The experimental multiprocessor system MEMSY was built to validate the
concept of a scalable multiprocessor architecture based on local shar
ed memory. The main application areas are scientific computations with
high demand for processing power and memory capacity. In designing th
e hardware architecture the extensive use of standard components and f
ault tolerance were prerequisites. The programming model of MEMSY is c
ustom made, reflecting its true hardware structure, whereas the operat
ing system is a Unix extension. In massively parallel systems with its
complexity and large number of components the chance of a single or m
ultiple failure is no longer negligible. Therefore, redundancy, reconf
igurability and diagnosis techniques have been incorporated at the des
ign stage itself and not as a subsequent add-on.