Optimal Server Prediction in Data Center Networks using Machine Learning Techniques

Data Center Networks (DCN) are the major support in managing traffic on the internet. In a DCN, connecting large-scale servers is a tedious task. Irrespective of the data center architectures such as a fat tree, VL2, camcube, and Bcube, there exists a need to regulate the flows towards each server. Further, there are diverse applications in a DCN such as web search, which is latency-sensitive, and database update task which is throughput sensitive. Hence, it is essential to provide an efficient server load balancing strategy that predicts an optimal server based on the incoming flow size. We have proposed a prediction strategy based on machine learning techniques. The method is designed as a dynamic server selection strategy wherein the varying server metrics are considered as well as the current server status is inputted for the prediction module. The key parameters considered for prediction are the flow size and the current server weightage. The prediction part is experimented with the most popular machine learning mechanisms which includes SVM, Logistic Regression, and KNN. The experimentation is performed using python code and analyzed using various metrics such as accuracy and confusion matrix. The results are compared and depicted in graphs. From the analysis, it is observed that SVM outperformed the other two techniques. Meanwhile, the machine learning-based technique shows tremendous improvement over the threshold-based methods for elephant flows in DCN.