Development Of Advanced Computer Hardware And Software

随着个人计算机、工作站和服务器硬件价格的不断降低和性能的不断提高，计算化学研究人员能够开发和使用更强大的软件来研究各种理论问题，并对生化过程进行更丰富的模拟。我们所做的主要努力包括开发和改进适合于大分子模拟的并行计算技术，以及构建有效执行该技术所需的硬件。后一项任务包括评估新的硬件技术，以确定利用并行化代码的最具成本效益的方法。 目前的项目包括： -Lobos，使用现成PC硬件的高性能计算机器。 -开发并行QM/MM方法，包括复制/路径(线性缩放方法)。 -开发和支持CHARMM计算化学软件。 -通过一种新的空间分解算法来提高CHARMM的并行性能。 -电子显微镜应用的EMAP方法的扩展。 -实现了用于远程相互作用计算的IPS方法。 -开发用于膜系统模拟的2D IPS方法。 -改进了自制导朗之万动力学仿真算法。 Lobos(货架上的许多盒子)系统是Beowulf？风格的高性能计算系统，使用现成的硬件，通过运行Red Hat Linux操作系统的高速网络连接。自1997年以来，已使用商用PC设计和构建了六个Lobos计算机集群。从历史上看，与传统超级计算供应商的系统相比，这些系统的性价比提高了10倍以上。早期的LOBOS系统证明了这种方法的有效性，并在成功管理这类系统的技术开发中起到了至关重要的作用。 Lobos V在可靠性和性能方面都比以前的系统有了实质性的改进。在2004年12月，我们通过添加62个新节点完成了这个集群，这意味着研究人员现在不必花费那么多时间等待计算资源可用。这将导致更快的模拟吞吐量和更容易的软件开发(由于对作业的反馈更快)。从Lobos IV开始，群集使用了高速Myrnet互连，而在Lobos V中，我们越来越多地使用这些互连来实现我们现有软件的更高性能。此外，在过去一年中，我们用更可靠的线速管理型交换机升级了千兆以太网基础设施。这使得可以更有效地使用非Myrinet集群节点来运行并行作业，并减少了网络瓶颈造成的问题，其结果是提供了更好的研究计算环境。 群集还特别适合于超大型应用程序，因为每个计算节点都有自己的磁盘(最大限度地提高分布式I/O)，并且其工作负载可以配置为独立于其他节点。作为复制/路径QM/MM方法发展的一部分，我们创造了一种新的方法来并行化这些大规模的量子计算部分，使得？从而实现了高效的算法。我们已经开始将这项技术集成到CHARMM中，并在Lobos V上进行测试。

With the continuing price reductions and performance increase in personal computer, workstation, and server hardware, computational chemistry researchers are able to develop and use more powerful software to study various theoretical problems and to conduct richer simulations of biochemical processes. The primary efforts we have made include the development and refinement of parallel computing techniques suitable for macromolecular simulation and the construction of the hardware required to efficiently execute it. The latter task includes the evaluation of new hardware technology to ascertain the most cost effective methods of utilizing parallelized codes. Current projects include: - LoBoS, high performance computing machine using off the shelf PC hardware. - Development of parallel QM/MM methods including Replica/Path (linear scaling methods). - Development and support of the CHARMM computational chemistry software. - Increase the parallel performance of CHARMM via a new spatial-decomposition algorithm. - Expansion of the EMAP method for electron microscopy applications. - Implementation of IPS method for long range interaction calculation. - Development of 2D IPS method for membrane system simulation. - Improvement in Self-Guided Langevin Dynamics simulation algorithm. The LoBoS (Lots of Boxes on Shelves) system is an implementation of a ?Beowulf? style high performance computing system using off the shelf hardware connected via high speed networks running the Red Hat Linux operating system. Since 1997, six Lobos computer cluster have been designed and constructed using commodity PCs. Historically, these systems have provided a greater than 10-fold improvement in price/performance when compared with systems from traditional supercomputing vendors. The early LoBoS systems demonstrated the effectiveness of this approach and were essential in the development of techniques to successfully manage these kinds of systems. LoBoS V is a substantial improvement over previous systems in terms of both reliability and performance. In December, 2004, we completed this cluster with the addition of 62 new nodes, meaning that researchers now do not have to spend as much time waiting for computing resources to become available. This leads to faster simulation throughput and easier software development (due to quicker feedback on jobs). Starting with LoBoS IV, the clusters have used high-speed Myrinet interconnects and with LoBoS V we are increasingly using these to achieve higher performance with our existing software. Furthermore, over the past year we have upgraded the gigabit Ethernet infrastructure with a more reliable, wire speed managed switch. This has allowed for more efficient use of non-Myrinet cluster nodes to run parallel jobs and has reduced problems caused by network bottlenecks, the result of which is a better research computing environment. Clusters are also uniquely suited to very large applications since each compute node has its own disk (maximizing distributed I/O) and its workload may be configured to be independent of other nodes. As part of the development of the Replica/Path QM/MM method, we have created a novel way to parallelize these large scale quantum part of these calculations, making an ?embarrassingly parallel? and thus highly efficient algorithm. We have begun integrating this technique into CHARMM and testing it on LoBoS V.