Software fault-tolerance in the Pluribus

Over the past decade, the decreasing cost of minicomputer components has encouraged the use of multiprocessor techniques in the design of high-speed, cost-effective computer systems. Multiprocessor architectures have two principal advantages over conventional single-processor designs. First, a multiprocessor system can achieve greater computational speed through parallelism in its task structure. A second advantage is that multiprocessors realize greater reliability through redundancy of processing elements. This paper discusses several aspects of multiprocessor reliability as it applies to the Bolt Beranek and Newman (BBN) Pluribus system. l Recent advances in processor technology, including the development of inexpensive LSI-based microprocessors, have further increased the cost advantages of multiprocessor systems and have led to the consideration of multiprocessor architectures involving very large numbers of processors. Early examples of multiprocessor systems designed around minicomputers include the C.mmp system at Carnegie-Mellon~,3 and the BBN Pluribus. Each of these early systems has prompted the development of larger microprocessorbased systems, such as the Carnegie-Mellon Cm* project and the BBN "Butterfly" machine.