Details
Description
In our cluster, ResouceManager stuck twice within twenty days. Yarn client can't submit application. I got jstack info at second time, then found the reason.
I analyze all the jstack, I found many thread stuck because can't get LinkedBlockingQueue.putLock. (Note: Sorry for limited space , omit the analytical process)
The reason is that one thread hold the putLock all the time, printEventQueueDetails will called forEachRemaining, then hold putLock and readLock. The AsyncDispatcher will stuck.
Thread 6526 (IPC Server handler 454 on default port 8030): State: RUNNABLE Blocked count: 29988 Waited count: 2035029 Stack: java.util.concurrent.LinkedBlockingQueue$LBQSpliterator.forEachRemaining(LinkedBlockingQueue.java:926) java.util.stream.AbstractPipeline.copyInto(AbstractPipeline.java:482) java.util.stream.AbstractPipeline.wrapAndCopyInto(AbstractPipeline.java:472) java.util.stream.ReduceOps$ReduceOp.evaluateSequential(ReduceOps.java:708) java.util.stream.AbstractPipeline.evaluate(AbstractPipeline.java:234) java.util.stream.ReferencePipeline.collect(ReferencePipeline.java:499) org.apache.hadoop.yarn.event.AsyncDispatcher$GenericEventHandler.printEventQueueDetails(AsyncDispatcher.java:270) org.apache.hadoop.yarn.event.AsyncDispatcher$GenericEventHandler.handle(AsyncDispatcher.java:295) org.apache.hadoop.yarn.server.resourcemanager.DefaultAMSProcessor.handleProgress(DefaultAMSProcessor.java:408) org.apache.hadoop.yarn.server.resourcemanager.DefaultAMSProcessor.allocate(DefaultAMSProcessor.java:215) org.apache.hadoop.yarn.server.resourcemanager.scheduler.constraint.processor.DisabledPlacementProcessor.allocate(DisabledPlacementProcessor.java:75) org.apache.hadoop.yarn.server.resourcemanager.AMSProcessingChain.allocate(AMSProcessingChain.java:92) org.apache.hadoop.yarn.server.resourcemanager.ApplicationMasterService.allocate(ApplicationMasterService.java:432) org.apache.hadoop.yarn.api.impl.pb.service.ApplicationMasterProtocolPBServiceImpl.allocate(ApplicationMasterProtocolPBServiceImpl.java:60) org.apache.hadoop.yarn.proto.ApplicationMasterProtocol$ApplicationMasterProtocolService$2.callBlockingMethod(ApplicationMasterProtocol.java:99) org.apache.hadoop.ipc.ProtobufRpcEngine$Server$ProtoBufRpcInvoker.call(ProtobufRpcEngine.java:528) org.apache.hadoop.ipc.RPC$Server.call(RPC.java:1070) org.apache.hadoop.ipc.Server$RpcCall.run(Server.java:1040) org.apache.hadoop.ipc.Server$RpcCall.run(Server.java:958) java.security.AccessController.doPrivileged(Native Method)
I analyze LinkedBlockingQueue's source code. I found forEachRemaining in LinkedBlockingQueue.LBQSpliterator may stuck, when forEachRemaining and take are called in different thread.
YARN-8995 introduce printEventQueueDetails method, "eventQueue.stream().collect" will called forEachRemaining method.
Let's see why? "put.png" shows that how to put("a"), "take.png" shows that how to take()。Specical Node: The removed Node will point itself for help gc!!!
The key point code is in forEachRemaining, we see LBQSpliterator use forEachRemaining to visit all Node. But when got item value from Node, will release the lock. If at this time, take() will be called.
The variable 'p' in forEachRemaining may point a Node which point itself, then forEachRemaining will be in dead loop. You can see it in "deadloop.png"
Let's see a simple uni-test, Let's forEachRemaining called more slow than take, the problem will reproduction。uni-test is MockForDeadLoop.java.
I debug MockForDeadLoop.java, and see a Node point itself. You can see pic "debugfornode.png"
Environment:
OS: CentOS Linux release 7.5.1804 (Core)
JDK: jdk1.8.0_281
Attachments
Attachments
Issue Links
- is duplicated by
-
YARN-10643 Fix the race condition introduced by YARN-8995.
-
- Resolved
-
- relates to
-
YARN-10643 Fix the race condition introduced by YARN-8995.
-
- Resolved
-