Sensitivity analysis of computer systems and their queueing network models

Conventional analytical models of computer systems typically assume that the different model parameters are independent. Experimental evidence produced on a system running UNIX shows that many such parameter independence assumptions are false. Significant interdependencies can exist between unexpected system parameters. Sensitivity analysis experiments also indicate that change in certain system parameters can result in significant changes to other system parameters. These results indicate that a good understanding of parameter interdependencies and sensitivities is a prerequisite to constructing a good system model. A comprehensive experimental sensitivity analysis resulting from a wide range of controlled experiments on a UNIX system is described. Parameters which were varied include the CPU workload demand, the workload memory requirements, the number of customers, the terminal workload demand, the size of main memory, and other system parameters. An analogous analytic sensitivity analysis is conducted and the results compared. The results provide insight into the complex interactions between various system parameters and their impact upon system performance.