Reinforcement learning-driven task migration for effective temperature management in 3D noc systems

The advent of multi-core systems necessitates effective thermal and reliability control strategies to improve system dependability. The rise in power density and heat hotspots in multi-core systems presents considerable problems to reliability and performance. Current methodologies frequently lack scalability and do not account for long-term dependability effects. Notwithstanding its multiple benefits, 3D stacking elevates the power density per unit area of the chip, hence raising the chip temperature and introducing new problems. The rise in temperature will result in diminished dependability and performance decline, thus necessitating the construction of thermal management algorithms for these systems. This study presents an algorithm for this goal that is based on task migration. Choosing the migration destination for tasks on hot cores is a Complete-NP problem that can be addressed via heuristic approaches. We have employed Reinforcement Learning in the proposed strategy for this purpose. In selecting the migration location, we have also taken into account the migration overhead alongside the core temperature. The evaluation findings demonstrate that this strategy can decrease the maximum chip temperature by as much as 31% for the core with the highest task load, while its effect on performance is minimal.