Spider: Parallelizing Longest Prefix Matching with Optimization for SIMD Instructions

摘要

Longest prefix matching (LPM) is a fundamental process in IP routing used not only in traditional hardware routers but also in modern software middleboxes such as the applications of Network Function Virtualization. However, the performance of recent LPM methods in software routers is insufficient for high-speed packet processing such as two or more 100 Gbps throughput. To improve the performance of LPM, we propose Spider, a new LPM method that achieves a fully parallelized LPM procedure using single instruction, multiple data (SIMD) instructions in a CPU. The evaluation shows that the proposed method has 1.8-1.9 times faster LPM performance compared with the state-of-the-art methods in this study area. We describe the Spider's lookup procedure fully parallelized by SIMD instructions and the design of the routing table efficiently processed by the procedure. We also report the following three evaluations: (1) The effect of parallelization by SIMD instructions on the performance of Spider; (2) the scalability of Spider with the number of CPU cores; and (3) the performance comparison with the previous methods in terms of randomly generated and real-trace traffic patterns.