Research On Workload Forecasting And Performance Interference-aware VM Scheduling

With rapid development of cloud computing technologies and applications,the number and scale of cloud data centers have grown exponentially in recent years,resulting in an increasingly serious problem of high energy consumption and low efficiency.Virtualization technology is the key technology of cloud computing.VM Scheduling plays the major significant role in improving the overall energy efficiency of data centers.Improving the energy efficiency of the data center means increasing the resource utilization of the physical machines,but this can lead to resource contention among co-located virtual machines,thereby causing performance interference problems.The performance of virtual machine cannot be guaranteed,which will cause the violence of Service Level Agreement.Therefore,the performance of virtual machine and the energy efficiency of data center cannot achieve the optimal simultaneously.In addition,due to the frequent changes in the workload state of the virtual machine in cloud,the workload value obtained by monitoring may have a large lag.Having an accurate control of the workload changes contributes to the scheduling of virtual machine.To sum up,the current issues that need to be addressed are monitoring and predicting the workload,defining and modeling performance interference.Based on this,the strategy aims to reduce the energy consumption of the data center while ensuring the performance of the virtual machine.The main research contents of this paper are as follows:(1)This thesis investigates virtual machine workload forecasting and performance interference modeling,and then points out their shortcomings in real-time,accuracy and comprehensiveness.Based on sufficient research,a method named Attentive CNN-LSTM Resource Utilization Forecasting(ACRUF for short)is proposed.This method takes into account the load patterns of different resources,carries out feature engineering and model construction in a good manner,and better improves the forecasting effect of the load with large fluctuations in the multi-step forecasting.On this basis,a multi-dimensional resource based performance interference model is constructed to quantitatively model the performance interference of both virtual machine and physical machine.The concept of "persistent performance interference degree" is put forward to measure the interference degree of virtual machine and host in the next period of time.The load prediction and performance interference model will play an important role in scheduling strategy.(2)There exist certain problems concerning the robust and real-time in most of the current host overload detection algorithms.This thesis proposes a Continuing Interference based Host Overload Detection algorithm(CIHOD).First,the ACRUF method is used to predict the utilization of future time steps,then the continuous performance interference degree of the host will be calculated based on the multi-dimensional performance interference model,finally the overloaded host is detected.The experimental results show that CIHOD algorithm can effectively reduce the SLA violation rate and migration times.(3)In order to ensure the performance of VM while reducing the energy consumption of cloud data centers,this thesis proposes a strategy named Interference Aware VM Scheduling(IAVMS).The strategy first performs host overload detection based on CIHOD.Then,the VM with the highest load correlation will be selected for migration based on the DTW distance based Maximum Correlated VM Selection algorithm(DTW-MC),which will better alleviate the contention of overloaded resources.Next,an Interference Energy-efficient Trade-off Host Selection algorithm(IETHS)is proposed to maintain a candidate host list according to the performance interference degree and then select the host with optimal energy-efficiency,which ensures that the host runs efficiently while reducing the effects of performance interference.Meanwhile,the strategy will manage the state of VM and host,make sure the underloaded host will be shut down in time after the VM is destroyed or migrated in order to reduce energy consumption.The experimental results on Multi RECloudsim platform which extend the function for multi-resource simulation show that,compared with other strategies,IAVMS can effectively reduce SLA violations(SLA are improved by at least 7.03%),and the number of migration is the least.Besides,a good trade-off with energy consumption optimization is realized.