go-v.orgdoi

Performance Models for Embedded Software Product Lines

Birte Kristina Friesel · osnaDocs (Osnabrück University) · 2025

  • DOI 10.48693/670
  • Publisher Osnabrück University
  • Type dissertation
  • Issued 2025

Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.

osnaDocs (Osnabrück University) · Osnabrück University · dissertation · 2025

Abstract

Software product lines deal with functional properties (features) of configurable software systems. However, these systems also have non-functional properties such as latency, energy usage, or memory footprint. Performance models formalize knowledge about those properties and their relation to individual features; they are often crucial to building competitive products. While the literature offers a variety of performance modelling methods for conventional product lines, these do not address configurable device drivers and hardware components.

Moreover, many approaches either rely on time-consuming and error-prone manual annotations, or result in models that are so complex that engineers cannot gain insights by analysing them. This thesis covers interpretable performance models for configurable embedded systems, including software product lines, configurable hardware components, and hybrid product lines that combine both. It focuses on automation for the entire life cycle, ranging from unattended data acquisition over machine learning methods for model generation to performance-aware product line configuration.

In doing so, it shows that the disjoint Product Line Engineering and Internet of Things communities address similar challenges, and combines and extends aspects from both to obtain accurate and interpretable performance models for hybrid and hardware-centric software systems. Its key contribution is the Regression Model Tree data structure and machine learning method. Regression model trees resemble the structure of feature models, but exclusively rely on benchmark data for model generation. They can be used with software product lines, configurable hardware components and hybrid product lines, and can be learnt – and understood – even if no feature model is available.

An evaluation on eight product lines and product line-like system components shows that regression model trees are more accurate and less complex than other tree-based modelling methods when applied to hybrid product lines and hardware components. In addition, this thesis contributes a method for energy measurement automation even if no out-of-band synchronization methods are available, an analysis of whether performance models should be part of feature models or kept separate, a product line perspective on configurable hardware components and device drivers, and case studies that apply regression model trees to real-world product lines.

The analysis finds that performance models should be kept separate; regression model trees follow this approach. The case studies cover both manual model analysis and tool-assisted performance-aware configuration of Kconfig-based product lines.

Open the resolved locator · 10.48693/670

Metadata licence: OpenAlex metadata (CC0)