View article

Adding adaptation capabilities to existing distributed systems is a major concern. The question addressed here is how to retrofit existing systems with self-healing, adaptation and/or selfmanagement capabilities. The problem is obviously intensified for “systems of systems” composed of components, whether new or legacy, that may have been developed by different vendors, mixing and matching COTS and “open source” components. This system composition model is expected to be increasingly common in high performance computing. The usual approach is to train technicians to understand the complexities of these components and their connections, including performance tuning parameters, so that they can then manually monitor and reconfigure the system as needed. We envision instead attaching a “standard” feedbackloop infrastructure to existing distributed systems for the purposes of continual monitoring and dynamically adapting their activities and performance.(This approach can also be applied to “new” systems, as an alternative to “building in” adaptation facilities, but we do not address that here.) Our proposed infrastructure consists of multiple layers with the objectives of probing, measuring and reporting of activity and state within the execution of the legacy system among its components and connectors; gauging, analysis and interpretation of the reported events; and possible feedback to focus the probes and gauges to drill deeper, or–when necessarydirect but automatic reconfiguration of the running system.