FWIW I am on Ubuntu 10.04 with all patches applied and could not reproduce the above problem when I tried. The version of python-psycopg2 from the standard repository is 2.0.13. Not an iron-clad guarantee because something else could be at play but at least a data point. I also drove the application a lot harder from ab with 10x number of requests to try and reproduce but still nothing.
The fact it doesn't fail when processes=5 and threads=1 agrees well with the GIL problem - without threads there is no need to hold the GIL.