Publications

Detailed Information

Techniques for Ease of OpenCL Programming : OpenCL의 프로그래밍 용이성 향상 기법

Cited 0 time in Web of Science Cited 0 time in Scopus
Authors

김정현

Advisor
이재진
Major
공과대학 전기·컴퓨터공학부
Issue Date
2016-02
Publisher
서울대학교 대학원
Keywords
OpenMPOpenCLease of programminghigh performanceclustersheterogeneous systemsprogramming modelacceleratorsbenchmarks
Description
학위논문 (박사)-- 서울대학교 대학원 : 전기·컴퓨터공학부, 2016. 2. 이재진.
Abstract
OpenCL is one of the major programming models for heterogeneous systems. This thesis presents two limitations of OpenCL, the complicated nature of programming in OpenCL and the lack of support for a heterogeneous cluster, and proposes a solution for each of them for ease of programming.

The first limitation is that it is complicated to write a program using OpenCL. In order to lower this programming complexity, this thesis proposes a framework that translates a program written in a high-level language (OpenMP) to OpenCL at the source level. This thesis achieves both ease of programming and high performance by employing two techniques
data transfer minimization (DTM) and performance portability enhancement (PPE). This thesis shows the effectiveness of the proposed translation framework by evaluating benchmark applications and the practicality by comparing it with the commercial PGI compiler.

The second limitation of OpenCL is the lack of support for a heterogeneous cluster. In order to extend OpenCL to a heterogeneous cluster, this thesis proposes a framework called SnuCL-D that is able to execute a program written only in OpenCL on a heterogeneous cluster. Unlike previous approaches that apply a centralized approach, the proposed framework applies a decentralized approach, which gives a chance to reduce three kinds of overhead occurring in the execution path of commands.
With the ability to analyze and reduce three kinds of overhead, the proposed framework shows good scalability for a large-scale cluster system. The proposed framework proves its effectiveness and practicality by compared to the representative centralized approach (SnuCL) and MPI with benchmark applications.

This thesis proposes solutions for the two limitations of OpenCL for ease of programming on heterogeneous clusters. It is expected that application developers will be able to easily execute not only an OpenMP program on various accelerators but also a program written only in OpenCL on a heterogeneous cluster.
Language
English
URI
https://hdl.handle.net/10371/119178
Files in This Item:
Appears in Collections:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share