Dissertation Defense Announcements

As heterogeneous parallel architectures grow increasingly complex, achieving high performance and effectively teaching parallel programming have become more challenging. Benchmark suites are powerful tools for illustrating and evaluating optimization techniques in practical, performance-critical scenarios. But commonly used parallel benchmark suites (e.g., SPEC OMP and Rodinia) are primarily designed for performance assessment purposes only. They are not intended for performance optimization training or educational instruction in parallel programming. Furthermore, their complexity in configuration and deployment often limits their accessibility, reducing their practical utility for researchers and students in educational settings. To address these limitations, this dissertation presents two novel benchmark suites, NeoRodinia and CUDAMicroBench, that support not only performance evaluation, but also the exploration of optimization strategies. These suites are further augmented with educational features, such as integration with large language models (LLMs) for optimization guidance and interactive, browser-accessible execution environments.

NeoRodinia features a structured three-level parallelization model (P1, P2, P3) across CPU worksharing, GPU offloading, SIMD, and tasking. It provides standardized execution workflows, automated performance evaluation scripts and visualization tools. Additionally, NeoRodinia integrates AI-assisted analysis, allowing LLMs to offer optimization recommendations and debugging insights. CUDAMicroBench is a modular microbenchmark suite targeting key GPU optimization challenges such as memory hierarchy usage, warp divergence, and concurrent kernel execution, serving as a practical reference for GPU performance tuning.

In addition to benchmark-based contributions, this dissertation advances parallel programming education by introducing the Interactive OpenMP Programming book. By employing deliberate prompt engineering strategies, it effectively leverages large language models (ChatGPT-4, Gemini Pro 1.5, and Claude 3) to enhance the quality, relevance, and pedagogical value of the generated content. Delivered via a Jupyter-based environment, it enables real-time experimentation with OpenMP constructs, promoting hands-on learning and deeper understanding.

Collectively, these contributions form a unified educational infrastructure for modern parallel computing. By combining benchmarking, structured optimization guidance, and LLM-driven interactive learning, this work bridges performance engineering and pedagogy, providing a scalable and adaptable solution for educators and learners in today's heterogeneous HPC landscape.

This study examines the relationship between Enterprise Risk Management (ERM) maturity and firm performance, with an emphasis on how organizational complexity, industry type, and board engagement may moderate this relationship. Although prior research has broadly acknowledged ERM’s role in mitigating risks, few studies explore how ERM maturity contributes directly to firm performance outcomes. Additionally, the literature often treats ERM as a static process rather than a dynamic capability that evolves and adapts within different organizational structures and industry-specific environments. This gap highlights an important opportunity to examine ERM maturity as a strategic asset that enhances decision-making, optimizes resource allocation, and drives sustainable competitive advantage.

Drawing on the Resource-Based View (RBV), Agency Theory, and Contingency Theory, this study addresses these gaps through a quantitative survey of risk professionals from 111 publicly listed companies. By focusing on how ERM maturity impacts firm performance and controlling for variables such as firm size, market volatility, and leverage, this research seeks to provide nuanced insights into ERM's role as a driver of resilience and adaptability. The findings will contribute to the literature by offering a more comprehensive view of ERM maturity’s influence across diverse organizational and industry contexts, ultimately providing practical guidance for companies aiming to refine their risk management capabilities in pursuit of competitive advantage.

Flooding is one of the most frequent and destructive natural disasters, posing significant risks to communities, infrastructure, and economies worldwide. In North Carolina, diverse flood types—driven by varied topographic and climatic conditions—necessitate adaptable forecasting methods. Traditional flood prediction models rely on hydrological and meteorological data but often struggle to incorporate the complexity of different flood types within a single framework. This research explored the application of machine learning (ML) techniques in flood forecasting, specifically evaluating how different ML models handle varying data requirements across different regions with flood types.
Focusing on the Upper Haw River and Cape Fear River watersheds in North Carolina, this study employed a Convolutional Neural Network (CNN) to predict flood stages using hydrometeorological data, including rainfall, elevation, distance from the river, soil type, land use/cover, wind speed, wind direction, soil water volume, lag feature, and gauge stages. Rather than comparing ML accuracy against traditional hydrological models, this study examined how ML models can adapt to diverse flood conditions and data constraints. The results indicate that CNN-based models effectively capture spatial dependencies and patterns, providing valuable insights into the role of different input features, such as lag effects and rainfall distribution, in flood prediction.
A Command Line Interface (CLI) was developed to enable real-time interaction with the model, enhancing its usability for decision-makers. The study highlighted the strengths and limitations of ML-based forecasting, demonstrating its potential while identifying areas requiring further refinement, such as incorporating additional meteorological variables and real-time data. Aimed to evaluate the feasibility of using ML for statewide flood prediction to encompass a range of flood causes, this research also contributed to determining the boundaries of coastal and piedmont regions and the type of data requirements to develop flood forecasting models in these regions.

Partly interval-censored failure time data are common in clinical and epidemiological studies, where the failure time of interest is either exactly observed or known to lie within a specific interval. Additionally, two-phase sampling designs are often employed to measure covariates for a subset of participants, reducing study costs. This paper addresses regression analysis of partly interval-censored failure time data under stratified semiparametric transformation models, incorporating time-dependent covariates subject to: (i) missingness due to two-phase sampling and (ii) measurement errors during observation. We propose a maximum weighted likelihood estimation method and develop an EM algorithm for implementation. A weighted bootstrap approach is introduced for variance estimation, and the asymptotic properties of the proposed estimator are established. Extensive simulation studies demonstrate the method’s satisfactory finite-sample performance, and its practical utility is illustrated through an application to data from the HIV prevention trials HVTN-703/HVTN-704.