Control Theory for Model-based Performance-driven Software Adaptation

Abstract

Self-adaptive techniques have been introduced in the last few years to tackle the growing complexity of software systems, where a major complexity factor leans on their dynamic nature subject to sudden and unpredictable changes that can heavily impact on the software architecture quality. Nonfunctional models, as generated from architectural descriptions of software, have been proven as effective instruments to support designers meeting non-functional requirements since the early architectural phases. However, such models still lack of intrinsic support for adaptable software. Goal of this paper is to extend the modeling capabilities to the case of software adaptation aimed at satisfying performance requirements. In particular, we illustrate how control theory can solve the problem of keeping within pre-defined ranges the indices of a Queueing Network model (such as queue length) through software adaptation actions (such as replacing software services with less resource-demanding ones), while the model is subject to disturbances (such as workload and/or operational profile variations). For this goal we first introduce a library of Modelica components that represent Queueing Network (QN) elements with adaptable parameters (that can be used as knobs for adaptation actions). Then we use such components to build "adaptable" QN models that are subject to disturbances. Finally, in the same framework, we introduce controllers that drive the QN adaptation. We demonstrate the soundness of our approach on a simple representative example in two ways: (i) on one end, we provide a formal proof of controller performance guarantees, and (ii) on the other end we show the sensitivity over time of software adaptation actions to different (types and intensities of) disturbances.

What Can Control Theory Teach Us About Assurances in Self-Adaptive Software Systems?

“…Similar models have been presented for various entities that contain queues [41,42] [20, 16,2]. They were derived from the dynamics of the queues and from experiments.…”

“…Formally, the stability is proven by necessary and sucient conditions and, given a model, one can use Bode and Nyquist plots [3] to study the stability. Examples of stability studies in control of software and computing systems have been presented in [21,2]…”

“…Commonly, the stability is designed in the frequency domain (Z or S transforms) and implies solving an algebraic equation that will give the controller equations. Recent work on controller design for stability is presented by Cortellesa et al [2] for a performance case study. 6.5.3 Performance In addition to stability, a controller has to achieve certain performance metrics: rise time, overshoot, settling time, steady error or accuracy [3,6], as illustrated in Example 2.…”

“…One important decision when building a model is the sampling rate, that is, the frequency with which the states, outputs and commands are monitored and processed. In terms of our models (1)- (2), what is the dierence between time k and k + 1? The meaning of 1 above was the next sample but after how many microseconds or minutes is the next sample?…”

Software Engineering Meets Control Theory

“…These contributions have been especially influential for performance management and resource allocation; however, new trends in self-adaptive software introduce new software models and a variety of quantitative and functional requirements beyond the scope of those works. More recently, methodological approaches for performance control [2,58] and the design of self-adaptive operating systems [48] have been proposed. Software Engineering and Autonomic Computing has also highlighted the centrality of feedback loops for adaptive systems [8,41].…”