MIMO-RLS-LQR Adaptive Governor for Multi-Cluster DVFS

This paper introduces a novel control-theoretic adaptive governor for Dynamic Voltage and Frequency Scaling (DVFS) in modern heterogeneous multi-cluster processors. Standard Linux scaling governors, such as schedutil, fundamentally rely on heuristic look-up approaches that struggle to maintain optimal performance-thermal trade-offs under dynamic, highly variable workloads. To resolve this limitation, we present an internally closed loop framework that models multi-cluster frequency domain interaction as a multi-variable problem. The architecture integrates an online Multi-Input Multi-Output Recursive Least Squares (MIMO-RLS) estimator for real-time, non-stationary system identification, combined with a periodic Linear Quadratic Regulator (LQR) optimal control system computed over a six-dimensional state vector. Robust multi-layer safety thresholds are introduced to handle transient instabilities in userspace actuation. Experimental evaluations conducted on Raspberry Pi 5 (Broadcom BCM2712) and Orange Pi 5 Plus (Rockchip RK3588) platforms demonstrate measurable improvements in thermal stabilization (up to 5°C reduction in peak core temperatures) and systemic energy efficiency compared to default kernel behaviors, with acceptable execution time overhead.