ZeroDVFS: Zero-Shot LLM-Guided Core and Frequency Allocation for Embedded Platforms

TL;DR

ZeroDVFS presents a model-based hierarchical MARL framework for thermal- and energy-aware DVFS and task-to-core allocation on embedded multi-core platforms. It combines accurate environment modeling with two collaborating RL agents and a novel LLM-based semantic feature extraction pipeline to enable zero-shot deployment on unseen workloads and platforms. The approach yields up to $7.09\times$ energy efficiency and $4.0\times$ better makespan than Linux ondemand, with first decisions $8{,}300\times$ faster than exhaustive table profiling and sub-10 ms inference thereafter. Cross-platform experiments show promising generalization from TX2 to Orin NX and RubikPi, while transfer gaps can be mitigated with small-scale fine-tuning; LLM features offer substantial benefits for zero-shot generalization and scalable deployment in dynamic embedded settings.

Abstract

Dynamic voltage and frequency scaling (DVFS) and task-to-core allocation are critical for thermal management and balancing energy and performance in embedded systems. Existing approaches either rely on utilization-based heuristics that overlook stall times, or require extensive offline profiling for table generation, preventing runtime adaptation. We propose a model-based hierarchical multi-agent reinforcement learning (MARL) framework for thermal- and energy-aware scheduling on multi-core platforms. Two collaborative agents decompose the exponential action space, achieving 358ms latency for subsequent decisions. First decisions require 3.5 to 8.0s including one-time LLM feature extraction. An accurate environment model leverages regression techniques to predict thermal dynamics and performance states. When combined with LLM-extracted semantic features, the environment model enables zero-shot deployment for new workloads on trained platforms by generating synthetic training data without requiring workload-specific profiling samples. We introduce LLM-based semantic feature extraction that characterizes OpenMP programs through 13 code-level features without execution. The Dyna-Q-inspired framework integrates direct reinforcement learning with model-based planning, achieving 20x faster convergence than model-free methods. Experiments on BOTS and PolybenchC benchmarks across NVIDIA Jetson TX2, Jetson Orin NX, RubikPi, and Intel Core i7 demonstrate 7.09x better energy efficiency and 4.0x better makespan than Linux ondemand governor. First-decision latency is 8,300x faster than table-based profiling, enabling practical deployment in dynamic embedded systems.

Figures (10)

• Figure 1: Performance metrics (CPU utilization, makespan, energy consumption, branch misses, context switches, and average temperature) for NVIDIA Jetson TX2 running the FFT benchmark parallelized through OpenMP API. Results highlight different responses of performance metrics to changes in different frequencies and selected cores.
• Figure 2: MARL design: (a) Hierarchical action selection architecture and (b) reward function definitions.
• Figure 3: The simplified design of model-based RL for temperature- and energy-aware core allocation for multi-core processors.
• Figure 4: Comparison of (a) profiler data prediction using the one dimensional convolution model and (b) temperature prediction of 4 cores. Both use sensor data as ground truth on Intel Core i7 8th gen.
• Figure 5: Comparison of (a) makespan and (b) energy consumption over 100 training episodes on BOTS FFT benchmark with input size 262144 on Jetson TX2. ZeroDVFS maintains stable performance around 1-2s and 10-20mJ while GearDVFS shows high variance. Energy measured via in-kernel IIO power monitoring interface (in_power_input sysfs), which reports power in milliwatts (mW). Energy computed as $E = \sum P_i \cdot \Delta t_i$ where $\Delta t_i$ is the sampling interval (10ms). Y-axis units are millijoules (mJ).