Re: simulating large models

[Paul Johnson]

I waited for the dust to clear on this thread because I was eager to see 
what answers emerge.

I don't think the number of agents will cause any unexpected scalability 
problems if you write the model in Objective-C. In C you can allocate 
memory and free it explicitly, and you are not dependent on a garbage 
collector to do the right thing.

In Java, I personally do encounter several scalability problems that I 
have not been able to understand.  The size of the JVM gets huge and 
many of the tricks that I expect to help don't make any difference. For 
example, converting IVARS that are common among a class of agents into 
static Class variables has no effect. Converting doubles to single 
precision floats and changing ints to small and so forth yields no 
noticeable advantage for me.

I don't think these problems I have in scaling with Java are due to 
Swarm, but rather Java itself and my relative lack of incentive to learn 
my around in it.  Over the last year and a half, Marcus Daniels has 
posted some memos about how to deal with some Swarm related Java memory 
problems and users indicate they work pretty well.  One problem was that 
the Java garbage collector never could figure out when to drop things, 
and there were a couple of other issues that, frankly, flew past my ears 
and never caused me to look further.  I have a collection of those notes 
somewhere.

Now, in terms of scheduling agents and the Swarm activity library, I 
think you will see huge differences in performance depending on how you 
structure your simulation.  If you have dynamic scheduling, I see some 
major performance differences across theoretically equivalent designs. 
It is BY FAR faster to just write a loop over all the agents and tell 
them to do something at each time step than it is to employ 
createActionForEach or createFActionForEach or whatever, especially if 
you want randomized order of traversal.  Going at it the "loop way" does 
not, however, interleve actions of diverse sets of agents in a 
meaningful way, and if you need that kind of thing, you will see a 
performance hit that will frustrate you a lot.  I have worked on that 
question some and talked about it at Swarmfest in the context of 
alternative approaches for dynamic scheduling. There are occassions 
where you really do need the hierarchical dynamic schedules, and 
especially the use of randomization in them has a pretty big performance 
hit.

http://lark.cc.ukans.edu/~pauljohn/ResearchPapers/Presentations/Swarmfest02/DynamicScheduling/

That reminds me I made a promise to develop a couple of methods for 
collections to speed up use of randomized collection indexes and I will 
do it.  I also am trying to understand some elements of design in the 
Activity library to see if there is a way I can make the think I called 
decentralized dynamic scheduling go faster.


Juan A. Rodriguez wrote:
> Hi there,
>
> Does anyone have experience with simulating large models with swarm?
> Is swarm prepared to distributed simulations across several
> processors/machines?
> We're  thinking of large market models which may include hundreds (or even
> thousands) of customers (agents) and
> we foresee scalability problems.
> Perhaps we should use some high performance simulator, but we're not sure
> about it.
> Any hint or suggestion will be welcome.
> Thanks.
>
>
> Juan A. Rodríguez-Aguilar, PhD
> Senior Researcher

--
Paul E. Johnson                       email: address@hidden
Dept. of Political Science            http://lark.cc.ku.edu/~pauljohn
University of Kansas                  Office: (785) 864-9086
Lawrence, Kansas 66045                FAX: (785) 864-5700


                 ==================================
  Swarm-Support is for discussion of the technical details of the day
  to day usage of Swarm.  For list administration needs (esp.
  [un]subscribing), please send a message to <address@hidden>
  with "help" in the body of the message.