Institutional Repository
Technical University of Crete
EN  |  EL



My Space

Design and implementation of a multi-FPGA acceleration system for large-scale population genomics analyses based on linkage disequilibrium

Bozikas Dimitrios

Full record

Year 2017
Type of Item Diploma Work
Bibliographic Citation Dimitrios Bozikas, "Design and implementation of a multi-FPGA acceleration system for large-scale population genomics analyses based on linkage disequilibrium", Diploma Work, Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών, Πολυτεχνείο Κρήτης, Chania, Greece, 2017
Appears in Collections


Modern sequencing technologies have contributed to the creation and rapid expansion of DNA databases, already numbering thousands of whole genomes. The astonishing rate at which genomic data are being collected, combined with the fact that it has outpaced Moore’s law, establishes the necessity of the development of novel tools, capable of conducting large-scale genomics analyses efficiently. This work addresses the computational challenges inherent to the analysis of linkage disequilibrium (LD) levels in large-scale datasets. LD is a statistic that quantifies the non-random association between alleles at different genomic locations. While it contributes to a wide variety of genomics and genetics analyses, the compute- and memory-intensive operation of counting set bits (population count) in large vectors, required for the estimation of LD, hinders the efficient use of modern CPUs for suchanalyses, mainly due to the lack of a vectorized population counter. To overcome this obstacle, we present a novel hardware architecture for the calculation of pairwise LD scores based on reconfigurable logic. The proposed accelerator exploits the ability of reconfigurable machines to be programmed at the hardware level. The effective use of multiple levels of parallelism, combined with the efficient manipulation of the data structures on memory through the transformation of the memory layout, result in high throughput capabilities for the estimation of LD in arbitrarily large datasets. The architecture is, subsequently, mapped onto a high-performance heterogeneous computing platform that enables the parallel cooperation of 4 reconfigurable devices, while, simultaneously, providing a high-speed memory interface. The implemented accelerator is evaluated for analyses of simulated genomic data of varying sizes, through its comparison with corresponding state-of-the-art parallel software implementations, achieving speedups between 6.35X and 134.93X, depending on the dataset size. Concerning real-world analyses, such as scanning the 22nd chromosome of the human genome, the accelerator is capable of potentially achieving quintupled throughput when compared to highly optimized reference software running on multiple cores.

Available Files