[ 
https://issues.apache.org/jira/browse/HADOOP-3120?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12583155#action_12583155
 ] 

Pete Wyckoff commented on HADOOP-3120:
--------------------------------------

I am setting it up on a non-hadoop test server, just to ensure it operates the 
way I think it should.  You are right, if we went from 10 microseconds/call to 
something more like ~0.5-1 microsecond that should probably do it -- The 10 
microseconds is measured from a call to LOG.info with a couple of strings and 2 
ints to format (standalone program with idle disk and about a million calls). 
The 0.5 is the same call with info logging disabled and just guestimating the 
overhead of the call to AsyncAppender. Probably can have #s on asyncappender 
performance maybe early next week.
-- pete


> Large #of tasks failing at one time can effectively hang the jobtracker 
> ------------------------------------------------------------------------
>
>                 Key: HADOOP-3120
>                 URL: https://issues.apache.org/jira/browse/HADOOP-3120
>             Project: Hadoop Core
>          Issue Type: Bug
>         Environment: Linux/Hadoop-15.3
>            Reporter: Pete Wyckoff
>            Priority: Minor
>
> We think that JobTracker.removeMarkedTaks does so much logging when this 
> happens (ie logging thousands of failed taks per cycle) that nothing else can 
> go on (since it's called from a synchronized method) and thus by the next 
> cycle, the next waves of jobs have failed and we again have 10s of thousands 
> of failures to log and on and on.
> At least, the above is what we observed - just a continual printing of those 
> failures and nothing else happening on and on. Of course the original jobs 
> may have ultimately failed but new jobs come in to perpetuate the problem.
> This has happened to us a number of times and since we commented out the 
> log.info in that method we haven't had any problems. Although thousands and 
> thousands of task failures are hopefully not that common.

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

Reply via email to