Nevertheless, I am willing to investigate the problem. It does not
look to me like a hware problem, seeing how it happens reasonably
consistently at the same point depending on the amount of memory
used.
I will sprinkle some traces to see if I can narrow the problem area
down.
As I said, the memory is rather busy. The 16MB increased memory
is used as a memory buffer for a large sort, so it is fully written
and read during this phase. Of course, the problem may be triggered
by the extra memory load but maybe it is not related to the sort
process itself.
Most of the base memory that the program uses (16MB always) is a
database cache which is again very busy. The last 1MB is the main
program.
So a 1MB program uses a 16MB db cache (malloced in 2KB blocks) while
sorting a large file (about 200MB) using another 16MB sort workarea
(malloced as one 16MB piece).
I will test it with the non-SMP kernel to see if it really matters.
later.
-- Eyal Lebedinsky (eyal@eyal.emu.id.au)