[ https://issues.apache.org/jira/browse/YARN-8382?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Hu Ziqian updated YARN-8382: ---------------------------- Description: As Jiandan said in YARN-6525, NM may delete Cgroup container file timeout with logs like org.apache.hadoop.yarn.server.nodemanager.util.CgroupsLCEResourcesHandler: Unable to delete cgroup at: /cgroup/cpu/hadoop-yarn/container_xxx, tried to delete for 1000ms we found one situation is that when we set *yarn.nodemanager.sleep-delay-before-sigkill.ms* bigger than yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms, the cgroup file leak happens *.* One container process tree looks like follow graph: bash(16097)───java(16099)─┬─\{java}(16100) ├─\{java}(16101) {{ ├─\{java}(16102)}} {{when NM kill a container, NM send kill -15 -pid to kill container process group. Bash process will exit when it received sigterm, but java process may do some job (shutdownHook etc.), and may exit unit receive sigkill. And when bash process exit, CgroupsLCEResourcesHandler begin to try to delete cgroup. So when *yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms* arrived, the java processes may still running and cgourp/tasks still not empty and cause a cgroup file leak.}} {{we add a condition that *yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms* must bigger than *yarn.nodemanager.sleep-delay-before-sigkill.ms* to solve this problem.}} was: As Jiandan said in YARN-6525, NM may delete Cgroup container file timeout with logs like org.apache.hadoop.yarn.server.nodemanager.util.CgroupsLCEResourcesHandler: Unable to delete cgroup at: /cgroup/cpu/hadoop-yarn/container_xxx, tried to delete for 1000ms we found one situation is that when we set *yarn.nodemanager.sleep-delay-before-sigkill.ms* bigger than yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms, the cgroup file leak happens *.* One container process tree looks like follow graph: {\{bash(16097)───java(16099)─┬─{java}(16100) }} {\{ ├─{java}(16101) }} {{ ├─\{java}(16102)}} {{when NM kill a container, NM send kill -15 -pid to kill container process group. Bash process will exit when it received sigterm, but java process may do some job (shutdownHook etc.), and may exit unit receive sigkill. And when bash process exit, CgroupsLCEResourcesHandler begin to try to delete cgroup. So when *yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms* arrived, the java processes may still running and cgourp/tasks still not empty and cause a cgroup file leak.}} {{we add a condition that *yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms* must bigger than *yarn.nodemanager.sleep-delay-before-sigkill.ms* to solve this problem.}} > cgroup file leak in NM > ---------------------- > > Key: YARN-8382 > URL: https://issues.apache.org/jira/browse/YARN-8382 > Project: Hadoop YARN > Issue Type: Bug > Components: nodemanager > Environment: we write an container with a shutdownHook which has a > piece of code like "while(true) sleep(100)" . when > *yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms <* > *yarn.nodemanager.sleep-delay-before-sigkill.ms , cgourp file leak happens; > when* *yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms >* > ** *yarn.nodemanager.sleep-delay-before-sigkill.ms, cgroup file is deleted > successfully*** > Reporter: Hu Ziqian > Assignee: Hu Ziqian > Priority: Major > > As Jiandan said in YARN-6525, NM may delete Cgroup container file timeout > with logs like > org.apache.hadoop.yarn.server.nodemanager.util.CgroupsLCEResourcesHandler: > Unable to delete cgroup at: /cgroup/cpu/hadoop-yarn/container_xxx, tried to > delete for 1000ms > > we found one situation is that when we set > *yarn.nodemanager.sleep-delay-before-sigkill.ms* bigger than > yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms, the > cgroup file leak happens *.* > > One container process tree looks like follow graph: > bash(16097)───java(16099)─┬─\{java}(16100) > > ├─\{java}(16101) > {{ ├─\{java}(16102)}} > > {{when NM kill a container, NM send kill -15 -pid to kill container process > group. Bash process will exit when it received sigterm, but java process may > do some job (shutdownHook etc.), and may exit unit receive sigkill. And when > bash process exit, CgroupsLCEResourcesHandler begin to try to delete cgroup. > So when *yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms* > arrived, the java processes may still running and cgourp/tasks still not > empty and cause a cgroup file leak.}} > > {{we add a condition that > *yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms* must > bigger than *yarn.nodemanager.sleep-delay-before-sigkill.ms* to solve this > problem.}} > -- This message was sent by Atlassian JIRA (v7.6.3#76005) --------------------------------------------------------------------- To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org