Investigation on JVM crash without core dump (by JIT compiler)

Question

Hi, All
I posted "JVM crash without core dump due to CompilerThread1" couple months ago.
http://forum.java.sun.com/thread.jspa?threadID=5253434
I would like show what we found, and ask couple questions
1) The reason of JVM crash is "CompilerThread0/1 wanted to allocate more native memory and eventually exceeded the limit of 32-bit Linux (on Redhat, the limit of VIRT is 3G)
2) After we lower the heap size (to reduce process size) and added process size monitor (track process size every minute) and JIT compilation log (-XX:+PrintCompilation),
We found sometimes JVM process jumped more than 800M when compiling one method. with the following log
Total time for which application threads were stopped: 0.1997400 seconds
5828 xxx.xxx.SomeClass::someMethod (1507 bytes)
5828 COMPILE SKIPPED: out of nodes during split (not retryable)
549002.449: [GC [PSYoungGen: 517897K->28390K(551296K)] 942855K->453348K(1229952K), 0.0726350 secs]
if we saw "COMPILE SKIPPED: out of nodes during split (not retryable)", the process size of JVM (VIRT/RES) always jumped 800M~1000M, and sometimes the memory get reverted in 30mins to couple hours, and sometimes it lasted forever. (so before we have larger footprint, this jump will kill JVM, right now, if the jump lasts forever, any more allocation on top of it also can kill jvm (much rare))
3) This is definitely a bug of JVM, because
it only happened on server mode, not client mode
it is random, that method can be compiled success on other JVM or next restart (we have 20 JVMs)
the method is not that complicate, like 100 lines, (bigger method gets compiled success)
this only happen on PROD environment, we can't reproduce it locally or QA (the method is always compiled success)
It mostly like under some condition (maybe node space is not enough), to compile that method will trigger JVM trying to allocate much more native memory.
4) We are going to disable this method by ".hotspot_compiler" to fix it, right now we are using lower footprint, JVM dies rarely.
+To understand more about this, I have some questions about JIT compiling+
I saw same method are compiled more than once from JIT log, like the method caused our problem, it only happens on 2nd time compilation.
Is it that JIT compiler will recompile the method with deeper optimization level some time, and more optimization it uses, the more memory it requires?
(like gcc has -O2 -O3)
Thanks!
Neo

Answer 1

Today, one JVM crashed again, with
{Heap before gc invocations=4188:
PSYoungGen      total 540864K, used 487958K [0x8ba50000, 0xb1250000, 0xb1250000)
  eden space 467328K, 100% used [0x8ba50000,0xa82b0000,0xa82b0000)
  from space 73536K, 28% used [0xa82b0000,0xa96d5850,0xaca80000)
  to   space 71680K, 0% used [0xacc50000,0xacc50000,0xb1250000)
ParOldGen       total 1024000K, used 493477K [0x4d250000, 0x8ba50000, 0x8ba50000)
  object space 1024000K, 48% used [0x4d250000,0x6b439610,0x8ba50000)
PSPermGen       total 101760K, used 101240K [0x2d250000, 0x335b0000, 0x4d250000)
  object space 101760K, 99% used [0x2d250000,0x3352e1f8,0x335b0000)
405290.711: [GC [PSYoungGen: 487958K->17397K(544192K)] 981435K->518248K(1568192K), 0.0769350 secs]
Heap after gc invocations=4188:
PSYoungGen total 544192K, used 17397K [0x8ba50000, 0xb1250000, 0xb1250000)
eden space 472512K, 0% used [0x8ba50000,0x8ba50000,0xa87c0000)
from space 71680K, 24% used [0xacc50000,0xadd4d480,0xb1250000)
to space 70208K, 0% used [0xa87c0000,0xa87c0000,0xacc50000)
ParOldGen total 1024000K, used 500851K [0x4d250000, 0x8ba50000, 0x8ba50000)
object space 1024000K, 48% used [0x4d250000,0x6bb6cf70,0x8ba50000)
PSPermGen total 101760K, used 101240K [0x2d250000, 0x335b0000, 0x4d250000)
object space 101760K, 99% used [0x2d250000,0x3352e1f8,0x335b0000)
Total time for which application threads were stopped: 0.0779620 seconds
Exception in thread "CompilerThread1" java.lang.OutOfMemoryError: requested 4522768 bytes for Chunk::new. Out of swap space?