Metamorphic relations for detection of performance anomalies

Metamorphic relations can be used to improve performance testing by comparing successive runs of the software under test. We examine one such metamorphic relation for page load times, which we used to discover and repair a race condition in the Adobe Experience Platform Launch Tag Manager. Histograms for page load times had different modalities, which alerted us to the presence of the bug. We discuss the need for performance measures in addition to the popular mean and standard deviation. We describe two approaches to automatically determine modality: Gaussian Mixture Models and the Silverman Test for Multimodality. Metamorphic relations which involve these performance measures can be used to alert engineers to the presence of performance anomalies.