Dynamic routing strategy for embedded distributed architectures

Abstract

The number of processors integrated in embedded platforms is expected to grow and reach thousands of cores in the near future. Manycore architectures gained a large interest over the years but the problem remains in scaling the control fabric and the interconnection network. We present in this paper CEDAR, a Configurable Embedded Distributed ARchitecture, and its adaptive routing strategy based on ACO (Ant Colony Optimization). CEDAR offers a high degree of flexibility and can handle any interconnection topology. Routing paths for remote data transfers are defined at runtime and allow a homogeneous distribution of traffic, avoiding deadlocks and contentions. We show that flexibility generates little overhead for exploring paths, which decreases for large amounts of data transfers. CEDAR is convenient for implementing irregular applications with high computational complexities.