GAMT: A fast and scalable IP lookup engine for GPU-based software routers
Yanbiao Li, Dafang Zhang, Alex X. Liu, Jintao Zheng · Architectures for Networking and Communications Systems · 2013
Recently, the Graphics Processing Unit (GPU) has been proved to be an exciting new platform for software routers, providing high throughput and flexibility. However, it is still a challenging task to deploy some core routing functions into GPU-based software routers with anticipatory performance and scalability, such as IP address lookup. Existing solutions have good performance, but their scalability to IPv6 and frequent updates are not so encouraging. In this paper, we investigate GPU's characteristics in parallelism and memory accessing, and then encode a multibit trie into a state-jump table. On this basis, a fast and scalable IP lookup engine called GPU-Accelerated Multi-bit Trie (GAMT) has been presented. According to our experiments on real-world routing data, based on the multi-stream pipeline, GAMT enables lookup speeds as high as 1072 and 658 Million Lookups Per Second (MLPS) for IPv4/6 respectively, when performing a 16M traffic under highly frequent updates (70, 000 updates/s). Even using a small batch size, GAMT can still achieve 339 and 240 MLPS respectively, while keeping the average lookup latency below 100 μs. These results show clearly that GAMT makes significant progress on both scalability and performance.