Optimizing Data Intensive Flows for Networks on Chips

Abstract

A novel framework is proposed to find efficient data intensive flow distributions on Networks on Chip (NoC). Voronoi diagram techniques are used to divide a NoC array of homogeneous processors and links into clusters. A new mathematical tool, named the flow matrix, is proposed to find the optimal flow distribution for individual clusters. Individual flow distributions on clusters are reconciled to be more evenly distributed. This leads to an efficient makespan and a significant savings in the number of cores actually used. The approach here is described in terms of a mesh interconnection but is suitable for other interconnection topologies.