VotePipe: Efficient Heavy Hitter Detection in Programmable Data Plane

Heavy Hitter Detection (HHD) is of crucial importance in various applications, such as load balancing, traffic engineering, and DDoS attack detection. The emergence of programmable switches provides novel and effective solutions to heavy hitter detection by offloading network measurement tasks to the data plane. However, restrictions of the hardware architecture make it difficult for most HHD algorithms to be deployed on real devices. In this paper, we propose VotePipe, an HHD algorithm which completely deployed on the data plane of P4 programmable switches and circumvents the hardware restrictions. It adopts a flow count decay mechanism based on “flow age” to kick out outdated flows timely. Besides, it evicts small flows efficiently and minimizes replacement operations with a flow filtering and updating mechanism based on voting. We implement and deploy VotePipe on a real device and use 5 real Internet backbone network datasets and an Ixia traffic generator to test and compare the performance of VotePipe with 3 classic or state-of-the-art HHD algorithms that can be implemented on real machines. The results show that with acceptable memory consumption, VotePipe can achieve better accuracy and throughput than other algorithms, while running more stable and efficiently.