Performance monitoring infrastructure in Shrimp multicomputers

Efficient, suitably detailed performance information is of crucial importance to the developers and users of parallel systems. The ability to evaluate system designs and higher-level software performance is key for furthering parallel processing research. With this in mind, the authors developed performance monitoring infrastructure for two different generations of Shrimp multicomputers. This paper discusses the two monitoring platform designs. The differing structures of the two Shrimp machines call for different monitoring approaches. In Shrimp-C, they developed a hardware monitoring board, while in Shrimp-M, their monitoring support is implemented primarily as firmware running on the network interface board. In both cases, they show how the designs offer detailed, useful performance information with very little run-time system perturbation.