A novel approach to improve the performance of interconnection networks with hot-spots

Abstract

Congestion in interconnection networks due to the presence of hot spots is an important and difficult problem that occurs in parallel machines. This problem has been studied in depth and different solutions for the case of multiprocessors with shared memory have been proposed. Current trends point towards the implementation of systems with physically distributed memory, either based on, message passing (multicomputers) or on a single shared memory address space (multiprocessors). Our paper is developed in this context. Up to now, proposals to improve the throughput of networks with hot-spots have focused on using virtual channels or adaptive algorithms. We present a novel solution based on reconfigurable networks. A reconfigurable network is one in which nodes can change their position depending on the communication pattern in order to diminish the congestion produced in the network and, therefore, increase its throughput. We studied this problem in two-dimensional k-ary n-cube networks using a deterministic routing algorithm and wormhole routing. In this paper the main features of a reconfigurable network are presented and the results obtained by simulation are shown. These results confirm that this technique is a very interesting one for systems with distributed memory, with applications to a great variety of problems.