CSR: Small: RUI: Task Mapping for Emerging Network Topologies

该项目旨在通过开发任务映射的新策略来提高下一代高性能计算（HPC）系统上大型科学模拟的性能，任务映射是将应用程序的特定部分分配给HPC系统中的每个许多处理节点。HPC系统上应用程序性能的一个关键决定因素是应用程序相关部分之间消息传递的速度。这又取决于连接处理节点的网络。任务映射有可能提高网络性能，方法是通过一种更均匀地在网络中分配消息的方式来安排应用程序的通信部分，从而防止应用程序的任何部分过载。PI先前已经表明，任务映射可以将当前HPC系统上的应用程序运行时间减少多达30%。更大的下一代系统需要新的算法，由于功率限制，必须使用新颖的网络拓扑进行节点间连接。所研究的具体网络拓扑是Dragonfly，其两种变体在商业系统中使用。蜻蜓将网络交换机组织成组，形成一个高基数的“虚拟交换机”。这允许在每对组之间建立直接连接。再加上组内交换机之间的高连通性，这保证了每一对节点都是通过短路径连接的。问题在于，如果每对组之间只有一个直接连接，那么这种连接就会成为潜在的瓶颈。该项目将开发任务映射算法，以平衡定位相关任务的目标，以利用组内连接，并需要在整个系统中分散任务，以便它可以同时利用许多组间链接。该项目还将开发节点分配算法，以支持不同大小的作业，为每个作业提供连接良好的节点，同时最大限度地减少作业之间的争用。通过提高下一代高性能计算系统的应用性能，该项目将有助于实现这些强大系统的全部潜力。此外，该项目将大量涉及本科生研究人员，他们将被培养成未来科学和工程领域的领导者。

This project aims to improve the performance of large scientific simulations on next-generation high-performance computing (HPC) systems by developing new strategies for task mapping, the assignment of specific parts of an application to each of the many processing nodes in an HPC system. A key determinant of application performance on HPC systems is the speed of message delivery between related parts of the application. This in turn depends upon the network connecting the processing nodes. Task mapping has the potential to improve network performance by arranging communicating parts of the application in a way that distributes messages more evenly through the network, preventing any part of it from becoming overloaded. The PI has previously shown that task mapping can reduce application running time on current HPC systems by up to 30%. New algorithms are needed for larger next-generation systems, which must use novel network topologies for the internode connections due to power limitations.The specific network topology studied is Dragonfly, two variations of which are used in commercial systems. Dragonfly organizes network switches into groups that form a high-radix "virtual switch". This allows a direct connection between every pair of groups. Together with high connectivity between the switches within a group, this guarantees that every pair of nodes is connected by a short path. The problem is that having only a single direct connection between each pair of groups makes that connection a potential bottleneck. The project will develop task mapping algorithms that balance the goal of localizing related tasks to exploit intra-group connectivity with the need to spread the job across the system so that it can utilize many inter-group links simultaneously. The project will also develop node allocation algorithms to support jobs of varying size, providing each job with nodes that are well-connected while minimizing contention between jobs.By improving application performance on next-generation HPC systems, the project will help realize the full potential of these powerful systems. In addition, the project will heavily involve undergraduate student researchers, who will be trained as future leaders in science and engineering.