I have an SML program which sometimes fails while trying to allocate an array
of 3 million ints, with the message
Run out of store - interrupting threads
and an exception.
This happens on one of the first substantial allocations in the program, near
the start of its run, so there should be space available - and it only happens
on occasional runs, and if it doesn't fail, then the rest of the program (which
does a lot more work) runs fine.
The exact size of the array being allocated doesn't seem to matter so long as
it's reasonably large, but despite this and the fact that it's so early in the
program, I haven't yet managed to make a simple test program with a similar
allocation pattern that fails.
The situation seems to be the same regardless of whether I use Poly/ML 5.7,
5.8, or current git.
I compared a successful run with an unsuccessful one based on logging in the
runtime. In both cases this allocation gets sent for a "quick GC", which
returns failure, followed by a full GC, which runs and then either succeeds or
fails to establish that there is enough memory available afterwards (via
CheckForAllocation).
In the "good" case, when this check comes to AllocHeapSpace, the test "if
(space->allocationSpace)" never succeeds and we fall out of the first loop with
currentAllocSpace == 0, defaultSpaceSize == 131072, spaceBeforeMinorGC ==
120608, and our request for minWords == 3000001 is successfully fielded in the
following conditional block.
In the "bad" case, when this check comes to AllocHeapSpace, the "if
(space->allocationSpace)" test succeeds once (available == 131070), then we
fall out of the first loop with currentAllocSpace == 131072, defaultSpaceSize
== 131072, spaceBeforeMinorGC == 120608, the following conditional is not
entered, and we return failure from the line at the end of the function,
causing the GC to return failure as well.
If I comment out the check
if (currentAllocSpace/* + minWords */ < spaceBeforeMinorGC)
then it apparently always succeeds, but I can see that this is no solution
because no GC is happening at all without it...
Any suggestions for what I might look at next?
Chris
_______________________________________________
polyml mailing list
polyml@inf.ed.ac.uk
http://lists.inf.ed.ac.uk/mailman/listinfo/polyml
