The impending advent of population-scaled sequencing cohorts involving tens of millions of individuals with matched phenotypic measurements will produce unprecedented volumes of genetic data. Storing and analysing such gargantuan datasets places computational performance at a pivotal position in medical genomics. In this thesis, I explore the potential for accelerating and parallelizing standard genetics workflows, file formats, and algorithms using both hardware-accelerated vectorization, parallel and distributed algorithms, and heterogeneous computing.

First, I describe a novel bit-counting operation termed the positional population-count, which can be used together with succinct representations and standard efficient operations to accelerate many genetic calculations. In order to enable the use of this new operator and the canonical population count on any target machine I developed a unified low-level library using CPU dispatching to select the optimal method contingent on the available instruction set architecture and the given input size at run-time. As a proof-of-principle application, I apply the positional population-count operator to computing quality control-related summary statistics for terabyte-scaled sequencing readsets with >3,800-fold speed improvements. As another application, I describe a framework for efficiently computing the cardinality of set intersection using these operators and applied this framework to efficiently compute genome-wide linkage-disequilibrium in datasets with up to 67 million samples resulting in up to >60-fold improvements in speed for dense genotypic vectors and up to >250,000-fold savings in memory and >100,000-fold improvement in speed for sparse genotypic vectors. I next describe a framework for handling the terabytes of compressed output data and describe graphical routines for visualizing long-range linkage-disequilibrium blocks as seen over many human centromeres. Finally, I describe efficient algorithms for storing and querying very large genetic datasets and specialized algorithms for the genotype component of such datasets with >10,000-fold savings in memory compared to the current interchange format.