Research On Complex Cloud Resource Feature Oriented Workload Prediction Methods

Cloud platform data centers are very important infrastructure for cloud computing,but many data centers have low resource utilization,large cluster servers,and huge energy consumption in the data centers,which causes high cost waste for cloud service providers.In order to make full use of cloud resources,virtualization technology is widely used.A cloud server can be virtualized into multiple relatively independent virtual machine instances.Forecasting the load of hosts and virtual machines can effectively promote the resource scheduling and load balancing of cloud platforms,and has become one of the effective methods to increase the resource utilization of cloud platforms.However,due to the impact of users submitting jobs randomly,the load has the characteristics of short-term mutation,nonlinearity,and weak periodicity.How to improve the performance of cloud platform load forecasting is of great significance.For this purpose,the following work has been done:(1)This paper analyzes the resource usage of current mainstream cloud platform data centers,and analyzes the virtual machine load and host load behavior of different cloud platform data centers,providing reference information for load forecasting.(2)In order to accurately analyze the load change mode of cloud virtual machines and improve the CPU load prediction performance of cloud virtual machines,the empirical mode decomposition and principal component analysis(EMPC)and the bidirectional gated recurrent unit based convolution(BCGRU)models of cloud virtual machines are proposed.Through empirical mode decomposition and principal component analysis of cloud virtual machine load mode decomposition,the characteristic fluctuation series of different scales are obtained;The convolution layer of the prediction model can fully extract the decomposed features,and improve the ability of the prediction model to learn the load change mode of cloud virtual machine through the bidirectional gated cyclic neural network bidirectional learning sequence forward and backward dependency.On the 2019 VM datasets generated by Microsoft Azure in the real cloud environment,single-step and multi-step prediction experiments are carried out to verify the effectiveness of the prediction method.(3)In order to improve the accuracy of host load prediction,an LSTM(Long Short Term Memory With Zoneout,LSTM-Z)host load prediction method based on Zoneout is designed and implemented.This method can adapt to the host load forecasting mode with volatility characteristics,and explore the optimal historical window weight vector in the iterative evolution process through genetic algorithm,which can make full use of historical data dependency and improve the accuracy of prediction.Single-step and multi-step prediction experiments are carried out on two real cloud platform data of Google Cloud and Alibaba Cloud to verify the effectiveness of this method.