Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.662013
Title: A statistical approach to performance evaluation of parallel systems with reference to chemical engineering
Author: Skilling, Neil
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 1995
Availability of Full Text:
Access through EThOS:
Full text unavailable from EThOS. Please try the link below.
Access through Institution:
Abstract:
Distribution memory multicomputers appear to offer a cost effective general purpose parallel computing resource. Unfortunately these multicomputers have not always delivered the processing performance promised from a summation of individual processor speeds. A lot of time and effort can be expended trying to close this performance gap. Detailed dynamic simulations of chemical processing equipment can be naturally and robustly modelled as a set of communicating sequential processes where the information flow accurately mirrors the material flow in the real equipment. These programs have a static, time invariant process graph which is suited to execution on a distributed memory MIMD machine. There are many factors that affect the performance of a parallel program. The programmer, usually with the aid of profiling tools, is faced with a trial and error tuning up process. This thesis addresses the issue of performance evaluation of parallel systems by presenting a methodology that enables rapid identification of performance limiting factors. In particular the study of static placement strategies as performance factors can be readily investigated for a range of programs. Through the use of standard statistical design of experiments, synthetic program graphs and a general purpose multiprocessor simulation system, placement strategies and other performance factors can easily be identified and their precise effect quantified. Through statistical analysis predictive performance models can also be constructed. The approach presented is general and can be applied to an arbitrary parallel program. Results are presented for a common class of parallel programs called structured spatial decomposition and for process systems simulations.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.662013  DOI: Not available
Share: