A parallel program development system for multi-core architectures

Component-oriented programming model was suggested for clusters of multi-cores systems [1], and in the topic of parallel program performance prediction, the accuracy of communication cost prediction plays a critical role in algorithm designs to maximize efficiency. In this project, we have explored...

Full description

Saved in:
Bibliographic Details
Main Author: Fu, Yong.
Other Authors: Stephen John Turner
Format: Final Year Project
Language:English
Published: 2010
Subjects:
Online Access:http://hdl.handle.net/10356/39953
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Component-oriented programming model was suggested for clusters of multi-cores systems [1], and in the topic of parallel program performance prediction, the accuracy of communication cost prediction plays a critical role in algorithm designs to maximize efficiency. In this project, we have explored a component-oriented parallel language P-COM2 at the SCE PDCC cluster, as well as communication models namely LogP and LogGP. Our goal is to predict the execution time of a parallel program involving intensive communications. Starting with assessing LogGP parameters for the cluster, together with practical concerns over an Ethernet-based cluster, two approaches were taken in model implementations; one captures the performance characteristics when message size is moderately long (e.g. a few KBytes), the other approach used a curve fitting technique to describe performance characteristics up to a much wider message range. Predicted commutation cost were very accurate based on the two approaches. Finally a case study of 2D FFT application was used to validate the model implementation; results obtained proved the execution time prediction error is within 9% for experimented 2D FFT problem sizes. We have demonstrated the accurate execution time prediction of parallel program by applying LogP and LogGP in a practical way, and we believe such practical approach of applying models made a closer step towards a realistic parallel program performance prediction and helps with parallel algorithm designs concerns.