guest ::
| Home > Publications database > Konzeption eines Softwaresystems zur Prognose von Benchmarking-Ergebnissen paralleler Programme im Scientific Computing |
| Home > Publications database > Konzeption eines Softwaresystems zur Prognose von Benchmarking-Ergebnissen paralleler Programme im Scientific Computing |
| Master Thesis | PreJuSER-5125 |
2009
Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag
Jülich
Please use a persistent id in citations: http://hdl.handle.net/2128/3666
Report No.: Juel-4298
Abstract: In today’s research environment scientific computing acts as an interface between theory and experiment. Large simulations are calculated on supercomputers. Therefore it is quite important for scientists how an application program behaves on a certain computer. But that is also true for the management and administrators of a computer center who want to meet the needs of their users. Especially during a procurement process for a new supercomputer it is fundamental to estimate the performance of that computer. Normally, benchmark suites are used to measure the performance of a supercomputer. But on a future computer these benchmarks cannot be run, so one is interested in having a prediction. In this master thesis, benchmarking is introduced in the context of supercomputing. Questions which benchmarking should answer are reviewed. Several possibilities of performing a prediction for benchmark results are shown. It seems that Artificial Intelligence is a promising method for this kind of prognosis and is investigated further. In detail, expert systems together with ontologies and artificial neural networks are taken into account and compared with each other. After a decision for one of these methods, it is used in an extensive test series. In the last part of this thesis the conception of a software system for prediction of application benchmarks on new computer systems is built on the performed tests.
|
The record appears in these collections: |
PDF
Fulltext by OpenAccess repository