Multiprocessors are widely used today as a way to achieve high performance using commodity microprocessors. Since different applications place different demands on the system, performance evaluation in the design phase of such systems is important. The most common method is the use of simulators. This thesis presents a simulator of a multiprocessor for throughput-oriented applications, using a distribution-driven simulation method. With this method hardware, and to some extent software, parameters can be studied. A case study is performed, focusing on performance issues for various characteristics of a multiprocessor implementation of a telecom server that today is a uniprocessor. The results show that the key bottlenecks for this particular system lies in parallelisation of the software, which is written with a uniprocessor environment in mind. Also, providing sufficient bus bandwidth is an important factor. Other issues investigated, such as scheduling techniques, three different architectural choices, and degree of memory interleaving, is found to have a smaller impact on performance.
@masterthesis{PlatformEvaMP_master,
title = {A Platform for Evaluation of Multiprocessors in Throughput-Oriented Systems},
author = {Warg, Fredrik},
year = {1999},
month = {05},
abstract = {Multiprocessors are widely used today as a way to achieve high performance using commodity microprocessors. Since different applications place different demands on the system, performance evaluation in the design phase of such systems is important. The most common method is the use of simulators.
This thesis presents a simulator of a multiprocessor for throughput-oriented applications, using a distribution-driven simulation method. With this method hardware, and to some extent software, parameters can be studied. A case study is performed, focusing on performance issues for various characteristics of a multiprocessor implementation of a telecom server that today is a uniprocessor.
The results show that the key bottlenecks for this particular system lies in parallelisation of the software, which is written with a uniprocessor environment in mind. Also, providing sufficient bus bandwidth is an important factor. Other issues investigated, such as scheduling techniques, three different architectural choices, and degree of memory interleaving, is found to have a smaller impact on performance.},
keywords = {Technology, computer engineering, computer architecture, performance, evaluation, distribution-driven simulation, multiprocessors},
school = {Luleå University of Technology},
url = {https://urn.kb.se/resolve?urn=urn:nbn:se:ltu:diva-45046},
note = {Publication data: https://warg.org/fredrik/publ/}
}