Hardware Multicast Routing Techniques for Scalable Parallel Computers

多播通信是并行计算机的一个基本特征。它使多个并行应用程序受益，并对共享内存、分布式内存和分布式共享内存系统非常有用。多播还有助于支持一类称为集体操作的通信模式。多播可以由硬件或软件方法来支持。然而，硬件方法比软件方法快得多。大多数并行计算机供应商不是没有意识到这一点，就是试图实现无死锁的硬件多播方案，但没有成功。本课题主要研究用于网格和基于交换机的并行计算机的高效硬件多播路由技术。本项目将解决几个实施细节，例如启动延迟、路由器延迟和多播。研究的主要目标将包括通过减少网格中多播所需的启动步骤来最小化总的通信延迟。对于基于交换的系统，本项目将研究支持高效无死锁多播的替代交换机设计。最后，将努力设计支持组播操作的低成本路由器。在这项研究中，将观察到两个主要限制因素。首先，算法的成本和复杂性将降至最低。其次，PI将专注于不仅支持无死锁的组播操作，而且还增强节点到节点单播通信性能的技术。

MLulticast communication is an essential feature in parallel computers. It benefits several parallel applications and is useful for shared memory, distributed memory, and distributed-shared memory systems. Multicasting also helps in supporting a class of communication patterns known as collective operations. Multicasting can be supported by either hardware or software approaches. However, the hardware approach is significantly faster than the software approach. Most vendors of parallel computers are either unaware of it or have made unsuccessful attempts at implementing deadlock-free hardware multicasting schemes. This project is concerned with efficient hardware multicast routing techniques for meshes and switch-based parallel computers. Several implementation details, such as start-up latency, router delays, and multiple multicasts, will be addressed in this project. The primary objectives of the research will include the minimization of the total communication latency by reducing the number of start-up steps required for multicasting in meshes. For switched-based systems, this project will examine alternative switch designs to support efficient deadlock-free multicasting. Finally, an effort will be made to design low-cost routers that support multicast operation. Two primary constraints will be observed during this research. First, the cost and complexity of the algorithms will be minimized. Second, the PI will focus on techniques that not only support deadlock-free multicast operations but also enhance the performance of node-to-node unicast communications.