CloudMach: Cloud Computing Application Performance Improvement through Machine Learning

Abstract

Cloud computing is rapidly becoming the standard through which enterprises of all sizes fulfill their computing infrastructure demands. This work aims at exploring the impact that machine learning algorithms can have on Cloud application behavior profiling and prediction. Although classic machine learning algorithms have been used in Cloud Computing context before, cutting-edge algorithms like deep learning (DL) and reinforcement learning (RL) are yet to be convincingly exploited for this specific problem. Despite being a revelation with fields like image processing and speech recognition, these algorithms (deep neural networks for instance) face adoption challenges outside certain topics. There is a high demand for timely research work that dissects these algorithms and develops novel techniques to facilitate seamless adoption for Cloud providers and clients. In this work, we evaluate the efficiency of machine learning algorithms in the Cloud context by applying them to a large scale application resource utilization data set (TU Delft Bitbrains traces). The objective is to design a Cloud application behavior prediction technique based on machine learning predictors. Any improvement on prediction precision has direct impact on key performance indicators for both Cloud providers and Cloud tenants/clients. Experimental results show the potential of our approach to improve Cloud resource scheduling in a Cloud data center.