Towards an Intelligent Speculative Software-Defined Networking

Abstract

Software-Defined Networking (SDN) separates the control and data planes, allowing better programmability of the control plane to predict, route, and schedule traffic at the data plane. As a more flexible approach, Reactive SDN installs the right flow rule dynamically when a new flow arrives. This helps respond to application dynamics, making Reactive SDN a strong candidate for low-latency applications. Low-latency applications like online gaming and AR/VR have become very popular these days. However, they require millisecond-level response times for an acceptable quality of experience. A limitation of Reactive SDN is that it necessitates a miss upon the arrival of a new flow, causing a Packet-in message to be sent from the switch to the SDN controller, increasing overall delay. To meet millisecond-level delays, reducing the miss rate by predicting the arrival of flows and pre-installing the necessary flow rules dynamically is necessary. Reinforcement Learning (RL) is an approach where agents interact with an environment by making actions to receive rewards to accomplish tasks like forecasting. RL has potential for predicting flow arrivals, where agents can predict the arrival of flows that may not have been seen before. We show that the RL agents can learn and speculatively install the unseen flow rules to avoid latency from reactive installations. We propose an SDN design called ‘speculative’ to overcome Reactive SDN limitations for applications requiring fast responses. The contributions of our work are: 1) We present a Speculative SDN framework that incorporates RL to predict and install never-seen-before flows, reducing control latency. 2) We design a reward function to improve the RL agents’ prediction accuracy for the best set of flows to install. 3) We develop a priority policy when selecting a flow for removal. 4) We use spatial locality information to assist agents when speculating unseen flows. 5) We evaluate the framework using real traffic traces across various metrics.