Wide-Area Host Clustering for Network Performance Monitoring

Network performance monitoring is an active area in networking research. An open problem of network monitoring research is where to put the monitoring host. A possible way is to cluster the hosts that need to be monitored based on their performance perception, and then put the monitoring node at the representative host of each cluster. In this class project, I experiment three different clustering methods: K-Means, Gaussian Mixture Model, and a kernel-based clustering algorithm. A kernel-based model selection method is used to choose the number of clusters and PCA and Kernel PCA are used for Singular Vector Decomposition. Round Trip Times (RTTs) collected at 104 sites are used as the performance measurements. Experiment results show that clustering hosts is a feasible approach to choose monitoring hosts.