A Layered Approach to the Characterisation of Parallel Systems for Performance Prediction
G.R. Nudd, E Papefstathiou, Y. Papay, Timothy J. Atherton, Christopher T. Clarke, Darren J. Kerbyson, A. Stratton, Ridha Ziani, Mohamed Jamal Zemerly · 1993
An approach to the characterisation of parallel systems using a structured layered methodology is introduced here. This approach is based upon two similar techniques which have been suggested for use in the modelling of computer systems. Conventional modelling techniques rely on workload parameters obtained from an existing system and the development of a system model. For software under development this is a hindering factor for performance evaluation and performance prediction. Software Performance Engineering (SPE) offers a solution to this problem by using a software execution model, in addition to the system model. The use of SPE for parallel software has several disadvantages:complex nature of the SPE system model, non re-usablemodel components, and lack of analytical methods to evaluate the system model. A novel layered approach for characterisation is presented here, using separate hardware, parallel paradigm, and application layers to overcome these disadvantages. Although this method can be used in modelling, it is primarily targeted towards characterisation studies due to it's hardware independent nature. The layered approach is illustrated, and results obtained, using an image processing benchmark.