[Python-3000] the future of the GIL
talin at acm.org
Thu May 10 04:21:34 CEST 2007
Greg Ewing wrote:
> Giovanni Bajo wrote:
>> using multiple processes cause some
>> headaches with frozen distributions (PyInstaller, py2exe, etc.), like those
>> usually found on Windows, specifically because Windows does not have fork().
> Isn't that just a problem with Windows generally? I don't
> see what the method of packaging has to do with it.
> Also, I've seen it suggested that there may actually be
> a way of doing something equivalent to a fork in Windows,
> even though it doesn't have a fork() system call as such.
> Does anyone know more about this?
I also wonder about embedded systems and game consoles. I don't know how
many embedded microprocessors support fork(), but I know that modern
consoles such as PS/3 and Xbox do not, since they have no support for
virtual memory at all.
Also remember that the PS/3 is supposed to be one of the poster children
for multiprocessing -- the whole 'cell processor' thing. You can't write
an efficient game on the PS/3 unless it uses multiple processors.
Admittedly, not many current console-based games use Python, but that
need not always be the case in the future, and a number of PC-based
games are using it already.
This much I agree: There's no point in talking about supporting multiple
processors using threads as long as we're living in a refcounting world.
Thought experiment: Suppose you were writing and brand-new dynamic
language today, designed to work efficiently on multi-processor systems.
Forget all of Python's legacy implementation details such as GILs and
refcounts and such. What would it look like, and how well would it
perform? (And I don't mean purely functional languages a la Erlang.)
For example, in a language that is based on continuations at a very deep
level, there need not be any "global interpreter" at all. Each separate
flow of execution is merely a pointer to a call frame, the evaluation of
which produces a pointer to another call frame (or perhaps the same
one). Yes, there would still be some shared state that would have to be
managed, but I wouldn't think that the performance penalty of managing
that would be horrible.
More information about the Python-3000