An Efficient Sampling and Classification Approach for Flow Detection in SDN-Based Big Data Centers

Abstract

Software defined networking (SDN) provides flexible management for datacenter networks with the flow-level control. Such the fine-grained management, however, consumes large amount of bandwidth between data and control planes, which results in the bottleneck in the scalability of SDN-based datacenters. elephant and mouse phenomenon suggests that there are only very few elephant flows that carry the majority of bytes in datacenters so that it can improve management efficiency to detect and reroute elephant flows while leaving mice flows in data plane leveraging wildcard flow table in OpenFlow. Unfortunately, existing mechanisms for elephant flow detection suffer from high bandwidth consumption and long detection time. In this paper, we propose an efficient sampling and classification approach (ESCA) with the two-phase elephant flow detection. In the first phase, ESCA improves sampling efficiency by estimating the arrival interval of elephant flows and filtering out redundant samples using a filtering flow table. In the second phase, ESCA classifies samples with a new supervised classification algorithm based on correlation among data flows. The mathematical analysis proofs our ESCA outperforms related schemes. Extensive experiment results on real public datacenter traces further demonstrate that our ESCA can provide accurate detection with less sampled packets and shorter detection time.