Dissertation
Enhancing soft error resilience in HPC systems through performance variation analysis and cross-layer optimization
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 2025
DOI: 10.25820/etd.008223
Abstract
Soft errors, caused by radiation and transient faults, are becoming increasingly prevalent in modern HPC systems, leading to silent data corruptions (SDCs) that compromise result integrity. While many hardware-level techniques have been developed to mitigate these errors, software-level solutions remain critical due to their flexibility and portability across architectures.
Among these, instruction duplication has emerged as a widely used and architecture-agnostic technique for soft error detection. By duplicating instructions and checking their consistency at runtime, it offers a general-purpose way to detect errors without requiring hardware modifications. However, existing research has largely focused on evaluating its SDC coverage, while two crucial aspects remain underexplored: (1) the true protection effectiveness after compilation, where transformations across abstraction layers may silently invalidate software-level protection guarantees, and (2) the runtime performance variation introduced by duplication under different workloads and programs.
In this dissertation, I focus on addressing these two limitations.First, in ISSRE’23, I systematically characterize how duplication-induced performance varies depending on instruction patterns and hardware features such as branch prediction and register pressure, identifying factors of high and low overhead. Second, in SC’23, I analyze the root causes of protection deficiencies by examining instruction duplication across the LLVM IR and binary levels, revealing semantic mismatches and cross-layer inconsistencies that degrade fault coverage. Finally, in DSN’24, I propose a fast, low-overhead protection technique that allows efficient and selective protection while preserving strong error detection at the assembly level. Together, these efforts contribute a comprehensive understanding and a set of optimizations for instruction duplication, aiming to improve both practical error coverage and runtime efficiency in high-performance computing systems.
Details
- Title: Subtitle
- Enhancing soft error resilience in HPC systems through performance variation analysis and cross-layer optimization
- Creators
- Zhengyang He
- Contributors
- Guanpeng Li (Advisor)Muchao Ye (Committee Member)Tianyu Zhang (Committee Member)Weiran Wang (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Computer Science
- Date degree season
- Autumn 2025
- DOI
- 10.25820/etd.008223
- Publisher
- University of Iowa
- Number of pages
- xiii, 62 pages
- Copyright
- Copyright 2025 Zhengyang He
- Language
- English
- Date submitted
- 11/30/2025
- Description illustrations
- Illustrations, graphs, charts, tables
- Description bibliographic
- Includes bibliographical references (pages 58-62).
- Public Abstract (ETD)
High-performance computing (HPC) systems power some of the world s most critical technologies ; from climate modeling to space exploration. However, as these systems grow more powerful, they also become more sensitive to subtle hardware faults. These small errors, known as soft errors, can silently alter a program s results without causing it to crash, making them especially dangerous.
This dissertation explores ways to make HPC software more resilient to such faults. It focuses on improving software-level error detection techniques by identifying their hidden weaknesses and enhancing their coverage across multiple levels of the computing stack. Through a series of practical experiments and new tools, this work demonstrates how to better protect scientific applications from invisible errors while keeping performance overhead low.
The goal of this research is to ensure that as we push the boundaries of science and engineering with advanced computing, we can also trust the results these systems produce. This work contributes toward building more reliable and trustworthy HPC software for future technologies and discoveries.
- Academic Unit
- Computer Science
- Record Identifier
- 9985135345702771
Metrics
1 Record Views