DREAM: Distributed Reinforcement Learning Enabled Adaptive Mixed-Critical NoC

Abstract

Applications of different criticality sharing the same System-on-Chip (SoC) platform are increasing in popularity to reduce overall cost. Spatial and temporal isolation techniques are utilized to reduce inter application influence and to ensure real-time requirements are met. Spatial isolation involves partitioning communication resources and such partitions can result in irregular topologies. It is desirable that the on-chip interconnect on such systems support communication within all possible partition shapes using efficient routing techniques. To improve flexibility, adaptivity and reliability in such systems, it is desirable to incorporate topology agnostic routing algorithms which can compute optimal routes at runtime. For this purpose, we present a Distributed Reinforcement learning Enabled Adaptive Mixed-Critical Network-on-Chip (DREAM NoC) and supporting framework. DREAM is a distributed NoC which uses a topology agnostic reinforcement learning enabled routing algorithm based on the Ant Colony optimization (ACO) metaheuristic. We propose the DREAM framework which comprises of runtime discovery of paths and selection of optimal routes over time based on traffic fluctuations. We compare the performance against other topology agnostic algorithms under uniform random traffic and application traffic of a MPEG4 video decoder. The results show that the presented technique has upto 63% decrease in latency and 25% increase in throughput for certain irregular topologies under uniform random traffic scenario.