MRI-R2 : Acquisition of a Heterogenous Terascale Shared Campus Computing Facility

该奖项是根据2009年《美国复苏和再投资法案》(公法111-5)资助的。提案#：09-59382PI(S)：费舍尔，罗伯特·T；考尔斯，杰弗里；W.，戈特利布；西格尔，卡纳，高拉夫；王，成研究所：马萨诸塞大学-达特茅斯北达特茅斯分校，MA 02727-2300标题：磁共振成像-R2：收购异质Terascale共享校园计算设施项目建议：该项目，收购包含图形处理单元(GPU)的异质万亿级并行计算机集群，服务于跨学科和多学科的数学家和计算科学家和工程师群体，以及他们的本科生和研究生。这一共享的校园研究仪器利用了马萨诸塞州大学达特茅斯分校现有的没有GPU的小规模集群，产生了一个大小超过四倍的综合系统。具体地说，建议的集群包含32个8路节点(256个核心)，外加每个节点一个NVIDIA Tesla GPU，总共32个GPU。该集群的目的不是作为一般的计算资源，而是提议中的研究人员小组开发用于科学计算的高性能算法所必需的高度专业化的工具。研究人员将利用每个节点中整合的GPU来开发基于混合并行和多核架构的计算。研究人员将把这些GPU加速的解算器抽象成一个开源的GPU科学库(GPUSL)，它将通过一组定义良好的接口与上面的模拟代码完全分开。来自达特茅斯大学物理、数学、海洋科学和土木工程系的教职员工组成了一个科学计算小组，该小组将成为共享计算资源的核心管理和用户小组。该小组的研究重点从巨型分子云的演化(使用FLASH程序)到海岸海洋模拟(使用FVCOM程序)，再到不可压缩湍流对流流体的模拟。该小组的活动包括研究合作、每周一次的团队会议，以分享数值算法方面的专业知识，为学生提供共同建议，以及开发一个多学科的博士项目。该集群将安装在位于马萨诸塞州达特茅斯大学的校园信息技术服务(CITS)数据中心，并由CITS企业系统管理团队直接支持。该团队将负责用户磁盘空间的预防性维护、会话备份和恢复。机器的使用和分配将由一个研究用户组委员会处理，该委员会将负责向寻求计算机时间的教职员工征求建议书，并负责平衡用户群中教职员工、研究生和本科生的研究需求。广泛影响：该集群的可用性允许培训下一代高性能计算专家，并支持目前正在开发的一个新的跨学科计算科学博士项目。研究生受益于在多学科环境中为各种具有挑战性的物理问题开发尖端并行和GPU加速算法的机会。作为NSF资助的本科生计算科学(CSUMS)项目的一部分，该集群还直接使本科生教育受益，该项目旨在帮助本科生为计算数学的持续研究体验做准备，并让他们参与其中。这种方法是招收有前途的高三学生，然后指导一年级和二年级的学生学习计算科学的研究课题，最后通过精心挑选的研究项目指导高年级学生。此外，马萨诸塞州达特茅斯大学有一个成功的传统，通过其大学现在迈向抽象推理和思维(START)计划，将高等教育机会扩大到代表不足的学生。在其40年的历史中，该计划在将STEM学生的机会扩大到妇女和代表不足的少数族裔方面发挥了重要作用；特别是最近一个学期的Start学生中，约有25%-30%是代表不足的少数族裔。对于科学界来说，在这一使能集群上开发的所有计算机代码，包括GPU软件库，都将通过一个专门的网站发布，立即影响到全世界使用社区代码Flash和FVCOM的2500多名研究人员。

"This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5)." Proposal #: 09-59382PI(s): Fisher, Robert T.; Cowles, Geoffrey; W., Gottlieb; Sigal, Khanna, Gaurav; Wang, ChengInstitution: University of Massachusetts-Dartmouth North Dartmouth, MA 02727-2300Title: MRI-R2: Acquisition of a Heterogeneous Terascale Shared Campus Computing FacilityProject Proposed :This project, acquiring a heterogeneous terascale parallel computer cluster incorporating graphics processing units (GPUs), services an inter- and multi-disciplinary group of mathematicians and computational scientists and engineers, and their undergraduate and graduate students. This shared campus research instrument leverages an existing small-scale cluster without GPUs at UMass-Dartmouth resulting in a combined system of more than four times the size. Specifically, the proposed cluster contains 32 8-way nodes (256 cores) plus one NVIDIA Tesla GPU per node for a total of 32 GPUs. The cluster is not intended to be a general computing resource, but rather highly specialized instrument necessary for the development of high-performance algorithms for scientific computation by the group of researchers in the proposal. Researchers will take advantage of the incorporation of the GPUs in each node to develop hybrid parallel and many-core-architecture-based computing. The investigators will abstract out these GPU-accelerated solvers into an open source GPU Scientific Library (GPUSL), which will be cleanly separated from the overlying simulation codes through a set of well-defined interfaces. Faculty members from the Physics, Mathematics, Marine Science and Civil Engineering departments at UMass-Dartmouth have formed a scientific computing group that will become the core management and user group of the shared computational resources. The research foci of the group range from Evolution of Giant Molecular Clouds (using the FLASH code) to Coastal Ocean Modeling (using the FVCOM code) to the Simulation of an Incompressible Turbulent Convective Fluid. Activities of this group include research collaborations, weekly team meetings for the sharing of expertise on numerical algorithms, co-advising of students, and a development of a multi-disciplinary doctoral program. The cluster will be housed in the Campus Information Technology Services (CITS) data center at UMass-Dartmouth, and directly supported by the CITS Enterprise Systems Administration Team. This team will be responsible for preventive maintenance session backup and recovery of user disk space. The access to the machine and its allocation will be handled by a Research User Group Committee which will be responsible for soliciting proposals from faculty members seeking computer time and will also be responsible for balancing the research needs of the faculty, graduate, and undergraduate students in the user base.Broader Impacts: The availability of this cluster allows for the training of the next generation of high performance computing experts, and supports a new interdisciplinary Ph.D. program in computational science, currently in development. Graduate students benefit from the opportunity to develop cutting-edge parallel and GPU-accelerated algorithms for a variety of challenging physical problems, in a multi-disciplinary setting. The cluster also directly benefits undergraduate education, as part of the NSF-funded Computational Science for Undergraduate Mathematics Students (CSUMS) project, which aims to prepare undergraduate students for, and engage them in, a sustained research experience in computational mathematics. The approach is to recruit promising high school seniors, then guide freshman and sophomores students through research topics in computational science, and culminate by mentoring seniors through carefully selected research projects.In addition, UMass-Dartmouth has a successful tradition of broadening higher-education access to underrepresented students through its College Now Steps Toward Abstract Reasoning and Thinking (START) program. Over its 40-year history, the program has played a significant role in broadening STEM access to disadvantaged students to women and underrepresented minorities; in particular, approximately 25% - 30% of START students in recent semesters are underrepresented minority persons of color. For the scientific community, all computer codes developed on this enabling cluster, including the GPU Software Library, will be released via a dedicated website, immediately impacting more than 2500 researchers worldwide who use the community codes FLASH and FVCOM.