A high-performance, memory-based interconnection system for multicomputer environments

The design and operation of a very-high-performance, memory-mapped interconnection system called Merlin is discussed. The design can be effectively utilized to interconnect processors in a wide variety of environments, ranging from closely-coupled, dedicated systems to distributed workstations. The system provides a uniform approach to parallel programming which is independent of interconnection topology, processing elements, and languages. By using dynamically mapped reflective memory, the system can support selectively shared virtual memory regions. This approach permits user-selected information to be shared at high speeds and low latency. There is no software involvement in the actual sharing of information, and the system overlaps computation and communication automatically, word-by-word, to the extent possible.<
>