Reinforcement Learning Enabled Multi-Layered NoC for Mixed Criticality Systems
Nidhi Anantharajaiah, Fabian Lesniak, Tanja Harbaum, Juergen Becker · 2023
Increasingly multiple applications of different criticality are sharing the same System-on-Chip (SoC) platform to reduce overall cost. With the increase in number of cores and complexity of such mixed-critical platforms, the on-chip interconnect is becoming a critically shared resource. It is desirable if such a mixed-critical Network-on-Chip (NoC) is able to adapt to the requirements of different applications to improve overall performance while ensuring guarantees are met. To reduce inter application influence, multi-layered topologies have been investigated in recent years to decouple traffic of different criticality. In this paper, we propose a reinforcement learning based topology agnostic routing for multi-layered topologies. Such an adaptive routing transmits packets based on criticality over different network layers. The algorithm determines paths based on local network information and past history to discover better routes over time. We compare the proposed routing algorithm to static routing like XY and an adaptive congestion aware routing targeting multi-layered NoCs. The results show an increase of upto 42% in throughput and upto 75% decrease in latency when compared with static XY routing. When compared against adaptive congestion aware routing algorithm, the results show an increase of upto 29% in throughput and upto 65% decrease in latency.