Family-Based Performance Analysis of Variant-Rich Software Systems

Abstract

We study models of software systems with variants that stem from a specific choice of configuration parameters with a direct impact on performance properties. Using UML activity diagrams with quantitative annotations, we model such systems as a product line. The efficiency of a product-based evaluation is typically low because each product must be analyzed in isolation, making difficult the re-use of computations across variants. Here, we propose a family-based approach based on symbolic computation. A numerical assessment on large activity diagrams shows that this approach can be up to three orders of magnitude faster than product-based analysis in large models, thus enabling computationally efficient explorations of large parameter spaces.

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

“…We are most closely related to recent previous work by some of this paper's authors [25], where PAADs were introduced for the first time. In [25], the objective was still to exploit commonalities across variants by building a symbolic 150% model solution.…”

“…In [25], the objective was still to exploit commonalities across variants by building a symbolic 150% model solution. From a modeling viewpoint, there are two major limitations in [25], imposed by certain syntactic restrictions on the PAAD which are removed in this paper: (i) activities are assumed to take exponentially distributed service times; (ii) and they are assumed to be performed by computational units without parallelism (i.e., single servers).…”

“…The assumptions of [25] led to an underlying PAAD semantics based on Jackson queuing networks (e.g., [6]). The parametric analysis was directly performed on the traffic equations of the network.…”

“…The parametric analysis was directly performed on the traffic equations of the network. In this paper we are essentially inspired by an observation from [25] that parametric analysis is possible if the problem is linear with a closed-form solution (such as matrix inversion for the traffic equations). Despite using similar definitions of PAADs and 150% model construction, from a technical viewpoint the present contribution is starkly different: (i) with Coxian-distributed multi-server stations, the traffic equations cannot be reused directly for a family-based analysis; (ii) to overcome this issue we turned to an ODE representation of the system; (iii) the ODEs are nonlinear, yet we show that their traffic equations, defined here for the first time to are knowledge, are linear.…”

“…For this we are inspired by software product line engineering as a highly successful and powerful approach to designing systems exhibiting a high degree of variability, due, for instance, to the number of possible configurations which are possible [29]. Following previous work of some of this authors [25], our reference model is a Performance Annotated Activity Diagram (PAAD) supported by a delta-oriented (cf. [30]) language.…”