Performance of a Message-Based Multiprocessor
Sanguinetti · Computer · 1986
Multiprocessor computers have been commercially available for over 20 years in a variety of forms. Typically, loosely coupled systems provide good throughput at the expense of flexible resource sharing, while tightly coupled systems provide flexible resource sharing at the cost of degraded throughput. Most tightly coupled multiprocessor systems degrade markedly when the number of CPUs reaches three or four. That is, for each additional CPU, the system performance improves by some fraction of the original CPU (usually 0.8-0.9 for the second one), and that fraction decreases as the number of CPUs increases. The measurements that are reported here were performed in order to characterize the performance of a tightly coupled multiprocessor, the Elxsi 6400, throughout the range of 1 to 10 CPUs, which was the implementation limit at the time. This system has a message-based architecture with several design features intended to enhance multiprocessor performance with the goal of achieving linear performance improvements as CPUs are added. This set of experiments was intended to answer the question of whether or not this type of system could achieve, or possibly exceed, linear performance improvement.