Analysis of communication and computation overlap in accelerated programs

Accelerated computing has revolutionized a broad range of industries by applying Graphics Processing Units (GPU(s)) to optimize workloads for efficient performance. Any application that is designed to run on such specialized devices has a typical flow that begins and ends with data movement between Central Processing Unit (CPU) and GPU. Previous research has shown that this data movement is indeed a major bottleneck with even most optimized applications hitting only 30% of the peak performance [1]. While most recent efforts have been to manually tune communication to improve performance, in this work, we present three different general optimization techniques to introduce overlap between communication and computation. The optimized model thus developed on a stencil application is evaluated against the baseline model using various performance metrics including time and throughput. Sets of experiments performed on two different GPUs reported that a certain combination of these optimizations result in a maximum speedup of 6.5 and a 160% increase in bandwidth utilization when compared to the baseline model in Compute Unified Device Architecture (CUDA).

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Yellapragada, Sushma

Contributors dc:contributor

• Snir, Marc

• Graphics Processing Units (GPU(s))
• Compute Unified Device Architecture (CUDA)

Statement dc:rights

• Copyright 2022 Sushma Yellapragada

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/115786

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F115786