Adaptive Optimization Techniques for High-Performance Computing
The Pennsylvania State University
The Pennsylvania State University, 2024
@article{gudukbay2025adaptive,
title={Adaptive Optimization Techniques for High-Performance Computing},
author={Gudukbay Akbulut, Gulsum},
year={2025}
}
The dataset sizes and computing needs of increasingly prevalent high-performance computing (HPC) applications have grown exponentially over the last decade. Moreover, modern computing architectures are evolving with different paradigms, and accelerators have become indispensable parts of computing. Consequently, the imperative for performance optimization for HPC applications and intelligent resource management for evolving architectures has become paramount, particularly within the ones that leverage Machine Learning/Deep Learning (ML/DL). This dissertation aims to optimize performance and resource management in modern computing architectures for high-performance computing applications by targeting it from different angles, which are characterizing and analyzing applications, intelligent job scheduling, dynamic architecture reconfiguration, and compiler optimizations. Toward this objective, this dissertation consists of four methods from different computation layers that target performance optimization and resource management and answer the following questions: As compute platforms are evolving to meet the needs of state-of-the-art HPC applications, understanding the characteristics of these applications running on emerging architectures is crucial: (i) How could we efficiently comprehend the microarchitectural characteristics of applications from different areas and characterize them? As multi-GPU hardware platforms are widespread, finding intelligent job scheduling for HPC applications hosted by these platforms is important: (ii) How to efficiently and dynamically schedule jobs to GPUs in multi-GPU platforms to ensure fewer context switches and higher performance in time-critical applications? Memory accesses dominate the end-to-end execution times of HPC applications due to increased computational power demand and dataset size. So, there is a need to optimize the number of data accesses: (iii) How could we leverage a compiler-guided strategy to decrease costly data accesses and increase the performance of multithreaded applications? As renewable energy is becoming popular among HPC hardware platforms, optimization for energy-harvesting (EH) processors is vital: (iv) How to decrease instantaneous energy utilization of EH processors to allow more forward progress using dynamic reconfiguration of microarchitectural resources? All these methods can be utilized collectively as a boost for HPC applications running on intelligently managed resources of modern architectures.
January 27, 2025 by hgpu