Home Forums MapReduce HDP 1.3 JobTracker hitting Out of Memory

This topic contains 11 replies, has 5 voices, and was last updated by  Koelli Mungee 10 months, 3 weeks ago.

  • Creator
    Topic
  • #38131

    We are using HDP 1.3, and JT is Version: 1.2.0.1.3.0.0-107.
    We are seeing a steady growth of the JobTracker heap , and finally it hits Java heap space error.

    The below are the cluster specs;
    Nodes – 4
    Map Tasks – 32
    Reduce Tasks – 12
    Job Submissions a day = 2000 (around 100 jobs out of this runs with 250+ maps tasks)
    Job Tracker heap = 2G
    mapred.job.tracker.retiredjobs.cache.size=30

    With some good amount of jobs,the JT is lasting hardly for 3-4 days.
    Looking at the heap, there are like many instances ~8K of DFSClient which gets me suspicous.

    Any help appreciated.

Viewing 11 replies - 1 through 11 (of 11 total)

The topic ‘HDP 1.3 JobTracker hitting Out of Memory’ is closed to new replies.

  • Author
    Replies
  • #41715

    Koelli Mungee
    Moderator

    Thanks for the update Vivek.

    Jason, yes, MAPREDUCE-5351 exists in HDP 1.3.0 and shipped in HDP 1.3.2

    -Koelli

    Collapse
    #41548

    Hi All,

    Please find my JOBTRACKER OPTS;
    HADOOP_JOBTRACKER_OPTS=”-server -XX:ParallelGCThreads=8 -XX:+UseConcMarkSweepGC -XX:ErrorFile=/var/log/hadoop/$USER/hs_err_pid%p.log -XX:NewSize=200m -XX:MaxNewSize=200m -Xloggc:/var/log/hadoop/$USER/gc.log-`date +’%Y%m%d%H%M’` -verbose:gc -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintGCDateStamps -Xmx4096m -Dhadoop.security.logger=INFO,DRFAS -Dmapred.audit.logger=INFO,MRAUDIT -Dhadoop.mapreduce.jobsummary.logger=INFO,JSA ${HADOOP_JOBTRACKER_OPTS}”

    I have verified that the problem is with MAPREDUCE-5351, with below two params the JT is no longer having any issues.

    keep.task.files.pattern
    dummytextthatnevermatches

    keep.failed.task.files
    true

    The JT is still running with > 10K job submissions , out of which 3K is having around 250 Map Tasks each. And heap used so far is 360MB :)

    Koelli, Thanks for letting me know this.I am on HDP 1.3.0 Let me checkout HDP 1.3.2. Since we are already on production need to see how to upgrade and test.

    Thanks
    Vivek

    Collapse
    #41312

    Jason Morse
    Participant

    We just had an oom on our job tracker as well. Is this a bug in 1.3.0?

    Collapse
    #40310

    Koelli Mungee
    Moderator

    Hi Vivek,

    Thanks, please let us know how you have verified that it is the same issue. We did ship https://issues.apache.org/jira/browse/MAPREDUCE-5351 in HDP 1.3.2. I believe you are on HDP 1.3.0.

    Can you also paste in the HADOOP_JOBTRACKER_OPTS?

    Regards
    Koelli

    Collapse
    #40271

    Hi,
    So even with all these conf set we are still hitting the OOM. Is this incident reported
    To me from the heap dump this looks like https://issues.apache.org/jira/browse/MAPREDUCE-5351
    (I am verifying this)
    Will there be any HDP release which will have this fix ?

    Thanks
    Vivek

    Collapse
    #39739

    abdelrahman
    Moderator

    Hi Vivek,

    There is no implication on mapreduce.job.restart.recover = true that I know off. And the configuration ” mapred.jobtracker.complete.maximum.” is a typo.

    Thanks
    -Abdelrahman

    Collapse
    #39385

    Hi Abdelrahman,
    Thanks for the suggestion , but looks like out defaults are of the value that you have mentioned. The below are the confs
    mapred.jobtracker.retirejob.interval = 21600000
    mapred.job.tracker.retiredjobs.cache.size = 30
    mapred.jobtracker.completeuserjobs.maximum = 5
    mapred.job.tracker.jobhistory.lru.cache.size 5

    Sorry I couldn’t figure out the value for the property mapred.jobtracker.complete.maximum.
    Is there any implication on mapreduce.job.restart.recover = true (default value)

    Thanks
    Vivek

    Collapse
    #39363

    abdelrahman
    Moderator

    Hi Vivek,

    Please configure mapred.jobtracker.retirejob.interval to no more that one day 21600000 (24x60x60x1000) or even less. And mapred.jobtracker.complete.maximum should be set to a low number as well. In addition, mapred.jobtracker.completeuserjobs.maximum should be set to no more than 5 as this property is on a user level.

    Thanks
    -Abdelrahman

    Collapse
    #39227

    Anyone can suggest anything on this ?

    Collapse
    #39181

    I am using manual. The initial Xmx was 2G, for which it hit OOM. Increasing it to 4G also results in an OOM.
    There are tens and thousands of DFSClients lying in the Jobtracker heap.

    Collapse
    #38242

    Robert Molina
    Moderator

    Hi Vivek,
    Are you using ambari installation or manual? If manual see if increasing the Xmx for jobtracker in the hadoop-env.sh file helps.

    Kind Regards,
    Robert

    Collapse
Viewing 11 replies - 1 through 11 (of 11 total)