A light-weight approach for online state classification of self-organizing parallel systems

Abstract

The growing complexity of future heterogeneous and parallel computing systems is addressed by Organic Computing principles, employing so-called Self-X features for autonomous adaptation and optimization. Here, one major problem is the fact that individual system components only have knowledge about their own states and is therefore lacking the global picture; as a result, each component is unable to determine whether given constraints or requirements are met, whether an optimization cycle should be triggered or not. Even worse, a local instance cannot evaluate the outcome of such optimization cycles and therefore is unable to rate whether the measures taken resulted in a global improvement or not. In order to solve this problem, we present a novel rule-based approach for online system-state evaluation and classification. The rules used for system evaluation are derived during runtime from the information provided by a dedicated, distributed monitoring infrastructure. An important feature of this approach is its capability to self-adapt, i.e., the monitoring infrastructure can adapt the rules to react to given requirements and/or changed system behavior. The proposed method is light-weight to be efficiently employed in self-organizing parallel manycore systems. © 2011 Springer-Verlag.