On Performance of Modified Torus Interconnection Networks

Abstract

Parallel processing systems are the most powerful tools to real time applications that requires large processing of data. In recent years of advancement within the technology, a remarkable transition has been observed. Important components of these systems include; processing elements, memory modules and interconnection networks. The main purpose of interconnection network is to build up connection between these components, as it corresponds to the performance of parallel computing system. Three attributes characterize an interconnection network namely: topology, routing algorithm and flow control. On the other hand, a topology is differentiated by its topological properties such as degree, diameter, average distance, bisection width, scalability and fault tolerance. In terms of exchanging data among the components of parallel processing system, several mechanistic approaches are used. Although it is simple to use and implement, the main issue is its fault tolerance. Pertaining to the various problems and research gaps such as communication system, scalability, efficiency etc. within the interconnection network, topologies, and related factors have been discussed. Briefly, the objectives of the present thesis were to design efficient topologies with improved properties. We have proposed four topologies such as modified X-Torus, Center concentrated X-Torus, Hexagonal X-Torus and Modified Diagonal Torus in this thesis. All these topologies were tested using OMNeT++ simulator under various traffic patterns such as uniform, bit complement, neighbor, tornado and hotspot. First three topologies are compared with Mesh, Torus, and X-Torus and lastly with Mesh, Torus, D-Mesh and D-Torus. The quality of services which were obtained from simulation process were Average Throughput, Average End to End Delay and Average Hop Count. newline