An Investigation into Reproducibility and Performance in Bioinformatics Software: A Case Study of BLAST+ and Floating-Point Arithmetic

This report is on the reproducibility and performance of bioinformatics software, with a specific focus on versions of the widely-used Basic Local Alignment Search Tool (BLAST+) suite. The core challenges addressed are the inconsistencies arising from floating-point arithmetic implementations across different C/C++ compilers and hardware architectures, alongside the identification and analysis of software performance bottlenecks. Any investigation of this problem should be focused on three primary areas: first, an empirical testing and documentation of the reproducibility of BLAST+ outputs generated under varying compilation and execution environments, with particular attention to floating-point mathematical discrepancies; second, the identification and characterization of performance bottlenecks within the BLAST+ codebase using established profiling tools and analysis; and third, an exploration and preliminary evaluation of optimization strategies, including code reordering, the use of alternative mathematical functions, and any application of machine-specific instruction sets (e.g., SIMD). These recommendations are based on the goals of reliability of scientific results derived from bioinformatics tools and the computational costs associated with large-scale biological sequence analysis. This report further discusses the advantages of open-source development paradigms and delves into the technical intricacies of floating-point arithmetic, including considerations for 32-bit versus 64-bit builds and the historical context of legacy software, that underpin the identified challenges.