High-performance global optimization and signal processing on multicore, GPU, and hybrid architectures

计算现在被认为是当代科学的三大支柱之一，与理论和实验并列。拟议的研究将通过开发图形处理单元(GPU)、多核计算机系统和GPU/多核混合体系结构的潜力来解决全球优化和信号处理相关领域的复杂问题，从而促进跨学科计算科学。许多不同领域的研究人员正在探索GPU，它最初是为了加快计算机图形的简单算术运算而开发的，用于通用计算。这项研究将集中于识别和利用数据和任务级别的并行性，以解决需要处理大量数据的优化和信号处理问题。优化将应用于多维尺度，这是信息可视化和地理空间问题分析的重要组成部分，以及粒子群技术，将用于校准大型数学模型，如空间计量经济学中常见的模型。全局优化也将用于盲源分离，这对噪声抵消助听器的发展具有重要意义。此外，新架构上的信号处理可以通过有效地分析来自传感器网络和气象站的流数据来促进地理空间研究。这些计算领域具有重要的社会和经济意义，因为改进的高维全局优化方法将促进解决重要科学和工程问题所需的高保真模拟和模型。利用新架构的信号处理算法可以更有效地解决生物医学和实时应用中的问题。这项研究的目标是调查和分析多核/GPU算法的能力和局限性，这些算法可以整合到模拟和建模、生物医学信号处理和地理空间数据分析中的应用中。由于大多数台式机和笔记本电脑现在都可以使用多核芯片和GPU，因此开发这些功能变得越来越重要。

Computation is now considered as one of the three pillars of contemporary science, along with theory and experimentation. The proposed research will contribute to interdisciplinary computational science by exploiting the potential of graphics processing units (GPUs), multicore computer systems, and hybrid GPU/multicore architectures to solve complex problems in the related areas of global optimization and signal processing. Many investigators in diverse fields are exploring GPUs, originally developed to speed simple arithmetic operations for computer graphics, for general purpose computation. This research will concentrate on identifying and exploiting data- and task-level parallelism for optimization and signal processing problems that require the processing of large amounts of data. Optimization will be applied to multidimensional scaling, an important component in information visualization and analysis for geospatial problems, and to particle swarm techniques, which will be used to calibrate large mathematical models, such as those that are common in spatial econometrics. Global optimization will also be used in blind source separation, which has implications for the development of noise-cancelling hearing aids. Furthermore, signal processing on new architectures can facilitate geospatial research by enabling efficient analysis of streaming data from sensor networks and weather stations. These computational areas have great societal and economic importance, as improved high-dimensional global optimization approaches will facilitate high-fidelity simulations and models needed to solve important scientific and engineering problems. Signal processing algorithms utilizing the new architectures allow more efficient solutions to problems in biomedicine and real-time applications. The goal of the proposed research is to investigate and analyze the capabilities and limitations of multicore/GPU algorithms that can be incorporated into applications in simulation and modeling, biomedical signal processing, and in geospatial data analysis. As most desktop and laptop computers are now available with both multicore chips and GPUs, exploitation of these capabilities is increasingly relevant.