FEA-Sketch: flow entries assisted sketch for heavy flow detection in software-defined networking

Download PDF
Open Webpage

Xiaocan Wu, He Huang, Yang Du, Yu-e Sun, Yifan Han

Published: 2025, Last Modified: 14 Jan 2026Sci. China Inf. Sci. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Software-defined networking decouples the control plane from the data plane to enable centralized flow-level network management, while requiring periodically collecting traffic statistics from the data plane to enforce optimal management. As one of the most important traffic measurement tasks, heavy flow detection has received wide attention for its providing fundamental statistics in various practical applications. Existing studies have proposed sketch-based detection solutions to address the mismatch problem between massive traffic and limited high-speed memory resources for measurement in the data plane. However, they overlook the potential of integrating the flow table, where each entry simultaneously enforces forwarding rules for specific flows and records flow statistics into the sketch design, leading to redundant measurement between the flow table and sketch and being unable to utilize their statistics to jointly enhance estimation accuracy. We propose flow entries assisted sketch (FEA-Sketch) in this work, which employs a differentiated flow recording strategy to record flow statistics jointly using the flow table and sketch for memory-efficient and computationally efficient heavy flow detection. We also propose an optimization-based estimation algorithm to accurately recover per-flow sizes for the flows that only have aggregated statistics due to the sharing of entries in the table (or counters in the sketch). We extend the FEA-Sketch to the distributed measurement setting with a hop-based collaborative measurement strategy, which reduces the measurement workload on switches across the network by avoiding redundant measurements. The experimental results on real Internet traces show that the accuracy of heavy flow detection is improved up to 1.95 times, and the bias of flow size estimation is improved up to 2.99 times, demonstrating that integrating flow entries can significantly improve the performance of heavy flow detection.