Performance evaluation of the Caltech hypercube multiprocessing system

A Caltech hypercube has been characterized using known methods for obtaining performance figures from a MIMD computer. Speed-up, efficiency and s_1/2 parameter have been evaluated versus the physical problem dimension. The considered problem, the solution of the discretized Laplace equation with assigned boundary potential, permits variation of the computation/communication load ratio, varying the number of spatial points assigned to each processor. The P hypercube processors were mapped on a square, two-dimensional lattice, in order to obtain an Illiac IV-like configuration, and each of them was assigned to a square portion of the problem with N*N/P spatial points. The problem was studied on a simulator kindly supplied from Caltech and installed on a VAX-11/730 with VMS 3.4. The solution was implemented twice, using two different communication protocols. Tests were carried on directly on a Mark II 8 MHz 32 nodes machine at Caltech and run times were measured. Finally, the results were fitted in order to obtain experimental expressions for the various involved times; from these expressions, after an agreement check with the measured data, performance figures were evaluated.