Deep Learning-based Selection Strategy of Polling Switch for Low-cost and High-accuracy Flow Statistics Collection in SDN

The key challenge of network monitoring and measurement in software-defined networking (SDN) is how to design a high-accuracy flow statistics collection (FSC) scheme with minimum polling cost. At present, there are two kinds of FSC mode, such as push-based FSC and pull-based FSC. As for the most studied pull-based mode, in response to the challenge, there are some FSC strategies, such as sampling, task decomposition and minimum polling switches. However, most of them collect flow statistics with partial collection, linear fitting and known flow characteristics. With the advantages of deep learning in data coverage, nonlinear fitting and knowledge discovery, we proposed a deep learning-based selection strategy of polling switch for low-cost and high-accuracy flow statistics collection in SDN. In our strategy, we select the switches that are intensively traversed by the top-k shortest link paths with bandwidth load balancing as polling switches. Moreover, to minimize the communication cost of the FSC on unknown flow, the deep learning technique is used to predict the flow forwarding path, and the switch in the predicted path whose polling communication cost is smaller than that of any existing polling switch is selected as the new polling switch. The experimental results show that, compared with PFP (per-flow polling) and Greedy, the FSC average delay of our strategy is reduced by 30.2026% and 14.2042%, the communication cost is reduced by 60.0494% and 35.7943%, and the non-redundant flow ratio is increased by 13.9883% and 8.7765%, respectively.