Energy-aware acceleration on GPUs: Findings on a bioinformatics benchmark

• Published in: Sustainable computing informatics and systems 2018-12, Vol.20, p.88-101
• Main Authors: Pérez, J., Rodríguez, A., Chico, J.F., López-Rodríguez, D., Ujaldón, M.
• Format: Article
• Language: English
• Subjects: FDT; GPU; HPC; NiftyReg; Power-aware; Q-norm; Speed-up
• ISSN: 2210-5379
• DOI: 10.1016/j.suscom.2018.01.001

•Power usually goes up whenever time shortens, but time reduction uses to be wider than the average increase.•Higher speed-up factors correspond to newer technology, where transistor gate shrinks, thus providing us more power-efficient hardware over time.•GPUs do not abuse of frequency in latest generations. They try to keep it relaxed, thus influencing power consumption in a very positive way.•GPUs may increase throughput and save energy at the same time, as our work has extensively proven. This paper performs a complete study on performance and energy efficiency of biomedical codes when accelerated on GPUs (Graphics Processing Units). We have selected a benchmark composed of three different building blocks which constitute the pillars of four popular biomedical applications: Q-norm, for the quantile normalization of gene expressions, reg_f3d, for the registration of 3D images within the NiftyReg library, bedpostx (from the FSL neuroimaging package) and a multi-tensor tractography for the analysis of diffusion images. We try to identify (1) potential scenarios where performance per watt can be optimal in large-scale biomedical applications, and (2) the ideal GPU platform among a wide range of models, including low power Tegras, popular GeForces and high-end Titans. Experimental results conclude that data locality and arithmetic intensity represent the most rewarding ways on the road to high performance bioinformatics when power is a major concern.