Numerical Simulation of Laminar Diffusion Flames

                               Craig C. Douglas

                            IBM Research Division,
                      Thomas J. Watson Research Center,
                                P. O. Box 218,
                     Yorktown Heights, NY 10598-0218, USA
                                     and
                       Department of Computer Science,
                               Yale University,
                              P. O. Box 208285,
                        New Haven, CT 06520-8285, USA.

                                Alexandre Ern

                    Department of Mechanical Engineering,
                               Yale University,
                              P. O. Box 208286,
                        New Haven, CT 06520-8286, USA
                                     and
                      CERMICS, ENPC, La Courtine, 93167,
                            Noisy-le-Grand Cedex,
                                   FRANCE.

                              Mitchell D. Smooke

                    Department of Mechanical Engineering,
                               Yale University,
                              P. O. Box 208286,
                        New Haven, CT 06520-8286, USA.


                                   ABSTRACT

Not too long ago, anyone wanting to solve large science or engineering
problems had to first get access to a supercomputer costing millions of
dollars.  Quite recently, a new breed of relatively inexpensive work stations
became widely available.  These machines have scalar peak speeds of 30--275
megaflops with ones on the horizon of 400 or more (which compares rather
favorably with vector supercomputers of not so long ago).  While these rates
are only seen for simple problems like dense matrix--matrix multiplication,
the rates seen for many problems are quite high.

In this article, we describe a class of problems which can now be solved on
machines individuals can afford to own rather than just on ones costing
millions of dollars.  Of course, the problem with using a single one of these
machines is that the option of connecting a collection of machines together or
buying a parallel version of the work station becomes more and more
tantalizing.

In fact, during the course of two years we did all of the above.  We started
on a single machine with a 100 megaflop peak rate (an IBM RISC System/6000
model 560 computer).  Then we used a farm of the IBM's, an IBM SP1, and
finally an SP2.  Due to a nice feature of the communications' library we used
(EUIH), the executables worked on the ethernet at Yale or on the fast switches
in the SP1/SP2's without either recompiling or relinking.