Abstract:

Traffic bursts are common in networks, which have a significant impact on quality of user experience. In the case of traffic bursts, huge volumes of packets can overwhelm the physical links in a short time duration(i.e., milliseconds), resulting in congestion and frequent packet loss. However, traditional routing schemes are either traffic oblivious such as OSPF, which can’t adapt to real-time traffic changes, or centralized control such as linear programming, which can’t efficiently react to traffic bursts due to slow computation. To address this problem in a practical and efficient approach, a novel intelligent routing algorithm based on machine learning (ML) was proposed. On the one hand, the proposed algorithm can leverage the promising modelling ability of machine learning to learn the implicit clue of routing decision. On the other hand, the proposed algorithm enjoys the ultralow processing latency benefited from the fast inference of ML, thus speeding up the reaction to traffic bursts. Experiments on two open-source datasets demonstrate that the proposed scheme can reduce utilization of bottleneck link by 13%～70%, compared with the baselines.

Key words: internet routing algorithm, traffic burst, machine learning, deep reinforcement learning