Chiarulli D M, Levitan S P, Melhem R G, Qiao C
Appl Opt. 1994 Mar 10;33(8):1528-37. doi: 10.1364/AO.33.001528.
Hybrid optoelectronic computing structures are required for providing the information processing capabilities for the next generation of computing and communications systems. Reconfigurable optoelectronic interconnection networks are networks constructed of optical waveguides in which messages are switched or routed by means of optoelectronic devices. For these networks, the dichotomy between the bandwidth of the optical channels that carry messages and the performance of the electronic controllers and decoders that determine the routing and destination of those messages is a significant bottleneck. We introduce a class of routing algorithms for reconfigurable networks that is designed to bridge this gap in optical versus electronic performance. The algorithms are based on a new control paradigm that exploits the locality in multiprocessor communication streams to reduce the control latency inherent in reconfigurable interconnection structures. In addition, we show that this problem maps directly to the problem of page replacement in a virtual-memory hierarchy. Thus our solution is well suited to networks for multiprocessor applications.
下一代计算和通信系统需要混合光电计算结构来提供信息处理能力。可重构光电互连网络是由光波导构成的网络，其中消息通过光电器件进行交换或路由。对于这些网络，承载消息的光通道带宽与决定这些消息路由和目的地的电子控制器及解码器性能之间的差异是一个重大瓶颈。我们为可重构网络引入了一类路由算法，旨在弥合光与电子性能之间的这一差距。这些算法基于一种新的控制范式，该范式利用多处理器通信流中的局部性来减少可重构互连结构中固有的控制延迟。此外，我们表明这个问题直接映射到虚拟内存层次结构中的页面替换问题。因此，我们的解决方案非常适合用于多处理器应用的网络。
