SQLaw: Detecting Bugs in GPU Database Management Systems via Rule-Based Differential Execution

Database Management Systems (DBMSs) are essential for managing structured data. To meet the increasing performance requirements for complex, large-scale data management and analysis, GPU DBMSs have been introduced to enhance processing and query execution speeds. Despite the growing interest in GPU DBMSs and the inherent presence of bugs, there has been no systematic effort, to our knowledge, to detect bugs in GPU DBMSs. To this end, we design SQLaw, an innovative and comprehensive framework that combines offline rule learning with an online interpreter incorporating mutation for efficient and general GPU-related bug detection. The offline rule learning component automatically extracts differential execution rules, which are used to guide the synthesis of configuration and query statements for testing. The online interpreter with mutation ensures the generalization of these statements. We evaluated SQLaw on three major GPU DBMSs. Our extensive evaluations demonstrate that SQLaw outperforms current state-of-the-art approaches by up to 2.22$\times$ in the number of bugs detected within 24 hours. Additionally, SQLaw detected 51 previously unknown GPU-related bugs, of which 37 have been confirmed or fixed by developers.