Scaling Size and Parameter Spaces in Variability-Aware Software Performance Models (T)

Abstract

In software performance engineering, what-if scenarios, architecture optimization, capacity planning, run-time adaptation, and uncertainty management of realistic models typically require the evaluation of many instances. Effective analysis is however hindered by two orthogonal sources of complexity. The first is the infamous problem of state space explosion -- the analysis of a single model becomes intractable with its size. The second is due to massive parameter spaces to be explored, but such that computations cannot be reused across model instances. In this paper, we efficiently analyze many queuing models with the distinctive feature of more accurately capturing variability and uncertainty of execution rates by incorporating general (i.e., non-exponential) distributions. Applying product-line engineering methods, we consider a family of models generated by a core that evolves into concrete instances by applying simple delta operations affecting both the topology and the model's parameters. State explosion is tackled by turning to a scalable approximation based on ordinary differential equations. The entire model space is analyzed in a family-based fashion, i.e., at once using an efficient symbolic solution of a super-model that subsumes every concrete instance. Extensive numerical tests show that this is orders of magnitude faster than a naive instance-by-instance analysis.