As announced I let the machine run till today - with no more crashes. The distributed GC interval was set to 4 hours. I believe we will be able to use this workaround in our production environments. Later this year we may upgrade to Java 6 which did not show the erratic behavior. Moreover, as we do not really know which combination of individual factors cause the bug to surface at all, we cannot even know for sure if there is a hardware configuration out in the field which would fail in the first place.
Nevertheless, thanks for all the input and suggestions. As a roundup I will just let the machine run memtest86 for some time to make sure this is not a memory problem. However I do not believe it will show any problems for reasons stated earlier.