Research On Stochastic Model-based Analysis Of Cloud Service Availability

With the rapid development of cloud computing,more and more critical applications are migrating to cloud data centers,and various services are provisioned to customers through cloud data centers.Infrastructure as a Service(Iaa S)is one of the basic cloud services,which is provisioned to customers over the network in the form of virtual machines(VMs)and/or virtual links(VL).Virtual machines are deployed on physical servers,and each virtual machine is assigned with different number of virtual cores,memory and amount of storage to satisfy various service requirements from customers.Service delays,infrastructure failures and security incidents can occur in large and heterogeneous cloud data centers.Availability,as a measure of the ability of an authorized entity to access,providing services effectively,or recovering from a system failure and attack,is becoming the main focus of customers to choose cloud providers.At the same time,availability analysis also helps service providers to optimize system design and reduce construction costs,providing guidance for building efficient and effective cloud data platforms.With the widespread deployment of Iaa S cloud services,both the customers and the cloud service providers are increasingly demanding cloud service availability analysis.This dissertation aims to investigate large-scale,heterogeneous,dynamic and vulnerable cloud data centers.We explore steady state availability and transient availability analysis models that can reflect the Iaa S cloud data center behaviors in a approximately accurate way,and study the calculation method and sensitivity analysis method of the availability measures of various Iaa S service based on the scalable model.The main research work and contributions are as follows:(1)An availability analysis model for cloud services under multiple resource types is proposed.This work aims to take into account the quality of the cloud services(that is,whether the cloud services are good or not),to establish an analysis model for the single physical server under the assumption that the physical server will not fail,and to evaluate the cloud service availability under various requests by using the job immediate service probability and average completion time as the measure.First,we analyze the service characteristics of cloud data center with multiple types of resources.For each type of physical resources,the resource amount of different tenants can be different and follow a general distribution.Then,a novel analytical model for cloud service availability evaluation was developed using Continuous Time Markov Chain(CTMC).The methods of calculating the steady-state availability are developed.Finally,numerical analysis and simulation are used to verify the accuracy of the model under various parameter settings.The availability analysis model established in this study is more comprehensive than the single resource type model,which can reflect the actual operations of cloud service and improve the accuracy of availability evaluation.(2)Monolithic and scalable availability analysis models of Iaa S cloud data centers under different repair policies are constructed.The aim of this work is to compare physical servers availability of a large scale and heterogeneous data center under different repairing policies.We establish monolithic and scalable analysis models for each policy,and then evaluate their steady-state availability under different operating policies.The metrics include the average number of physical hosts available and the system denial-of-service time as the measure.First,we analyze the characteristics of Iaa S cloud data center,such as large-scale,heterogeneous and dynamic migration,etc.Then,we propose a monolithic availability analysis model of Iaa S cloud data centers and analysis of the effects of different repair policies and system parameters on cloud data center availability.Meanwhile,to overcome the limitation of the monolithic analysis model,we establish a scalable availability analysis model.Finally,experiments are carried out under different parameter setting to verify the approximate accuracy of the scalable models.We also apply scalability models to compare the system availability under the two repair policies,and then to verify the impact of different repair policies and repair capabilities on the availability and construction costs of cloud services.The monolithic and scalable models established in this study are more comprehensive and reflect the actual cloud environment,which can improve the accuracy of availability analysis,and compare the advantages and disadvantages of different repair policies.(3)A sensitivity analysis method for a hierarchical scalable model is proposed.For the scalable model,the parameters are in different sub-models.The proposed method can analyze the impact of these parameters on system availabity together.The proposed method achieves this goal by solving the sensitivity of the parameters in each sub-model and then analyzing the sensitivity level of the parameters involved in each sub-model.According to the sensitivity analysis,the parameters with less influence are removed.Finally,several sub-models are applied to determine the most significant parameters for system availability.This method is used to verify the degree of influence of different repair policies and system parameters on availability.(4)A transient availability analysis model for cloud services is proposed.For Iaa S cloud data centers in critical application areas,both steady-state and transient availability analysis is necessary.This work aims for cloud service transient availability analysis of a scenario,where allows an attacker to attack the target systems with different intrusion rates.The average rates of data leakage occurring and data leakage completion after the intrusion are various.We develop a survivability model to capture the target system behaviors and intruder behaviorss from the first intrusion of the computer to the defense mechanism,and to re-develop the formula for calculating key metrics.The simulation results verify the approximate accuracy of the model and formulas.This study can help network administrators trade-off system losses and protection costs.