An analysis of processor resource models for use in performance prediction

With the increasing sophistication of both software and hardware systems, methodologies to analyse and predict system performance is a topic of vital interest. This is particularly true for parallel systems where there is currently a wide choice of both architectural and parallelisation options; and where the costs are likely to be high. Performance data is vital to a diffuse range of users including; system developers, application programmers and tuning experts. However, the level of sophistication and accuracy required by each of these users is substantially different. In this paper characterisation based technique is described (considering both application and hardware resources) which addresses these issues directly. Initially, a framework is described for characterisation based approaches. A classification scheme is used to illustrate differences in the level of sophistication and detail in which the underlying resource models are specified. Finally, verification is provided of the characterisation techniques applied to several application kernels on a MIMD system. The performance predictions and error bounds resulting from the level of resource specification are also discussed