Parallel Computing to Speed up Whole-Genome Bayesian Regression Analyses

Bayesian multiple regression methods, which are widely used for whole-genome analyses, are computationally intensive. However, the computing time for these methods is linear in the amount of data. Thus, parallel programming techniques can be used to speed up computing time.

In a widely used Markov chain Monte-Carlo (MCMC) strategy, a single-locus mixed model equation is solved to sample the effect for each locus. Here, the most time consuming calculation is a dot product of a vector of length n, the number of observations. This dot product is split up and done in parallel on multiple cores; results from each core are combined to get the dot product. This same strategy is also applied to other calculations of order n. When n was a million, with 120 cores a speedup of about 58 times was achieved.

When n is any larger than p, the number of markers, sampling marker effects is more efficient by setting up the full mixed model equation for all loci. Once the full mixed model equation is formed, the dot product required to sample a marker effect is of order p rather than n. However, setting up the full mixed model equation involves dot products of order n. Thus, these dot products for setting up the full mixed model equations are done in parallel. When n was a million, with 2400 cores a speedup of about 1279 times was achieved.