Swing Filter: A Low-Overhead Filter with Larger Filtering Range for Network Traffic Measurement

Zhaojie Wang, He Huang, Yu-e Sun, Hanwen Zhang, Guoju Gao, Haibo Wang, Shigang Chen

Abstract: Traffic measurement provides indispensable information to many applications in improving network performance. However, the limited on-chip resources face great challenges in measuring millions of flows simultaneously with high accuracy, and the highly skewed traffic distribution further worsens the performance. Although filtering the vast majority of small flows in advance can help to improve the estimation performance of large flows, the existing filters have limitations in filtering range and processing overhead. This paper proposes an efficient filter with a flexible and extended filtering range for network traffic measurement. One key to our design is the use of signed counters whose values swing in positive and negative directions to cancel out small flows, thereby enlarging the filtering range. We show that the proposed filter is highly effective in filtering small flows, with lower memory overhead and processing over-head than the existing work. It supports various measurement tasks and provides a guaranteed bound on misreport rate. We implement our filter on P4 and a NetFPGA-equipped prototype, and conduct extensive experiments based on real-world Internet traces. Experimental results show that the proposed filter reduces the flow-size estimation error by an order of magnitude and achieves 1.69 times higher throughput than SOTA.

External IDs:dblp:conf/infocom/Wang0SZG0C25