December 10, 2025Journal of Circuits Systems and Computers0 citations

Reinforcement-Learning-based Multi-Objective Constrained Optimization of Task Mapping for 3D MPSoC Systems

TLTiantian LiZhejiang University of Technology LLLiangyu LinZhejiang University of Science and Technology BCBin CaoBeijing University of Posts and Telecommunications

Key Points

Optimization improved energy consumption and performance in 3D MPSoC systems, indicating strong efficiency called for.
Energy consumption and performance metrics were enhanced through a novel multi-objective optimization approach based on reinforcement learning.
Implementation of a Markov decision process developed task mapping solutions, measured across various DAG applications to validate effectiveness.
This intelligent method highlights the flexibility of optimization strategies in complex 3D MPSoC environments, suggesting broader applicability in embedded systems.

Abstract

3D MPSoC system is a very promising design which largely improves the performance by stacking more processors vertically and lowering the communication delay with the limited chip size. However, the higher integration density also results in much more severe energy and thermal issues, which are equally important for embedded real-time applications. Existing optimization works either aim at 2D MPSoC systems which cannot be simply extended to 3D MPSoC systems due to their distinctive and complex structure, or only focus on partial of the three objectives (performance, energy consumption and temperature) and adopt sequential optimization, which cannot sufficiently explore the optimization space due to the close relationship among the objectives. Further, existing traditional mathematical or evolution-based heuristic methods have failed to effectively and efficiently deal with the complex and large-scale optimization problem in 3D MPSoC systems. To this end, we propose a more intelligent multi-objective constrained task mapping optimization approach with DVFS technique based on reinforcement learning specific for 3D MPSoC systems. Based on the adopted comprehensive models, we first construct the task mapping problem as a Markov decision process, then learn the mapping policy based on reinforcement learning, and finally generate the optimal task mapping solution leveraging the well-trained neural network. Extensive experiments conducted on real-world DAG applications with different settings verify the superiority of the proposed intelligent approach from perspectives of generalizability, scalability and flexibility. Three valuable conclusions are also obtained to provide insights for the application of this approach.

Ask AI

Helpful

Bookmark

View Full Paper