Application uses a lot of memory under RHEL 6

Environment

• Red Hat Enterprise Linux (RHEL) 6
  • system upgraded from RHEL 5.6 to RHEL 6.4
  • system upgraded to RHEL 6
  • java application running on RHEL 6

Issue

• Application uses lot of memory on rhel6.3 as well as 6.4. It works fine under 5.6. Problems with unacceptably large memory growth (both RSS and VSZ) . Is part of this problem is likely due to the new memory arena allocator including in glibc 2.10 and later referenced on Content from sourceware.org is not included.sourceware.org and Content from www.ibm.com is not included.IBM developerworks

  • With the default transparent huge page(THP) support enabled, memory RSS would grow for the application [compared to a RHEL 5 build on a RH5 box of the same software].
  • Disabling transparent huge page support seemed to bring the memory back in line after setting the kernel command line option [appending transparent_hugepage=never] or dynamically setting echo never > /sys/kernel/mm/redhat_transparent_hugepage/enabled
  • In some cases, even with THP [Anon Huge pages] disabled, the virtual memory would balloon [compared to a rhel5 build on a rhel5 box of the same software]
  • Enabling transparent hugepages and setting the arenas seems to have the virtual memory growth.
  • Setting MALLOC_ARENA_MAX environment variable to a 4 to start, we did some preliminary testing and it appeared to improve the virtual memory growth.[ +5 or 10% wouldn't be quite a concern, but 10 times as large getting into 250G
  • Virtual memory (VM) range causes some concern about the kernel's management of the Virtual memory (VM).

• While testing a Java app under RHEL 6 using jdk 7u5_x64, we ran into a situation where VM usage increased a lot. We did some research and based on Content from issues.apache.org is not included.HADOOP-7154 we set MALLOC_ARENA_MAX=4. That brought our VM usage back to levels we would expect.
  We read KB article 61778 which helped, but we're looking for more information about how an arena functions. We are currently testing value of 4, which seems to have desired effect, but we are not certain why? Some google articles note that value of 1 should not be used, but no clear explanation as to why.

• We are witnessing JBoss instances process size getting larger and larger (using a lot more physical heap than is configured). But this only seems to happen on RHEL servers. What could cause this and what is the best way to analyse something like this

Resolution

Red Hat is still working to establish what the root cause of the issue may be.

In the interim, the following articles may be of interest

• A multi-threaded application not seem to use Transparent HugePages in RHEL-6
• Why do multithreaded applications use significantly more virtual memory on a more recent version of Red Hat Enterprise Linux
• Understanding malloc_arena_max

Diagnostic Steps

• Depending on how quickly the memory bloats, capture pmap for the process at preset intervals e.g daily to get a profile of how quickly the memory grows
• If capturing the information on RHEL 5 and RHEL 6 systems use the following commands to see whether there are higher numbers of bigger anonymous mappings on one system compared to another

 grep anon pmap.out.rhel5 | awk '{ print $3 }' | sort -n
 grep anon pmap.out.rhel6.3 | awk '{ print $3 }' | wc -l

• If running a java application see the article Why does the JVM consume more memory than the amount given to -Xmx? to ensure you have an accurate computation of the process size

• See How to track OS application memory leaks.

• See How to identify JVM native memory segment allocations on OpenJDK / Oracle JDK.

• See Understanding memory leaks - using Valgrind (memcheck).

This solution is part of Red Hat’s fast-track publication program, providing a huge library of solutions that Red Hat engineers have created while supporting our customers. To give you the knowledge you need the instant it becomes available, these articles may be presented in a raw and unedited form.