Boosting the Performance of Sketch-Based Traffic Measurement for Programmable Data Center Networks: A Machine Learning Approach

Abstract

In modern data center networks, fine-grained traffic information is crucial for ensuring cloud service quality and conducting routine network management. Sketch algorithms are widely applied in programmable data center networks for fine-grained traffic measurement due to their ability to provide guaranteed error bounds with low computational and memory overhead. However, the severely limited resources of off-the-shelf programmable network devices mean that even simple sketches can exhibit significant measurement errors. Furthermore, sophisticated sketch-based schemes are challenging to implement and deploy on such hardware. To address this dilemma, this paper introduces a fine-grained, accurate, and generic traffic measurement framework leveraging machine learning techniques. The framework utilizes an ML-based encoder-decoder architecture to infer precise per-flow statistics from raw sketch data captured in the data plane, thereby enhancing traffic status monitoring. To effectively discern flow characteristics and their collision dynamics within the sketch, we have developed two distinct encoder architectures: the Single-flow Single-layer Collision (SSC) encoder, based on Deep Neural Networks (DNN), and the Multi-flow Multi-layer Collision (MMC) encoder, built upon Graph Attention Networks (GAT). Extensive experiments with real and synthesized traffic traces demonstrate that our framework significantly outperforms direct sketch queries, reducing measurement errors by $1\!\sim \! 2$ orders of magnitude while maintaining efficiency.