How to reliably take Java Heap Dumps?

Question

My team is running into difficulties when trying to take good heap dumps triggered by OutOfMemoryErrors. For specific reasons we are currently taking the dumps with jmap called from a bash script instead of using the HeapDumpOnOutOfMemoryError flag. We're using a 64-bit 1.6 JVM with a heap size around 3 GB. Our heap dumps fail 90% of the time (guesstimate).

Is there anything we can do to improve our odds of getting a clean heap dump we can use to troubleshoot memory problems? I have read that jmap had major issues in Java 1.4 but that those issues should be mostly addressed now.

Solution 1

Which is your operating system? (I can't add comments).

For Solaris we get better results first forcing a core dump (gcore <pid>) and then attaching jmap to the core dump file (jmap -heap:format=b <path to java bin> <path to core>)

gcore is a *nix utility to generate an image of a running program. See link.

Solution 2

we have a JSP that queries ManagementFactory.getThreadMXBean() and produces a report. May not be useful when the app has crashed, but if you poll every minute or so, you'll get an idea of what's happening.

More info here.

Solution 3

you could monitor your application via jmx from the outside. when you know some metrics which indicate an upcoming OutOfMemory, you could trigger a jmap run before the exception is thrown.

Solution 4

This is a fairly old question, but i will put answer with the hope that somebody may find this useful.

jmap has a -F option (force). This has proven to not work that well in the past for me. If you are to use the -F option, I would recommend that you also specify the java.io.tmp directory as a part of the jmap command. There was an issue with JVM version 1.6.22 where the jmap utility did not work properly because of a temp directory setting.

You can also try to take a core dump via gdb. Once you have the core, jmap can convert the core into a heap dump.

Solution 5

Thanks to you all for your suggestions.

What we wound up doing is writing a script to actively monitor the garbage collection logs. In our experience, back-to-back Full GC's almost always precede an OOM, so our script detects this event, gracefully removes the server from the load balancing pool, and forces the heap dump. This has greatly increased our effectiveness.

View more solutions

Comments

[phoebus]

I nominate this question for "most unintentionally disgusting-sounding".

[karlcyr]

Hah- I thought about making it intentionally disgusting-sounding but I'm new here and I wasn't sure how the community would take that :).

[karlcyr]

Thanks Christian- is jmap more likely to be reliable before the error is thrown?

[user2987902]

jmap will still need some time to get you a heap dump. but you will get a full heapdump as long as your jvm/tomcat is mostly responsible.

[user2987902]

tried it with gdb on linux and it works great.

[djangofan]

which JDK has "gcore" ? mine, Sun 32 bit jdk for linux 1.6.0.20 doesnt have it.

[fglez]

Edited with gcore clarification.