annala%neuro.usc.edu@usc.edu (A J Annala) (09/04/89)
I am beginning to study possible solutions to the problem of solving large systems of differential equations (linear & nonlinear ode's & pde's) as part of an effort to use computer modeling as an aid for understanding how biological neural networks (e.g. the cerebellum) operate. Several people have attempted to analyze possible means to accelerate such computations by using either vector supercomputers or various regular interconnected homogeneous processor arrays. In my case, I have a large collection of machines available for use in my simulation work including Aliant, Convex, Cray, DEC, Gould, Pyramid, Silicon Graphics, and Sun systems. I have available means for automatically starting up servers, handling remote procedure calls, transferring data (integers, floating point numbers, etc) between machines in a common format, and coordinating assignment of tasks from a central dispatcher task running on my own workstation. So far, I have only used this assembly of machines for ray tracing. However, I want to use this large computing capacity to do rapid computations for my modeling work. The problem, however, is that I have found no literature references on building models using pde's and/or ode's with portions of the computations being performed with varying degrees of precision on a heterogenous mixture of machines. Perhaps someone has already thought this problem through. Or perhaps I could interest one of the readers of this newsgroup to write up their thoughts on solving this problem. In any case, a simple solution taking into account use of the native floating point format and computing structure of individual processors with sharing of information among processors in a standard format would enable one to tap into the vast unused capacity of workstations spread throughout most campuses and across the country. In my case, the ability to make effective use of the local workstations (of which there are hundreds with Motorola 68020 and 68881 chips as well as high speed risc etc) may outperform the available capacity of our local supercomputer center. Any advice you would have to offer on corrdinating many different machines in a single modeling experiment would be much appreciated. Please send email resposes to: "annala%neuro.usc.edu@usc.edu" Thanks, AJ Annala, USC Neuroscience Program