Realizing a Proactive, Self-Optimizing System Behavior within Adaptive, Heterogeneous Many-Core Architectures

For maintaining high performance and minimizing power consumption, adaptive, heterogeneous many-core architectures can be adapted at runtime to changing environmental requests or conditions as well as to changes resulting from the dynamics of the workload itself. However, the huge complexity of such architectures makes their optimization very challenging at runtime. This challenge is therefore addressed within this paper by an Organic Computing approach for realizing a proactive, self-optimizing system behavior within adaptive, heterogeneous systems using a light-weight Learning Classifier System and a Run Length Encoding Markov predictor. The first realizes a self-optimizing behavior, freeing the user from the burden of optimizing the system manually, and the latter captures the system behavior, permits prediction of future system states, and therefore permits exploiting regular behavior for further improving the overall system performance. Using the use case of optimizing the overall system performance, results showed that the proactive, self-optimizing system achieved a performance improvement of 11.3% in comparison to a non-optimizing system.