Problem with large select - PostgreSQL starts eating memory/disk

I am having a problem with an INSERT into a table from a SELECT which
joins over 15 tables none of which have over 3K tuples. I run it
multiple times with different values as part of my application and it
randomly locks up when executed with different parameters. When I say
lock up, I mean that it never finishes executing and that memory usage
creeps up to the maximum (according to "top" - 1 GB from ~120 MB
during normal operations) and it starts taking up the disk space for
the partition with the data files - /var/lib/pgsql (goes from ~100 MB
to ~12 GB over 45 minutes). It finally dies when there is no disk
space left. If I drop the database and recreate, this goes back to 120
MB.

We are running PostgreSQL 7.1.3 compiled on Red Hat 7.1 which runs on
a dual Pentium III 1 GHz box with a RAID 5 setup.

I would suspect that it is a problem with our query but the odd part
is that it will succeed for a few iterations (from 2 to 20) before
locking up. We have done all the normal optimizations/tuning including
creating indices on all the columns in the join and setting the OS
parameters according to the PostgreSQL documentation.

I looked every where for other examples of this behavior and could not
find anything. I wondered if it was an issue with PostgreSQL splitting
processes over the CPUs but have found numerous examples of success
running PostgreSQL on multiple CPU servers and there should be no
reason why a database couldn't split its processes over multiple CPUs.

Has anyone ever come across this problem?