A hybrid heuristic-based tuned support vector regression model for cloud load prediction

Cloud computing elasticity helps the cloud providers to handle large amount of computation and storage demands in an efficient manner. Proactively provisioning cloud workload is essential in order to keep the cloud utilization and service-level agreement at an acceptable level. Problems such as new virtual machine start-up latency, energy minimization and efficient resource provisioning, requires to predict resource demands for a few minutes ahead. Since the Cloud workloads have a very dynamic nature, CPU/memory usage varies considerably in the cloud. Also, existing prediction methods have considerable prediction error and erroneous results. So we propose a novel tuned support vector regression (TSVR) scheme that carefully selects three SVR parameters by a hybrid genetic algorithm and particle swarm optimization method. A chaotic sequence is devised into the algorithm to improve prediction accuracy and simultaneously avoid premature converging. To demonstrate the prediction accuracy of our TSVR model, we conduct a simulation study using Google cloud traces. The simulation results show that the proposed TSVR model achieves better prediction performance than conventional models in terms of standard metrics.