Development of a low-cost 64 bit cluster supercomputer for engineering and physics simulations

摘要建议编号：CTS-0321170首席研究员：犹他州州立大学J.T.豪泽这笔赠款专注于结合两项前沿技术，即双处理器AMDOpteron系统和基于千兆以太网的平面邻居网络(FNNS)，以开发设计独特的集群超级计算机。除了提供位内存地址空间外，皓龙还采用了一种新的内存架构，可避免通常与共享内存多处理器相关的内存性能下降。同样，使用FNN互连模式可以实现单交换机延迟，尽管其节点比低成本网络交换机的端口多。FNN还提供了针对运行在所提出的集群超级计算机上的应用程序的通信模式来优化网络特性的可能性。该项目将提供一种替代市场上可获得的高性能集群网络解决方案的低成本替代方案，如Myrinet和Quadrics。设计和优化这种网络的软件将免费提供，以便其他研究小组可以设计和建立自己的FNN集群。与其他现有的存储系统设计相比，建议的系统还将为新设计的大型应用存储系统的性能提供有价值的见解。将进行的研究包括计算流体动力学、生态水力学、聚变等离子体、地球物理流体动力学和计算化学。为了满足对高性能计算技能学生日益增长的需求，工程学院将为高级本科生和初级研究生开设一门面向工程应用的并行计算课程。这门课程以及为期一周的集群计算研讨会也将向数学、物理、化学和计算机科学的学生开放。此外，拟议的集群超级计算机将加强与附近国家实验室(如爱达荷州国家工程和环境实验室(INEEL))的现有关系。密歇根州立大学的教师和研究生正在与INEEL工程师合作开发计算密集型CFD模型。这一合作努力带来的知名度将进一步提高吸引更多学生和教师的目标。

ABSTRACTProposal No. CTS-0321170Principal Investigator: J. T. Hauser, Utah State UniversityThis grant focuses on combining two leading-edge technologies, dual-processor AMDOpteron systems and Flat Neighborhood Networks (FNNs) based on Gigabit Ethernet, to develop a uniquely designed cluster supercomputer. In addition to providing a 64-bit memory address space, the Opteron uses a new memory architecture that should avoid the memory performance degradation normally associated with shared-memory multiprocessors. Similarly, using a FNN interconnection pattern makes it possible to achieve single-switch latency despite having more nodes than a low-cost network switch has ports. FNNs also provide the possibility to optimize the networking characteristics toward the communication patterns of applications running on the proposed cluster supercomputer. This project will provide a low-cost alternative to commercially available high performance cluster networking solutions such as Myrinet and Quadrics. The software to design and optimize such a network will be made available for free, so that other research groups can design and build their own cluster with FNNs. The proposed system will also provide valuable insight in the performance of a newly designed memory system for large applications compared to other available memory system designs.The research to be conducted includes computational fluid dynamics, ecohydraulics, fusion plasmas, geophysical fluid dynamics and computational chemistry. To meet the rising demand for students skilled in high-performance computing, the College of Engineering will introduce a course in parallel computing for engineering applications targeted at senior level undergraduate and beginning graduate students. This course as well as a one-week workshop on cluster computing will also be open to students in mathematics, physics, chemistry and computer science.In addition, the proposed cluster supercomputer will enhance current relationships with nearby national laboratories such as the Idaho National Engineering and Environmental Laboratory (INEEL). USU faculty and graduate students are developing very computationally intensive CFD models in conjunction with INEEL engineers. The visibility arising from this cooperative effort will further enhance the goal of attracting more students and faculty.