DeepFlow

Abstract

Providing fine grained traffic measurement is crucial for many network management tasks such as traffic engineering, anomaly detection, traffic accounting, and load balancing. Software-defined networks can potentially enable fine-grained measurement by providing statistics for each forwarding rule of an OpenFlow-enabled switch. However, providing fine grained traffic measurement in a scalable fashion in hardware switches poses significant challenges due to the limitations in the size of the TCAMs that fit only a relatively small number of rules compared to the number of active flows in the network. In this paper, we present DeepFlow, a framework for scalable software-defined measurement that relies on an efficient mechanism that a) adaptively detect the most active source and destination prefixes in the network, b) collects fine-grained flow-size measurements for the most active prefixes and coarse grained for the less active ones, and c) uses historical measurements in order to train a cloud-based Deep Learning model that can be used to provide short-term predictions whenever exact flow counters cannot be placed at a switch due to its limited resources. Thus the number of fine-grained flows measured can increase significantly without the need to use other flow sampling solutions that loose accuracy. An extensive experimental evaluation using a prototype implementation and real network traces shows that DeepFlow can provide very high accuracy for estimating flow sizes at various aggregation levels.