On multicast for dynamic and irregular on-chip networks using dynamic programming method

Abstract

Mutlicast is an intrinsic communication pattern in emerging applications including the Internet service, bio-inspired computing, online data analysis, etc. Providing hardware multicast largely boosts system performance and reduce power consumption for these applications running on many-core systems. However, many-core systems suffer from dynamically changing topologies, which can be caused by traffic isolation, power management and faults. Links and routers may be removed from a subnetwork or added to a subnetwork, this imposes the routers to alter the routing paths accordingly to make the routing energy efficient. Furthermore an energy efficient multicast scheme that fits to any topology is required in this scenario. Most existing fault tolerant routings cannot detect and found energy efficient communication paths effectively for both unicast and multicast. In this work, a lightweight network couples with on-chip routers is used to propagates topological information. The shortest path between any node pair is also computed in this network using dynamic programming method. Unicast packets are routed along a shortest path to its destination, and multicast packet copies at each router is minimized based a rule called minimal replication to reduce multicast link occupation and energy consumption. Simulation results show that the proposed NoC routes 96% packets to their destinations if there exists a path, and demonstrate 40% lower latency and 30% power consumption compared with Stochastic communication. The extra hardware cost to build optimal multicast path is estimated to occupy less than 5% of the total area of a router.