Bug Prediction of SystemC Models Using Machine Learning

Abstract

In system-on-chip design, resources for verification is limited by time-to-market and cost. In order to allocate verification resources effectively, managers need to rely on their experience backed by design related metrics. However, often there are also other aspects of development process, such as bug history and developer information that can improve the effectiveness of verification. Software bug prediction is a machine learning (ML)-based technique which predicts whether a given software module is bug-prone by using product and process metrics of the module. Therefore, it can help direct verification effort, reduce costs, and improve the quality of software. Although there is a plethora of work in software bug prediction, no such work exists for SystemC. We propose an ML-based software bug prediction solution for verification of SystemC models used in virtual prototypes that takes into account system level design metrics and demonstrate its effectiveness on several open source system level designs. We find that 96% of modules could be correctly predicted as buggy or clean.