Low-cost Elephant Flow Detection And Load-balanced Scheduling In Software Defined Networks

Software defined networking(SDN)provides flexible management for datacenter networks with flow-level control.“The elephant and mouse phenomenon” suggests that there are only very few elephant flows that carry the majority bytes in datacenters,and most flows are the delay-sensitive small flows.Thus,fine-grained management of SDN results in frequent interactions between the data plane and the control plane,as well as on-time delivery of small flows.On the other hand,the mixed routing of elephant and mice flows can cause some links to be blocked.Therefore,it can improve management efficiency to detect and reroute elephant flows while leaving mice flows in data plane leveraging wildcard flow table in Open Flow.However,existing mechanisms for elephant flow detection suffer from high bandwidth consumption,long detection time and needs for specialized facilities.In addition,the existing routing strategy only for the elephant flow scheduling flow control,while sacrificing a small flow,resulting in small flow is congested.In this paper,we focus on how to solve these problems,and propose an efficient sampling and classification approach ESCA,to achieve fast and accurate detection of elephant flows without additional network overhead.Based on the detection result,we apply different scheduling strategy to elephant and mice flows.The main contributions of this paper are as follows:· We propose an efficient sampling and classification approach ESCA,for realtime and accurate detection of elephant flows.The edge switches take samples from the wildcarded flow table and send them to the controller for classification decisions.By combining the similarity between flows,we improved the C4.5 classification algorithm to improve the classification accuracy up to 12%.ESCA has lower network overhead compared to similar methods.· We propose an optimization method to find the time intervals and interval lengths in which the elephant flows are dense.Different sampling frequencies are used in different time intervals to minimize the sampling cost.We designed a filtering flow table based on Open Flow to filter out redundant samples belonging to the same flow.Experiments show that our scheme reduces the sampling cost by 75% compared with the traditional sampling scheme.· Based on the elephant flow detection result,we propose DiffTE,a hybrid flow scheduling mechanism,to provide different routing schemes for elephant and mice flows to improve network link utilization and reduce congestion.We introduce the blocking islands mechanism into the elephant flow path searching,which can reduce the searching cost while guaranteeing the available bandwidth.On the basis of the equivalent multi-path algorithm,we consider the link utilization,and design the dynamic weighted multipath routing algorithm for the small flow aggregation route.Finally,we evaluate our elephant flow detection mechanism ESCA and the hybrid flow scheduling mechanism Diff TE by simulating a data center network on Mininet.The performance of elephant flow detection is evaluated,including the accuracy,detection cost,detection time,network performance such as throughput,delay and link utilization,and the results are compared with Hedera,s Flow and ECMP.