Threadpool item mailboxes design problem

[Charles Hixson]

On 04/14/2013 07:32 PM, Chris Rebert wrote:
>
> On Apr 14, 2013 4:27 PM, "Charles Hixson" <charleshixsn at earthlink.net 
> <mailto:charleshixsn at earthlink.net>> wrote:
> >
> > What is the best approach to implementing actors that accept and 
> post messages (and have no other external contacts).
>
> You might look at how some of the existing Python actor libraries are 
> implemented (perhaps one of these might even save you from reinventing 
> the wheel):
>
> http://www.pykka.org/en/latest/
> http://www.kamaelia.org/Docs/Axon/Axon.html
> https://pypi.python.org/pypi/pulsar
>
> Kinda old:
> http://candygram.sourceforge.net/contents.html
> http://osl.cs.uiuc.edu/parley/
>
Candygram looks interesting.  I'd forgotten about it.  The others look 
either a bit limited (in different ways), or overly general, with the 
costs that that brings.  I'll need to study Candygram a bit more.  
However, even Candygram seems to have a RAM centric model that I'd need 
to work around.  (Well, the mailbox synchronization must clearly be RAM 
centric, but the storage shouldn't be.)
>
> > So far what I've come up with is something like:
> > actors     = {}
> > mailboxs = {}
> >
> > Stuff actors with actor instances, mailboxes with 
> multiprocessing.queue instances.   (Actors and mailboxes will have 
> identical keys, which are id#, but it's got to be a dict rather than a 
> list, because too many are rolled out to disk.)  And I'm planning of 
> having the actors running simultaneously and continually in a 
> threadpool that just loops through the actors that are assigned to 
> each thread of the pool.
> <snip>
> > It would, however, be better if the mailbox could be specific to the 
> threadpool instance, so less space would be wasted.  Or if the queues 
> could dynamically resize.  Or if there was a threadsafe dict.  Or... 
>  But I don't know that any of these are feasible.  (I mean, yes, I 
> could write all the mail to a database, but is that a better answer, 
> or even a good one?)
>
> My recollection is that the built-in collection types are threadsafe 
> at least to the limited extent that the operations exposed by their 
> APIs (e.g. dict.setdefault) are atomic.
> Perhaps someone will be able to chime in with more details.
>
If list operations were threadsafe, why would multiprocessing.queue have 
been created?  I don't recall any claim that they were.  Still, I've 
found an assertion on StackOverflow that they are...at least for simple 
assignment and reading.  And the same would appear to be true of dicts 
from that post.  This *does* require that either the index be constant, 
and the stored value be constant, or that the code section be locked 
during the access.  Fortunately I'm intending to have id#s be 
unchangable, and the messages to be tuples, and thus constant.

OTOH, to use this approach I'll need to find some way to guarantee that 
removing messages and posting messages don't occur at the same time.  So 
that still means I'll need to lock each access.  The answer that seems 
best is for each thread to have a mailbox that cells within the thread 
post and read messages from.  This will automatically deal with internal 
to thread synchronization.  Then I'll need a mailman thread that...

This seems a promising approach, that avoids the problem of fixed length 
queues, but I'll still need to do a lot of synchronization.  Still, it's 
a lot less, and each thread would be locked for shorter amounts of time.

-- 
Charles Hixson

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/python-list/attachments/20130415/09c6e692/attachment.html>