Multicomputer Technology Initiative

Introduction

This project is funded jointly by De Beers, and the South African Department of Trade and Industry. The project has its origin in the following classic problems associated with massive computing requirements:

Highly specialised software support, unique to a specific processor.
Rapid obsolescence of hardware and software.
Highly inflexible solution i.e. cannot be adapted for increased load or for mode changes in the case of real time systems.
Difficult and unpredictable solutions associated with bus oriented designs.

It seems that the optimum solution is a flexible hardware, probably based on crossbar technology, supported by overarching
layers of software implemented in standard languages such as `C'. The support software should allow the crossbar to be reconfigured `on the fly' to handle different load requirements. In addition, low level, processor specific libraries should be available to extract maximum performance from specific processors. These libraries should preferably be available from third party sources, leaving the system integrator with the job of writing high level code.

In addition, a high level simulation environment is required. In this environment, system level software should be prepared and tested with a high level of confidence that the target system will perform to specification. It is not required that the simulator would run in real time. To some extent, the flexible, scalable nature of a crossbar configuration allows errors of judgement in algorithm load to be accomodated.

This project has many of the requirements of the USA based initiative known as RASSP, and the many Beowulf initiatives around the world.

The project assumes that to meet the rapid prototyping requirement, that commercial technology will be utilised. During 1998, a cluster of 8, 350MHz Pentium class machines (one with a dual processor) was purchased, and is now running Linux. A PBIL algorithm to design antenna arrays was successfully tested using the PVM (Parallel Virtual Machine) libraries. Work is now directed towards testing some general purpose image processing algorithms specified by our sponsors. During 1999, this work will continue, as well as looking at implementing Message Passing Interface (MPI).

Currently the eight machines are linked together with an 8 port, 100MHz EtherNet hub. During 1999, we will investigate using ATM network cards and router for linking the machines. It is anticipated that special management software will be utilised to ensure permanent ATM circuits between sub-clusters of the machines, to optimise performance for algorithms where significant inter-processor data flow needs to take place.

to MTI home page