Performance Comparison of Deep VM Workload Prediction Approaches for Cloud

With the exponential growth of distributed devices, the era of cloud computing is continued to expand and the systems are required to be more and more energy-efficient with time. The virtualization in cloud manages a large-scale grid-of-servers to efficiently process the demands while optimizing power consumption and energy efficiency. However, to ensure the overall performance, it is critical to predict and extract the high-level features of the future virtual machines (VMs). To predict its load deeply, this paper investigates the methods of a revolutionary machine-learning technique, i.e., deep learning. It extracts the multiple correlation among VMs based on its past workload trace and predicts their future workload with high accuracy. The VM workload prediction helps the decision makers for capacity planning and to apply the suitable VM placement and migration technique with a more robust scaling decision. The effectiveness of deep learning approaches is extensively evaluated using real workload traces of PlanetLab and optimized with selection of model, granularity of training data, number of layers, activation functions, epochs, batch size, the type of optimizer, etc.