This thesis presents an approach for the design time analysis of energy efficiency for static and self-adaptive software systems.
The quality characteristics of a software system, such as performance and operating costs, strongly depend upon its architecture. Software architecture is a high-level view on software artifacts that reflects essential quality characteristics of a system under design. Design decisions made on an architectural level have a decisive impact on the quality of a system. Revising architectural design decisions late into development requires significant effort. Architectural analyses allow software architects to reason about the impact of design decisions on quality, based on an architectural description of the system. An essential quality goal is the reduction of cost while maintaining other quality goals. Power consumption accounts for a significant part of the Total Cost of Ownership (TCO) of data centers. In 2010, data centers contributed 1.3% of the world-wide power consumption. However, reasoning on the energy efficiency of software systems is excluded from the systematic analysis of software architec ... mehrtures at design time. Energy efficiency can only be evaluated once the system is deployed and operational. One approach to reduce power consumption or cost is the introduction of self-adaptivity to a software system. Self-adaptive software systems execute adaptations to provision costly resources dependent on user load. The execution of reconfigurations can increase energy efficiency and reduce cost. If performed improperly, however, the additional resources required to execute a reconfiguration may exceed their positive effect.
Existing architecture-level energy analysis approaches offer limited accuracy or only consider a limited set of system features, e.g., the used communication style. Predictive approaches from the embedded systems and Cloud Computing domain operate on an abstraction that is not suited for architectural analysis. The execution of adaptations can consume additional resources. The additional consumption can reduce performance and energy efficiency. Design time quality analyses for self-adaptive software systems ignore this transient effect of adaptations.
This thesis makes the following contributions to enable the systematic consideration of energy efficiency in the architectural design of self-adaptive software systems: First, it presents a modeling language that captures power consumption characteristics on an architectural abstraction level. Second, it introduces an energy efficiency analysis approach that uses instances of our power consumption modeling language in combination with existing performance analyses for architecture models. The developed analysis supports reasoning on energy efficiency for static and self-adaptive software systems. Third, to ease the specification of power consumption characteristics, we provide a method for extracting power models for server environments. The method encompasses an automated profiling of servers based on a set of restrictions defined by the user. A model training framework extracts a set of power models specified in our modeling language from the resulting profile. The method ranks the trained power models based on their predicted accuracy. Lastly, this thesis introduces a systematic modeling and analysis approach for considering transient effects in design time quality analyses. The approach explicitly models inter-dependencies between reconfigurations, performance and power consumption. We provide a formalization of the execution semantics of the model. Additionally, we discuss how our approach can be integrated with existing quality analyses of self-adaptive software systems.
We validated the accuracy, applicability, and appropriateness of our approach in a variety of case studies. The first two case studies investigated the accuracy and appropriateness of our modeling and analysis approach. The first study evaluated the impact of design decisions on the energy efficiency of a media hosting application. The energy consumption predictions achieved an absolute error lower than 5.5% across different user loads. Our approach predicted the relative impact of the design decision on energy efficiency with an error of less than 18.94%. The second case study used two variants of the Spring-based community case study system PetClinic. The case study complements the accuracy and appropriateness evaluation of our modeling and analysis approach. We were able to predict the energy consumption of both variants with an absolute error of no more than 2.38%. In contrast to the first case study, we derived all models automatically, using our power model extraction framework, as well as an extraction framework for performance models. The third case study applied our model-based prediction to evaluate the effect of different self-adaptation algorithms on energy efficiency. It involved scientific workloads executed in a virtualized environment. Our approach predicted the energy consumption with an error below 7.1%, even though we used coarse grained measurement data of low accuracy to train the input models. The fourth case study evaluated the appropriateness and accuracy of the automated model extraction method using a set of Big Data and enterprise workloads. Our method produced power models with prediction errors below 5.9%. A secondary study evaluated the accuracy of extracted power models for different Virtual Machine (VM) migration scenarios. The results of the fifth case study showed that our approach for modeling transient effects improved the prediction accuracy for a horizontally scaling application. Leveraging the improved accuracy, we were able to identify design deficiencies of the application that otherwise would have remained unnoticed.