Measured performances on vectorization and multitasking with a Monte Carlo code for neutron transport problems
✍ Scribed by Yves Chauvet
- Publisher
- Elsevier Science
- Year
- 1985
- Tongue
- English
- Weight
- 450 KB
- Volume
- 37
- Category
- Article
- ISSN
- 0010-4655
No coin nor oath required. For personal study only.
✦ Synopsis
This paper summarized two improvements of a real production code by using vectorization and multitasking techniques. After a short description of Monte Carlo algorithms employed in our neutron transport problems, we briefly describe the work we have done in order to get a vector code. Vectorization principles will be presented and measured performances on the CRAY iS, CYBER 205 and CRAY X-MP compared in terms of vector lengths.
The second part of this work is an adaptation to multitasking on the CRAY X-MP using exclusively standard multitasking tools available with FORTRAN under the COS 1.13 system. Two examples will be presented.
The goal of the first one is to measure the overhead inherent to multitasking when tasks become too small and to define a granularity threshold, that is to say a minimum size for a task.
With the second example we propose a method that is very X-MP oriented in order to get the best speedup factor on such a computer.
In conclusion we prove that Monte Carlo algorithms are very well suited to future vector and parallel computers.