Transaction mix performance models
Terence Kelly · 2005
This poster describes a simple model of application-level performance as a function of workload. The model is intuitive, easy to apply, and requires little knowledge of the application. The model can be used for performance anomaly detection: identifying relatively rare cases where workload does not explain performance. Knowing when workload explains performance well vs. poorly can help to distinguish between true performance faults and mere overload. This in turn can inform our choice of performance analysis and debugging tools, and can also suggest remedial measures. Extensive empirical results demonstrate that the proposed method accurately models performance in three large distributed commercial production applications serving real customers. Furthermore it flags as anomalous an episode of a subtle performance bug in one of these applications.