DeepTLE: Learning Code-Level Features to Predict Code Performance before It Runs

Abstract

With the continuous expansion of the software market and the updating of the maturity of the software development process, the performance requirements of software users are becoming increasingly prominent. Performance issues are essentially related to the source code. For solving the same problem, different programmers may write completely different "correct" code with the same functionality but have different performance. Most online judge system on programming make use of automated grading systems, usually rely on test results to quantify the correctness and performance for the submitted source code. However, traditional dynamic testing takes a lot of time, and the discovery of performance problems is usually after the fact even for those small scale programs. Therefore, we proposed DeepTLE which is used to effectively predict the performance of submitted source code before it runs. DeepTLE can automatically learn the semantic and structural features of the source code. In order to verify the effect of our approach, we applied it to the source code collected from the program competition website to predict if the source code would be time limit exceed or not without running its test cases. Experiment results show that our method can save 96% of the time cost compared to the dynamic testing, and the accuracy of the prediction reaches 82%.