PacketUsher: Exploiting DPDK to accelerate compute-intensive packet processing

Abstract

Many compute-intensive network applications such as application-layer traffic generator, Deep Packet Inspection(DPI) and web servers are widely deployed on commodity PC for the reason of flexibility and cheap price. However, how to improve their performance on general purpose OS is challenging due to the high packet I/O related overheads. This paper presents PacketUsher, a high-performance packets processing framework to remove these performance bottlenecks. In building PacketUsher, we constructed a DPDK wrapper as the underlying packet I/O engine to accelerate packet transmission, and utilized the strategies of zero copy, batch processing and parallelism to improve packet processing. Through RFC2544 benchmark, we demonstrate that DPDK wrapper has excellent packets transmission capability. As a case study of PacketUsher, we design and implement a commercial application-layer traffic generator. The experiment results show that the FPS (Flow Per Second) value of our traffic generator over PacketUsher is more than 4 times of that over standard Linux platform. By comparison, the FPS value over PacketUsher is about 3 times of that over existing methods (Netmap and PF_RING).