Multi-objective biogeography-based optimization and reinforcement learning hybridization for network-on chip reliability improvement

Abstract

Reliability is increasingly a major concern in network-on-a-chip (NoC) design, alongside increased performance demands from new applications and the need for continued miniaturization of silicon technology. In this article, we look at the task migration mechanism, used to recover from permanent processing element (PE) failures in NoCs, by remapping tasks performed on faulty cores to spare ones.An innovative reliability-aware task mapping technique is presented, based on a hybridization between Multi-Objective Optimization (MOO) and Reinforcement Learning (RL). It takes place in two steps. In the first, a set of optimal remapping solutions for different failure scenarios is generated at design-time, using a Biogeography-Based Multi-Objective Optimization algorithm, while considering communication energy and migration costs. In the second step, an artificial neural network agent is trained to select the best remapping solution, from those generated at design-time, to recover from execution failures at run-time.Experiments were carried out to evaluate our technique for different sizes of networks and on different benchmarks. The results obtained show that the technique based on the hybridization MOO_RL brings a great improvement in the reliability of the NoC and achieves a good compromise between reliability and performance. It also guarantees a reduction of the overhead caused by the storage space of the remapping solutions, compared to the existing solutions.