[YARN-5449] nodemanager process is hung, and lost from resourcemanager

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.2.0
• Component/s: nodemanager

Environment

The os version is 2.6.32-573.8.1.el6.x86_64 GNU/Linux
The java version is jdk1.7.0_45
The hadoop version is hadoop-2.2.0

• Assignee: Unassigned
• Reporter: Shurong Mai
• Created: 29/Jul/16 09:49
• Updated: 29/Apr/19 03:37

Description

The nodemanager process is hung(is not dead), and lost from resourcemanager.
The nodemanager's log is stopped from printing.
The used cpu of nodemanager process is very low(nearly 0%).
GC of nodemanager jvm process is stopped, and the result of jstat(jstat -gccause pid 1000 100) is as follows:
S0 S1 E O P YGC YGCT FGC FGCT GCT LGCC GCC
0.00 100.00 95.06 24.08 30.46 3274 623.437 7 5.899 629.335 No GC G1 Evacuation Pause
0.00 100.00 95.06 24.08 30.46 3274 623.437 7 5.899 629.335 No GC G1 Evacuation Pause
0.00 100.00 95.06 24.08 30.46 3274 623.437 7 5.899 629.335 No GC G1 Evacuation Pause
0.00 100.00 95.06 24.08 30.46 3274 623.437 7 5.899 629.335 No GC G1 Evacuation Pause
0.00 100.00 95.06 24.08 30.46 3274 623.437 7 5.899 629.335 No GC G1 Evacuation Pause
0.00 100.00 95.06 24.08 30.46 3274 623.437 7 5.899 629.335 No GC G1 Evacuation Pause

The nodemanager jvm process is also accur this problem using CMS garbage collector or g1 garbage collector.

The parameters of CMS garbage collector are as following:
-Xmx4096m -Xmn1024m -XX:PermSize=128m -XX:MaxPermSize=128m -XX:+UseConcMarkSweepGC -XX:+CMSConcurrentMTEnabled -XX:ConcGCThreads=4 -XX:+UseCMSCom pactAtFullCollection -XX:CMSFullGCsBeforeCompaction=8 -XX:ParallelGCThreads=4 -XX:CMSInitiatingOccupancyFraction=70

The parameters of g1 garbage collector are as following:
-Xmx8g -Xms8g -XX:PermSize=128m -XX:MaxPermSize=128m -XX:+UseG1GC -XX:MaxGCPauseMillis=1000 -XX:G1ReservePercent=30 -XX:InitiatingHeapOccupancyPercent=45 -XX:ConcGCThreads=4 -XX:+PrintAdaptiveSizePolicy