A Neural Approach To Predicting CPU Utilization Of Virtual Machine

Cloud services have grown rapidly in recent years,which provide high flexibility for cloud users to fulfill their computing requirements on demand.Many companies have migrated their workloads to cloud service platforms such as Microsoft Azure and Amazon Web Services.Under the pressure of market competition,cloud service suppliers have to provide attractive features to the customers while saving their platform costs.To wisely allocate computing resources in the cloud,it is inevitably important for cloud service providers to be aware of the potential utilization of various resources in the future.Specifically,when it is foreseen that the demands for resources will increase,cloud providers could prepare more physical hosts to meet the growth of future demands in time.Similarly,when the demands of virtual machine(VM)resources are predicted to experience a declining trend,cloud managers could stop allocating new resources and migrate the underloaded VMs properly so that the idle physical hosts can be shut down to avoid waste of resources and improve the lifetime of equipment.Among various resources for VM workloads,CPU utilization is one of the most important indicators since it has a great impact on the total cost of the cloud service.This paper focuses on predicting CPU utilization of VMs in the cloud.We conduct empirical analysis on Microsoft Azure's VM workloads and identify important conceptual characteristics of CPU utilization among VMs,including short-term,medium-term and long-term characteristics.Through the correlation analysis between VMs,we find that similarity exists between different VMs.We propose a neural network method,named Temporal Component Networks(TCopoNet),to model the observed conceptual characteristics with expanded network depth for CPU utilization prediction.TCopoNet consists of three component networks,which are composed of original networks such as GRU,LSTM,and residual network,to capture features at different frequencies and uses fully-connected layers to model deep feature interactions.We conduct extensive experiments to evaluate the effectiveness of our proposed method and the results show that TCopoNet consistently outperforms baseline methods in various metrics including RMSE,MAE and MAPE.