Experiments: Analytical Models
Accuracy The resource profiles should be unbiased even if the workload at the respective resources varies. We have already addressed this central requirement in the experiments presented in section 4.3.4. However, during those load tests the resources were at all times exclusively used by a single service. Furthermore, the profiling workload consisting of x concurrent services invocations is not comparable to the workload during regular operations. In contrast, in load tests for the validation of performance models realistic scenarios with multiple concurrently active services and varying workload can be simulated. We claim that if performance models parameterized with consumption estimates from the resource profiles accurately predict system performance (utilization, response time, throughput) during those load tests, the accuracy of the resource profiles is also sufficient for cost allocation.
Capacity Planning In contrast to arbitrarily chosen cost allocation keys, the resource profiles should bridge the gap between business forecasts and IT resource requirements. The major advantage of an analytical performance model is that once has it has been validated it can be readily used for capacity planning. It enables capacity planners to conduct “what-if” analyses and thus anticipate the effects of changes in the workload composition, hardware configuration or system architecture.
KeywordsLoad Test Processor Utilization Capacity Planning Queueing Network Queueing Network Model
Unable to display preview. Download preview PDF.