Uploaded image for project: 'Hadoop Common'
  1. Hadoop Common
  2. HADOOP-4666

Launch reduces only after a few maps have run in the Fair Scheduler

    XMLWordPrintableJSON

Details

    • Improvement
    • Status: Closed
    • Major
    • Resolution: Fixed
    • None
    • 0.20.0
    • None
    • None
    • Reviewed

    Description

      It makes no sense to schedule reduces for a job before its maps have started running. As an initial fix, we will wait until a certain percent have run (likely 10%). In the future it would be good to choose the time to wait based on amount of map output data as well - launching reducers that will mostly be idle is not helpful. Average amount of map output bytes per mapper is easy to compute using counters in JobInProgress.

      Attachments

        1. fairscheduler-reduce-launch-condition.patch
          5 kB
          Matei Alexandru Zaharia

        Activity

          People

            matei@eecs.berkeley.edu Matei Zaharia
            matei Matei Alexandru Zaharia
            Votes:
            0 Vote for this issue
            Watchers:
            6 Start watching this issue

            Dates

              Created:
              Updated:
              Resolved: