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

Area of Science:

• Computer Engineering
• Reliability Engineering
• Thermal Management

Background:

• Multi-core systems face reliability and performance challenges due to increasing power density and heat hotspots.
• 3D stacking exacerbates these issues by increasing chip temperature, demanding advanced thermal management solutions.
• Existing thermal control strategies often lack scalability and long-term dependability considerations.

Purpose of the Study:

• To develop an effective thermal management algorithm for multi-core systems, focusing on enhancing system dependability.
• To address the challenge of task migration for hot cores, a Complete-NP problem, using heuristic approaches.
• To minimize chip temperature and mitigate performance degradation in advanced computing architectures.

Main Methods:

• A novel algorithm based on task migration is proposed for thermal management.
• Reinforcement Learning (RL) is employed as a heuristic approach to solve the task migration destination problem.
• The algorithm considers both core temperature and migration overhead in selecting migration locations.

Main Results:

• The proposed strategy significantly reduces the maximum chip temperature, achieving up to a 31% decrease for the core with the highest task load.
• The impact of the task migration strategy on system performance is minimal.
• The approach demonstrates effectiveness in managing thermal challenges in high-density computing environments.

Conclusions:

• Reinforcement Learning-based task migration is a viable and effective strategy for thermal management in multi-core systems.
• The algorithm successfully balances thermal reduction with performance preservation.
• This approach offers a scalable solution for improving the dependability of modern high-performance computing systems.