hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Tsz Wo (Nicholas), SZE (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HADOOP-3762) Task tracker died due to OOM
Date Wed, 16 Jul 2008 21:21:31 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-3762?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12614110#action_12614110

Tsz Wo (Nicholas), SZE commented on HADOOP-3762:

For example, calls FileSystem.get("hdfs://host:8020/", conf), then it looks up the cache with
username (may be null since user might not have logged in yet) and the uri "hdfs://host:8020/"
as a key.  Suppose it fails, it will create a file system object (depending on the FileSystem
class, user might login) and put it to the cache with a new key.  The uri in the new key will
be "hdfs://host" (no port) since 8020 is the DEFAULT_PORT.  Then, if there is another call
to FileSystem.get("hdfs://host:8020/", conf), it won't hit the cache.

> Task tracker died due to OOM 
> -----------------------------
>                 Key: HADOOP-3762
>                 URL: https://issues.apache.org/jira/browse/HADOOP-3762
>             Project: Hadoop Core
>          Issue Type: Bug
>            Reporter: Runping Qi
>            Assignee: Tsz Wo (Nicholas), SZE
>            Priority: Blocker
>             Fix For: 0.18.0
>         Attachments: 3762_20080715.patch, 3762_20080715b.patch, 3762_20080715c.patch,
> When running about 100 moderate jobs on a small cluster (with 19 Task Trackers),
> the task trackers all died due to OOM.
> I got a chance to dump the jstack strace of a task tracker before it died.
> Its image size was close 4GB!
> I saw 1200+ threads of DFSClient.LeaseChecker.
> Clearly we have a severe resource leakage problem!

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message