Efficient Flow Table Caching Architecture and Replacement Policy for SDN Switches

Abstract

Software-defined networks (SDN) rely on flow tables to forward packets from different flows with different policies. To speed up packet forwarding, the rules in the flow table should reside in the forwarding plane as much as possible to reduce the chances of consulting the SDN controller, which is a slow process. The rules are usually cached in the forwarding plane with a Ternary Content Addressable Memory (TCAM) device. However, a TCAM has limited capacity, because it is expensive and power-hungry. As a result, wise caching of a subset of flow rules in TCAM is needed. In this paper, we address two related issues that affect caching efficiency: rules to be cached and rules to be replaced. For the first issue, caching an active rule hit by a flow may need to cache inactive rules due to rule dependency. We propose a two-stage caching architecture called CRAFT, which reduces inactive rules in cache by cutting down long dependent chains and by partitioning rules with massive dependent rules into non-overlapping sub-rules. For the second issue, unawareness of the flow traffic characteristics may evict heavy hitters instead of mice flows. We propose RRTC to address this issue, which is a rule replacement policy taking the real-time network traffic characteristics into consideration. By recognizing the heavy hitters and protecting their matching rules in TCAM, RRTC performs better than least recently used(LRU) policy in terms of cache hit ratio. Simulation results show that our combined rule caching and replacement framework outperforms previous work considerably.