SDCI: HPC: Improvement: Infrastructure for Multi-Node Manycore Computing

当今多核计算面临的主要挑战之一是使主流程序员能够访问并行性。在今天?在这个时代，几乎每台计算机都包含一个以GPU形式出现的多核处理器，但我们还没有成功地构建出让整个计算社区都能使用多核处理所需的原语和技术。应对这一挑战的一个解决方案是构建封装通用编程模式和习惯用法的库。库是独立的，因此很容易添加到现有项目中；可以轻松升级；并且非常适合像我们这样的学术团体进行开发和维护。这些库在CPU端广泛支持高性能计算，但在多核世界中却很少。我们的CUDPP （CUDA Data-Parallel Primitives）库广泛应用于GPU计算社区，尽管仅在单节点（主要是非hpc）系统上。我们相信下一代OpenCL编程环境是多核编程的未来，因此我们未来的工作目标是这个标准。我们的计划是扩展和支持CUDPP用于高性能计算社区，该社区越来越多地采用多核处理器作为核心计算引擎。我们将在CUDPP和基于opencl的CLDPP库中添加单节点和多节点原语，我们认为这些原语对多核高性能计算社区的发展至关重要。我们还计划为许多核心库定义最佳实践，不仅要构建一个库，还要构建帮助其他人构建库的工具。该方案的智力优势在于开发和改进了针对多核计算环境的核心数据结构和算法，特别是针对高性能计算环境。这个项目的其他有趣成果包括多节点数据结构和算法，以及构建多核心库的工具的软件工程。像它的前身CUDPP一样，这项工作产生的库将是开源的，并在多核编程社区中广泛使用。优化技术和工具也将得到广泛的应用。我们将与OpenCL联盟和我们的行业合作伙伴合作，将库纳入OpenCL标准，直接影响整个多核社区。PI还将继续通过多核计算引入本科生进行研究，包括与我们的工业合作伙伴合作，在这个项目上举办100名暑期代码工程师。

One of the major challenges facing manycore computing today is to make parallelism accessible to the mainstream programmer. In today?s era, where nearly every computer contains a manycore processor in the form of the GPU, we have not yet successfully built the primitives and techniques that we require to make manycore processing readily available to the entire computing community.One solution to this challenge is the construction of libraries that encapsulate common programming patterns and idioms. Libraries are self-contained and thus easily added to existing projects; can easily be upgraded; and are well-suited for development and maintenance by academic groups such as ours.High-performance computing is extensively supported by such libraries on the CPU side, but there are few in the manycore world. Our CUDPP (CUDA Data-Parallel Primitives) library is widely used in the GPU computing community, albeit only on single-node (largely non-HPC) systems. We believe the next-generation OpenCL programming environment is the future of manycore programming and thus target our future work toward this standard.Our plan is to extend and support CUDPP for use in the high-performance computing community, which is increasingly adopting manycore processors as core computational engines. We will add single- and many-node primitives to CUDPP and the OpenCL-based CLDPP library, primitives we consider to be vital to the growth of the manycore HPC community. We also plan to define best practices for manycore libraries and to build not just a library but also tools to help others build libraries.The intellectual merit of this proposal lies in the development and improvement of core data structures and algorithms targeted to manycore computing environments in general and HPC computing environments specifically. Other interesting outcomes from this project include multi-node data structures and algorithms and the software engineering of tools to build manycore libraries.Like its predecessor, CUDPP, the library that will result from this work will be open-sourced and widely used in the manycore programming community. The optimization techniques and tools will also find widespread use. We will work with the OpenCL consortium and our industry partners to bring libraries into the OpenCL standard, directly impacting the entire manycore community. The PI will also continue to introduce undergraduates to research through manycore computing, including collaborating with our industrial partners to host Google Summer of Code engineers on this project.