Re: [Swarm-Modelling] Question about Distributed Simulations

[Lou Gross]

Gary,
    Good to know that you are trying this out. Below is a response 
on parallelization that I sent to this list in 2001 in response
to a question from Alan Su - my points are still valid I believe. 
I have archived a long thread of various || questions to this group 
from 1997-2004 which I can send you if you want. 
    Regarding BlueGene, Kirk Jordan (Deep Computing contact person
at IBM who is working with several biology groups to develop methods
for applications using BlueGene) and I were talking about issues 
similar to those I mention below two weeks ago. I am sure he would 
be interested in discussing your use of BlueGene. Our group here
would too as we are doing a variety of multi-scale || models
particularly tied into spatial optimization for natural resource
management. We are using ORNL clusters for most of this. See
www.tiem.utk.edu/gem for the various recent papers.
    Cheers,
       Lou

Louis J. Gross
Professor of Ecology and Evolutionary Biology
    and Mathematics
Director, The Institute for Environmental Modeling
University of Tennessee - Knoxville
President, UTK Faculty Senate
Past-President, Society for Mathematical Biology (www.smb.org)
address@hidden
http://www.tiem.utk.edu/~gross/
http://atlss.org/ (ATLSS Project Home Page)
http://www.tiem.utk.edu/bioed/ (Quantitative Life Sciences Education)

-------------------------------------------------------------------

Alan,
     We have published a few papers on parallelization issues for
IBM's - list is below. The other main collection of papers on
this are by Jim Haefner at Utah State in Biology. As you will 
see from reading our papers, getting speedup from multiple
processors is non-trivial for true IBM's (there's alot of people 
who misuse this term in the literature, essentially using it for
a model of an individual that they then expose to different 
conditions). It is the interactions between agents that make 
parallelization a real bear for agent-based models, and so 
we have stuck to grid-partitioning as the most feasible method.
With new funding from NSF, we intend to work on adaptive 
grid partitioning algorithms for these types of models, which we
feel are necessary for doing spatial control. However, we 
expect that this will require the fast backplanes of SMPs and
will not be feasible with typical slow interconnects of clusters.
As you are no doubt aware, data transport bandwidth simply cannot keep 
up with processor speeds these days, and this is a real limitation
for parallel methods due to load balancing. 
     There is an entirely different reason we use for suggesting
that parallelization of IBMs is worth considering, separate from
the issue of potential speedup. This is the argument that the
way you think about models and order of interaction in a parallel
simulation should be quite differnt from the way you think about 
it for serial implementations. Parallelization changes the model
and for ecological systems in which concurrency does really happen, 
this may provide more "realism" than any serial approach can. 
     I am attaching a text file archive of a swarm modeling thread
on parallelization that may be useful - trash this if you've seen 
it. If you run across other people working on this, I'd appreciate 
it if you'd let me know about it. 
    Cheers,
       Lou  Wed, 24 Oct 2001

Luh, H.-K., C. Abbott, M. Berry, E. J. Comiskey, J. Dempsey, L. J. Gross
(1997)  Parallelization in a spatially-explicit individual-based model (I) -
Spatial data Interpolation. Computers and Geosciences 23: 293-304.

Abbott, C. A., M. W. Berry, E. J. Comiskey, L. J. Gross and H.-K. Luh (1997)
Computational models of white-tailed deer in the Florida Everglades.  IEEE
Computational Science and Engineering 4:60-72.

Mellott, L. E., M. W. Berry, E. J. Comiskey, and L. J. Gross (1999) The 
design and implementation of an individual-based predator-prey model for
a distributed computing environment. Simulation Theory and Practice 7:47-70.