ASoCS: An Architecture Concept for Self-Optimizing Parallel and Distributed Computer Systems

Application optimization usually means optimization for a given, static computer architecture. In such systems, the programmer optimizes towards given objective functions with respect to the existing system. This scenario changes dramatically when dealing with recongurable systems. Here, the underlying system architecture may change and be changed during runtime. Such changes might be both, measures of self-healing or a direct result of self-optimization. In any case, it is not possible to optimize in a classic sense anymore since the objective functions might change with each reconguration. Optimization therefore be-comes an effort of nding a best compromise between sometimes contradictory objective functions and the current system conguration. In this paper we outline an architecture- and application-independent framework for self-organizing par-allel computer systems. Main focus of this framework is to provide a novel, uniform approach to modeling Self-X systems. Core of this framework is a layer model with exible, decentralized monitoring and adaptive components on each system layer. Monitoring components feed an adaptive planning stage, which evalu-ates this data against given objective functions and instruments the adaptive components to perform required system updates. The suitability and applicability of this framework is demonstrated by successful adaptation of two exam-ple scenarios: data locality optimization serves as a software-only application example, whereas a dedicated Software-dened Radio architecture is presented as an example for an architecture based on the framework sketched in this paper.