On large-scale multiprocessors, access to common memory is one of the key performance limiting factors. The shared-memory performance depends not only on the characteristics of the memory hierarchy itself, but also upon the characteristics of the memory address streams and the interaction between the two. We present a technique for multiprocessor workload construction and a family of artificial kernels, called MAD-kernels, to systematically investigate the behavior of the memory hierarchy. The measured performance is independent of any particular application or algorithm. The proposed methodology is demonstrated on two commercial shared-memory systems.

The authors propose a method to characterize the performance of multiprocessor systems at the level of a single grain, called a unit grain, of execution. The characterization is via experimental measurement of individual components of performance. The authors introduce a family of artificial workload kernels, called SAD-kernels, as an effective tool for measuring this performance. The usefulness of these kernels lies in their ability to selectively assess a given shared-memory multiprocessor along several performance dimensions that can be controlled by the person performing the evaluation. The proposed methodology was demonstrated by measuring and comparing the performance of two commercial shared-memory machines currently in use.>

A comprehensive benchmark workload generation and performance characterization methodology is described for multiprocessors supporting the shared-variable computational paradigm. The method can be tailored to meet the selective assessment needs of each individual situation. The approach is based on characterizing a unit grain of computation to generate a desired benchmark workload, and using a family of workload emulation kernels to systematically investigate the effect of each parameter in the workload on the multiprocessor performance. The resultant characterization is independent of any particular application or algorithm.>