This project develops science and technology for ensuring performance stability of large-scale software systems, such as resource management mechanisms used in data centers. The term "stability" is used here in a control-theoretic sense that applies to dynamical systems and roughly means freedom from divergence a range of desired system states. Performance stability refers to stability of performance parameters, such as latency, response time, service throughput, utilization, cache hit ratio, or timeout rate. The project develops the foundations and tools necessary to ensure software stability, and to diagnose and undo root causes of unstable behavior when it occurs in deployed systems. A significant contribution lies in exploring rules and guidelines that, if obeyed, allow reasoning about software stability in a compositional manner, such that stability of composite systems can be inferred from stability of components. Compositional stability analysis of software performance is facilitated by advances in control theory such as Passivity Theory and the Theory of Positive and Dissipative Systems. These advances offer a wealth of results on stability and compositionality for a restricted category of non-linear systems that fits software models.
Performance stability challenges have been largely overlooked in software design. They are not typically manifest in small systems, but grow with the size and complexity of systems. The trend towards more software consolidation and outsourcing of computing services entails more complexity, more layering, and more interactions among various resource management mechanisms, making it harder to anticipate side-effects, and more likely there will be stability problems. The project improves the current understanding of the design, execution, and management of large systems that exploit feedback mechanisms to achieve performance and robustness objectives. Educational activities include incorporation of project elements into several courses taught by the PI, and involvement of undergraduate students in the research. Outreach activities include an on-campus "Feedback Computing Day", a tutorial on feedback computing to be offered in conjunction with a major research conference, and efforts by the PI to recruit students from under-represented groups. Dissemination activities include documentation of results of the research in a book by the PI, and efforts to transition technology through collaborators in industry.
