Integrating Organic Computing Mechanisms into a Task-based Runtime System for Heterogeneous Systems

Abstract

Modern computer architectures feature a high degree of parallelism and heterogeneity. They are deployed in different fields of application, which leads to different constraints and optimization goals. Additionally, the current and future system states have to be considered, making dynamic and proactive adaptations necessary. Organic Computing techniques offer a solution for this. A combination with runtime systems, which control execution and monitor the system, sensibly provides reduction in system complexity, efficient resource usage and the ability to dynamically adapt. To enable Organic Computing in runtime systems, we first study heterogeneous systems in different fields of application. As we identified dependability as a major concern, we study symptom-based fault detection, a light-weight technique to detect faults. We develop a mechanism based on rule-based machine learning to consider the identified requirements and constraints and dynamically balance contradicting optimization goals. Additionally, we present a scheduling mechanism to globally optimize several instances of a runtime system and show first results.