Tuning the genomic evaluation system of Holstein-Friesian cattle

Abstract

The aim of this study was to compare the CPU time of two most widely used Fortran compilers – the Absoft and the Intel, in the context of two programs used in the genomic evaluation system of dairy cattle - a program for re-coding of the raw genotyping data of single nucleotide polymorphisms and a program for the estimation of additive genetic effects of single nucleotide polymorphisms. The data set used for the analysis corresponded to the Polish national routine genomic evaluation for non-return rate of heifers and comprised genotypes of 46,267 single nucleotide polymorphisms for each of 2904 evaluated bulls, which resulted in 130,393,459 non-missing genotype records. Programs were compiled under several different compilation options using the Absoft (release 2015) and the Intel (release 2019) compilers. The CPU times of both programmes significantly differed between compilers. In the program for genotype re-coding the Intel compiler always resulted in shorter running times than the Absoft compiler. More variation was observed in CPU times of the program for effect estimation, for which the Absoft compiler resulted in both, the shortest and the longest computing times. A more detailed run-time memory consumption visualisation of single runs of the fastest Absoft and Intel compiled program executions demonstrated that the former compiler allows for the utilisation of more of the available memory, which in the case of a memory intensive program for effect estimation resulted in a faster execution time. In conclusion, an overall performance of the Intel compiler was better than this of the Absoft compiler. Even though significant differences in CPU time were observed, the Intel compiled execution times were less dependent on the pre-imposed compilation parameters. However, it is important to note that for programs performing a large number of arithmetic evaluations it is possible to tune the Absoft compiler to outperform the Intel compiler. The key compilation parameters are those, which allow for lower floating point accuracy and dedicated optimization of mathematical operations. Although blind use of such options is not recommended, in the case of genomic evaluation the math and floating point based optimisations did not affect the convergence of the solver nor the estimates of single nucleotide polymorphisms’ effects.