UC Davis

Scientific applications rely heavily on floating-point computations to model complex phenomena in physics, chemistry, biology, and other scientific domains. However, floating-point arithmetic is inherently imprecise and can lead to numerical inconsistencies that affect the correctness and reliability of scientific software. Additionally, optimizing these applications to efficiently utilize modern high-performance computing resources is a significant challenge.This dissertation addresses these challenges through four complementary approaches using program analysis and testing techniques to facilitate debugging and optimization of scientific applications. First, we present Ciel, a tool that uses bisection search and precision enhancement for enhanced isolation of compiler-induced numerical inconsistencies in heterogeneous code, helping developers identify the expressions that cause unexpected numerical behavior. Evaluation demonstrated 99.4% precision across 330 GPU programs and real-world mini-apps. Second, we introduce CIGEN, a tool utilizing a multi-phased approach, combining sampling, input clustering, and function optimization techniques to find input ranges with compiler-induced numerical inconsistencies. On 175 GNU Scientific Library functions, CIGEN found 53.4% more functions with high inconsistencies than prior work. Third, we present FloatGuard, a dynamic testing approach for detecting floating-point exceptions in AMD GPUs. FloatGuard successfully detected floating-point exceptions in 89.7% of evaluated programs (507 out of 565). Finally, we develop Muppet, a framework to optimize the performance of OpenMP applications via source-based code transformation and mutation testing. Muppet achieved performance improvements in 75.9% of evaluated programs with a maximum speedup of 3.57x.Together, these contributions provide developers and researchers with practical tools to improve the reliability and performance of scientific applications. Our evaluation on real-world scientific codes demonstrates the effectiveness of these approaches in detecting, localizing, and mitigating numerical issues while maintaining computational efficiency.