Genomics GPUs and next generation computational statistics

描述（由申请人提供）：随着遗传学的计算需求呈指数级增长，人们越来越担心传统 CPU 能否提供所需的计算能力。并行计算已经被吹捧了好几年了，但大规模并行 CPU 计算机非常昂贵，而且仅限于几个国家中心。图形处理单元 (GPU) 提供了一种更便宜、更分布式的解决方案。数百个这样的单元被制造在一张卡上，并且几张卡可以安装在台式计算机中。因此，目前存在的廉价硬件有望使许多基本算法加速一百倍。 GPU 供应商的预测表明，这些设备的计算能力和多功能性将在未来十年快速增长。因此，软件开发是阻碍 GPU 开发的主要障碍。该提案针对的是现代计算链中的这个薄弱环节。通过一系列的示范项目和低级软件库的制作，我们希望能够催化GPU在遗传学领域的普及。具体项目包括：1）eQTL作图，2）QTL作图的方差分量模型，3）基因型和单倍型构建，4）种族混合的估计，5）通过RNA-Seq技术发现亚型，6）遗传景观和谱系的计算，7）从随机多重图构建基因网络，以及8）设计用于数据挖掘的新并行算法。高维优化是支持所有这些应用程序的共同主线。我们之前的优化研究已经证明了四个基本思想的有效性，即惩罚估计、坐标下降、MM（最大化-最小化）原理和参数分离。这些想法也推动了并行计算。在 GPU 上实施我们的演示项目将需要生成在计算统计中具有相当普遍价值的子例程。我们打算向开源社区发布我们的工具箱库，包括 C/C、Fortran 和 R 软件包装器。这可能会产生乘数效应，改善健康和物理科学领域许多学科的计算环境。根据该提案产生的所有其他应用程序将免费分发给科学界。我们生产和分发具有优质文档的可用软件的记录表明了我们对这一理念的承诺。

DESCRIPTION (provided by applicant): With computational demands in genetics growing exponentially, concerns are rising whether traditional CPUs can deliver the needed computing power. Parallel computing has been touted for several years, but massively parallel CPU computers are enormously expensive and limited to a few national centers. Graphics processing units (GPUs) offer a far cheaper and more distributed solution. Hundreds of these units are fabricated on a single card, and several cards fit inside a desktop computer. Thus, cheap hardware currently exists that promises a hundred-fold speedup of many basic algorithms. Projections from the vendors of GPUs suggest that these devices will grow rapidly in computational power and versatility over the next decade. Thus, software development is the main hurdle hindering the exploitation of GPUs. This proposal targets this weak link in the chain of modern computing. Through a series of demonstration projects and the production of low-level software libraries, we hope to catalyze the spread of GPUs in genetics. The specific projects include: 1) eQTL mapping, 2) variance component models for QTL mapping, 3) genotype and haplotype construction, 4) estimation of ethnic admixture, 5) isoform discovery through RNA-Seq technology, 6) computation of genetic landscapes and clines, 7) construction of gene networks from random multigraphs, and 8) design of new parallel algorithms for data mining. High-dimensional optimization is a common thread enabling all of these applications. Our previous research on optimization has demonstrated the efficacy of four fundamental ideas, namely, penalized estimation, coordinate descent, the MM (majorization-minimization) principle, and separation of parameters. These ideas also propel parallel computing. Implementation of our demonstration projects on GPUs will require the production of subroutines of considerable general value in computational statistics. We intend to release our toolbox libraries to the open source community, including C/C++, Fortran, and R software wrappers. This may lead to a multiplier effect that will improve the computing climate in many disciplines throughout the health and physical sciences. All other application programs produced under this proposal will be freely distributed to the scientific community. Our record of producing and distributing usable software with superior documentation shows our commitment to this philosophy.