482 EVOLUTION OF COMPUTER BUILDING BLOCKS
performance of multi-cluster Cm* configurations to predict the performance degradations associated with intercluster references. Figure 22 shows the performance of Cm* on two different versions of the PDE program for both single-cluster and multi-cluster configurations. Note that nearly negligible degradation was achievable, particularly in method 4, which is an asynchronous version of the PDE specifically designed to cope with processors of
Figure 22. Single- and multiple-cluster execution time.
varying run times. The small degradation in going from the one cluster configuration to the multi-cluster configuration gives considerable hope that hierarchical switching structures like the one used in Cm* can provide very nearly the performance of much more expensive switching structures that give uniformly fast access to all of physical memory.
CONCLUDING REMARKS
The major accomplishment of the Cm* project has been to bring an experimental multi-microprocessor system to an operational state, and to demonstrate that almost-linear speedup can be achieved with several applications. Moreover, there have been no serious bottlenecks or deficiencies in the processor/memory bus structure that preclude configurations with 100 or more processors.
Many aspects of Cm*, and multi-micro processors in general, require further investigation. Our own plans call for considering alternative memory mapping and interprocess control architectures, developing a large application system on Cm* to test larger configurations, and integrating a practical I/O system into the Cm* structure.
As other multi-microprocessors become operational and competing solutions are found to some of the problems currently facing multiprocessors, the relative merit of the Cm* organization will be put into much better perspective. A comparison of alternate multiprocessor organizations is especially important in the initial stages when most investigations are necessarily empirical, and no one solution may claim optimality.
APPENDIX: DESCRIPTION OF THE BENCHMARK PROGRAMS
Five programs from different application areas were used in the initial performance evaluation of the Cm* system. Four of these pro grams are described here, and the HARPY speech recognition program is described in Jones et al., [1978]. More detailed descriptions of these programs may be found in Fuller et al., [1977].
Partial Differential Equations, a Numerical Application
This is the solution to Dirichlet's problem of Laplace's partial differential equation (PDE) by the method of finite difference. This program solves the PDE:
