[Stackless] Multi-CPU Actor Based Python

Timothy Baldridge tbaldridge at gmail.com
Tue Nov 18 21:33:24 CET 2008


> Hm.  Are you forgetting that "objects" in Python are all by reference.
> And therefore it is pretty much impossible to "contain" them.
>
> Imagine I create a message:
>
>  msg = [ [1,2], [3,4] ]  # Notice msg is a reference to a tree of 7 objects
>

That's a good point, and one I hadn't really thought of. Erlang gets
around this issue by doing a "deep copy" and can at times pay a
penalty for it. I see three other options how this may work:

1) make all mutable types objects. This way lists would look like
objects, and therefore would get their own tasklet and message loop.
This would be very inefficient.

2) make lists a special type that is implemented internally with a
highly efficient locking mechanism.

3) make lists single threaded, and when passed outside of the owning
thread( eg. foo(self.barlist) ) convert them into #1 on the fly.

I understand that I'm suggesting something a bit deeper than "a simple
hack". And I know this would break allot of code, but I'm hoping that
with some modifications, a framework could be developed that would be
as good (if not better) than Erlang in terms of performance, while
keeping the extremely elegant OOP-ness of Python.

As some background to this project, I just got done studying Erlang
for a couple months, and while I love allot of the concepts of the
language, such as cheap threads, and message passing, I hate the
syntax and the fact that there isn't a good way to move data around in
the program.

I think we should look at this as a higher approximation of the real
world than what we are used to with OOP. Right now, I am typing at my
computer, as I press a key, I'm passing a message to the machine. It
responds by passing information back trough the display. At no time do
I open it up and read the data as it is being calculated. The entire
world is based off of this concept of passing "messages" in between
objects. Not in-between functions as in Erlang, or sharing memory as
in current Python implementations, but passing messages in between
objects. If we can get a framework to match this model, I think it
could be quite usable.

I'm going to develop the stackless version of this, and try to migrate
it to a VM-per-CPU model. At that point I'll see what the performance
hit is, and how well it performs against Erlang.

Keep the questions coming I'm very interested in getting feedback.




More information about the Stackless mailing list