An Efficient Update Mechanism for GPU-Based IP Lookup Engine Using Threaded Segment Tree

Abstract

Abstract. Recently, the Graphics Processing Unit (GPU) has been used to deploy high-speed software routers. On this platform, designing an efficient IP lookup engine is still a challenging task, especially when taking into account the comprehensive performance under frequent updates. Existing solutions either fail in dealing with update overhead, or can not provide stable throughput. In this paper, we first propose the Threaded Segment Tree, a novel tree-like structure. On this basis, we then present a fast IP lookup engine with an efficient parallel update mechanism. According to our experiment results on real-world data, the proposed mechanism reduces the memory accesses on the GPU and the overall update overhead by at least 82.5% and 89.6% respectively. Moreover, it also ensures the lookup engine provides stable throughput under highly frequent updates, which only decreases by less than 1% even though update frequency increases to 100, 000 updates/s.