Research On VM Workload Forecasting And Resource Fragmentation-aware Scheduling

As an important technology to support data center energy saving,virtual machine scheduling has been widely used in industry.Designing an efficient virtual machine scheduling algorithm to improve resource usage is still one of the key concerns of academia and industry.However,the cost of these algorithms,such as causing SLA to become high,cannot be ignored.In order to better solve these problems,in addition to accurate monitoring,predicting the load changes of virtual machines(that is,CPU,memory,and I/O resource utilization)is also necessary,because this allows better resource allocation(including virtual machine scheduling),so as to finally achieve the purpose of optimizing energy consumption.To sum up,it is a problem to be solved to break down the monitoring and forecasting of resource usage and use it as a basis to reasonably schedule virtual machines to improve some indicators(such as the degree of resource fragmentation).The main contributions made by this article are as follows:(1)To fill the research gap of the existing forecasting strategy for multi-resource workload in public data sets,after a comparative analysis of multiple forecasting models,a method named VM Utilization Forecasting based on LSTM(VUFL for short)is proposed.This method takes into account the current input characteristics of the system and the underlying correlation inside the history workload sequence,and can be used for virtual machine utilization prediction after offline training.(2)This thesis investigates virtual machine scheduling algorithms,and then analyzes and compares the advantages and disadvantages of these algorithms,such as algorithm complexity,robustness,resource fragmentation,and so on.Based on sufficient research,in order to improve resource waste and ensure service quality,this thesis proposes a multi-step Resource Usage Prediction-based dynamic Threshold Adaption algorithm(RUPTA).The algorithm first predicts the utilization of M future time steps and then obtain the utilization of N time steps.These multiple past and future data were used to calculates the median absolute deviation,which is one of the core parameters to calculate the host's upper threshold.(3)Under the framework capable of monitoring and forecasting,this thesis proposes a space partition model-based host selection algorithm(SPMHS)to optimize resource fragmentation.This algorithm aims to balance the utilization rate of each resource of the host as much as possible,and reduce the number of active hosts.The space partition model(abbreviated as SPM)manages three priority queues during the target host selection phase,and each queue contains potential candidate target hosts.According to the space partition model,only the hosts calculated based on its specific formulas and satisfying certain constraints can be added to the corresponding priority queue.The queue with the highest priority is allowed to be preferentially selected as the target host,because the hosts that can join the queue mean that the resources in each dimension are nearly used up.If the highest-level queue is empty,the next-level queue is traversed.And so on.If there are multiple candidates in the queue at the same level,a specific distance formula is used to continue screening.All scheduling experiments in the cloud environment are conducted on Multi RECloud Sim which is an extended version of Cloud Sim.The experimental results show that,compared with other control algorithms,SLA and fragmentation are improved by at least 8.89% and 2.05%,and the number of migrations is the least.Besides,a good trade-off with energy consumption optimization is realized.