Learn to Recover: Deep Reinforcement Learning for Failure Recovery in Large Networks

Abstract

Proactive link recovery approaches have been approved to be efficient by pre-calculating and installing backup routing paths in advance. Once a link failure happens, the switches on data plane directly change to next available pre-calculated path, saving time from communications and path re-calculations. However, existing approaches sacrifice flexibility of traffic management and unable to optimize network performance. In this paper, we propose Learn2Recover, a proactive link recovery technique that prioritizes pre-calculated backup paths such that failure recovery rate and real-time network performance can be optimized at the same time. Specifically, we propose a deep reinforcement learning (DRL) based algorithm, called DRL-PathPRI, to learn priorities of backup paths via interacting with real network traffics in a simulator. We further propose a framework, called Learn2Recover, to scale DRL-PathPRI to large networks, which partitions the network into clusters, routes Inter-Cluster flows with an efficient designed algorithm with effective capacity evaluation, and applies DRL-PathPRI within each cluster independently and parallelly. We evaluate on real networks with both real and simulated traffics. Experiment results have shown Learn2Recover superiors the baseline methods in success rate and link utilization. It improves around 29.0% success rate and 20.4% network throughput compared to the second best method.