Files in this item

FilesDescriptionFormat

application/pdf

application/pdf9305506.pdf (7MB)Restricted to U of Illinois
(no description provided)PDF

Description

Title:Performance Evaluation and Modeling Techniques for Parallel Processors
Author(s):Dimpsey, Robert Tod
Doctoral Committee Chair(s):Iyer, R.,
Department / Program:Electrical Engineering
Discipline:Electrical Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:Ph.D.
Genre:Dissertation
Subject(s):Engineering, Electronics and Electrical
Abstract:This thesis addresses the issue of application performance under real operational conditions. A technique is introduced which accurately models the behavior of an application in real workloads. The methodology can evaluate the performance of the application as well as predict the effects on performance of certain system design changes. The constructed model is based on measurements obtained during normal machine operation and captures various performance issues including multiprogramming and system overheads, and contentions for resources.
Methodologies to measure multiprogramming overhead (MPO) are introduced and illustrated on an Alliant FX/8, an Alliant FX/80, and the Cedar parallel supercomputer. The measurements collected suggest that multiprogramming and system overheads can significantly impact application performance. The mean MPO incurred by PERFECT benchmarks executing in real workloads on an Alliant FX/80 is found to consume 16% of the processing power. For applications executing on Cedar, between 10 and 60% of the application completion time is attributable to overhead caused by multiprogramming. Measurements also identify a Cedar FORTRAN construct (SDOALL) which is susceptible to performance degradation due to multiprogramming.
Using the MPO measurements, the application performance model discussed above is constructed for computationally bound, parallel jobs executing on an Alliant FX/80. It is shown that the model can predict application completion time under real workloads. This is illustrated with several examples from the Perfect Benchmark suite. It is also shown that the model can predict the performance impact of system design changes. For example, the completion times of applications under a new scheduling policy are predicted. The model-building methodology is then validated with a number of empirical experiments.
Issue Date:1992
Type:Text
Description:157 p.
Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 1992.
URI:http://hdl.handle.net/2142/71979
Other Identifier(s):(UMI)AAI9305506
Date Available in IDEALS:2014-12-16
Date Deposited:1992


This item appears in the following Collection(s)

Item Statistics