Parallel Processing in FDS

[Kevin McGrattan <kevin.mcgrattan@nist.gov>]

All

We've received several inquiries about running FDS in
parallel, that is, running FDS
on more than one CPU. We are currently working on a version
of FDS that uses MPI
(Message-Passing Interface), a fairly well-established set
of libraries that allow
two or more computers to share information during a
calculation. We have the code
running on our Linux machines connected with a fast
(gigabit) network. This type of system is often called a
Beowulf cluster, although there are several different
definitions. At some point, we will release the code for
beta testing, but some
forewarning is necessary. Running FDS in parallel will not
be as easy as running it
on a single processor. The reason is that the MPI libraries
have to be installed by
the user, and a seamless network of machines has to be
established. By ``seamless''
we mean that the machines have to be able to communicate
without any password
prompting, network slowdowns, etc. Linux and Unix are far
better operating systems
for this type of work, but it is possible to run in parallel
on a network of 
Windows PCs. We're working on this right now, so we'll keep
our fingers crossed.

In any event, for those of you who might be interested in
using a parallel version
of FDS in the future, you will have to install the MPI
libraries on your machine(s).
There are several implementations of MPI, including MPICH, a
product of Argonne
National Labs, and LAM/MPI, a product of Indiana University
and the University of
Notre Dame. There are others, and this is not an endorsement
of either. We use LAM
on our Linux cluster, but LAM will not run on a Windows
cluster (MPICH and others
will). There is some pain involved in getting MPI to work,
and we would expect the
users to undergo this pain without much help from us. We're
not experts in parallel
programming. We have implemented into FDS a relatively
simple set of MPI constructs
(MPI-1) that should run under anyone's implementation of
MPI, much like our Fortran
should compile and run on any machine with the appropriate
Fortran libraries installed. But remember ``should'' has a
different meaning than ``will.''

Otherwise, keep us apprised of any bugs or problems with the
present (v 3.1) version
of FDS. Make sure you set the ``Reply to'' in your email to
your working address.
Too often I write a lengthy reply to someone's question,
only to have my mail bounced
back to me by the server to which I am writing. I'm too lazy
to search around for
an alternate address, so make sure that ``Reply to'' will
get the message to you.
By the way, would the user in Germany who wrote this morning
send me another 
message. I deleted your email after replying, then my email
was bounced back.

Kevin

-- 
Kevin McGrattan
National Institute of Standards and Technology
100 Bureau Drive, Stop 8663
Gaithersburg MD 20899-8663

Telephone: (301) 975 2712
FAX:       (301) 975 4052