April 1996This is a preprint of a paper intended for publication in a journal or proceedings. Since changes may be made before publication, this preprint is made available with the understanding that it will not be cited or reproduced without the permission of the author.
PREPRINT
AbstractAs the performance of individual processing elements within parallel processing systems increases, increased capability to communicate information between individual processor and memory elements is required. Since the limited performance of today's electronic interconnects will likely prevent the system from achieving its ultimate performance, there is great interest in using fiber optics to improve interconnect communication. Many groups have considered approaches based on WDM, star-coupled fiber optics for moderate size multiprocessors. Here we propose a fiber optic transceiver approach to such systems that can provide low latency, high bandwidth channels using a robust multimode fiber technology. We use instruction-level simulation to quantify the bandwidth, latency, and concurrency requirements that enable a multiple bus-type optical interconnect based on such transceivers to scale to 256 nodes, each operating at GFLOPS performance. Our key conclusion is that scalable performance, to ≈100 GFLOPS, is achievable for scientific application kernels using a small number of wavelengths (8 to 32), one optical bus receiver per node, and achievable optoelectronic bandwidth and latency requirements.2