CBL - Campus del Baix Llobregat

Projecte llegit

Títol: KPI-Aware Path Computation and Dynamic Provisioning of Deterministic Services

Estudiants que han llegit aquest projecte:

Director/a: SPADARO, SALVATORE

Departament: TSC

Títol: KPI-Aware Path Computation and Dynamic Provisioning of Deterministic Services

Data inici oferta: 28-02-2025     Data finalització oferta: 31-10-2025


Estudis d'assignació del projecte:
    DG ENG AERO/SIS TEL
Tipus: Individual
 
Lloc de realització: EETAC
 
Segon director/a (UPC): PAGÈS CRUZ, ALBERT
 
Paraules clau:
Deterministic services, KPI
 
Descripció del contingut i pla d'activitats:
The main aim of the project proposal will be to design and implement a tool based on theoretical approaches to compute E2E KPIs to assure deterministic services with pre-defined requirements.
 
Overview (resum en anglès):
Current research on beyond-5G/6G systems and networks addresses business cases requiring deterministic guarantees, such as Industry 4.0/5.0 and autonomous guided vehicles. These applications impose strict Key Performance Indicator (KPI) control to ensure predictable service performance, controlled jitter, and near-zero packet loss, driving the rise of Deterministic Networking concepts that add deterministic capabilities to network infrastructures. In general, optical networks are regarded as enablers of latency and jitter-bounded communications for their intrinsic deterministic nature, exemplified in Industry 4.0 frameworks where optical transport layers enable remote smart factory control, reducing CAPEX and OPEX. However, guaranteeing stringent KPIs across hybrid infrastructures that combine transparent optical transport with Ethernet edges is non-trivial: queue interactions, traffic bursts, and link failures invalidate static provisioning. This thesis addresses this challenge through complementary contributions that build a predictive, KPI-aware control framework that reconciles deterministic guarantees with operational agility.
First, a new analytical foundation is established, providing closed-form stochastic models to predict latency, jitter, and loss distributions within milliseconds. Second, these models are embedded into a stateless, Software Defined Networking (SDN)-style Path Computation Element (PCE) equipped with a network-wide impact-assessment check, preventing KPI regressions for incumbent services. Third, the framework's efficacy is demonstrated in two diverse, high-stakes use cases-a smart-factory industrial IoT network and a fly-by-light avionics backbone-showing significant gains in service acceptance and resilience. Fourth, the PCE is enhanced with machine learning; a Deep Sets surrogate provides sub-millisecond KPI estimation, while a Light Gradient Boosting Machine (LightGBM) ensemble forecast transient KPI evolution, enhancing the controller's proactive capabilities.
The results demonstrate that rigorous KPI guarantees can coexist with millisecond-class agility across diverse applications. The framework achieves up to 60% improvement in service acceptance for industrial networks and enables 30-ms restoration of critical flight-control services following failures, while reducing physical resource requirements. A key finding reveals that hybrid analytical-ML approaches provide optimal synergy, combining the interpretability and level of assurance of formal models with the speed and foresight required for proactive control.
This thesis delivers a complete, model-driven control framework that enables deterministic performance to be achieved dynamically and efficiently in complex, hybrid-technology networks, establishing the foundation for next-generation time-sensitive applications.

© CBLTIC Campus del Baix Llobregat - UPC