Fine-Grained Active Queue Management in the Data Plane with P4

Abstract

Traditional source-based congestion control methods cannot meet the complex network requirements. The proposal of active queue management algorithm can alleviate the pressure of congestion from the network level. However, most of the algorithms lack fine-grained analysis of queue, which fails to quickly alleviate congestion and will affect some innocent data flows. With the rapid development of programmable data plane, the fine-grained analysis of queue in data plane becomes a promising way to alleviate congestion. In this paper, a Fine-Grained Active Queue Management (FG-AQM) scheme is proposed to achieve rapid congestion avoidance and network performance optimization. In the proposed scheme, different hash algorithms and registers are used to analyze and determine the target flows which have great impacts on network performance. And proportional integral controller is used to calculate the packet dropout probability according to the queue delay and jitter. Combined with the output of the proportional integral controller and the target flows, FG-AQM achieves dynamic adjustment of packet dropout probability to achieve congestion avoidance. We implement FG-AQM on programmable switch and evaluate the proposed scheme against the state-of-the-art AQM solutions. Extensive simulation results show that FG-AQM can effectively deal with the data flow causing congestion and improve the throughput by 34% (compared with P4-RED) and 22% (compared with P4-PI2) on average in microburst scenarios.