Research On Software Aging Prediction In Android System

Software aging is known as the gradual resource consumption,performance degradation,failure rising,and eventually system crash,which is caused by the accumulation of aging related bugs.Studying on the software aging is of great significance to improve the system reliability and user experience.Nowadays Android devices have become necessities of life,nevertheless,performance anomalies in Android system,such as hangs and crashes,affect user experience seriously.Software aging in Android system is one of the essential reasons to trigger the performance problems,but few researches have been done on the prediction of software aging in Android system.Based on two aging indicators,i.e.,Page Fault(PF)and Launch Time(LT),this thesis investigates the effectiveness of models using machine learning algorithms for software aging prediction in Android system.The main works of this thesis are as follows:1)In the stage of data preprocessing and feature extraction,the indicators of system performance on the network,virtual memory and system load are additionally considered in this thesis,and the algorithm of K-means is used to remove the outliers produced in the start-up phase of devices.In addition to the traditional indicator,i.e.,LT,this thesis introduces an aging indicator,i.e.,PF,and analyzes the correlation between the indicators of system performance and the aging indicators for feature extraction.2)In the stage of aging model training and prediction,firstly,this thesis investigates the effectiveness of models using three machine learning algorithms(i.e.,Decision Tree,Support Vector Machine,and Deep Belief Network)in the field of software aging prediction in Android system by using the indicator of LT to label the data.Experimental results show that the models using the decision tree and the support vector machine to predict the software aging have better effectiveness,and when data volume in each state,i.e.,Health State,Sub-health State,and Aging State,reaches 5000 levels,the accuracy of the model using the deep belief network is comparable to other two models.In addition to the LT,this thesis uses the PF to label the data and employs the models using three machine learning algorithms to predict software aging in Android system.Experimental results show that the models using the three machine learning algorithms can predict the software aging in Android system well,which also means the PF can be used as an indicator of software aging prediction.Moreover,this thesis combines the LT and PF to label the data,and retrains the models.Experimental results show that when the smaller points,used for aging states partition,are selected,the variance of models using the Support Vector Machine and Deep Belief Network is improved about 33%,and 66% separately,compared with the variance of models using one of the aging indicators to label the data.This thesis provides new ideas and references for the research on the software aging prediction in Android system in terms of outlier processing,the selection of machine learning algorithms and aging indicators.