Learning-Based Peak Temperature Optimization for Multicore Pipelined Hard Real-Time Systems

Abstract

Thermal-timing coupling in multicore pipelined systems makes it difficult to reduce peak temperature while satisfying hard real-time constraints. Periodic Thermal Management (PTM), which cyclically switches cores into low power mode, offers a deterministic and analyzable control strategy for thermal management in multicore processors. While existing PTM algorithms based on threshold or heuristic methods simplify the problem into a linearly constrained optimization for theoretical analysis, thereby neglecting the inherently nonlinear nature of the original optimization problem. To address this, we propose Autonomous Learning PTM (ALPTM), an online framework based on TD3 that formulates PTM optimization as a continuous-action Markov decision process, explicitly modeling the nonlinear coupled dynamics between thermal evolution, service curves, and queueing behavior. Experiments on representative streaming applications demonstrate that ALPTM consistently preserves hard real-time correctness and achieves significantly lower peak temperatures compared to existing PTM-based methods.