I am following up on my own thread because I found the culprit!
I turns out that the JVM was leaking native memory (that's why profiling
the heap usage did not show anything, nor tracking file descriptors).
The process virtual size (VSZ in top) was growing slowly but surely,
until the OS start
We are running into OOM errors that we think are related to this:
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=4705373
Our setup:
JDK 1.5.0_08b3 on Debian Linux (2.4.21SMP)
tomcat 5.5.17
We have 3 webapps in the tomcat instance with fairly similar WEB-INF/lib
content
(struts + spring +