Growing complexity in many current computers makes performance evaluation and characterization both increasingly difficult and increasingly important. For parallel systems, performance characterizations can be especially difficult to obtain, since the hardware is more complex, and the simulation time can be prohibitive.

This technical report describes the design, implementation, and case studies of a performance monitoring system for the SHRIMP multicomputer. This system is based on a hardware performance monitor which combines several features including multi-dimensional histogram generation, trace generation, and sophisticated triggering and interrupt capabilities.

Demonstrated in the case studies is the direct measurement of an implicit form of interprocessor communication implemented on the SHRIMP multicomputer.