On Mon, May 27, 2013 at 12:19 AM, Ian Kelly <ian.g.ke...@gmail.com> wrote: > 7) Since the program being tested does basically nothing except start > and exit threads, the extra 40% probably represents the overhead of > all that starting and stopping, which would be done outside the GIL.
To test this, I tried running the script in Python 2.7 in Linux with the print statements removed and verified that it was using about 135% of the CPU. However, top also told me that only about 95% of that was user processes; the other 40% was kernel usage. The 40% doesn't seem to be threading overhead though, because I tried adding large xrange loops to slow down the thread creation process and it had no effect on the stats. Then I tried running the same program in Python 3.2, and I got the more expected 100% CPU usage with minimal kernel time. So I'm thinking now that the extra 40% may actually be overhead induced by the GIL. If that's the case then wow, the old GIL really did suck. -- http://mail.python.org/mailman/listinfo/python-list