Use this URL to cite or link to this record in EThOS:
Title: Engineering the performance of parallel applications
Author: MacDonald, Neil Blair
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 1996
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Parallel computing platforms are widely used to run scientific applications. The vast majority of these applications are programmed in an explicitly parallel style. Often the performance of a parallel application is considered only after implementation - in the guise of performance debugging and tuning. Performance engineering approaches incorporate this into the design phase, using performance data to inform design decisions. This thesis is concerned with performance engineering of parallel applications. Performance engineering requires accurate predictive models of application performance. The accuracy of micro-analysis techniques for predicting the execution time of sequential code is investigated on a number of representative uniprocessor platforms. This approach is extended to SPMD (Single Program Multiple Data) parallel programs written in a message-passing style using collective communication operations. The approach is used to predict the execution time of commonly occurring parallel application structures, and its accuracy is assessed on a number of representative parallel platforms. Reasoning about the performance of parallel applications in the absence of contention is straightforward; situations in which all communication serialises can be analysed with a little more sophistication. Reasoning about the effects of contention between these two extreme cases is difficult. Furthermore, allowing point-to-point message-passing operations destroys the assumption of synchrony used to analyse SPMD programs using collective communications. The complexities introduced by these issues inhibits informal reasoning about performance properties of parallel systems. A formal framework for reasoning about the performance of parallel systems is developed, based on a timed process algebra - Eager Timed CCS. Methods for automatically analysing the performance of Eager Timed CCS models are developed and extended to handle abstract Eager Timed CCS models in which time can be represented symbolically. The techniques allow the derivation of parametric expressions for the execution time of models.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available