Aging-Aware Runtime Management for Many-Core Systems

Transistor scaling has enabled the fabrication of devices implementing many-core architectures, but has introduced challenges such as the need to limit power consumption and increase reliability. These factors can have a significant impact on the degradation of devices through time and, as a consequence, managing aging effects has become an important design requirement. While dynamic power and runtime resource management are generally used for efficient execution and to control workloads to achieve optimal performance while satisfying power constraints, they can also be effective tools to improve the lifetime of the system.

This thesis presents an aging-aware runtime management approach is proposed to address the combined effects of traffic congestion and task execution on reliability degradation in many-core systems. The method relies on a predictive fitness-driven mapping and remapping strategy, where observed energy consumption in task computation and communication activity guide decisions to balance workload distribution and reduce stress on vulnerable components and the total system aging. By redistributing tasks accordingly, our approach ensures that remapping not only maintains application performance but also mitigates aging effects that accumulate during execution.

To address system aging and evaluate proposed management strategy, a dedicated simulation environment \textbf{CASTLE} is developed using a multi-agent model that captures the dynamic interactions of components existed in many-core systems in computation and communication induced aging phenomena. The proposed aging-aware runtime management strategy is evaluated and experimented in \textbf{CASTLE} simulation platform by applying different size of systems with typical applications, to be proved that it can enable more effective and efficient mitigation of aging effects caused by heavy computational load or traffic. Results demonstrate that incorporating predictive temperature modelling into runtime management significantly improves mitigation of aging effects, extending system lifetime without compromising application performance and showing the potential of runtime-aware strategies for ensuring the long-term reliability of future many-core systems.