The increasing power density of modern data centres makes thermal management a structural constraint on sustained computational performance. Conventional approaches typically rely on temperature-centric single-loop regulation or on optimisation-based supervisory schemes. Although effective in specific application contexts, such solutions often do not address in a structured manner the coordination between fast on-chip power control and slower cooling-capacity modulation, leading to overcooling, unnecessary energy expenditure, or avoidable performance limitations. This thesis proposes an Advanced Regulatory Control (ARC) perspective for data centre thermal–performance management. The underlying premise is that thermal safety and cooling efficiency can be achieved through explicit actuator allocation and time-scale separation, rather than through complex optimisation frameworks. To this end, a control-oriented model of a liquid-cooled CPU is developed, capturing the coupled electrical–thermal dynamics and the nonlinear interaction between power generation and heat removal. Based on this model, a coordinated multi-loop architecture is designed in which Dynamic Voltage and Frequency Scaling (DVFS) acts as a fast thermal protection layer, while coolant flow regulation operates as a slower capacity modulation layer. A supervisory constraint on the temperature rise across the water block and a bias term based on thermal margin complete the structure. The resulting scheme adopts established ARC principles—cascade interaction, override logic, and hierarchical tuning—adapted to the context of computing systems. A structural equilibrium analysis shows that the admissible operating region of the system is intrinsically constrained by actuator limits and by the nonlinear coupling between electrical and thermal domains. This analysis indicates that cooperative regulation between DVFS and coolant flow is not merely a design preference, but a structural necessity imposed by system physics. Simulation results confirm that the proposed architecture enforces prescribed thermal constraints, enables tracking of the governor-requested frequency whenever physically admissible, and reduces overcooling compared with fixed-flow strategies. These results are obtained using proportional–integral regulators and an explicit bandwidth hierarchy, ensuring practical deployability and compatibility with existing industrial control practice. The work demonstrates that Advanced Regulatory Control can provide a structured and robust methodological foundation for data centre thermal–performance management, bridging process control principles and modern computing infrastructures.
La crescente densità di potenza dei moderni data centre rende la gestione termica un vincolo strutturale al mantenimento delle prestazioni computazionali nel tempo. Gli approcci convenzionali si basano tipicamente su regolazioni a singolo anello centrate sulla temperatura oppure su schemi supervisori basati su ottimizzazione. Sebbene efficaci in specifici contesti applicativi, tali soluzioni spesso non affrontano in modo strutturato il coordinamento tra il controllo rapido della potenza on-chip e la modulazione più lenta della capacità di raffreddamento, con conseguenti fenomeni di sovraraffreddamento, spreco energetico o limitazioni evitabili delle prestazioni. La presente tesi propone una prospettiva di Advanced Regulatory Control (ARC) applicata alla gestione termo-prestazionale dei data centre. L’idea di fondo è che sicurezza termica ed efficienza del raffreddamento possano essere conseguite mediante un’esplicita allocazione degli attuatori e una separazione delle scale temporali, piuttosto che attraverso schemi di ottimizzazione complessi. A tal fine viene sviluppato un modello orientato al controllo di una CPU raffreddata a liquido, in grado di descrivere le dinamiche accoppiate elettrico–termiche e l’interazione non lineare tra generazione di potenza e rimozione del calore. Sulla base di tale modello viene progettata un’architettura multi-anello coordinata, nella quale il Dynamic Voltage and Frequency Scaling (DVFS) svolge il ruolo di livello rapido di protezione termica, mentre la regolazione della portata del fluido refrigerante opera come livello più lento di modulazione della capacità di raffreddamento. Un vincolo supervisore sulla differenza di temperatura attraverso il water block e un termine di bias basato sul margine termico completano la struttura. Lo schema risultante riprende principi consolidati dell’ARC — interazione a cascata, logica di override e taratura gerarchica — adattandoli al contesto dei sistemi di calcolo. Un’analisi strutturale dell’equilibrio evidenzia come la regione operativa ammissibile sia intrinsecamente vincolata dai limiti degli attuatori e dall’accoppiamento non lineare tra dominio elettrico e termico. Tale analisi mostra che la regolazione cooperativa tra DVFS e portata del refrigerante non rappresenta una scelta progettuale arbitraria, bensì una necessità strutturale imposta dalla fisica del sistema. I risultati di simulazione confermano che l’architettura proposta mantiene i vincoli termici prescritti, consente il tracciamento della frequenza richiesta dal governor quando fisicamente ammissibile e riduce il sovraraffreddamento rispetto a strategie a portata fissa. Tali risultati sono ottenuti mediante regolatori proporzionale–integrali e una gerarchia esplicita delle bande, garantendo implementabilità e compatibilità con le pratiche industriali esistenti. Il lavoro dimostra come l’Advanced Regulatory Control possa costituire una base metodologica strutturata e robusta per la gestione termo-prestazionale dei data centre, integrando principi del controllo di processo con le esigenze delle moderne infrastrutture di calcolo.
Advanced regulatory control for thermal-performance management in high-power data centre computing nodes
DIONIGI, FEDERICO
2025/2026
Abstract
The increasing power density of modern data centres makes thermal management a structural constraint on sustained computational performance. Conventional approaches typically rely on temperature-centric single-loop regulation or on optimisation-based supervisory schemes. Although effective in specific application contexts, such solutions often do not address in a structured manner the coordination between fast on-chip power control and slower cooling-capacity modulation, leading to overcooling, unnecessary energy expenditure, or avoidable performance limitations. This thesis proposes an Advanced Regulatory Control (ARC) perspective for data centre thermal–performance management. The underlying premise is that thermal safety and cooling efficiency can be achieved through explicit actuator allocation and time-scale separation, rather than through complex optimisation frameworks. To this end, a control-oriented model of a liquid-cooled CPU is developed, capturing the coupled electrical–thermal dynamics and the nonlinear interaction between power generation and heat removal. Based on this model, a coordinated multi-loop architecture is designed in which Dynamic Voltage and Frequency Scaling (DVFS) acts as a fast thermal protection layer, while coolant flow regulation operates as a slower capacity modulation layer. A supervisory constraint on the temperature rise across the water block and a bias term based on thermal margin complete the structure. The resulting scheme adopts established ARC principles—cascade interaction, override logic, and hierarchical tuning—adapted to the context of computing systems. A structural equilibrium analysis shows that the admissible operating region of the system is intrinsically constrained by actuator limits and by the nonlinear coupling between electrical and thermal domains. This analysis indicates that cooperative regulation between DVFS and coolant flow is not merely a design preference, but a structural necessity imposed by system physics. Simulation results confirm that the proposed architecture enforces prescribed thermal constraints, enables tracking of the governor-requested frequency whenever physically admissible, and reduces overcooling compared with fixed-flow strategies. These results are obtained using proportional–integral regulators and an explicit bandwidth hierarchy, ensuring practical deployability and compatibility with existing industrial control practice. The work demonstrates that Advanced Regulatory Control can provide a structured and robust methodological foundation for data centre thermal–performance management, bridging process control principles and modern computing infrastructures.| File | Dimensione | Formato | |
|---|---|---|---|
|
Executive_Summary__Federico_Dionigi.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
805.29 kB
Formato
Adobe PDF
|
805.29 kB | Adobe PDF | Visualizza/Apri |
|
Tesi_Federico_Dionigi.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
9.18 MB
Formato
Adobe PDF
|
9.18 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/251271