Nehalem and GC

Today I the pleasure of setting up a benchmark on a duel 8-core Nehalem machine. Funny thing was, the owner of the box thought there were two quad-core chips in the machine but a little investigation demonstrated otherwise.

I started running the application using the default CentOS JRE (1.6.0_09). I test ran the bench for 5 minutes and as is customary, I was watching server GC activity during the run. When the run stopped I figured GC would stop also but it didn't, it kept on running and running. So, I let it run just to see how long it would run for. After 20 minutes I returned to the machine to find that GC had just stopped.

I then upgraded to the 1.6.0_16 and re-ran the bench. The difference was huge! I've not done analysis on the logs but all I can say is that there were far few full GC pauses. Not only that, GC stopped almost immediately after the benchmark completed. Good new because what I figured to be a problem with the application turned out to be a problem with the JVM.

Just exactly what was the problem was, I'm not sure. I do know that some where after _09 and before _14, the throughput collector became NUMA aware. It's made me wonder if running older versions of JVMs that are not Nehalem (NUMA) aware maybe a bit buggy when running on that class of machine? I don't really know. What I do know if the difference in GC behavior is stunning! Long story short, if you are using Nehalem, make sure you are using and up to date VM.