CFlam: Cost-effective flow latency monitoring system for software defined networks

Flow latency monitoring is a fundamental task in network measurement. The development of software defined networking enables flexible flow latency monitoring in the control plane. Existing approaches mainly focus on direct probe-based latency measurement, which has a high measurement overhead, especially in high accuracy monitoring systems. In this paper, we revisit software defined latency monitoring framework and explore a cost-effective approach named CFlam to produce flow latency results. We observe large scale latency monitoring generates too many duplicate probe packets. Based on this observation, we attempt to measure only a very small subset of active flows to infer the rest flow latencies. We formulate the monitoring flow selection problem by an algebraic model, and develop an efficient algorithm to generate the optimal probe flow set. We implement and deploy CFlam on a SDN testbed to verify its feasibility and performance. We conduct experiments on a public-available network topology with real packet traces collected from a data center. Experiment results demonstrate that our scheme generates accurate flow latencies at minimum measurement overhead.