Resumen de Paralelizando aplicacions de genètica de poblacions
With the increasing availability of genome-scale data for genetic research, molecular population geneticists need to work with more complex models, which cannot be done in a time-fashion using the standard coalescent methods. This scenario led to the development of several alternative numerical simulation applications. Despite the ever-increasing access to High Performance Computing (HPC) clusters in the academy, it is not being leveraged in the field of population genetics. Parallelizing existing applications is hard to achieve by developers without a comprehensive understanding of the HPC, and new applications only take advantage of multiprocessing capabilities from a single computer.
This thesis proposes a technique to parallelize coalescent applications and effectively use all the available processing power from an HPC cluster. We use a strategy to reduce the intra- node communications in the message-passing paradigm. This solution allows for getting better scalability for coalescent applications that require generating millions of replicas. As a result, population geneticists can use the standard coalescent tools for running Approximate Bayesian Computation (ABC) analysis without relying on less accurate applications.
We have evaluated our strategy parallelizing the de facto standard coalescent application and run experiments at genome-scale in a real HPC cluster. We have obtained significant performance gains in tuning different aspects of our approach, leading to a 4x speedup over our initial parallelization, which accounted for a 50x speedup over the reference coalescent application.
