Methods And Key Techniques Of Reliability Time Series Prediction For Service-composition-based System Of Systems

For the state's major strategic needs,the demand for large-scale complex software systems in China's aerospace,transportation,energy and other fields has been increasingly highlighted.A service-oriented System of Systems(or SoS),considers a System as a Service,aims at constructing a robust and value-added complex system by outsourcing external component systems through service composition.Service composition technology has become an important and feasible way to build SoS.The concept of service has become an important idea to build large-scale complex software system.Nevertheless,the constructed loosely coupled service-oriented SoS operates under dynamic and uncertain running environments.Any change within a single system may even cause cascading failures,which may make the whole system unable to work.Therefore,the question of how to guarantee the performance of the constructed system is of significant importance for a service-oriented SoS.Online reliability time series prediction,which aims at predicting the reliability time series in near future for the component systems of a service-oriented SoS,arises as a grand challenge in SoS research.In particular,the prediction emphasizes proactive guidences for the component system selection in constructing and maintaining a stable service-oriented SoS.Specifically,the online reliability time series prediction faces the following three challenges.First,the component systems run under a dynamic and variety execution environment,which makes it difficult to identify time-dependent regularities in reliability time series.Second,the quality assurance application of a service-oriented SoS desires a high prediction accuracy,which needs to cover a time interval containing multiple time points especially.Third,with the exception of response time,throughput,and reliability,the performance parameters of a component system,which can be used for predicting the online reliability,is difficult to be collected from client-side evaluations.Moreover,the large volume of online services coupled with high variability of Quality of Service(QoS)makes the velocity of data cumulating increase rapidly.It is more challenging when the historic data,which can be used to capture the evolution regularities of component system's temporal evolution,has cumulated to a certain scale.The online reliability prediction should deal with the following big data challenges.First,the large volume of historic data.Second,uncertain evolution of component system's execution parameters.Third,complexity of the solution space in building the prediction model.Relevant existing works,including System of Systems,reliability prediction for Service-oriented Systems,and the online failure prediction approaches for traditional computer systems,cannot systematically address the above prediction and big data challenges.To deal with the prediction challenges and make online reliability time series prediction,different machine learning methodologies have been investigated in this paper.In particular,Probabilistic Graphical Model(PGM)-,multi-steps trajectories Dynamic Bayesian Networks(multi DBNs)-,and Convolutional Neural Networks(CNN)-based prediction approaches are proposed.To demonstrate the application,the paper also presents the application scenarios and software framework on how to apply the proposed prediction methods.Specifically,the main contributions in this paper are summarized as follows.(1)The performance metric of reliability for the component systems in a service-oriented SoS has been systematically studied.The reliability for SoS component systems is defined.A performance-aware Probability of Failure on-Demand(or paPoFod)metric has been proposed to count the failure rate of a component system during a certain time span.As for paPoFoD,the failure responses and performance anomalies are regarded as failures.The paPoFoD is used in exponential reliability function to calculate the reliability of a component system.(2)The time series of application-level system performance parameters(i.e.,response time and throughput)are used to describe the running states of a component system,during a specific time span.Two independence assumptions(concerning 1st-order Markov process assumption and conditional independence assumption)are proposed to describe the uncertain evolution of a component system's system parameters time series.These provide important fundamentals for the online reliability inference.(3)Time series motifs are integrated with Probabilistic Graphical Models(PGMs),henceforth proposed a motifs-based Dynamic Bayesian Networks(or m_DBNs)model.The online reliability time series prediction approach based on m_DBNs is proposed.(4)To learn the cumulative effects of the deviations from the previously continuous multiple steps'm_DBNs predictions,a multi-step trajectories DBNs(or multi_DBNs)model and the corresponding prediction approach are proposed in this paper.(5)To deal with the big data related online reliability time series prediction challenges,deep learning approaches are investigated in this paper.A Convolutional Neural Networks(CNN)-based online reliability time series prediction approach is also proposed for the Big service-oriented systems.(6)Extensive experiments over real-word large-scale Web services have been conducted to compare the performance and investigate the effectiveness of the proposed m DBNs-,multi DBNs-,and CNN-based prediction approaches.The experimental results demonstrate the high prediction accuracy,better robustness,and higher convergence speed than other well-known approaches consistently.(7)Application scenarios(i.e.,TripPlanner and Shipping Management System)for the proposed approaches are presented.The software architecture of optimal service selection and proactive fault-tolerant service selection are proposed for the application scenarios,respectively.In sum,this paper presents a significant contribution to the research and application of software quality assurance for Service-Oriented Systems and the advance of SoS technologies.The proposed methodologies will provide some support for the development and application of large-scale complex software systems in China.