OK, the problem has receded now, thanks for your input, much appreciated. The confusing thing was the JVM JMX beans showing the GC times and counts for PS Scavenge and PS MarkSweep painted a totally false picture. After reading the document suggested by PeterJ, and reading on about Ergonomics of the JVM I arrived at having only one JVM option set: -XX:MaxGCPauseMillis=500 We now let the JVM manage memory itself, as discussed here: http://java.sun.com/j2se/1.5.0/docs/guide/vm/gc-ergonomics.html The other problem was quickly highlighted once all the GC'ing was eliminated; our half hourly DB backup was gumming up the system. Thanks again to everyone for taking the time. PeterJ, I had already read your presentation, and I would like to add that I considered implementing something based on your suggestions, but instead of using the GC debug information, getting the info from the JMX beans. The advantages being that i) we already have a graphing solution ready to plug in the output (cacti), and ii) I could turn graphing of GCs on and off without a restart. However, I now have to conclude that the JMX beans do not paint such a complete or accurate picture as the GC debug output. View the original post : http://www.jboss.com/index.html?module=bb&op=viewtopic&p=4172034#... Reply to the post : http://www.jboss.com/index.html?module=bb&op=posting&mode=reply&a...