Self-awareness in heterogeneous, adaptive many-core architectures enabling proactive, self-optimizing systems

Adaptive, heterogeneous many-core architectures, such as the Digital on-Demand Computing Organism (DodOrg), consisting of hundreds of cores that can be reconfigured with dedicated functionality at runtime, can provided high computing power as well as high energy efficiency, when compared to homogeneous many-core architectures. Although promising high computing power, the high complexity of adaptive many-core architectures pose the challenge of management and efficient utilization to developers and administrators. Hence, the goal of this thesis is hiding this complexity by applying Organic Computing principles, exploiting self-x properties, which include self-optimization and self-configuration, and thereby easing efficient utilization and management. This thesis presents a novel, holistic approach for realizing a self-optimizing as well as a proactive system behavior within heterogeneous, adaptive many-core architectures and addresses the following challenges. The first challenge is the realization of self-awareness within the observer component. Selfawareness is the basic property of self-organizing systems, which describes the ability of a system to determine, evaluate, and classify its current state. This challenge includes the development of a dedicated, scalable monitoring infrastructure for coordinated, cooperative, continuous, and system-wide system observations, which has to be capable of processing monitoring data in real-time and handle the heterogeneity of the underlying hardware. The second challenge is realization of a self-optimizing system behavior. In self-organizing systems, the system should autonomously learn at runtime, which optimization to perform in which situation for optimizing the overall system state. Finally, the last challenge is achieving a proactive system behavior. Proactive systems can predict future system states and can initiate system changes in advance for optimizing the system or for avoiding bad or harmful system states. Basis of the proposed approach is a flexible, hierarchical monitoring infrastructure for sustained monitoring of an entire adaptive, heterogeneous computing system, performing data aggregation and filtering on hardware level and therefore reducing the amount of data that must be analyzed on higher levels. A novel, light-weight, rule-based approach addresses the realization of self-awareness, providing the system with the capability to evaluate its current system state autonomously. Based on self-awareness, a proactive, self-optimizing system behavior is realized by exploiting the regular behavior of the DodOrg-Architecture. Within this thesis, a Learning Classifier System (LCS)-based approach, a reinforcement learning-based method, addresses the challenge of realizing a self-optimizing behavior and a Run Length Encoding (RLE) Markov predictor the challenge of creation of a predictive model for prediction of future system states. The entire approach was evaluated using a many-core simulation infrastructure and benchmarks from the MiBench benchmark suite. Achieved results demonstrate that the LCT-based self-optimizing system shows a performance improvement of 9.3% in comparison to the nonoptimizing many-core system. Finally, using the proactive approach, the performance improvement increases to 11.3% in comparison to the non-optimizing system.