Automatically Reconfigurable Optical Network for HPC System Based on Deep Reinforcement Learning

High-density optical interconnects enable high-bandwidth, low-latency and energy-efficient communication. Also, in order to enhance the adaptability of High Performance Computing (HPC) networks for traffic variation, optical networking reconfiguration capability is a promising alternative to traditional fixed interconnection architecture. We propose a reconfigurable optical network for HPC systems that can automatically switch between different topologies controlled by a Deep Reinforcement Learning (DRL) agent. By varying the topology of our network, we evaluate it using an Advantage Actor-Critic (A2C) and a Deep Deterministic Policy Gradient (DDPG) algorithm-based scheme. Results show that our prototype achieves up to 1.7x performance improvement.