Performance Analysis of Massively Parallel Architectures

Cube based networks have received much attention over the past decade since they offer a rich interconnected structure with a number of desirable properties such as low diameter, high bisection width, lesser complexity and Cost. Among them the hypercube architecture is widely used network for parallel computer system due to its low diameter. The major drawback of hypercube based architectures is the difficulty of its VLSI layout. Several variations of hypercube have also been reported which are designed by considering a specific topological property. Nevertheless, no particular topology claims to have better performance with all the desirable topological properties. In this paper the performance analysis of various interconnection networks is presented. The performance is compared by considering cube type architectures as well as linear type architectures on different parameters such as degree, diameter, bisection width, scalability and cost etc. The Analysis indicates that cube based architectures have a rich interconnected structure with high cost and complexity. On the other hand linear type architectures are scalable, simpler and better in terms of cost and complexity. The comparative study suggests the various aspects to the design of new multiprocessor architectures.