Software Performance Modelling in Spacetime

Almost any practical software system today is configurable. A configurable software system provides a set of configuration options to adjust and optimize its functional and non-functional properties. In particular, a system's performance behavior often depends in intricate ways on individual configuration options as well as interactions among them. This complex behavior renders tasks such as identifying performance-critical options, deploying performance-optimal configurations, and pinning down reasons for performance regression challenging. Clearly, any practical performance modelling technique needs to take the dimension of configurability into account (space dimension). But this is only one side of the coin. Software systems evolve and so do their (configuration-dependent) performance behaviors (time dimension).

In this talk, I will review recent and ongoing work on modelling performance of configurable software systems. While early approaches concentrated on modelling performance across the configuration space, more recent approaches incorporate the evolution of the software system in question, effectively modelling the system's performance behavior in spacetime. Besides discussing the key ideas in this area, I will highlight challenges that arise from the interplay between configuration sampling and performance learning, the uncertainty that is inherent in performance measurement and behavior, and the interaction of workload-dependent and configuration-dependent behavior.