Feedback Controlled Software Systems

Abstract

Software systems generally suffer from a certain fragility in the face of disturbances such as bugs, unforeseen user input, unmodeled interactions with other software components, and so on. A single such disturbance can make the machine on which the software is executing hang or crash. We postulate that what is required to address this fragility is a general means of using feedback to stabilize these systems. In this paper we develop a preliminary dynamical systems model of an arbitrary iterative software process along with the conceptual framework for stabilizing it in the presence of disturbances. To keep the computational requirements of the controllers low, randomization and approximation are used. We describe our initial attempts to apply the model to a faulty list sorter, using feedback to improve its performance. Methods by which software robustness can be enhanced by distributing a task between nodes each of which are capable of selecting the best input to process are also examined, and the particular case of a sorting system consisting of a network of partial sorters, some of which may be buggy or even malicious, is examined.