A two-layer distributed MPC approach to thermal control of Multiprocessor Systems-on-Chip. (May 2022)

• Record Type: Journal Article
• Authors: Tilli, Andrea; Garone, Emanuele; Conficoni, Christian; Cacciari, Matteo; Bosso, Alessandro; Bartolini, Andrea
• Is Part Of: Control engineering practice. Volume 122(2022)
• Publication Date: 2022-05
• Subjects: Distributed MPC -- Many-core systems -- Thermal modeling; Automatic control -- Periodicals; 629.89
• DOI: 10.1016/j.conengprac.2022.105099
• Languages: English
• ISSNs: 0967-0661

Abstract

Abstract: Next-generation Multiprocessor, or Multicore, Systems-on-Chip offer very high computing performance at the expense of a very high power density unevenly distributed on the chip. The hot spots thus generated represent a significant source of performance and reliability degradation, as well as power consumption increase. In recent years, run-time thermal control strategies have been developed to deal with this issue by acting on some "computational knobs" (e.g., clock frequencies and supply voltages). In this context, schemes based on Model Predictive Control (MPC) are particularly suitable due to their capability to deal with constraints explicitly. In this paper, we first discuss relevant properties for the design of predictive controllers for thermal systems. Starting from the Partial Differential Equation (PDE) describing heat diffusion in a solid, we prove meaningful feasibility properties that can be leveraged for constraint reduction. We then present a procedure to derive approximated but effective modular thermal models intended to build an efficient distributed MPC. Finally, a two-layer control solution is proposed to maximize performance while preserving feasibility despite model approximations. The effectiveness of this approach is validated through extensive and realistic numerical simulations. Highlights: A distributed controller is proposed for the optimal thermal capping of MPSoCs. Control feasibility conditions are studied exploiting a PDE thermal model. A hysteresis controller is designed for decentralized ultimate thermal capping. A distributed MPC strategy is employed for performance optimization. The resulting two-layer solution is validated for relevant simulation benchmarks.