Re: [Swarm-Modelling] Something New.. ?

[Matteo Morini]

   Dear Simon,

   we (a few colleagues and me) may have something to get you started:
we've been working on our spare time (which means not on a regular
basis, nor systematically) on a project called Nemote, which is a
java-based *sort of* workload manager and job distributing system. The
original idea was to run Swarm code (it was an inherently parallel task,
based on a genetic algorithm) on multiple CPUs. In a sense we had
'legacy' code (objective-C only, monolythic swarm processes) we wanted
to interact with something outside of itself.  The part which may
perhaps interest you is the use of tcp/ip sockets to have Swarm
processes passing data back and forth.
        Bill's objection about a performance hit is beyond question, and
that's why we had to think of clever ways to keep network traffic (in
your case that would be whichever std *nix i/o traffic) low, which in
the end was just a matter of packing data ("jobs") in batches to send 
around once in a while. A frequent and tight interaction between your 
"legacy" code and a swarm process would very likely slow things down to 
a crawl.
	Feel free to contact me and to take the conversation off-list, since 
most subscribers, old acquaintances of mine, are probably fed up with 
this fossil piece of work :-)

---Matteo

Bill Northcott wrote:

> On 15/03/2005, at 10:48 AM, Simon Bland wrote:
>
> > I sent this to the tech list, but haven't had any response, so I'm
> >  asking again here, I hope that's ok. ... The project is basically
> > to create an environment for doing simulations using agents that
> > are already created and don't meet any real standard or
> > specification, and aren't required to be related in their 
> > implementation (except in that they use standard I/O in the *NIX
> > sense for control/output).
> >
> > What I'm currently looking at doing is to create a Swarm module
> > that would act as an interpretation layer between a legacy agent
> > and Swarm's agent API. ... Do people feel that Swarm would be an
> > appropriate system to develop a legacy agent module for? If not, do
> > you know of a system that would be more appropriate to the task?
>
>
> I saw this on the other list but did not really understand it.
>
> What is your motivation?
>
> The communication between agents and the rest of the simulation code
>  is pretty intense and my first thought is that using *NIX i/o
> instead of Objective-C messages would be horribly slow.
>
> Bill Northcott
>
> _______________________________________________ Modelling mailing
> list address@hidden 
> http://www.swarm.org/mailman/listinfo/modelling