[ https://issues.apache.org/jira/browse/HADOOP-5659?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12718665#action_12718665 ]
Matei Zaharia commented on HADOOP-5659: --------------------------------------- Hi Dhruba, A question about this feature: How do you want to decide when to kill the wellcare job? Do you want the tasks specified as a %, as a fixed number, or what? Matei > Fair share schduler may support preemption only with a specific pool > -------------------------------------------------------------------- > > Key: HADOOP-5659 > URL: https://issues.apache.org/jira/browse/HADOOP-5659 > Project: Hadoop Core > Issue Type: Improvement > Components: contrib/fair-share > Reporter: dhruba borthakur > > There are a set of jobs that helps to keep the cluster resources being used > optimally. For example, there are data sets that are made of a multiple files > in a directory. These part-xxx files could be concatenated to a relatively > few files (to reduce memory pressure on the namenode). Also, there are files > that could be compressed more efficiently (e.g. bzip2) to reduce save on disk > usage. These are kind of system-wellcare jobs that should run only if it does > not impact any other "real" user of the cluster. On an idle cluster, these > wellcare jobs should use all availale system resources. When a real user > submits a job, the wellcare job(s) should be pre-empted. If a scheduler can > support pre-emption only for jobs in a specified pool, then I can submit > these well-care jobs to that special pool. Real user's jobs will never get > pre-empted;but the wellcare jobs can get pre-empted as soon as there is > resource contention. If a task of well-care jobs is pre-empted more than a > configured max, the entire wellcare job will fail.. that this is the > behaviour I want. The wellcare jobs would run in idle slots as long as all > user-submitted jobs have been satisfied, but would be preempted as soon as > user jobs require any of those slots. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.