Using experimental data to improve the performance modelling of parallel linear algebra routines

Abstract

The performance of parallel linear algebra routines can be improved automatically using different methods. Our technique is based on the modellisation of the execution time of each routine, using information generated by routines from lower levels. However, sometimes the information generated at one level is not accurate enough to be used satisfactorily at higher levels. Therefore, a remodelling of the routines is performed by using (applied appropriately) polynomial regression. A remodelling phase is proposed, and analysed with a parallel matrix multiplication. © 2008 Springer-Verlag Berlin Heidelberg.