Date of Award

8-2013

Document Type

Thesis

Degree Name

Master of Science (MS)

Legacy Department

Computer Engineering

Committee Chair/Advisor

Smith, Melissa C

Committee Member

Birchfield, Stanley T

Committee Member

Ligon, Walter B

Abstract

General Purpose Graphics Processing Units (GPGPUs) have leveraged the performance and power efficiency of today's heterogeneous systems to usher in a new era of innovation in high-performance scientific computing. These systems can offer significantly high performance for massively parallel applications; however, their resources may be wasted due to inefficient tuning strategies. Previous application tuning studies pre-dominantly employ low-level, architecture specific tuning which can make the performance modeling task difficult and less generic. In this research, we explore the GPGPU design space featuring the memory hierarchy for application tuning using regression-based performance prediction framework and rank the design space based on the runtime performance. The regression-based framework models the GPGPU device computations using algorithm characteristics such as the number of floating-point operations, total number of bytes, and hardware parameters pertaining to the GPGPU memory hierarchy as predictor variables. The computation component regression models are developed using several instrumented executions of the algorithms that include a range of FLOPS-to-Byte requirement. We validate our model with a Synchronous Iterative Algorithm (SIA) set that includes Spiking Neural Networks (SNNs) and Anisotropic Diffusion Filtering (ADF) for massive images. The highly parallel nature of the above mentioned algorithms, in addition to their wide range of communication-to-computation complexities, makes them good candidates for this study. A hierarchy of implementations for the SNNs and ADF is constructed and ranked using the regression-based framework. We further illustrate the Synchronous Iterative GPGPU Execution (SIGE) model on the GPGPU-augmented Palmetto Cluster. The performance prediction framework maps appropriate design space implementation for 4 out of 5 case studies used in this research. The final goal of this research is to establish the efficacy of the regression-based framework to accurately predict the application kernel runtime, allowing developers to correctly rank their design space prior to the large-scale implementation.

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.