PhD Final Oral Exam: Quazi Ishtiaque Mahmud

PhD Final Oral Exam: Quazi Ishtiaque Mahmud

Oct 29, 2026 - 5:00 PM
to , -

Optimizing Parallel Programs using Graph-Guided Learning and Structural Properties.

Program optimization is important to leverage the high-performance computing infrastructure. Understanding program structure is important for program optimization as different optimization strategies may be applicable for programs having different architectures and patterns. This dissertation investigates how program structure and learning-based techniques can enable accurate, efficient, and automated parallel code optimization for modern heterogeneous computing systems. It introduces PerfoGraph, an LLVM-IR-based graph representation that captures fine-grained program structure, and PerfoGraph+, which incorporates dynamic data dependencies to improve performance optimization with fewer training samples. This study explores the combination of graph neural networks (GNNs) and large language models (LLMs) to detect parallelism and generate correct, optimized OpenMP code (AutoParLLM). In addition to that, this study also employs multi-modal contrastive learning (ConTraPh) across source code and program graphs to predict optimal parallel configurations for OpenMP, OpenACC, and OpenCL without exhaustive execution. Extending these approaches to GPU programming, the study also proposes CUDAPerf, which uses reinforcement learning with execution-based and structural rewards to generate correct, high-performance CUDA kernels. Collectively, these contributions demonstrate how program structure can guide parallelism detection, code generation, configuration selection, and iterative optimization. The dissertation establishes a unified direction toward automated, scalable, and performance-aware code optimization across CPUs, GPUs, and heterogeneous computing architectures. 

Committee: Ali Jannesari (major professor), Myra Cohen, Liyi Li, Qi Li and Samik Basu