Design and simulations of a serial-link interconnection network for a massively parallel computer system

Abstract

This paper presents a study of an interconnecting architecture between an array of processing elements and shared memory modules in a massively parallel computer system. This architecture is called Serial Link Interconnection Network (SLIN). SLIN is based on ultrahigh-throughput (more than one gigabits per second) bi-directional serial links. These links connect serially ported processing elements and memory modules through a network of switches. The interconnection network places the outer columns of the switches in a tree topology while the innermost columns are linked point to point. The results of the SLIN computer models' simulations are analyzed. The findings show the advantages of SLIN architecture in terms of cost-effectiveness and link utilization.<>