Optimální konfigurace FETI řešičů v HPC

Abstract

The main objective of this thesis was the performance prediction of Total FETI and Hybrid Total FETI methods implemented in ESPRESO library. The secondary objective was the investigation of energy requirements of ESPRESO for a planned consumption model.

The performance model was created by generalized linear regression, using R-language for implementation of the software needed to process measured data. The energy consumption was investigated using MERIC and RADAR tools implemented under READEX project in Horizon2020 programme.

The model describing performance is useful for estimations of optimal settings, although the fit is not very precise for larger values. MERIC and RADAR were used to evaluate energy savings for multiple hardware and application parameters.

The performance model will be implemented into the ESPRESO library, so it will be able to estimate optimal settings for minimal run-time without demanding prior tests. Similarly, the consumption model will be assembled using results obtained by MERIC and RADAR and later implemented into the ESPRESO too, making it able to estimate not only optimal settings for minimal run-time, but for the minimal energy consumption too.