Answers are grounded in this paper and cite their sources.
Introduction. Nodes in distributed edge clusters operating on backup power require reductions in power demand while preserving acceptable application performance. This study proposes a model-guided method to finding a power-efficient CPU operating frequency within a user-defined performance-loss limit, using standard Linux metrics and Running Average Power Limit (RAPL) energy measurements. Materials and Methods. The proposed approach uses the CMOS dynamic-power model to guide the selection of candidate frequency caps. Standard Linux metrics describe CPU utilization, operating frequency, and temperature where available. RAPL energy differences over measured time intervals provide average power for the selected hardware domain, while workload measurements establish performance loss. The search aims to minimize measured domain power subject to performance constraints. The method and model were evaluated against three different CPU models in OS Linux environment. Results. Preliminary characterization indicates that frequency capping does reduce domain power and energy per unit of work while retaining acceptable performance for workloads. Benefits and performance penalties vary with the workload and core configuration, but demonstrating visible power consumption reduce by maintaining acceptable performance per-watt, without blindly limiting RAPL domains. Conclusions. Non-synthetic (CPU bound, memory bound, cache bound and AI) workloads measurements demonstrate that algorithm selects optimal CPU core frequency based on the provided performance penalty threshold. Results were validated and presented for three different CPU models that are used in edge nodes: Raptor Lake, Meteor Lake and Zen5. During research with implemented distributed monitoring service as a part of research process, a potential firmware defect for voltage curve value was discovered on Raptor Lake platform which requires separate investigation.
No takes yet. Share an insight, caveat, or question.
V. V. Torshyn (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: