Re: [computer-go] Event-driven programming

[david doshay <ddoshay@xxxxxxxxxxx>]

It is true that to date we have used the power of the cluster to
1) do parallel generation of initial moves for possible lookahead
	paths, and then
2) run a number of those paths in parallel.

However, we developed the system with an eye towards doing
something a bit closer to what Peter is trying. We are thinking
about sending messages to workers that are almost like your
objects: they each can bring a different perspective to solving
one part of the problem ... like you say, different kinds of move
generators or pattern detectors. They would not all always get
used, i.e. a joseki pattern database would get used at one part
of the game but not others; the worker that checks the result of
a ladder only gets assigned to a cpu when needed, etc.

We just are not that far yet.

At this time it is easy to test the infrastructure in the context of
Gnu Go. We will see how far we get with the next steps, which
are the harder ones.

David

On Thursday, January 15, 2004, at 09:21  PM, Peter Drake wrote:

> I think what we have mind is different from David Doshay's MPI stuff.  
> We're not trying to parallelize to speed things up.  We are working on 
> a serial machine (okay, technically mine is dual-processor, but I 
> don't think about that) and finding ways to avoid some of the 
> computation by updating only the objects that might change.  With 
> luck, we can design the system to it's easy to plug in new components 
> (detectors, patterns, move suggestors, whatever) without having to 
> modify the existing code.
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go