IMITATION may be the sincerest form of flattery, but it could also be the key to a flexible parallel computer architecture that can outperform its competitors. Researchers at the NEC Research Institute in Princeton, New Jersey, are using a combination of optics and electronics to build a system that can be reconfigured to behave as if it were any of today鈥檚 conventional parallel computers. The result should be a machine that can be instantly reconfigured to suit a particular application, thereby making it run as fast as a computer that was specifically designed for the job.
While personal computers have a single chip acting as a central processor, parallel computers divide up complicated tasks between many processors. In computer graphics, for example, each processor can work on a different part of the image simultaneously to generate the final image very quickly. But to cooperate properly, the processors must 鈥渢alk鈥� to each other, so the architecture of the system is very important. The processors must be arranged so that messages passed between them have as short a journey as possible. The longer the trip, the slower it is and the more distorted the message is likely to be when it arrives.
Computer architects have come up with many solutions to this problem, including structures known as the mesh, the tree and the hypercube. Different configurations are faster at particular types of calculations, such as converging repetitive algorithms or solving large numbers of equations.
Advertisement
But one of NEC鈥檚 researchers, Eugen Schenfeld, realised that whatever the structure, each processor talks most often to only a handful of close 鈥渃olleagues鈥� 鈥� perhaps as few as three or four. Schenfeld decided that, as long as switching within 鈥渃lusters鈥� of close-colleague processors was fast, the links between clusters could be slower without affecting the overall speed of the computer.
In 1992, Schenfeld brought in Ian Redmond, who had previously worked in optical computing at Heriot-Watt University in Edinburgh, Scotland. Together they have designed a system that connects four separate processor boards using arrays of tiny lasers. Each board has 16 processors that are connected to each other electronically.
Each processor is also attached to four lasers. These lasers provide connections with each of the four boards, including the board where the processor sits. The light travels through a passive system of lenses, mirrors and beamsplitters until it arrives at the processor鈥檚 counterpart on the appropriate board. For instance, if processor 14 on board 1 turns on laser number 3, the signal goes to processor 14 on board 3. If necessary, the message can then be routed electronically to another board-3 processor.
By assigning a particular 鈥渋dentity鈥� to each processor, the system can imitate different computer architectures. For instance, by choosing the right subgroup of processors to be on the same board, or to have direct optical links, the hypercube architecture can be 鈥渕apped鈥� onto the NEC hardware as shown (see Diagram).
The system is called the Mapped Inter-connection Cached Architecture and, because the laser arrays are a relatively mature technology, the researchers expect rapid progress. Although the current demonstration only includes 64 processors, Schenfeld aims to have a working computer incorporating between 500 and 1000 processors within three years.
