It's quite possible that's the problem. I'll re-run the tests over night and collect the run times according to the JobTracker.
If I want to test the patch in HADOOP-4780, should I pull down branch-0.19 and go from there? This is not a production environment, so I'm not worried about data loss or other problems. But I could certainly wait if 0.19.2 or 0.20.0 are expected to be released soon, if that would make more sense. -Sean On Tue, Mar 3, 2009 at 10:14 PM, Amar Kamat <[email protected]> wrote: > Runping Qi wrote: > >> Could it be the case that the latter jobs ran slower because the tasks >> took >> longer time to get initialized? >> If so, you may hit >> https://issues.apache.org/jira/browse/HADOOP-4780 >> >> Runping >> >> On Tue, Mar 3, 2009 at 2:02 PM, Sean Laurent <[email protected] >> >wrote: >> >> >> >>> Hrmmm. According to hadoop-defaults.xml, >>> mapred.jobtracker.completeuserjobs.maximum defaults to 100. So I tried >>> setting it to 1, but that had no effect. I still see each successive run >>> taking longer than the previous run. >>> >>> 1) Restart M/R >>> 2) Run #1: 142.12 (secs) >>> 3) Run #2 181.96 (secs) >>> 4) Run #3 221.95 (secs) >>> 5) Run #4 281.96 (secs) >>> >>> >> Yeah. May be its not the problem with the JobTracker. Can you check (via > job history) what is the best and the worst task runtimes? You can analyze > the jobs after they complete. > Amar
