Time-Series Data Regression Modeling Method for Efficient Operation of Virtual Environments

In recent years, efforts have been made to reduce the number of servers by virtualizing servers to improve their utilization rate. In this approach, it is necessary to predict and control the CPU utilization of all the virtual servers because the performance of the virtual servers may deteriorate due to the over-committed state in which the servers are allocated more CPUs than their own CPU resources. In this study, we discuss a regression modeling method for time-series data to generate a general-purpose deep-learning prediction model of the CPU utilization of virtual servers. After exploring methods, we confirmed that the number of data used during retraining could be reduced by extracting the time series data by the length required for training and using the data randomly after subdivision.