Collective Communication in Wormhole-Routed Massively Parallel Computers

超级计算行业现在由并行架构主导。在许多这样的系统中，为了实现更好的可伸缩性，内存在物理上分布在处理器之间，这要求节点通过网络进行通信。为了减少通信延迟，许多系统都采用虫洞路由交换策略，在这种策略中，消息通过单一目的地或集合管道传递。通信操作可以是点对点的，它涉及单个源和单个目的地，或者是集体的，其中有两个以上的进程参与。集体通信操作在科学计算中特别重要，其中大型数据阵列分布在节点的本地存储器上，但通过单个操作进行操作。本研究计划探讨虫洞路由网路的集体通讯运作设计。该研究旨在探索下一代虫洞路由网络通信子系统设计中的架构和算法替代方案。该项目大致可分为四个分区。首先，这两种模型都将被用于开发新的算法，用于集体行动，除了那些已经在研究中的算法。目标平台将包括直接网络和基于虫洞路由交换机的互连。其次，生成的算法将用于支持更高级别的并行编程结构，如虚拟进程拓扑、并行I/O操作和分布式共享内存。第三，由此产生的通信服务将用于提高并行应用程序的性能，特别是并行数值算法。最后，新模型的使用将在虫洞路由系统中进行研究，这些系统服务于并行计算以外的目的，例如高速局域网。使用的方法将是对商业虫洞路由系统的分析、模拟和实验研究。本研究的意义在于为并行应用的通信需求提供了新的解决方案。除了更好地理解在虫洞路由系统中设计和使用集体操作时必须解决的问题和问题外，该项目还将以新协议、新算法的原型实现和新应用软件的形式产生切实的结果，所有这些都将公开供其他研究人员使用。

The supercomputing industry is now dominated by parallel architectures. In many such systems, memory is physically distributed among processors in order to achieve better scalability, requiring nodes to communicate through a network. In order to reduce communication latency, many systems have the wormhole-routing switching strategy, in which messages are pipelined through and a single destination, or collective. Communication operations may be either point-to-point, which involves a single source and a single destination, or collective, in which more than two processes participate. Collective communication operations are particularly important in scientific computing, where large data arrays are distributed over the local memories of the nodes, but are manipulated through single operations. This research project addresses the design of collective communication operations for wormhole-routed networks. The research is intended to explore architectural and algorithmic alternatives in the design of communications subsystems for next generation wormhole-routed networks. The project can be divided approximately into four subareas. First, both models will be used to develop new algorithms for collective operations in addition to those already under investigation. Target platforms will include both direct networks, as well as wormhole-routed switch-based interconnects. Second, the resulting algorithms will be applied in order to support higher-level parallel programming constructs, such as virtual process topologies, parallel I/O operations and distributed shared memory. Third, the resulting communication services will be used to improve the performance of parallel applications, in particular parallel numerical algorithms. Finally, the use of the new models will be investigated in wormhole-routed systems that serve purposes other than parallel computing, such as high speed local area networks. The methods used will be analysis, simulation, and experimental studies on commercial wormhole-routed systems. The significance of this research is that it will provide new solutions to the communication needs of parallel applications. Besides providing a better understanding of the issues and problems that must be addressed in design and use of collective operations in wormhole-routing systems, this project will produce tangible results in the form of new protocols, prototype implementations of new algorithms, and new applications software, all of which will be made publicly available for use by other researchers.