Cloud-Based Clusters: Multivariate Optimization Techniques for Resource Performance Prediction
Jean Steve Hirwa, Ulysse Rugwiro, Mark Stammers, Chunhua Gu · 2016
In cloud systems, multivariate resource performance prediction is getting a greater attention, as it tends to accumulate more information rather than a univariate resource performance prediction from each host. The host here can be defined as a collection of different resources (i.e., cpu, memory usage and I/O) wherein multivariate data set are being produced. The goal of this work is to assess the theory of the several common models of multivariate machine learning methods, and to highlight the practical steps to take in order to fit those models to real data and evaluate the outcome. Additionally, we intend to optimize all machine learning approaches, so that in the end we can compare the optimized outcome with the previous non-optimized models. The Hill Climbing approach is taken into consideration to select the best-fit models (best fitting models). Each model is analyzed and evaluated according to its performance from the prediction errors. Furthermore, ensemble learning algorithms are applied to the best-fit of non-optimized and optimized models to improve performance. Lastly, we discuss the overall results.