[ https://issues.apache.org/jira/browse/FLINK-8526?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16347027#comment-16347027 ]
zhu.qing commented on FLINK-8526: --------------------------------- link there is more information in https://issues.apache.org/jira/browse/FLINK-8534 > When use parallelism equals to half of the number of cpu, join and shuffle > operators will easly cause deadlock. > --------------------------------------------------------------------------------------------------------------- > > Key: FLINK-8526 > URL: https://issues.apache.org/jira/browse/FLINK-8526 > Project: Flink > Issue Type: Bug > Components: DataSet API, Local Runtime > Affects Versions: 1.4.0 > Environment: 8 machines(96GB and 24 cores) and 20 taskslot per > taskmanager. twitter-2010 dataset. And parallelism setting to 80. I run my > code in standalone mode. > Reporter: zhu.qing > Priority: Major > Attachments: T2AdjActiveV.java, T2AdjMessage.java > > Original Estimate: 72h > Remaining Estimate: 72h > > The next program attached will stuck at some special parallelism in some > situation. When parallelism is 80 in previous setting, The program will > always stuck. And when parallelism is 100, everything goes well. According > to my research I found when the parallelism equals to number of taskslots. > The program is not fastest and probably caused network buffer not enough. How > networker buffer related to parallelism and how parallelism relate to > running task (In other words we have 160 taskslots but running task can be > far more than taskslots). > Parallelism cannot be equals to half of the cpu. > Or will casuse "java.io.FileNotFoundException". You can repeat exception on > your pc and set your parallelism equals to half of your cpu core. -- This message was sent by Atlassian JIRA (v7.6.3#76005)