FlowTracker: A refined and versatile data plane measurement approach

Abstract

To acquire per-hop flow level information, existing works have made significant contributions to offloading network measurement onto data center switches. Despite this, they still pose challenges due to increasingly complex measurement tasks and massive network traffic. In this paper, we introduce FlowTracker, a flow measurement primitive in the data plane. Our key innovation is a hash-based data structure with constant size and collision resolution, which allows fine-grained and real-time monitoring of various flow statistics. We have fully implemented a FlowTracker prototype on a testbed and used real-world packet traces to evaluate its performance. The results demonstrate FlowTracker’s efficiency under different measurement tasks. For example, with $\sim$0.5 MB of memory, FlowTracker can accurately estimate 98% heavy hitter out of 25K flows, with an average relative error of 1.28%. It also achieves 92.27% higher accuracy in packet delay estimation and 121.83% higher flow set coverage compared to competitors with only 64 KB of memory. Furthermore, FlowTracker imposes minimal overhead, requiring just $\sim$0.04% extra bandwidth for large-scale network processing. With these capabilities, FlowTracker can provide network operators with deep insights and efficient flow control of their networks.