Performance Model Extraction Using Kernel Event Tracing

Models are used in performance analysis when the analyst needs to be able to predict the effect of system changes that go beyond what can be measured. The model can be obtained from a combination of system knowledge and experimentation. This thesis addresses an experimental approach to obtaining layered queueing network (LQN) models of distributed systems. Kernel-level traces have the benefit that application instrumentation is not required, and communication with attached devices can be modeled, but they lack application context information. The research shows that modeling from Kernel traces is feasible in systems which communicate via TCP messages, including Java remote procedure calls. This covers most web-based systems. Systems using middleware pose special problems. The combination of Kernel and application-level tracing was included in some experiments. Tools are described that adapt the Kernel traces to SAME, and that extract CPU demand parameter calibration information.