[HBASE-16616] Rpc handlers stuck on ThreadLocalMap.expungeStaleEntry

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.2.2
• Fix Version/s: 1.4.0, 2.0.0
• Component/s: Performance
• Hadoop Flags: Reviewed
• Assignee: Tomu Tsuruhara
• Reporter: Tomu Tsuruhara
• Created: 12/Sep/16 13:43
• Updated: 11/Dec/18 21:55
• Resolved: 13/Sep/16 01:37

Description

In our HBase 1.2.2 cluster, some regionserver showed too bad "QueueCallTime_99th_percentile" exceeding 10 seconds.
Most rpc handler threads stuck on ThreadLocalMap.expungeStaleEntry call at that time.

"PriorityRpcServer.handler=18,queue=0,port=16020" #322 daemon prio=5 os_prio=0 tid=0x00007fd422062800 nid=0x19b89 runnable [0x00007fcb8a821000]
   java.lang.Thread.State: RUNNABLE
        at java.lang.ThreadLocal$ThreadLocalMap.expungeStaleEntry(ThreadLocal.java:617)
        at java.lang.ThreadLocal$ThreadLocalMap.remove(ThreadLocal.java:499)
        at java.lang.ThreadLocal$ThreadLocalMap.access$200(ThreadLocal.java:298)
        at java.lang.ThreadLocal.remove(ThreadLocal.java:222)
        at java.util.concurrent.locks.ReentrantReadWriteLock$Sync.tryReleaseShared(ReentrantReadWriteLock.java:426)
        at java.util.concurrent.locks.AbstractQueuedSynchronizer.releaseShared(AbstractQueuedSynchronizer.java:1341)
        at java.util.concurrent.locks.ReentrantReadWriteLock$ReadLock.unlock(ReentrantReadWriteLock.java:881)
        at com.yammer.metrics.stats.ExponentiallyDecayingSample.unlockForRegularUsage(ExponentiallyDecayingSample.java:196)
        at com.yammer.metrics.stats.ExponentiallyDecayingSample.update(ExponentiallyDecayingSample.java:113)
        at com.yammer.metrics.stats.ExponentiallyDecayingSample.update(ExponentiallyDecayingSample.java:81)
        at org.apache.hadoop.metrics2.lib.MutableHistogram.add(MutableHistogram.java:81)
        at org.apache.hadoop.metrics2.lib.MutableRangeHistogram.add(MutableRangeHistogram.java:59)
        at org.apache.hadoop.hbase.ipc.MetricsHBaseServerSourceImpl.dequeuedCall(MetricsHBaseServerSourceImpl.java:194)
        at org.apache.hadoop.hbase.ipc.MetricsHBaseServer.dequeuedCall(MetricsHBaseServer.java:76)
        at org.apache.hadoop.hbase.ipc.RpcServer.call(RpcServer.java:2192)
        at org.apache.hadoop.hbase.ipc.CallRunner.run(CallRunner.java:112)
        at org.apache.hadoop.hbase.ipc.RpcExecutor.consumerLoop(RpcExecutor.java:133)
        at org.apache.hadoop.hbase.ipc.RpcExecutor$1.run(RpcExecutor.java:108)
        at java.lang.Thread.run(Thread.java:745)

We were using jdk 1.8.0_92 and here is a snippet from ThreadLocal.java.

616:    while (tab[h] != null)
617:        h = nextIndex(h, len);

So I hypothesized that there're too many consecutive entries in tab array and actually I found them in the heapdump.

Most of these entries pointed at instance of org.apache.hadoop.hbase.util.Counter$1
which is equivarent to indexHolderThreadLocal instance-variable in the Counter class.

Because RpcServer$Connection class creates a Counter instance rpcCount for every connections,
it is possible to have lots of Counter#indexHolderThreadLocal instances in RegionServer process
when we repeat connect-and-close from client. As a result, a ThreadLocalMap can have lots of consecutive
entires.

Usually, since each entry is a WeakReference, these entries are collected and removed
by garbage-collector soon after connection closed.
But if connection's life-time was long enough to survive youngGC, it wouldn't be collected until old-gen collector runs.
Furthermore, under G1GC deployment, it is possible not to be collected even by old-gen GC(mixed GC)
if entries sit in a region which doesn't have much garbages.
Actually we used G1GC when we encountered this problem.

We should remove the entry from ThreadLocalMap by calling ThreadLocal#remove explicitly.

Issue Links

• relates to: HBASE-16146 Counters are expensive...