Predictive Flow Scheduling in Datacenter Based on Software-Defined Network

As a huge volume of data are generated everyday, data centers have invested a huge amount of resources to construct network infrastructures. To efficiently utilize the bandwidth of the networks, effective flow control mechanisms are mandates. Recent researches mainly focus on detecting elephant flows and allocating optimal transmission paths for them to increase the network utilization. However, the state-of-the-art approaches are still unable to precisely detect the elephant flows. This is because the traffic engineering usually considers only the average values, but pays little attention on the variance of the flows. To address the above mentioned critical issue, this paper proposes a data stream mining approach to boost the performance of elephant flows detection. We also propose a scheduling mechanism to avoid the congestion produced by the prediction error. Our system first extracts specific features from the first few packets of flows to build a stream mining model. Then, based on this model, our system is able to predict the flow demand and duration of the newly-generated flows in the network. The control layer in the Software Defined Network (SDN) is able to monitor the interval of the predicted flow. Therefore, our system can both consider mean and variance of the demand prediction results simultaneously to dynamically and reliably schedule the elephant flows, and the probability of network congestion is greatly reduced.