Research On Cloud Platform Fault Prediction Method Based On LSTM

With the continuous development and maturity of cloud computing,more and more enterprises tend to deploy their business systems on cloud platforms to reduce operating costs.The resources of the cloud platform are usually encapsulated in virtual machines.Failure of the virtual machine will cause the cloud platform to fail to provide services for users.As the scale of the cloud platform increases,virtual machines in the cloud platform are prone to failure and performance degradation when running,and the availability of cloud services has become the primary issue facing cloud computing.Based on the characteristics of the load in the cloud platform and the problems in the existing cloud platform fault prediction technology,this paper proposes a series of improved algorithms based on the traditional time series prediction model.The following work was done:In order to solve the problem that traditional point-value prediction models can not effectively communicate the load status of the cloud platform and the fixed parameter model is difficult to adapt to the dynamic load changes of the cloud environment,an LSTM algorithm based on adaptive confidence intervals is proposed,which is compared on the real Amazon data set.The experimental method proves that the proposed algorithm can achieve higher prediction accuracy.Based on the traditional confidence interval-based alarm algorithm,in the cloud environment,frequent task scheduling is prone to false alarms,and a fault early warning algorithm based on interval offset is proposed.The algorithm first defines the concept of interval offset,which indicates the degree of offset of the true value of the system relative to the confidence interval.Based on the fact that most of the failures of the cloud platform are accompanied by the gradual departure of some performance indicators of the virtual machine from the normal range,the average interval offset in the multi-step prediction is used to finally determine whether the system has failed,which improves The accuracy of the alarm.Aiming at the problems of traditional fault early-warning frameworks,such as high overhead and prone to data redundancy,a new cloud platform fault early-warning framework is designed and implemented.An online experimental platform was set up to simulate real cloud environment failures by injecting anomalies into the cloud platform,and using the proposed algorithm to predict the future health status of the cloud platform,the superiority of the proposed algorithm was demonstrated through comparative experiments.