Improving Bioinformatics Analysis of Large Sequence Datasets Parallelizing Tools for Population Genomics

摘要

Next-generation sequencing (NGS) technologies initiated a revolution in genomics, producing massive amounts of biological data and the consequent need for adapting current computing infrastructures. Multiple alignment of genomes, analysis of variants or phylogenetic tree construction, with quadratic polynomial complexity in the best case are tools that can take days or weeks to complete in conventional computers. Most of these analysis, involving several tools integrated in workflows, present the possibility of dividing the computational load in independent tasks allowing parallel execution. Determining adequate load balancing, data partitioning, granularity and I/O tuning are key factors for achieving suitable speedups. In this paper we present a coarse-grain parallelization of GH caller (Genotype/Haplotype caller), a tool used in population genomics workflows that performs a probabilistic identification process to account for the frequency of variants present between population individuals. It implements a master-worker model, using the standard Message Passing Interface (MPI), and concurrently and iteratively distributes the data among the available worker processes by mapping subsets of data and leaving the orchestration to the master process. Our results show a performance gain factor of 260x using 64 processes and additional optimizations with regard to the initial non-parallelized version.