Time Division Multiplexing of Optical Interconnection Networks

本研究的目的是研究、设计和评估多处理器应用中时分复用光互连网络的控制机制。 这样的机制将相对慢的网络控制与高光学数据传输带宽解耦。这项研究将评估时分复用的优点、能力和局限性。 具体而言，模型将开发指定的应用程序的通信要求，表达和测量通信负载中存在的通信局部性，并估计在多路复用互连网络的通信成本。 不同的控制协议将被模拟和测试从实际的并行应用程序收集的痕迹，并在合成的痕迹与通信局部性模型的帮助下产生的。集中式，分布式和流水线控制策略将实施和比较。 这个项目的动机是几个观察。首先，在没有复用的情况下，在网络中建立端到端连接的控制开销可能是禁止的。其次，应用程序表现出通信的局部性，这放大了多路复用的好处。第三，光路的高带宽允许多路复用，而对各个连接上的传输速率的影响最小。 最后，大规模并行程序的本质允许编译器使用当前的技术来确定满足大多数通信要求的多路复用序列。

The goal of this research is to study, design and evaluate control mechanisms for time-multiplexing optical interconnection networks in multiprocessor applications. Such mechanisms decouple the relatively slow network control from the high optical data transmission bandwidth. The research will assess the advantages, capabilities and limitations of time-division multiplexing. Specifically, models will be developed to specify the communication requirements of applications, to express and measure the communication locality present in communication loads, and to estimate the communication cost in multiplexed interconnection networks. Different control protocols will be simulated and tested on traces collected from actual parallel applications, and on synthetic traces generated with the help of the communication locality model. Centralized, distributed and pipelined control strategies will be implemented and compared. This project is motivated by several observations. First, without multiplexing, the control overhead of establishing end-to-end connections in a network may be prohibitive. Second, applications exhibit communication locality which amplifies the benefits of multiplexing. Third, the high bandwidth of optical paths allows for multiplexing with minimal effect on the transmission rate on individual connections. Finally, the nature of massively parallel programs allows compilers to determine, with current technology, the multiplexing sequences that satisfy most of the communication requirements.