smb@cs.columbia.edu said:
> run memtest+

I thought of that too, but what I'm seeing doesn't fit into
the usual pattern of memory corruption under load.
There is the "the hotter the box runs the better" effect,
and the programs affected are more I/O intensive then
CPU consuming. Well, actually, this mostly happened in
"make"-processes where data are streamed through pipes
a lot...
And, as said, never a single glitch if the processors don't idle.

best regards
Matthias