A survey of autonomic computing—degrees, models, and applications

Reviewer: Franz J Kurfess

Because my background is in artificial intelligence, I am quite familiar with the notion of autonomous systems; however, I had only a superficial understanding of autonomic computing. My curiosity was driven, to some degree, by the similarity of the two terms, but also by a desire to learn more about how systems can be designed to manage themselves better, thus decreasing the tedious aspects of system administration. Autonomic computing relies on autonomicity, the property of a self-managing system. Its use was inspired by biology, where the autonomic nervous system controls bodily functions not involving attention, such as unconscious reflexes. While there is some similarity to autonomy, which is used to describe autonomous systems, autonomicity puts more emphasis on the behavior of a system in an environment. In the context of intelligent agents [1], important aspects of an autonomous agent are its ability to complete a task without significant interference or guidance from a user or developer, and the ability of the agent to acquire its knowledge on its own, rather than having it implanted by the developer as prior knowledge. So, while both aspects rely on the ability of the system or agent to act independently, without outside supervision, autonomicity concentrates on activities that are internal to the system, whereas autonomy emphasizes the behavior of an agent in an environment. Huebscher and McCann give an excellent overview of the main issues in the relatively new area of autonomic computing, and trace the history of the field by discussing some of the influential approaches and systems. This is followed by a discussion of the main properties of self-management and a review of a reference model for autonomic control loops?the monitor, analyze, plan, execute, knowledge (MAPE-K) loop. The paper then addresses the applicability to multi-tier systems, such as e-commerce sites incorporating a Web server, an application server, and a database tier. Since autonomicity is not an all-or-nothing concept, different degrees of autonomicity can be observed. Huebscher and McCann analyze an approach that identifies different levels, ranging from basic (performed by humans) to fully autonomic (human intervention is only needed in response to changing business needs). They conclude that this approach is too narrow and the distinction between levels too vague. As an alternative, they suggest four elements of autonomicity: support, core, autonomous, and autonomic. While I agree with their critique of the level-based approach, I did not find the elements-based approach very helpful, either. This is enhanced by some confusion in the terminology: in the introductory paragraph, where they present their element-based approach, they specifically point out that these elements are not "levels"; yet, in the description of the autonomic element, they repeatedly refer to it as a level. This is followed by an overview of emerging applications of autonomic computing, with power management, distributed computing, and ubiquitous computing as the most prominent ones. Overall, the paper provides a good introduction to the topic of autonomic computing for readers with no or little familiarity with the topic. I would have liked to see more on fundamental issues such as limitations on self-observation and self-analysis, or the gap between the mental model of humans, textual descriptions, specifications, and implementations of systems. But this may well be a reflection of the relatively immature state of the field itself and not so much an omission by the authors. Online Computing Reviews Service