JVM doing frequent full and partial GC and still not freeing up any memory.

Question

Hi,
We are using sun JVM under Tibco Active Matrix Service Grid's OSGI container. We are observing a strange behaviour where JVM is doing full/partial GCs continuously without freeing up much memory. Old generation is getting filled up and the young generation is almost empty. We have a web service deployed on this JVM. All the requests to the service are stateless so all the objects created are short lived and should be immedietely available for GC. Please see below the heap size and the GC arguments which we are using.
# Increase the default maximum heap size.
java.heap.size.initial=4096M
java.heap.size.max=6144M
-server -XX:PermSize=128m -XX:MaxPermSize=256m -Xloggc:gc.log -verbosegc -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+UseParallelGC -XX:+MaxGCPauseMillis=100 -XX:GCTimeRatio=5 -XX:+UseTLAB -XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=/opt/home/tibcoadm/heapdump.log
We need to know why continuous full GC is not able to free up any memory in old generation ?

Answer 1

Did you try 121 VM instead of 117. GC algorithm in 12 is much better than
          117.
          - Prasad
          Kirk Fritsche wrote:
          > Hi,
          >
          > I am using WLS 4.5.1 service pack 10 on Solaris 2.7.
          > We have 8 servers on one machine running under java 1.1.7 with 128/196
          > min/max memory each.
          > We use statelesss session beans only.
          >
          > One test case we have uses 36 clients processing 200 interations of
          > various transactions.
          > On numerous occasions I have seen the percent free after garbage
          > collection slowly decrease.
          > Normally the servers have 86% free, but sometimes (on one server) this
          > goes down 1-6% each time GC occurs.
          > This continues until it gets down to 30-50% free after GC and then it
          > will go back up 90% free.
          >
          > The problem is that as it has less and less free, the execute queue
          > length grows and GC occurs
          > every 30-40 seconds instead of 2-3 min like the other servers.
          >
          > Some servers never get this, and we use round-robin on the cluster, so I
          > am not looking at our code.
          > Otherwise I would expect it to affect every server.
          > This problem seems to go from server to server for no particular reason.
          >
          > Has anyone had any experience with this?
          >
          > Thanks,
          >
          > Kirk