Research On Virtual Machines Performance Interference Prediction And Placement Techniques In Cloud Data Center

With the rapid development of virtualization techniques,modern data centers move into a new era of pay-as-you-go cloud in recent years.Despite numerous advantages such as high resource utilization,rapid service scalability,isolation,high availability and efficient cloud management,current virtualization techniques don't guarantee per-fect performance isolation among virtual machines sharing the same bare metal,which may lead to unstable user-perceived application performance and unpredictable ser-vice of quality in clouds.Some researchers recently have discussed the performance interference of collocated virtual machines.Most try to collect the virtualized system parameters in the virtual machine monitor and then predict the application performance by using machine learning methods,but still lack in-depth understanding of the inher-ent performance behavior of applications in virtual machines.Besides,the hypervisor and guest OSes usually run independent resource schedulers and are invisible into each other,thereby making accurately characterizing performance interference a non-trivial work.Therefore,understanding and modeling performance interference among collo-cated virtual machines is of utmost importance.From the perspective of user QoS guarantee and energy consumption optimization of cloud data center,this paper studies the problem of virtual machine performance in-terference prediction and performance-aware resource management in cloud data cen-ter.We propose an accurate virtual machine performance prediction model and an ef-ficient management mechanism of energy consumption and performance optimization.The main contributions of this paper are as follows:(1)We first conducted a lot of experimental research to explore the root cause and expressive mode of performance interference of virtual machines through analyzing the runtime state and parameters of different combinations of benchmark applications.Ex-periments show that,on one hand,VCPU binding can eliminate the floating overhead between multiple physical CPUs and significantly improve application performance;on the other hand,VM exits events,i.e.,the control transitions between the hypervi-sor and VMs,need thousands of CPU cycles on average and is the main reason for the performance interference of virtual machines.Based on the above experimental analysis,we further propose a KCCA-based VM performance interference prediction model in order to accurately depict the complex application performance behavior of virtual machines.The prediction model first collects the microarchitecture-independent application characteristics by Intel Pin and tracks the system level characteristics of vir-tualization layer based on Xentrace,and then proposes a virtual machine performance interference prediction method.The core of our method is to build multivariate cor-relations among application characteristics and performance metrics by using kernel canonical correlation analysis(KCCA),and then use these statistical relationships to estimate the level of application performance slowdown.Evaluation results prove that our model can estimate the performance interference effects under any combinations of workloads with a prediction error of less than 7.9%(2)Based on KCCA performance interference prediction model,this paper further optimizes the virtual machine resource management mechanism in cloud data center.To solve the online virtual machine placement problem,we propose a performance aware and utility-based online VM placement algorithm UtilityFit in order to optimize application performance and the least interference with collocated applications.Our evaluations show that UtilityFit outperforms two state-of-the-art algorithms,namely.BestFit and FirstFit,by up to 54.1%and 6.5%respectively.Besides,clouds need to consolidate the virtual machines periodically in order to optimize energy consumption.We then propose a virtual machine consolidation algorithm—LeastInterferenceFirst in order to balance the application performance and the cloud benefit.The simulation results show that LeastInterferenceFirst algorithm can effectively guarantee the QoS of users and simultaneously optimize the energy consumption in cloud data center.