Project Grant 2211538

This $350,000 Project Grant from the National Science Foundation Office of Advanced Cyberinfrastructure, under the Computer and Information Science and Engineering program (CFDA 47.070), will fund the development of a systematic approach to minimize compression error propagation in high-performance computing applications.

Specifically, the University of Iowa will receive funding from August 2022 through July 2025 to create an accurate and efficient fault injection infrastructure integrated with commonly used lossy compression algorithms. Researchers will also design a fine-grained method to characterize error propagation through program analysis and data dependencies. A predictive model using machine learning will be developed to select compression strategies that minimize error impact given a program and target compression ratio. The technique will be integrated with domain-specific metrics in real-world applications to demonstrate effectiveness in selecting compression options.

The goal is to help scientists automatically choose lossy compression methods with the lowest error effects based on their HPC programs and desired compression levels. This should significantly reduce data storage and I/O burdens on HPC systems through more effective use of lossy compression while preserving computation accuracy.