[
https://issues.apache.org/jira/browse/HADOOP-4780?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12654327#action_12654327
]
Joydeep Sen Sarma commented on HADOOP-4780:
-------------------------------------------
btw - i looked at the jobcache dirs on our machines. they aren't that big. a ls
-R | wc -l - indicates about 1.5K files per drive in mapred.local.dir/jobcache
- and there are 4 drives. i am having a hard time believing that we can spend
10 minutes or so in the current getDU implementation.
one thing i noticed is that there are lots of symlinks in the jobcache - and
that the getDu code does not distinguish between normal dirs and symlinks.
based on java documentation - the current code will follow through symlinks i
think. wondering whether not traversing symlinks will fix the problem? (take it
with a grain of salt - i reallly don't understand this code all that much)
> Task Tracker burns a lot of cpu in calling getLocalCache
> ---------------------------------------------------------
>
> Key: HADOOP-4780
> URL: https://issues.apache.org/jira/browse/HADOOP-4780
> Project: Hadoop Core
> Issue Type: Bug
> Components: mapred
> Affects Versions: 0.18.2
> Reporter: Runping Qi
> Attachments: 4780.patch
>
>
> I noticed that many times, a task tracker max up to 6 cpus.
> During that time, iostat shows majority of that was system cpu.
> That situation can last for quite long.
> During that time, I saw a number of threads were in the following state:
> java.lang.Thread.State: RUNNABLE
> at java.io.UnixFileSystem.getBooleanAttributes0(Native Method)
> at
> java.io.UnixFileSystem.getBooleanAttributes(UnixFileSystem.java:228)
> at java.io.File.exists(File.java:733)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:399)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at org.apache.hadoop.fs.FileUtil.getDU(FileUtil.java:407)
> at
> org.apache.hadoop.filecache.DistributedCache.getLocalCache(DistributedCache.java:176)
> at org.apache.hadoop.mapred.TaskRunner.run(TaskRunner.java:140)
> I suspect that getLocalCache is too expensive.
> And calling it for every task initialization seems too much waste.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
