Towards accessible, trustworthy high-performance approximate computing

The deceleration of hardware technology advancements challenges the continued acceleration of scientific simulations that drive discovery. While hardware has evolved toward GPU-dominated heterogeneous architectures, these advances alone cannot meet modern scientific computing requirements. This dissertation explores approximate computing (AC) as a complementary paradigm that trades controlled accuracy loss for substantial performance gains, addressing two fundamental barriers: easy access to AC techniques and trust that approximations meet accuracy requirements. To provide easy access, we develop declarative programming models that abstract implementation complexity while enabling automated exploration of accuracy-performance trade-offs. HPAC-Offload extends OpenMP offload to bring AC techniques to GPU applications, achieving up to 6.9x speedup with less than 10% error by adapting algorithms to GPU architectural constraints. HPAC-ML introduces a directive-based approach for integrating neural network surrogates into scientific applications, automating data collection, model training, and deployment to achieve up to 84x acceleration. While neural network surrogates offer impressive speedups, they generalize poorly to out-of-distribution (OOD) data, potentially producing arbitrarily incorrect results. To increase trust, we systematically evaluate OOD detection techniques through NNUEEHCS, which transforms evaluation from an intractable O(m x n) implementation effort to an O(m+n) plug-and-play workflow. Analysis of over 2,400 models reveals that data-centric approaches like Kernel Density Estimation achieve superior OOD detection on our data suite. Selectively deploying surrogates based on OOD detection creates systematic load imbalance in distributed simulations. To address this, Kombucha, a partial MPI implementation that equips MPI Python programs with Charm4Py load balancing. These components culminate in a blueprint for Trust-Aware Surrogate Systems (TASS) that selectively deploy neural acceleration based on OOD detection. Case studies demonstrate that successful deployment depends on OOD distribution patterns and application structure, with clustered OOD data enabling 3.85x speedup while uniform OOD distribution patterns yield marginal gains. This work establishes practical foundations for approximate computing in scientific applications, providing tools to explore approximation techniques and mechanisms to deploy them safely. By addressing accessibility and trustworthiness, we move toward enabling computational scientists to harness order-of-magnitude performance improvements while maintaining reliability required for scientific discovery in the post-Moore era.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Fink, Zane

Contributors dc:contributor

• Kale, Laxmikant V.
• Olson, Luke
• Misailovic, Sasa
• Parasyris, Konstantinos

• High-performance Computing
• Gpu
• Approximate Computing

Statement dc:rights

• Copyright 2025 Zane Fink

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129842

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129842