Research Of Data Race Detection Method Based On Deep Learning

The development of high-performance computing technology trends upward,with the popular wave of multi-core processors,and a golden age of parallel computing is expected.Although parallel programs have brought a lot of convenience to the development of computers,there are many parallel defects in parallel programs because of the uncertainty of the interaction between threads and events.Data race is one of the most prominent types of parallel defects.If not detected in time,the program will not function properly,or even more likely to crash,hence the urgent necessity to investigate the problem of data race detection.Data race detection is one of the hot topics in current research,which is mainly divided into program analysis-based and deep learning-based data race detection.The detection tools based on program analysis show more false positives and false negatives in the detection process.However,there are problems with single extracted features and low accuracy of data race detection methods based on deep learning.In order to address the problems of current research,we propose a deep learning-based data race detection method called SmartRacer in this paper.First,our approach extracts multi-level features from several real-world applications via a static analysis tool WALA to build the training set.Secondly,the sample data is vectorized and constructed.To judge the real data race,we employ the existing data race detection tool ConRracer and mark the sample data.We also leverage the Kmeans-SMOTE algorithm to make the sample data distribute balancing.Finally,the CNN-BiLSTM network is constructed and trained to detect data race.In the experiment,we verified the effectiveness of SmartRacer from several aspects.We selected 25 parallel programs from multiple benchmark suites for training sample data extraction and data race detection,and compared SmartRacer with the currently available deep learning-based detection method DeepRace with an accuracy improvement of 5.08%.In addition,SmartRacer was compared with the existing program analysis-based data race detection tools(e.g.Said,RVPredict and SRD),and the experimental results showed that the number of false positives and false negatives of SmartRacer was reduced.