Percolation routing in a three-dimensional multicomputer network topology using optical interconnection

Abstract

We investigate the network communication behavior of a three-dimensional (3D) multicomputer system using optical interconnection in which faulty nodes are left in place, a concept called "fail-in-place." We call this the percolation problem in which various amounts of missing nodes fixed in position in the network may have a dramatic effect on the network's ability to transport data effectively. As the number of failed nodes increases, data have to be rerouted through intermediate nodes creating potential "hot spots." These hot spots become the bottleneck that degrades performance. The ability to absorb rerouted data without ejecting it from the network is critical in massively parallel computing systems. Optical technology is a promising solution for internode communication with extraordinarily quick response time supporting enormous bandwidth. To adopt it in multiprocessor systems, efficient routing techniques are needed. We adapt self-routing strategies for all-optical packet routing in 3D mesh networks and investigate the percolation properties. To achieve percolation routing, we incorporate the features inherent in optics to achieve decoding and routing capability in real time. The objective is to develop a dynamic communication environment that adapts and evolves with a high density of missing units or nodes, and by employing analytical, experimental, and simulation methods, show that optical interconnection in a dense 3D system reduces considerably this percolation problem.