Automated design of self-adaptive software with control-theoretical formal guarantees

Abstract

Self-adaptation enables software to execute successfully in dynamic, unpredictable, and uncertain environments.Control theory provides a broad set of mathematically grounded techniques for adapting the behavior of dynamic systems. While it has been applied to specific software control problems, it has proved difficult to define methodologies allowing non-experts to systematically apply control techniques to create adaptive software. These difficulties arise because computer systems are usually non-linear, with varying workloads and heterogeneous components, making it difficult to model software as a dynamic system; i.e., by means of differential or difference equations. This paper proposes a broad scope methodology for automatically constructing both an approximate dynamic model of a software system and a suitable controller for managing its non-functional requirements. Despite its generality, this methodology provides formal guarantees concerning the system's dynamic behavior by keeping its model continuously updated to compensate for changes in the execution environment and effects of the initial approximation.We apply the methodology to three case studies, demonstrating its generality by tackling different domains (and different non-functional requirements) with the same approach. Being broadly applicable and fully automated, this methodology may allow the adoption of control theoretical solutions (and their formal properties) for a wide range of software adaptation problems.

Automated control of multiple software goals using multiple actuators

“…Our methodology makes two assumptions: 1) the user knows all available actuators and their limits; i.e., the maximum and minimum values they can assume; and 2) design-time tests can be performed that measure the effects actuator changes have on goals. The proposed methodology minimizes the number of tests to be performed -a big improvement with respect to prior work [15,16,49] -but some experimental data collection is required to build a model and synthesize the controller.…”

“…As seen in Figure 1b, this phase is the first step towards closing the feedback loop. Considering that k belongs to the interval [1, 100·v], this data collection lasts for 100 · v · ∆t time, which is O(v), a strong improvement over previous methods that sweep the entire parameter space [15,16]. Compared to prior work, the resulting models have less fidelity, but the synthesized MPC is robust to these modeling inaccuracies.…”

“…Compared to prior work, the resulting models have less fidelity, but the synthesized MPC is robust to these modeling inaccuracies. As in prior work [15], we update the model at runtime to capture variations.…”

“…The first automated modeling and controller synthesis approach has been proposed in [15]. It builds a locally linearized model of a SISO system collecting input-output measurements during system execution.…”

“…This combination of robustness and formal grounding has led to increased interest in developing self-adaptive software based on control theoretic techniques [46]. While several ad hoc approaches to control design have arisen, recent work proposes a general, automated methodology for creating formally robust software control [15]. However, this first approach can handle only a single, measurable goal (e.g., performance or reliability, but not both).…”