DeTAR: A Decision Tree-Based Adaptive Routing in Networks-on-Chip

Abstract

The deployment of heuristic algorithms is extensively utilized in the routing policy of Networks-on-Chip (NoCs). However, the escalating complexity and heterogeneity of multi-core architectures present a formidable task for human-designed efficient heuristic routing policies. Although recent works have exhibited that machine learning (ML) can learn efficacious routing policies that surpass human-designed heuristics in simulation, the intricate design and costly hardware overhead of ML-based routing algorithms preclude their practical application in NoCs. In this paper, we propose a Decision Tree-based Adaptive Routing algorithm, DeTAR, which is effective yet simple. The key insight of DeTAR is that routing decisions can be treated as selecting and prioritizing the key features among various NoCs’ metrics like free Virtual Channels (VCs), the buffer length, etc., that better affect the routing decision. This reveals a natural match between the adaptive routing algorithm and the Decision Tree (DT) model. We trained DeTAR from network behavior datasets and evaluated the DeTAR routing algorithm against existing routing algorithms. Our simulation results show that the average saturation throughput can be improved by up to $17.59\%$ compared with existing heuristic routing algorithms. Compared with the previous ML-based adaptive routing algorithm, the area of our routing logic is reduced by $88.95\%$ without significant performance degradation.