You are viewing a plain text version of this content. The canonical link for it is here.
Posted to yarn-issues@hadoop.apache.org by "Chandni Singh (JIRA)" <ji...@apache.org> on 2018/07/26 16:48:00 UTC
[jira] [Commented] (YARN-8508) GPU does not get released even
though the container is killed
[ https://issues.apache.org/jira/browse/YARN-8508?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16558571#comment-16558571 ]
Chandni Singh commented on YARN-8508:
-------------------------------------
This happens with a container that gets cleaned up before its pid file is created. To solve it, we need to release the resources at the end of \{{LinuxContainerExecutor.reapContainer()}} just like we do in \{{LinuxContainerExecutor.launchContainer()}},
{\{LinuxContainerExecutor.reLaunchContainer()}}, and \{{LinuxContainerExecutor.reacquireContainer}}.
Please see my explanation below:
Refer \{{container_e21_1532545600682_0001_01_000002}} in yarn8505.nodemanager.log
- 002 is launched but its pid file is not created
{code}
2018-07-25 19:08:54,409 DEBUG util.ProcessIdFileReader (ProcessIdFileReader.java:getProcessId(53)) - Accessing pid from pid file /.../application_1532545600682_0001/container_e21_1532545600682_0001_01_000002/container_e21_1532545600682_0001_01_000002.pid
2018-07-25 19:08:54,409 DEBUG util.ProcessIdFileReader (ProcessIdFileReader.java:getProcessId(103)) - Got pid null from path /.../application_1532545600682_0001/container_e21_1532545600682_0001_01_000002/container_e21_1532545600682_0001_01_000002.pid
{code}
- Since application is killed, 002 is killed by ResourceManager
{code}
2018-07-25 19:08:54,643 DEBUG container.ContainerImpl (ContainerImpl.java:handle(2080)) - Processing container_e21_1532545600682_0001_01_000002 of type CONTAINER_KILLED_ON_REQUEST
{code}
- The above triggers \{{ContainerLaunch.cleanupContainer()}} for 002. This happens before the pid file is created
{code}
2018-07-25 19:08:54,409 WARN launcher.ContainerLaunch (ContainerLaunch.java:cleanupContainer(784)) - Container clean up before pid file created container_e21_1532545600682_0001_01_000002
{code}
- \{{cleanupContainer}} invokes \{{reapDockerContainerNoPid(user)}}
{code}
2018-07-25 19:08:54,410 INFO launcher.ContainerLaunch (ContainerLaunch.java:reapDockerContainerNoPid(940)) - Unable to obtain pid, but docker container request detected. Attempting to reap container container_e21_1532545600682_0001_01_000002
{code}
- \{{reapDockerContainerNoPid(user)}} calls \{{exec.reapContainer(...)}}
{code}
2018-07-25 19:08:54,412 DEBUG docker.DockerCommandExecutor (DockerCommandExecutor.java:executeDockerCommand(89)) - Running docker command: inspect docker-command=inspect format=\{{.State.Status}} name=container_e21_1532545600682_0001_01_000002
2018-07-25 19:08:54,412 DEBUG privileged.PrivilegedOperationExecutor (PrivilegedOperationExecutor.java:getPrivilegedOperationExecutionCommand(119)) - Privileged Execution Command Array: [/.../hadoop-yarn/bin/container-executor, --inspect-docker-container, --format=\{{.State.Status}}, container_e21_1532545600682_0001_01_000002]
2018-07-25 19:08:54,530 DEBUG docker.DockerCommandExecutor (DockerCommandExecutor.java:getContainerStatus(160)) - Container Status: nonexistent ContainerId: container_e21_1532545600682_0001_01_000002
2018-07-25 19:08:54,530 DEBUG launcher.ContainerLaunch (ContainerLaunch.java:reapDockerContainerNoPid(948)) - Sent signal to docker container container_e21_1532545600682_0001_01_000002 as user hrt_qa, result=success
{code}
- The problem is that the \{{reapContainer}} in \{{LinuxContainerExecutor}} doesn't release the resources assigned to the container. The below code snippet that performs these tasks after the container completes doesn't happen at this point.
{code}
resourcesHandler.postExecute(containerId);
try {
if (resourceHandlerChain != null) {
LOG.info("{} POST Complete", containerId);
resourceHandlerChain.postComplete(containerId);
}
} catch (ResourceHandlerException e) {
LOG.warn("ResourceHandlerChain.postComplete failed for " +
"containerId: " + containerId + ". Exception: " + e);
}
}
{code}
- The launch of container fails after 4 minutes and only then the resources are released.
{code}
2018-07-25 19:12:09,999 WARN nodemanager.LinuxContainerExecutor (LinuxContainerExecutor.java:handleExitCode(593)) - Exit code from container container_e21_1532545600682_0001_01_000002 is : 27
2018-07-25 19:12:10,000 WARN nodemanager.LinuxContainerExecutor (LinuxContainerExecutor.java:handleExitCode(599)) - Exception from container-launch with container ID: container_e21_1532545600682_0001_01_000002 and exit code: 27
2018-07-25 19:12:10,000 INFO nodemanager.ContainerExecutor (ContainerExecutor.java:logOutput(541)) - Container id: container_e21_1532545600682_0001_01_000002
2018-07-25 19:12:10,003 INFO nodemanager.ContainerExecutor (ContainerExecutor.java:logOutput(541)) - Docker inspect command: /usr/bin/docker inspect --format \{{.State.Pid}} container_e21_1532545600682_0001_01_000002
2018-07-25 19:12:10,003 INFO nodemanager.ContainerExecutor (ContainerExecutor.java:logOutput(541)) - Failed to write pid to file /cgroup/cpu/.../container_e21_1532545600682_0001_01_000002/tasks - No such process
2018-07-25 19:12:10,003 INFO nodemanager.ContainerExecutor (ContainerExecutor.java:logOutput(541)) - Wrote the exit code 27 to /.../application_1532545600682_0001/container_e21_1532545600682_0001_01_000002/container_e21_1532545600682_0001_01_000002.pid.exitcode
2018-07-25 19:12:10,003 DEBUG container.ContainerImpl (ContainerImpl.java:handle(2080)) - Processing container_e21_1532545600682_0001_01_000002 of type UPDATE_DIAGNOSTICS_MSG
2018-07-25 19:12:10,004 DEBUG recovery.NMLeveldbStateStoreService (NMLeveldbStateStoreService.java:storeContainerDiagnostics(488)) - storeContainerDiagnostics: containerId=container_e21_1532545600682_0001_01_000002, diagnostics=[2018-07-25 19:08:49.283]Container Killed by ResourceManager
Container id: container_e21_1532545600682_0001_01_000002
Docker inspect command: /usr/bin/docker inspect --format \{{.State.Pid}} container_e21_1532545600682_0001_01_000002
Failed to write pid to file /cgroup/cpu/.../container_e21_1532545600682_0001_01_000002/tasks - No such process
{code}
> GPU does not get released even though the container is killed
> --------------------------------------------------------------
>
> Key: YARN-8508
> URL: https://issues.apache.org/jira/browse/YARN-8508
> Project: Hadoop YARN
> Issue Type: Bug
> Reporter: Sumana Sathish
> Assignee: Chandni Singh
> Priority: Major
>
> GPU failed to release even though the container using it is being killed
> {Code}
> 2018-07-06 05:22:26,201 INFO container.ContainerImpl (ContainerImpl.java:handle(2093)) - Container container_e20_1530854311763_0006_01_000001 transitioned from RUNNING to KILLING
> 2018-07-06 05:22:26,250 INFO container.ContainerImpl (ContainerImpl.java:handle(2093)) - Container container_e20_1530854311763_0006_01_000002 transitioned from RUNNING to KILLING
> 2018-07-06 05:22:26,251 INFO application.ApplicationImpl (ApplicationImpl.java:handle(632)) - Application application_1530854311763_0006 transitioned from RUNNING to FINISHING_CONTAINERS_WAIT
> 2018-07-06 05:22:26,251 INFO launcher.ContainerLaunch (ContainerLaunch.java:cleanupContainer(734)) - Cleaning up container container_e20_1530854311763_0006_01_000002
> 2018-07-06 05:22:31,358 INFO launcher.ContainerLaunch (ContainerLaunch.java:getContainerPid(1102)) - Could not get pid for container_e20_1530854311763_0006_01_000002. Waited for 5000 ms.
> 2018-07-06 05:22:31,358 WARN launcher.ContainerLaunch (ContainerLaunch.java:cleanupContainer(784)) - Container clean up before pid file created container_e20_1530854311763_0006_01_000002
> 2018-07-06 05:22:31,359 INFO launcher.ContainerLaunch (ContainerLaunch.java:reapDockerContainerNoPid(940)) - Unable to obtain pid, but docker container request detected. Attempting to reap container container_e20_1530854311763_0006_01_000002
> 2018-07-06 05:22:31,494 INFO nodemanager.LinuxContainerExecutor (LinuxContainerExecutor.java:deleteAsUser(828)) - Deleting absolute path : /grid/0/hadoop/yarn/local/usercache/hrt_qa/appcache/application_1530854311763_0006/container_e20_1530854311763_0006_01_000002/launch_container.sh
> 2018-07-06 05:22:31,500 INFO nodemanager.LinuxContainerExecutor (LinuxContainerExecutor.java:deleteAsUser(828)) - Deleting absolute path : /grid/0/hadoop/yarn/local/usercache/hrt_qa/appcache/application_1530854311763_0006/container_e20_1530854311763_0006_01_000002/container_tokens
> 2018-07-06 05:22:31,510 INFO container.ContainerImpl (ContainerImpl.java:handle(2093)) - Container container_e20_1530854311763_0006_01_000001 transitioned from KILLING to CONTAINER_CLEANEDUP_AFTER_KILL
> 2018-07-06 05:22:31,510 INFO container.ContainerImpl (ContainerImpl.java:handle(2093)) - Container container_e20_1530854311763_0006_01_000002 transitioned from KILLING to CONTAINER_CLEANEDUP_AFTER_KILL
> 2018-07-06 05:22:31,512 INFO container.ContainerImpl (ContainerImpl.java:handle(2093)) - Container container_e20_1530854311763_0006_01_000001 transitioned from CONTAINER_CLEANEDUP_AFTER_KILL to DONE
> 2018-07-06 05:22:31,513 INFO container.ContainerImpl (ContainerImpl.java:handle(2093)) - Container container_e20_1530854311763_0006_01_000002 transitioned from CONTAINER_CLEANEDUP_AFTER_KILL to DONE
> 2018-07-06 05:22:38,955 INFO container.ContainerImpl (ContainerImpl.java:handle(2093)) - Container container_e20_1530854311763_0007_01_000002 transitioned from NEW to SCHEDULED
> {Code}
> New container requesting for GPU fails to launch
> {code}
> 2018-07-06 05:22:39,048 ERROR nodemanager.LinuxContainerExecutor (LinuxContainerExecutor.java:handleLaunchForLaunchType(550)) - ResourceHandlerChain.preStart() failed!
> org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.ResourceHandlerException: Failed to find enough GPUs, requestor=container_e20_1530854311763_0007_01_000002, #RequestedGPUs=2, #availableGpus=1
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.gpu.GpuResourceAllocator.internalAssignGpus(GpuResourceAllocator.java:225)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.gpu.GpuResourceAllocator.assignGpus(GpuResourceAllocator.java:173)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.gpu.GpuResourceHandlerImpl.preStart(GpuResourceHandlerImpl.java:98)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.ResourceHandlerChain.preStart(ResourceHandlerChain.java:75)
> at org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor.handleLaunchForLaunchType(LinuxContainerExecutor.java:509)
> at org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor.launchContainer(LinuxContainerExecutor.java:479)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.launchContainer(ContainerLaunch.java:494)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.call(ContainerLaunch.java:306)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.call(ContainerLaunch.java:103)
> at java.util.concurrent.FutureTask.run(FutureTask.java:266)
> at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
> at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
> at java.lang.Thread.run(Thread.java:748)
> 2018-07-06 05:22:39,049 WARN launcher.ContainerLaunch (ContainerLaunch.java:call(331)) - Failed to launch container.
> java.io.IOException: ResourceHandlerChain.preStart() failed!
> at org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor.handleLaunchForLaunchType(LinuxContainerExecutor.java:551)
> at org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor.launchContainer(LinuxContainerExecutor.java:479)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.launchContainer(ContainerLaunch.java:494)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.call(ContainerLaunch.java:306)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.call(ContainerLaunch.java:103)
> at java.util.concurrent.FutureTask.run(FutureTask.java:266)
> at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
> at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
> at java.lang.Thread.run(Thread.java:748)
> Caused by: org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.ResourceHandlerException: Failed to find enough GPUs, requestor=container_e20_1530854311763_0007_01_000002, #RequestedGPUs=2, #availableGpus=1
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.gpu.GpuResourceAllocator.internalAssignGpus(GpuResourceAllocator.java:225)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.gpu.GpuResourceAllocator.assignGpus(GpuResourceAllocator.java:173)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.gpu.GpuResourceHandlerImpl.preStart(GpuResourceHandlerImpl.java:98)
> at org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.ResourceHandlerChain.preStart(ResourceHandlerChain.java:75)
> at org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor.handleLaunchForLaunchType(LinuxContainerExecutor.java:509)
> ... 8 more
> {code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org