ProDSPL: Proactive self-adaptation based on Dynamic Software Product Lines

Highlights

• ProDSPL combines proactive control and DSPL for self-adaptation of software systems.

• The reconfiguration service for DSPLs is formulated as an optimal control approach.

• An automatically learnt model of the system is used to anticipate future variations.

• Valid extended feature model configurations are encoded into linear constraints.

• ProDSPL reduces time and energy costs of reconfiguration, maintaining the quality.

Abstract

Dynamic Software Product Lines (DSPLs) are a well-accepted approach to self-adaptation at runtime. In the context of DSPLs, there are plenty of reactive approaches that apply countermeasures as soon as a context change happens. In this paper we propose a proactive approach, ProDSPL, that exploits an automatically learnt model of the system, anticipates future variations of the system and generates the best DSPL configuration that can lessen the negative impact of future events on the quality requirements of the system. Predicting the future fosters adaptations that are good for a longer time and therefore reduces the number of reconfigurations required, making the system more stable.

ProDSPL formulates the problem of the generation of dynamic reconfigurations as a proactive controller over a prediction horizon, which includes a mapping of the valid configurations of the DSPL into linear constraints. Our approach is evaluated and compared with a reactive approach, DAGAME, also based on a DSPL, which uses a genetic algorithm to generate quasi-optimal feature model configurations at runtime. ProDSPL has been evaluated using a strategy mobile game and a set of randomly generated feature models. The evaluation shows that ProDSPL gives good results with regard to the quality of the configurations generated when it tries anticipate future events. Moreover, in doing so, ProDSPL enforces the system to make as few reconfigurations as possible.

Introduction

The demand for self-adapting software systems has risen sharply, specially motivated by the growing importance of Cyber–Physical Systems (CPS) operating in complex and changing environments, and the increasing ubiquity possibilities of Information and Communication Technologies (ICT) (Tavčar and Horváth, 2019). Designing self-adaptive systems is a complex task, and has been of great interest in the last years, specially in areas such as mobile computing, robotics or ubiquitous computing (Maggio et al., 2017a). Runtime adaptation mechanisms sometimes are problematic since they tend to be unstable, inefficient and unreliable (Elkhodary et al., 2010). Therefore, in order to overcome some of these problems several approaches have been defined (Bashari et al., 2018, Pascual et al., 2015b, Almeida et al., 2014), being the Dynamic Software Product Lines (DSPL) (Carvalho et al., 2018) one of the most widely used. DSPL are a systematic engineering approach that uses Software Product Lines (SPLs) artefacts to model the dynamic variability of a system. DSPLs model the different adaptations of the system to context changes, either internal or external, as variation points. Examples of context variations that can trigger a system adaptation can be, the continuous variation of resource availability (e.g., battery or memory), the occurrence of system or environment failures, and also the fluctuations of user’s needs (Hallsteinsen et al., 2008, Abbas et al., 2020).

Adopting a DSPL approach requires specifying the dynamic variation points as part of a variability model and a reconfiguration service that performs the system adaptation when necessary. One of the most popular variability models is the feature model (FM), which, in the context of DSPLs, is used to specify a variability space model of a single system at runtime, without needing to enumerate all the possible system configurations. The reconfiguration service determines if the system needs to be adapted to satisfy the quality requirements under different execution contexts. Then, the reconfiguration service, containing the DSPL artefacts, will be continuously monitoring the context and finding out the best possible configuration after each context change at runtime. This new configuration should be the optimal one with respect to an objective criterion (e.g., minimize battery consumption while maintaining a prescribed quality of service). There are many reasoning techniques to select the suitable variant of a system, considering the current configuration, the context change and the DSPL artefacts.

Independently of the reasoning approach used to find the new configuration, DSPLs approaches implement decision strategies that are purely reactive: during the system execution the context is continuously monitored and when a context change occurs, the reconfiguration service react to this change by analysing if there is another configuration that fits better the current conditions trying to find an optimal or quasi-optimal solution for the current context (Capilla et al., 2014). In Cyber–Physical systems the strategy used to generate the new optimal or quasi-optimal configuration should be implemented at runtime, since the high number of unexpected context changes makes impossible to pre-load all possible reconfiguration plans. Dynamic strategies then require the generation of a valid configuration at runtime, and adapting the system accordingly, which can be both time- and resource-consuming tasks (in terms of energy, memory, computation, etc.) that can negatively affect the performance of battery-powered systems.

Therefore, when analysing different DSPL alternatives, the adaptation cost should also be considered in terms of resources involved in reconfiguration, such as energy consumption and also the time needed to analyse and perform an adaptation. Indeed, in some scenarios, the adaptation process can consume a significant amount of execution time. For example, in Wireless Sensor Networks, where most of the devices have usually limited computational resources, the cost of the reconfiguration is usually too high as it contributes to a faster degradation of the device batteries. In medical applications, it is simply not acceptable to stop – or delay – the normal functioning of the system for a reconfiguration, specially if there is a serious problem that should be fixed urgently. For this kind of scenarios, the preferred techniques are those that lessen the number of reconfigurations, but trying to maintain a good quality of service at the same time. Also, performing continuous adaptations makes the system more unstable, because, if the reconfiguration service makes the system jump from one configuration to another, this might reduce the user quality of experience. Then, in these situations, it would be preferable to adopt a proactive strategy in order to reduce the number of adaptations, making the system more stable. Proactive strategies consider not only the current context, but also the system expected evolution over time, so their adaptation solutions tend to stay valid for longer. But the price to pay when we apply a proactive strategy is that sometimes the calculated application configuration has a lower utility or quality compared to the optimal solution or to the solution provided by some reactive approaches. Therefore, the great challenge is to find a proactive strategy whose adaptation solutions are good enough to increase the lifespan of the system while reducing the number of required reconfigurations. By decreasing the number of runtime reconfigurations, the overall adaptation cost is reduced and also the system is more stable.

In this work, we propose ProDSPL, that is a self-optimizing approach that combines DSPL with a control-based proactive decision strategy for dynamically generating optimal configurations in a proactive way. Self-optimizing systems based on Proactive Controllers have been explored in the context of software systems adaptation with successful results (Maggio et al., 2012, Moreno et al., 2017, Angelopoulos et al., 2018), and in this work we explore how well it behaves in conjunction with DSPL artefacts.

ProDSPL formulates the problem of dynamic reconfiguration as a proactive controller (PC) that generates valid configurations of a DSPL at runtime. In the control theory field, this type of adaptation, which is known as Model Predictive Control (MPC), relies on dynamic models learnt from the system observation, and comes with a well-developed theory and myriads of successful applications (Camacho and Bordons, 2007).

The main contribution of this paper is the formulation of the DSPL reconfiguration service as a single-objective optimization problem over a prediction horizon subject to the linear constraints of the DSPL feature model following a proactive approach. This formulation is based on a mapping between extended feature models and linear constraints, which is part of our solution to combine proactive control with DSPL. Although previous works have proposed this kind of mapping (Noorian et al., 2017, Li et al., 2012), they only consider basic feature models. Feature models of this kind do not support numerical features and consider only simple cross-tree constraints (i.e., include or exclude), so they cannot be applied in several real case studies.

Both the time required to generate the configurations and the quality of the solutions found are crucial aspects for the success of a DSPL-driven reconfiguration at runtime. So, the performance and quality of the results obtained by the proactive controller ProDSPL are compared with DAGAME (Pascual et al., 2015b), a reactive approach that uses a DSPL approach with a genetic algorithm for the generation of feature model configurations at runtime. Taking into account the concept of optimality (Guo et al., 2011) (i.e., the ratio between the utility of the solution obtained by an algorithm and the utility of the optimal solution obtained using the exact method), DAGAME is able to obtain solutions with an optimality higher than 87.4% and execution times between 20 and 100 ms. In this paper, ProDSPL performance is evaluated using a simulator of a mobile app developed in Java, in order to make experiments and results reproducible by third parties. The comparison with the purely reactive approach shows good results with regard to the performance and the quality of the configurations generated. The results obtained support in practice our initial hypothesis about the use of a proactive approach. That is, our solution is more sustainable as it contributes to reduce the overall cost – in time and energy – of reconfiguration and leads to more stable systems, while maintaining a satisfactory quality of adaptation.

The remainder of the paper is organized as follows: Section 2 provides background on DSPLs, the decision-making process, and the model predictive control. Section 3 overviews the main activities of our approach, ProDSPL. Section 4 illustrates how to apply our proposal using a Mobile Game as a case study. The experimental results are presented, analysed and compared with DAGAME in Section 5, while the threats to the validity of our study are discussed in Section 6. Section 7 discusses related work. Finally, Section 8 presents some conclusions to the paper.