Font Size: a A A

Predicting QoS Of Virtual Machines Based On Bayesian Network

Posted on:2021-10-15Degree:DoctorType:Dissertation
Country:ChinaCandidate:J HaoFull Text:PDF
GTID:1488306230481294Subject:Automation Technology
Abstract/Summary:PDF Full Text Request
Infrastructure as a Service(IaaS)is a service mode of cloud computing.Relying on virtualization technology,IaaS service providers offer users with computing and storage resources in the form of virtual machines(VMs).During the period of renting,the Quality of Service(QoS)of VMs,such as response time,throughput,reliability,etc.,are guaranteed through Service Level Agreements(SLAs)signed between users and service providers,and then any violation of the SLAs will result in financial or reputational damage to the latter.The principle way to ensure the SLAs being reached is to predict the QoS of VMs accurately and then take the corresponding measures based on the prediction results timely.Therefore,in this thesis,we are to predict some quantifiable QoS indicators accurately,such as the response time and throughput of VMs,so as to avoid SLAs violations and then optimize the virtual resource allocation.However,there are some challenges in predicting the QoS of VMs.(1)The performance degradation of CPU,memory and I/O components cannot be directly observed,resulting in the difficulties in measuring performance degradation of the overall VMs and corresponding components accurately.Actually,performance degradation is one of the underlying environmental features affecting the QoS of VMs.(2)The features affecting the QoS of VMs include the features of software,hardware,configuration,and runtime environment.The features of software and hardware affect the QoS of VMs jointly with different influence degrees,even though the impacts from the same feature on different QoS indicators are different.In addition,there are performance interferences among multiple VMs deployed on the same physical host.Besides,due to the uncertainty,complexity and dynamics of the applications running on the VMs,it is difficult to evaluate the degrees of these performance interference.Only by effectively analyzing and quantifying the relationships among the features and the QoS,the prediction results can be accurate.(3)The changes of environment for VM deployment will cause the variation of the corresponding QoS.Thus,it is difficult for a static prediction model to capture the impacts from these dynamically changed features on QoS.Bayesian Network(BN)is a probabilistic graphical model that can express the uncertain dependencies among random variables,it uses a directed acyclic graph(DAG)to describe the dependencies among multiple variables directly,and then quantifies the dependencies degrees via conditional probability tables(CPT).Therefore,in this thesis,we resolve the above-mentioned difficulties based on BN.Under the premise of analyzing the uncertain dependencies and quantifying these impacts reasonably,we first measure the performance degradation degrees of CPU,memory,I/O and the overall VMs effectively.Then,we establish a model to predict the QoS of VMs accurately based on the performance degradation measurement results and the other environmental features.Finally,we further update the QoS prediction model based on transfer learning.The details are as summarized as follows.? In order to overcome the difficulties that the performance degradation of CPU,memory and I/O cannot be observed directly,we introduce hidden variables into a BN.The hidden variables are used to describe the features that cannot be observed directly,and then the directed edges can be used to express the dependencies among the hidden variables and the others.The expanded BN is called Virtual machine features and Performance Bayesian Network with hidden variables(VPBN).Based on the algorithm of probabilistic inferences with a BN,the performance degradation of CPU,memory,I/O and overall VMs can be measured accurately.Experimental results validate the efficiency of VPBN in measuring the performance degradation of different components and the whole VMs.In this study,we introduce hidden variables into a BN to discover the dependencies among the features and the performance of components effectively.By using the algorithm of inferences with a BN to calculate the dependencies degree among the variables in VPBN,the performance degradation can be inferred and the interpretability of VPBN model can be improved.? This thesis further proposes to effectively analyze and quantify the relationships among the features affecting the QoS of VMs,and then overcome the limitation that the QoS with the features not including in the training set cannot be predicted by BN directly.To this end,we introduce the XGboost classifier into BN and then further propose a Classparameter augmented BN(CBN).Firstly,multiple VM features are divided into different groups,and then the XGboost classifier is adopted to classify different feature configurations in each group.Based on the classification results and their corresponding QoS values,the parameters and structure of CBN can be constructed.When given the features of VMs,we incorporate the variable elimination(VE)algorithm to predict the QoS of VMs accurately.Experimental results verify the accuracy of CBN in predicting the QoS of VMs.In this study,we introduce XGboost classifier into BN,which solves the problem of CPTs combination explosion,and then avoids the failure of QoS prediction caused by the lack of VM features in the corresponding CPTs.? In order to characterize the impacts of a dynamically changing environment on the QoS,this thesis considers the idea of transfer learning and further proposes a method for dynamically updating the structure and parameters of the CBN,called CBNtransfer.CBNtransfer combines the idea of Boosting and the knowledges learned in the original CBN to update the parameters and structure of CBN effectively.Firstly,the forward sampling is adopted to simulate a part of data instances conforming to the original CBN structure and parameters.With new instances and weight updated functions,the weight of each instance could be changed dynamically.Thus,when the environment changes,the structure and parameters of the CBN could be dynamically updated according to the weighted instances,so as to effectively predict the QoS of VMs in the new environment.Experimental results validate the efficiency and accuracy of CBNtransfer in updating the structure and parameters of CBN.Based on the instance-based transfer learning framework TrAdaBoost,we first find out the instances similar to the newly added prediction tasks,so as to use the existing knowledges to facilitate the prediction of QoS of VMs in the new environment,and then effectively reduce the time and physical resources during CBN reconstruction.
Keywords/Search Tags:Virtual Machine, Quality of Service Prediction, Bayesian Network, XGboost, Transfer Learning
PDF Full Text Request
Related items