A component analysis based on serial results analyzing performance of parallel iterative programs
- Dalhousie Univ. (Canada)
This research is concerned with the parallel performance of iterative methods for solving large, sparse, nonsymmetric linear systems. Most of the iterative methods are first presented with their time costs and convergence rates examined intensively on sequential machines, and then adapted to parallel machines. The analysis of the parallel iterative performance is more complicated than that of serial performance, since the former can be affected by many new factors, such as data communication schemes, number of processors used, and Ordering and mapping techniques. Although the author is able to summarize results from data obtained after examining certain cases by experiments, two questions remain: (1) How to explain the results obtained? (2) How to extend the results from the certain cases to general cases? To answer these two questions quantitatively, the author introduces a tool called component analysis based on serial results. This component analysis is introduced because the iterative methods consist mainly of several basic functions such as linked triads, inner products, and triangular solves, which have different intrinsic parallelisms and are suitable for different parallel techniques. The parallel performance of each iterative method is first expressed as a weighted sum of the parallel performance of the basic functions that are the components of the method. Then, one separately examines the performance of basic functions and the weighting distributions of iterative methods, from which two independent sets of information are obtained when solving a given problem. In this component approach, all the weightings require only serial costs not parallel costs, and each iterative method for solving a given problem is represented by its unique weighting distribution. The information given by the basic functions is independent of iterative method, while that given by weightings is independent of parallel technique, parallel machine and number of processors.
- Research Organization:
- Front Range Scientific Computations, Inc., Boulder, CO (United States); US Department of Energy (USDOE), Washington DC (United States); National Science Foundation, Washington, DC (United States)
- OSTI ID:
- 223862
- Report Number(s):
- CONF-9404305-Vol.1; ON: DE96005735; TRN: 96:002320-0036
- Resource Relation:
- Conference: Colorado conference on iterative methods, Breckenridge, CO (United States), 5-9 Apr 1994; Other Information: PBD: [1994]; Related Information: Is Part Of Colorado Conference on iterative methods. Volume 1; PB: 203 p.
- Country of Publication:
- United States
- Language:
- English
Similar Records
Exploring parallel algorithms having no serial analogues
Krylov subspace acceleration of waveform relaxation