Re: parallelism!

[glen e. p. ropella]

Scott Christley writes:
 > I'm trying but I haven't quite figured out what MPI stands for (multiple
 > process interface?).

Message Passing Interface, I believe.

 > Its also unclear to me, probably because my understanding of Swarm is
 > minimal now, but what type of parallel models are you allowing/using?
 > 
 > Local memory and message passing with no shared memory?
 > 
 > What network topologies are you expecting to support?
 > 
 > What type of memory models are you expecting to support?

Ideally, we would use pure message passing with now shared
memory.  This should allow us to spread swarms over several
independent machines in a transparent network.  But, we will
have to have one hub machine that should synchronize all the 
processes on the other machines.

However, we may use some type of virtual shared memory.

 > Is the user expected to perform the problem decomposition and Swarm will
 > handle the distribution; is the user expected to specify both, or is Swarm
 > going to be able to perform both based upon common model "skeletons" or 
 > types.

Again, ideally, the user won't have to think about distributing
objects (in reality, of course, she will [grin]).  The idea is
to allow the user to program a model without thinking too much
about programming the computer[s].

 > I also see a different viewpoint when you discuss parallelism-1; which is
 > parallelism not of a single program, but of running many concurrent
 > simulations.  Presumably this is important because a researcher may perform
 > 1000 runs of a simulation and perform some statistical analysis on the 
 > results.

First off, none of this is true "parallelism".  That's a complete
misnomer.  It's really concurrency (in the abstract) and distributed
computing.  So, if we just accept the usage of the word "parallel"
to mean "doing two or more things at the same time", then we're ok. [grin]

But, semantics aside, ||-1 is important to a small extent in 
allowing a user to run a bunch of well-specified independent
Swarms at the same time.  The only thing this really saves is
time.  (There's a continuum between ||-1 and ||-2 and somewhere
in between the two, we could make use of the data being generated
by any one indep. swarm and make decisions to cancel some swarms
as dead-ends in a search or whatnot...)  And it will help us 
work an implementation of MPI into Swarm gently.  Other than
that it's pretty trivial.

glen


                  ==================================
   Swarm-Modelling is for discussion of Simulation and Modelling techniques
   esp. using Swarm.  For list administration needs (esp. [un]subscribing),
   please send a message to <address@hidden> with "help" in the
   body of the message.
                  ==================================