Parallel performance evaluation of general engineering applications
Engineering applications exist which have many expensive computational phases. They represent a class of programs that are difficult to parallelise because of the numerically dissimilar computational phases. Traditional performance analysis methods are difficult to apply to these codes. Two new overall performance metrics relevant to these codes are introduced. The first is a simple relation for overall speedup. The second quantifies the influence on overall code speedup of optimisations to individual phases. This relation is validated using a parallel nuclear accident simulation code. The interdependence of phase optimisation and the use of the performance relations for parallel optimisation is demonstrated using speedup data from this parallel code.
Bibliographic Reference: Paper presented: International Conference on Parallel Processing, St. Charles, Illinois (US), Aug. 14, 1991
Availability: Available from (1) as Paper EN 36346 ORA
Record Number: 199111211 / Last updated on: 1994-12-02
Original language: en
Available languages: en