Adaptive Sampling for OpenFlow Network Measurement Methods

Abstract

OpenFlow provides a statistics collection scheme for per-flow and aggregate metrics at a user-specified frequency. However, periodic polling for flow statistics cannot well balance the tradeoff between measurement accuracy and limited control channel bandwidth. To further discuss the resource/accuracy tradeoff, this paper extends OpenFlow protocol to add sampling action for each monitoring flow entry, and systematically sample the matching packets to infer the flow-level statistics. Although traffic sampling can somehow be error-prone, it will provide near-real-time measurements of flow dynamics and determine its accurate polling frequency, which the polling-based approach cannot achieve. In this paper, we propose a per-flow sampling solution to instruct the controller to adaptively adjust polling frequency, then evaluate it in the context of link utilization monitoring.