Tuning-Oriented Software Configuration Recommendation

With the increasingly complexity of modern software,the number of configuration options has been increasing rapidly,making the configuration search space exploding exponentially,which brings great challenges to performance tuning.Therefore,it is par-ticularly urgent to recommend critical configuration options for users and auto-tuning frameworks before performance tuning so as to reduce the search space.While improving software performance,such tuning might affect other quality attributes such as security and reliability.Given this challenge,we analyze and annotate the configuration options of a large number of open source software,and builds the first dataset which focus on which options affect performance and whether these options hurt other quality attributes.Based on the dataset,we designed and implemented Pre Tune,an automatic configuration recom-mandation framework.Through the semantic understanding and feature extraction of the configuration documents,Pre Tune recommends suitable configuration options according to different requirements for software quality.This work includes:1.Taking software performance as the main concern and taking into account of other quality attributes including safety,reliability and functionality,this work has manually analyzed software configuration documents of 16 software applications in 4 categories and constructed the first performance-oriented configuration tuning dataset;2.This work proposes an automatically dataset expansion method.Given that it is both labor-intensive and time demanding when constructing the dataset,Pre Tune expands the dataset previous annotated by applying FEAT,a frequent sequence pattern mining algorithms,to find the most frequent expression patterns in each category.These patterns are then used for dataset expansion.Experiments have proved that Pre Tune can reach an accuracy of 77.5% when expanding dataset with this method;3.This work has designed and implemented a configuration recommendation model for performance tuning.Based on the dataset constructed in the early stage,Pre Tune first utilizes patterns mined from the training configuration documents to highlight the critical information inside.With these critical information and configuration documents,Pr Tune then trains a configuraiton recommandation model based on the multi window size con-volutional neural network It also classified the impact of other software quality attributes,reducing the search space of performance tuning.Experiments show that the configu-ration recommendation model has an precision of 85.9% in recommending performance-related configuration options,and can predict the side effects of these options with overall precision of 79.6%.We apply PreTune to real software applications.And experiments show that it can provide critical configuration options in a short time according to the user’s requirements for quality attributes,including 84.3%(27/32)of those given by the existing work.Ad-ditionally,it can correctly predict extra 104 performance-related configuration options.Through experiments,we trigger a few scenarios where tuning of certain performance-related configuration options causes severe software quality declines,which further veri-fies the effectiveness of PreTune.