Acceleration of computations in AI REML for single-step GBLUP models

Abstract Text: The objective of this study was to evaluate the advantage of the YAMS package over the FSPAK package in average-information (AI) REML for single-step GBLUP models. Data sets from broiler and Holsteins were used in this study. (Co)variance components were estimated with the AIREMLF90 program which could switch YAMS and FSPAK for sparse operations. The YAMS package used the BLAS and LAPACK libraries using all the 16 cores on CPU. For a single-trait model applied to the data contained over 15,000 genotyped animals, FSPAK took over 4 hours to finish the first 5 rounds while YAMS took 20 minutes. For a 4-trait model applied to the same data set, FSPAK failed in the sparse factorization while YAMS took 5 hours to finish the first 5 rounds. The use of YAMS can dramatically increase speed and stability of AIREMLF90 for single-step GBLUP models.

Keywords: single-step GBLUP, supernodal methods, variance-component estimation