Effectiveness of Profiling-based Sketch Configuration: a Measurement Study

Network management relies on measuring a broad array of traffic metrics via the data plane for operational tasks (e.g., traffic engineering, attack detection, anomaly detection, and workload profiling). To collect each desired metric with a low data plane resource footprint, sketches have emerged as an attractive option. Sketch algorithms are sensitive to resource allocation, but existing methods for configuring sketch resources are not comprehensive. Some are not general enough, and others involve high computational costs. Moreover, when a network operator has diverse measurement tasks, these methods often only consider individual metrics and do not consider an ensemble of metrics. The profiling-based method can serve as an alternative approach to resource configuration, offering advantages such as generality, cost-effectiveness, and feasibility for an ensemble of metrics. However, due to the uncertainties introduced by network dynamics, the reliability of profiling-based measurements is still an important issue worth exploring. The goal of this paper is to conduct an extensive measurement study that evaluates the effectiveness of profiling-based sketch configuration. We observe from our measurement analysis that profiling-based configuration is practical and efficient as it is reliable even when traffic patterns change. We further identify a few challenging scenarios that need to adapt configurations to the proper profiles of varying traffic patterns.