Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!uflorida!gatech!hubcap!matloff%mole.Berkeley.EDU
From: matloff%mole.Berkeley.EDU@ucbvax.Berkeley.EDU (Norman Matloff)
Newsgroups: comp.parallel
Subject: Re: Distributed simulation
Message-ID: <5276@hubcap.clemson.edu>
Date: 26 Apr 89 20:31:51 GMT
Sender: fpst@hubcap.clemson.edu
Lines: 44
Approved: parallel@hubcap.clemson.edu

In article <5117@hubcap.clemson.edu> usha%FSU.BITNET@CUNYVM.CUNY.EDU writes:

>The following is a list of references on distributed simulation.  But
>this has not been updated for the past two years.

I'm glad you brought up the subject, because I would like to raise a
question.

Here is the setting.  It's discrete-event simulation, which many or most 
of your references are concerned with, on multiple-processor systems
(tightly or loosely coupled).  Let P be the number of processors and T
be the amount of run time needed to get the desired statistical accuracy.

The usual approach in these papers is to have different processors handle
different parts of this program.  E.g. one processor might manage the
event list, another might do the statistics collection, etc.  Or, in
process-oriented simulators like Simula or csim (or, for that matter,
Modula II applied to simulation), each processor might handle a different
process.

It has always seemed to me that these methods are inherently wasteful.
Each processor is going to have a substantial amount of idle time.

Moreover, it seems to me that the best method is also the simplest  --
just have all P processors run the uniprocessor version of the program,
but for T/P amount of run time, instead of T.  [Of course, each processor 
needs to use a different random number seed.]

This method gives 100% utilization for all processors, and except for
start-up transients, should give the same statistical accuracy as the
other methods, in fact BETTER accuracy, since in the other methods
time T is actually less than time T, due to idle periods.

Is anyone aware of any thorough work on this question, i.e. the question
of whether the complex methods are better or worse than the simple one?
I think there was a brief paper by Heidelberger on the effect of 
transients in this context a few years ago in one of the Winter 
Simulation Conferences, but it didn't go very far.

Some of the complex methods give a really elegant theory (e.g. see the
references to Chandy), but it still seems to me that the simple method
will give better performance.

   Norm