Workload modeling for parallel processing systems

The workload of a computer system is traditionally defined as all inputs from the user to the system. As the use of this real workload is usually impossible in most performance analysis studies, it is necessary to abstract from the real workload and to construct a test workload model. There are already quite established methodologies and techniques for workload modeling of uniprocessor systems, but these techniques cannot be applied directly to parallel processing systems (PPS) because of the different nature of the workload of a PPS. Therefore this paper addresses selected issues of a workload modeling methodology for PPS. The basic components that a workload model for a parallel system has to represent are identified and the influences of the objective of the analysis and the evaluation methods are discussed. Finally we investigate workload modeling in a performance oriented parallel software development cycle