An Efficient Per-Flow Size Measurement Framework Based on Dynamic Fuzzy Clustering Learning

Per-flow size measurement aims to count the number of packets or the total bytes associated with each flow, serving as a critical foundation for network management, anomaly detection, load balancing, and other essential network functions. However, due to the limited switch’s on-chip cache memory, most existing per-flow size measurement methods adopt counter sharing, which intermixes information from multiple flows. This introduces considerable noise and severely impacts the measurement accuracy, particularly for small to medium-sized flows. To address this challenge, this paper proposes a novel flow size measurement algorithm based on dynamic fuzzy clustering learning. The algorithm compresses and stores flow size information in on-chip memory, while selectively sampling a subset of traffic data for off-chip dynamic fuzzification processing. By applying dynamic fuzzy relational clustering, the sampled traffic data are grouped into multiple clusters, which are then used to train distinct models tailored to different flow categories. This approach significantly enhances the measurement accuracy for small and medium-sized flows. Experiments based on real traffic traces demonstrate that, under a constrained memory footprint of only 100KB, the proposed method reduces the average relative error by 70.92% compared to conventional techniques.