COMPUTAÇÃO PARALELA APLICADA À SELEÇÃO GENÔMICA VIA INFERÊNCIA BAYESIANA

Abstract

In genomic selection (GS), the data analysis using large number of genetic markers based on high computational demand from Bayesian models via Markov Chain Monte Carlo algorithms requires weeks or months to be finished. The parallel computing is a natural solution to this problem, since it splits an algorithm in several independent tasks that are simultaneously (in parallel) processed. It reduces the required computational time when compared with the traditional data processing approach. To demonstrate the importance of parallel computing in GS, its efficiency was compared with the standard sequential algorithm (traditional) by using the BayesCπ method. Two parallelization strategies were studied. The first one involved the analysis of multiple parallel MCMC chains, and the second one referred to the parallelization of the chain itself. The MPI library and OpenMPI package from the gfortran compiler were used for the parallel execution of these algorithms. Simulated data considering 10,000 markers and 4,100 individuals were used. The sequential algorithm was processed at 77.29 hours. The parallel multiple chains were 80% more efficient, while the second parallelization strategy presented an efficiency of 19%. In summary, the parallel computing was efficient and can be applied to GS.