random selection in all cases.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
