| With the popularity of Internet applications,online software services are affecting people’s lives all the time.They have greatly improved the current situation of people’s production and life.However,this also means that once these services are abnormal,they will directly cause huge losses to users and enterprises.Therefore,in order to ensure the stable operation of online services,it is very important to deeply study the root cause location technology of service exceptions.With the continuous acceleration of online service iteration and update speed,the continuous improvement of application complexity and the continuous popularization of microservice architecture,it brings great challenges to the research of service anomaly root cause location technology.In order to accurately locate the cause of abnormal services,it is necessary not only to analyze the abnormal services generated by the service layer itself,but also to further locate the abnormal root cause metric at the host level based on the complex mapping relationship between the service and its corresponding underlying host,that is,cross-layer correlation analysis and root cause location for the software system.At present,although the research on cross-layer oriented service exception root cause location algorithm has made great progress,there are still some problems.For example,the abnormal service location method based on the characteristics of performance indicators cannot accurately detect the abnormal changepoints;The root cause location method based on call chain has some problems,such as call chain data is not easy to obtain,and there are missing chains and wrong chains when building call chain;The root cause location methods based on service dependency graph mostly build service dependency graph for a single indicator,but fail to build a unified service dependency graph for multiple indicators,resulting in poor positioning effect and other problems.To solve the above problems,this thesis proposes a cross-layer oriented service exception root cause ranking algorithm.Firstly,at the service-oriented level,an abnormal service root cause location method based on service dependency graph is proposed.By analyzing the key performance indicators of each service,the root cause service that lead to a series of abnormal services are identified.Secondly,for the host level,on the basis of accurately locating the root cause of abnormal service,further analyze the complex correlation between service and host,propose the root cause location method of host exception indicator based on Prophet and kernel density estimation,deeply analyze the correlation between host level exception indicators,and identify the root cause indicator leading to current service exception.The research work of this thesis is as follows:(1)At the service-oriented level,service exception root cause location method based on service dependency graph(SDGraphRCL)is proposed.Firstly,in view of the problems that the root cause location information contained in a variety of monitoring indicators is not considered in the construction of the existing service dependency graph,and the time complexity of the optimization algorithm using Trace data pruning strategy is not considered,this thesis proposes a service dependency graph generation algorithm based on Trace data pruning,which uses a variety of service performance monitoring indicators to build a unified service dependency graph,A data pruning strategy based on Trace is proposed to reduce the number of additional independent testing steps during the construction of dependency graph,so as to reduce the time complexity of the algorithm.Then,aiming at the problems that the existing root cause service location methods based on random walk cannot carry out the reasoning of multiple indicators and causal analysis on the unified service dependency graph,and the reasoning accuracy is poor,this thesis improves the abnormal service correlation formula in the random walk algorithm,and based on this,improves the calculation formula of transition probability used to determine the direction of root cause reasoning,The reasoning weight balance formula is proposed to balance the abnormal correlation and randomness in the root cause reasoning direction in the process of causal reasoning,which improves the accuracy of the algorithm and the convenience of user parameter adjustment.Finally,simulation experiments are carried out for the proposed algorithm.Compared with the traditional root cause method of locating service exception,the algorithm has high accuracy in AC@1,AC@3,AC@5.(2)At the host-oriented level,a root cause location method of anomaly indicators based on Prophet and kernel density estimation(ProphetKdeRCL)is proposed.Based on the diagnosis results of the above abnormal service,further analyze the complex correlation between the abnormal service and its corresponding host.Firstly,in view of the problems that the existing indicators anomaly detection algorithms rely on manually setting the threshold or assuming that the data obey an ideal scene with certain distribution,this thesis proposes an anomaly boundary solution algorithm based on extreme value theory,which relies on rigorous extreme value theory to accurately estimate the value region with low probability,The calculation formula of abnormal low value boundary is proposed,which solves the problem that the existing extreme value based methods cannot detect anomalies in the minimum range.Then,in view of the poor applicability of the existing anomaly detection algorithms due to the trend and periodicity of the indicators at the host level,this thesis uses the prophet algorithm to fit the monitoring indicators of the mixed form,and proposes a changepoint update algorithm based on the anomaly boundary to solve the problem of inaccurate search of the changepoint of the prophet.Finally,aiming at the problem that the existing root cause location algorithm for the host layer does not consider the difference of the fluctuation degree of different indicators,this thesis measures the fluctuation degree of different indicators by calculating the probability of indicators fluctuation through kernel density estimation,and improves the PC causality analysis algorithm to analyze the dependence of delay through time window and intermediate structure,It solves the problem that the causal relationship cannot be learned due to the delayed time-series data and inconsistent indicator collection frequency,and improves the accuracy of cross-layer oriented service exception root cause ranking. |
