Re: parallel processing models

[Marcus G. Daniels]

>>>>> "SB" == Steve Brophy <address@hidden> writes:

SB> Our question: is the plan defined to enough detail that it would
SB> give us a place to start, and also might have some small enough
SB> pieces that we could develop and contribute?  Or are we better off
SB> looking to examples of how people have done model-specific
SB> parallization?

The easiest way to get parallelism is to spawn model runs with
different initial conditions to different processes, and to do this
on an SMP or cross-box multiprocessing system like MOSIX. 

For more control, MPI, PVM, or BSD sockets could be used to used as a
means for sending parameters to and collecting data from slave Swarm
models.  HDF5 also works with MPI, so it can be used to efficiently
collect data from multiple threads of execution.  And BSD sockets ought to
be fairly easy to make work with Lisp serialization.

Another way to get parallelism is to have libraries that have 
specialized support for it.  For example, a package like SME
(http://iee.umces.edu/SME3) can run diffusion across machines,
transparently handling the spatial boundary conditions, etc.  A person
might invent or adapt a library that did some domain-specific
task in a parallel way.

None of these things have much to do with Swarm as a multi-level
discrete-event simulation; both parallel parameter sweep and
library-specific parallelism can be done completely asynchronously of
general Swarm development.

As for work on making the activity library exploit concurrency
information, well, minimally, that would require someone to add logic
to the activity run loop to spawn threads (or whatever) for concurrent
actions.  It probably wouldn't be that hard to get *something* working,
but I suspect users would find it hard to debug and understand what
was happening to their models.  I believe activity visualization tools are 
crucial to making parallelism features useful.

All that is needed is enough money for one good full-time hacker.

                  ==================================
   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.