Re: [Pympi-users] pyMPI + threading

[Pat M. <pat...@gm...>]

On 11/17/06, Vinicius Petrucci <vpe...@ic...> wrote:
> "Python uses a global interpreter lock for thread safety. That global
> lock must be held by a thread before it can safely access Python
> objects."

The Python thread model is really pretty weak.  This is a weakness in
reference counted languages (the incref/decref's are critical sections).
One can think of Python threads as "co-routines"  even though each
thread is built on top of a real POSIX or Windows thread. You can have
as many as you want (upto the thread limit), but only one operates
at a time.  Every so many byte code operations, the threads context
switch.

A routine that takes a long time (like a blocking I/O call) is supposed to
play nice and release the global interpreter lock (and subsequently reacquire
it before returning to the interpreter).

With MPI, your mileage will vary.  An MPI implementation is allowed to put
restrictions on the use of MPI with threads.  A common restriction is to
force all MPI calls to be issued from the master thread.  A somewhat
weaker one is that all MPI calls must come from the *same* thread.
Weaker still is that only one MPI call can be extant at a given time (presumes
that the user provides locking).

So, to get back to Vinicius' question.  If your MPI supports threads
(e.g. OpenMPI),
then pyMPI will work fine with threads.  The Python global interpreter lock
will guarantee that only one thread is working with MPI at a time (unless you
write an extension that releases the global lock).   The bad news is that
the current configuration will not allow you to get much thread parallelism
going on.  If you have a blocking receive in one thread, it will hold the
global interpreter lock, so no other Python thread can make progress.

I have been thinking about modifying pyMPI to release the global interpreter
lock on entry to a pyMPI routine.  Then, one could write threaded routines
to have blocking calls.  To prevent simultaneous MPI calls, the pyMPI
routines would then need to acquire a pyMPI global lock before making
a call.

Bottom line:  pyMPI will work safely with threads, but the co-routine nature
of Python threads might not allow it to work the way you think.

Pat