Proactive Congestion Avoidance Mechanism with Attention based CNN

Abstract

Congestion avoidance has been a hot topic in the field of network transmission. In order to reduce the probability of congestion, it is an effective solution to dynamically adjust the flow table in the switch according to the time-varying network traffic. However, due to the peculiarity of flow burstiness, it always leads to an unexpected delay in congestion processing if directly using the traffic measurement. To this end, this paper proposes a proactive congestion avoidance mechanism with the attention based CNN (AttCNN-PCA). In particular, we first propose a multi-step convolutional neural network model with attention mechanism (AttCNN) to prevent potential congestion in advance. Then, a greedy-based heuristic algorithm is also proposed to quickly find the appropriate rerouting policy according to the predicted congestion traffic, which can effectively achieve congestion avoidance. Finally, we build a prototype system based on the barefoot tofino P4 switches, in which AttCNN-PCA is implemented. Based on the massive experiments, it shows that the proposed AttCNN-PCA solution effectively avoids the risk of congestion by reducing the maximum link utilization (MLU) by 11.3%-27.5% compared to the state-of-the-art solutions.