I want to figure out why JVM heap usage on Elasticsearch node is staying consistently above 80%. In order to do this, I take a heap dump by running
jmap.exe -heap:format=b 5348
(5348 is the Process ID). Then I can analyze the dump with VisualVM.
The problem is that jmap pauses the JVM while taking the dump, so the node is basically offline for around 5 minutes.
This article suggests a faster approach that relies on taking coredump with gdb on Linux. I already tried WinDbg, which created a core dump, but I couldn't use it in VisualVM.
Is there a similar approach for Windows? How one can take heap dumps in seconds, not minutes?
After you've taken the coredump by WinDbg, you need to extract the heap dump from it by running
jmap -heap:format=b "C:\Program Files\Java\...\bin\java.exe" core.mdmp
This could be done offline; no interaction with running Java process needed. Then you will be able to open the generated heap.bin in VisualVM.
Alternatively you may take the class histogram. It is produced a way faster than full heap dump.
jmap -histo <PID>
It shows you the list of classes whose instances occupy the most space in the heap. This information is often enough to get the idea of where's the memory lost.