Studying the performance properties of concurrent programs by simulation experiments on synthetic programs
Rosemary Candlin, Peter R. Fisk, Joe Phillips, Neil Skilling · ACM SIGMETRICS Performance Evaluation Review · 1992
We have developed a methodology for constructing performance models of different types of concurrent programs, and hence obtaining estimates of execution times on different multiprocessor machines. A given class of program is characterized in terms of a small set of parameters which summarise the behaviour of the program over time. Synthetic programs with selected sets of parameters can then be generated and their execution simulated on a model of some given parallel machine. By varying the parameters systematically, we can discover which factors most affect performance. Our approach has been to conduct factorial experiments from which we can obtain quantitative predictions of performance for arbitrary concurrent programs whose parameter values lie within the extreme factor levels, and whose synchronization behaviour conforms to one of a number of common patterns.