System performance in various computing systems is measured using various benchmarks. A benchmark allows users to observe a set of performance metrics of the system as a function of time and workload, and to determine if a performance metric has deviated or regressed. However, different regression analyzers are suitable for different metrics and finding accurate analyzers often requires substantial manual effort that needs to be repeated whenever a variable that impacts a performance metric changes. This disclosure describes techniques that obtain historical data (sourced from stable workloads) about the pattern of a performance metric and use the data to train a machine learning algorithm to analyze a performance metric and determine a suitable analyzer. The analyzer configuration is selected based upon classification of the metric as noisy or not noisy, and on what is suitable for the particular metric.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.