Deep Reinforcement Learning Driven Aggregate Flow Entries Eviction in Software Defined Networking

Abstract

Software-Defined Networking (SDN) separates control from network elements and logically centralizes it in SDN controller to provide global view and control of the network. Network elements, such as switches, only forward data using entries in the flow tables that are installed by the controller. The capacity of flow tables is limited and requires continuous management. Several studies have proposed eviction strategies to make space for new entries in the flow tables, but they assume 1:1 mapping between entries and incoming flows. This assumption is a major limitation, as in real networks many incoming flows can be handled by a single Aggregate Flow Entry (AFE). This paper handles this limitation by proposing Deep Reinforcement Learning (DRL) framework for eviction of AFEs. The proposed framework calculates the degree of AFEs (i.e., how many flows it entertains) along with other parameters to select AFE for eviction, where main objective is to minimize flow table overflows. The experiment results show that the proposed framework reduces the number of overflows by 37%, flow reinstallation by 87%, and the control signaling overhead by 45 % compared to the Random and Least Recently Used Algorithm (LRU).