The final performance model consists of several partial ones, each one created by a generalized linear regression, using R-language. This concept makes it possible to add new partial models easily to the final one. If needed, the final model can be made more precise like that by modeling more regions or expanded according to the new ESPRESO versions. These partial models are trained on the set of calibration measurements and their quality is subsequently evaluated both by investigation of residuals and 10-fold cross-validation testing. The complete, final performance model is designed for the prediction of the optimal settings of the solvers in a sense of optimal domain decomposition and the optimal number of MPI processes and OpenMP threads. The user only provides the size of the problem to be solved and the chosen FETI method.

Our tests show that this final model can predict optimal settings with an error smaller than 1s for a solver runtime that can vary between 3 and 50 seconds based on problem decomposition and parallelization setup.

Considering those results, the final model enables new users to utilize ESPRESO library as efficiently as possible with no previous experience required.

purchase the full-text of this paper (price £22)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £25 +P&P)