Optical centralized shared bus architecture for high-performance multiprocessing systems

Abstract

With the increasing demand for solving more complex problems, high-performance multiprocessing systems are attracting more and more research efforts. One of the challenges is to effectively support the communications among the processes running in parallel on the multiprocessors. Due to the physical limitations of electrical interconnects, interconnection networks impose a potential bottleneck limiting the overall performance. On the other hand, optics has many advantages as an interconnect technology. In this paper, benefits of optics are evaluated along with a comparison of two mainstream system topologies, shared bus and switched media. This analysis leads to an innovative interconnect architecture, optical centralized shared bus. The crucial design aspects of this architecture, including system organization, working principle, and conversion between free-space propagation and substrate-guided mode propagation by using volume holographic gratings, are delineated. To ensure the feasibility of using this architecture as high ­performance interconnection networks in real multiprocessing systems, a PCI implementation of the centralized shared memory multiprocessor system is proposed. In this prototype, the required connectivity is accomplished by using the optical centralized shared bus architecture. Some preliminary results are presented.