Python threading (was: Re: global interpreter lock not working as it should)
anton_shim at hotmail.com
Tue Aug 6 03:24:42 CEST 2002
I remembered something that changes my "coincidence" theory:
the Linux OS checks if it needs to reschedule after system calls and
(depending on if you have pre-emptive patch or 2.5+) on return from an
interrupt in entry.S
This makes the coincidences that allow another thread to grab the GIL as
1) the python thread currently holding the GIL depletes its time-slice
2) the current python thread calls pthread_cond_signal in this depleted
state BEFORE the OS notices it needs to be removed from the CPU by other
means. This means there can be no system calls between the time when the
process runs out of time and when it calls pthread_cond_signal. (In certain
cases this means no interrupts either) If the OS notices AND removes it
before it releases the lock and signals the waiting thread or threads back
onto the run-queue, it will get to go for another full timeslice.
I have a feeling that this coincidence is very likely.
But it still holds that a python thread will run a full time-slice which is
usually in the range of 1/10 of a second (110 + ms), but will try to release
the lock every 10 byte codes which on avg will is in the 10s of microseconds
range on my 1 Ghz P3.
SCHED_RR threads should also switch at the end of their timeslice. Only
SCHED_FIFO threads which have no concept of time-slices will never ever
release the GIL. I have not tested this with the semaphore version of
Another side-effect is that if the process it taken of the run-queue before
calling pthread_cond_signal, all other threads will be blocked for the
duration of its sleep and the duration of its next run.
If you're wondering what I changed, I had said before that the timer-tick
and the release have to coincide. This is wrong. The timer-tick only NOTICES
the depletion of the time-slice but will not take action on it *unless* you
have a patch or 2.5+ version which makes it check for need to reschedule on
an interrupt return. Even if this were true, it only happens 100 times a
Join the worlds largest e-mail service with MSN Hotmail.
More information about the Python-list