Project 59: Measuring the effectiveness of parallel architecture
Beijing 1989:
=============
Document: L. Fosdick and C.J.C. Schauble, "Simulation of Parallel
Computations", Proceedings of SOSS, Beijing, May 1989, pp 65-66.
This is a very current topic and several members of the group are
actively involved in the research related to this project.
Paul talked about the IBM RP3 and GF11 machines, and said that in the
near future they will be open to the public on an experimental basis.
Fosdick talked about the simulation of parallel computations (see SOSS
proceedings).
Rice and Houstis talked about their experiments with parallel
algorithms. With some colleagues from Purdue, they have made a number
of benchmarks over the past years on the speedup effectiveness of
various parallel machines. Their work is targeted toward programs for
solving partial differential equations which have not been "tuned" or
"targeted" to specific architectures but which are inherently highly
parallel. The parallel ELLPACK project involves development of expert
systems environment for the parallel processing of partial
differential equations around a library of parallel PDE solvers. The
current library includes finite element and difference discretization
modules, packages of parallel iterative solvers, a GE solver for full
matrices, etc. The system is currently running on NCUBE and SEQUENT
machines using an X11-window workstation as a front-end. They have
observed speedups that justify the use of such machines. For example,
in a 128 processor NCUBE they obtained speedups of 75 or 80 for
realistic codes, even though NCUBE has rather slow communication
facilities. Better performance was obtained on other machines.