OctoSketch: Enabling Real-Time, Continuous Network Monitoring over Multiple Cores

Sketching algorithms (sketches) have emerged as a resourceefficient and accurate solution for software-based network monitoring. However, existing sketch-based monitoring makes sacrifices in online accuracy (query time accuracy) and performance (handling line rate traffic with low latency) when dealing with distributed traffic across multiple cores. In this work, we present OctoSketch, a software monitoring framework that can scale a wide spectrum of sketches to many cores with high online accuracy and performance. In contrast to previous systems that adopt straightforward sketch merges from individual cores to obtain the aggregated result, we devise a continuous, change-based mechanism that can generally be applied to sketches to perform the aggregation. This design ensures high online accuracy of the aggregated result at any query time and reduces computation costs to achieve high throughput. We apply OctoSketch to nine representative sketches on three software platforms (CPU, DPDK, and eBPF XDP). Our results demonstrate that OctoSketch achieves about 15.6× lower errors and up to 4.5× higher throughput than the state-of-the-art.