Re: [Swarm-Modelling] ABMs on Graphical Processor Units

[Marcus G. Daniels]

Seth Tisue wrote:
>  Russell> I think the most useful outcome of this work would be a
>  Russell> library of data structures and algorithms with a C language
>  Russell> interface (with a JNI Java interface a relatively trivial
>  Russell> proposition on top of that) that provides useful building
>  Russell> blocks for ABMs. Such a library would be useful for all of the
>  Russell> above toolsets (except perhaps NetLogo, for which is not
>  Russell> obvious how you call the Java layer from logo).
>
> (NetLogo models may call out to arbitrary Java code.  This is documented
> at http://ccl.northwestern.edu/netlogo/docs/extensions.html .
Unless there are high level operators that express concurrency (e.g. map 
a function over a space, or merging discrete event schedules across 
agents), I don't think it will work very well to just layer on a 
library.   I think what's probably ideal is not a library but a compiler 
for an appropriate high level language for ABM.   Maybe write a MetaABM 
GCC or LLVM frontend?  Short of something like that, I think the only 
way to go is to write several tuned libraries for specific tasks (e.g 
spaces) which are also for specific architectures (assuming there are 
sensible interfaces from above). 

There are people working on adapting GOMP (GCC's OpenMP) for the Cell.  
One idea would be to extend GCC's Java compiler with some kind of 
OpenMP-like directives.   But for Java on Cell it's just such a big VM 
to slosh around.  I can't believe that would ever work well.  Unclear 
what kind of progress is being made on IBM's X10, but it doesn't sound 
to me like that project has displaced IBM's standard compiler suite (and 
Octopiler), or the Cell Superscalar project. 

Marcus