Algorithm-Hardware Co-Design for Performance-driven Embedded Genomics

Abstract

Genomics includes development of techniques for diagnosis, prognosis and therapy of over 6000 known genetic disorders. It is a major driver in the transformation of medicine from the reactive form to the personalized, predictive, preventive and participatory (P4) form. The availability of genome is an essential prerequisite to genomics and is obtained from the sequencing and analysis pipelines of the whole genome sequencing (WGS). The advent of second generation sequencing (SGS), significantly, reduced the sequencing costs leading to voluminous research in genomics. SGS technologies, however, generate massive volumes of data in the form of reads, which are fragmentations of the real genome. The performance requirements associated with mapping reads to the reference genome (RG), in order to reassemble the original genome, now, stands disproportionate to the available computational capabilities. Conventionally, the hardware resources used are made of homogeneous many-core architecture employing complex general-purpose CPU cores. Although these cores provide high-performance, a data-centric approach is required to identify alternate hardware systems more suitable for affordable and sustainable genome analysis. Most state-of-the-art genomic tools are performance oriented and do not address the crucial aspect of energy consumption. Although algorithmic innovations have reduced runtime on conventional hardware, the energy consumption has scaled poorly. The associated monetary and environmental costs have made it a major bottleneck to translational genomics. This thesis is concerned with the development and validation of read mappers for embedded genomics paradigm, aiming to provide a portable and energy-efficient hardware solution to the reassembly pipeline. It applies the algorithmhardware co-design approach to bridge the saturation point arrived in algorithmic innovations with emerging low-power/energy heterogeneous embedded platforms. Essential to embedded paradigm is the ability to use heterogeneous hardware resources. Graphical processing units (GPU) are, often, available in most modern devices alongside CPU but, conventionally, state-of-the-art read mappers are not tuned to use both together. The first part of the thesis develops a Cross-platfOrm Read mApper using opencL (CORAL) that can distribute workload on all available devices for high performance. OpenCL framework mitigates the need for designing separate kernels for CPU and GPU. It implements a verification-aware filtration algorithm for rapid pruning and identification of candidate locations for mapping reads to the RG. Mapping reads on embedded platforms decreases performance due to architectural differences such as limited on-chip/off-chip memory, smaller bandwidths and simpler cores. To mitigate performance degradation, in second part of the thesis, we propose a REad maPper for heterogeneoUs sysTEms (REPUTE) which uses an efficient dynamic programming (DP) based filtration methodology. Using algorithm-hardware co-design and kernel level optimizations to reduce its memory footprint, REPUTE demonstrated significant energy savings on HiKey970 embedded platform with acceptable performance. The third part of the thesis concentrates on mapping the whole genome on an embedded platform. We propose a Pyopencl based tooL for gEnomic workloaDs tarGeting Embedded platfoRms (PLEDGER) which includes two novel contributions. The first one proposes a novel preprocessing strategy to generate low-memory footprint (LMF) data structure to fit all human chromosomes at the cost of performance. Second contribution is LMF DP-based filtration method to work in conjunction with the proposed data structures. To mitigate performance degradation, the kernel employs several optimisations including extensive usage of bit-vector operations. Extensive experiments using real human reads were carried out with state-of-the-art read mappers on 5 different platforms for CORAL, REPUTE and PLEDGER. The results show that embedded genomics provides significant energy savings with similar performance compared to conventional CPU-based platforms.