High-performance packet classification on GPU

Abstract

Multi-field packet classification is a network kernel function where packets are classified and routed based on a predefined rule set. Recently, there has been a new trend in exploring Graphics Processing Unit (GPU) for network applications. These applications typically do not perform floating point operations and it is challenging to obtain speedup. This paper presents a high-performance packet classifier on GPU. We investigate GPU's characteristics in parallelism and memory accessing, and implement our packet classifier using Compute Unified Device Architecture (CUDA). The basic operations of our design are binary range-tree search and bitwise AND operation.We optimize our design by storing the range-trees using compact arrays without explicit pointers in shared memory. We evaluate the performance with respect to throughput and processing latency. Experimental results show that our approach scales well across a range of rule set sizes from 512 to 4096. When the size of rule set is 512, our design can achieve the throughput of 85 million packets per second and the average processing latency of 4.9 µs per packet. Compared with the implementation on the state-of-the-art multi-core platform, our design demonstrates 1.9x improvement with respect to throughput.