K-Medoids based clustering of PlanetLab's slice-centric data

This paper investigates the application of widely used K-Medoids based clustering algorithm on data collected through CoMon facility for the PlanetLab testbed. The averaged values of various metrics in passively collected slice-centric data has been considered for clustering purposes. Various groups of slices, depicting similar resource usage patterns have been identified in original data set. These clusters have been represented in reduced dimensional space formed by first two principal components of original data set. In order to capture variations in pattern of resource usage by various slices at a PlanetLab node, clustering of standard deviations of various metrics have also been carried out. It has been found that K-medoid based clustering can effectively split the original data space into various sub-spaces of different resource usage behaviour of slices. Thus, it can lead to better resource management and control in publicly available testbeds.