PROJ Service Migration in Cellular Networks
This project has three focus parts: Cellular Networks Testbed, LLMs, and optimization algorithms.
Cellular Networks Testbed
Image from: Real-Time Service Migration in Edge Networks: A Survey
The testbed architecture spans four tiers: Central Cloud, Regional MEC, Aggregation MEC, and Local MEC. Primary focus: Regional MEC and Aggregation MEC. Regional MEC, Aggregation MEC, and Local MEC should be deployed to my PVE Testbed.
Each MEC site (or per Metro (metropolitan area) / PoP (Point of Presence)) needs its own control plane because:
- Survivability: If WAN/backhaul drops, the site keeps running. A single, stretched cluster loses control-plane access and flakes.
- Latency/etcd constraints: Kubernetes control-plane (etcd) hates WAN latency/packet-loss; cross-site RTTs >~5–10 ms and jitter cause elections and outages.
- Blast radius & upgrades: Failures and rollouts stay local, enabling per-site upgrades.
- Regulatory / tenancy: Site-level isolation simplifies policy and compliance.
TODO Central Cloud
>200 km coverage | ≥50 ms RTT Global control
Cloud (Azure) High-level design: Azure Virtual WAN (Standard) with four hubs in a full inter-hub mesh. Regional spokes (AKS VNets) attach to their nearest hub; inter-hub routing provides global any-to-any.
Regions (paired for HA/DR):
- East US 2 (VA) — primary; paired with Central US (closest to UVA)
- Central US (IA) — DR for East US 2
- West US 3 (AZ) — west capacity/DR; paired with East US
- East US (VA) — additional east capacity and the formal pair for West US/West US 3
(All selected regions provide Availability Zones.)
TODO Regional MEC (e.g., Richmond PoP)
50–200 km coverage | RTT to Aggregation 15–30 ms Use cases: smart city, cloud gaming, content delivery Components: SMF/AMF/PCF (control plane) + Regional UPF
TODO Aggregation MEC (Hub PoP)
10–50 km coverage | RTT to Local: 10–20 ms Use cases: campus control, local CDN
OKD (3 master nodes) Components: SMF/AMF/PCF (control plane) + optional Aggregation UPF
TODO Local MEC
1–5 km coverage | UE↔UPF target: ≤5–10 ms Use cases: autonomous driving, industrial robots, real-time AR/XR
Components: Local UPF Deploy two OKD SNOs or MicroShift clusters (MEC-1: Campus South; MEC-2: Campus North)
N3 (gNB ⇄ UPF @ MEC): VLAN/VRF local to the site, low jitter N6 (UPF ⇄ campus/ISP): routed toward the PoP
TODO gNodeB (UEs)
-
Digital Twin
Must implement N2, N3, and optionally Xn. Focus on mmWave.
-
Physical RAN
CBRS band only
LLMs
Focus on multi-agent workflow design and LLM fine-tuning.
TODO LLM corpus
-
3GPP Standards
-
ETSI = European Telecommunications Standards Institute
Famous work includes ETSI MEC (edge computing) and the original ETSI NFV effort.
-
Kubernetes documentation
-
OKD documentation
TODO LLM Serving
Deploy Three LLMs to Regional MEC or Aggregation MEC:
- Custom LLM (based on Google Gemma) fine-tuned on UVA CS Slurm cluster.
- Codex CLI / Gemini CLI / Droid CLI
- Time-Series LLM (TBD)
https://ai-on-openshift.io/generative-ai/llm-serving/
TODO Retrieval Augmented Generation (RAG)
TODO The Multi-Agent Roster & Roles
| Agent | Location in Testbed | Primary Role | Key Inputs | Research Motivation |
|---|---|---|---|---|
| Mobility Predictor Agent (MPA) | Aggregation MEC / Local MEC (Near-RT RIC/O-RAN Layer) | Context Generation: Provides real-time prediction of UE handover and mobility patterns to anticipate service relocation. | Real-time Radio KPIs (RSRP, RSRQ), Handover/Xn/N2 events, UE location/velocity. | Proactive Migration: Essential for timely initiation of migration at the lowest latency tiers, ensuring QoE under high mobility. |
| MEC Resource Agent (RCA) | All Managed MEC Sites (Local, Aggregation, Regional) | Local State Reporting: Monitors the instantaneous resource utilization and available capacity of its local compute cluster (OKD/MicroShift) | CPU/Memory/GPU load, Available network bandwidth, K8s/OKD/MicroShift node metrics. | Survivability and Autonomy: Guarantees that every control-plane instance has local resource awareness, upholding isolation and independence |
| Migration Planner Agent (PLA) | Regional MEC and Aggregation MEC (Control Plane) | Decision-Making: Determines the optimal migration target, timing, and method based on its scope (Local → Local vs. Regional → Regional). | Aggregated Predictions (MPA data), Resource Availability (RCA reports), Service SLOs, Migration Cost Model. | Hierarchical/Decentralized Decision: Enables ultra-low-latency decision-making for local PoP movements and wide-area optimization, avoiding high Central Cloud RTT |
| State/Traffic Steering Agent (TSA) | Co-located with SMF/UPF | Execution & Cutover: Executes the migration by coordinating state transfer and updating the 5G Core traffic rules | PLA’s Decision (Target MEC ID), State Transfer Status, 5G Core N4/N11 APIs (for UPF/SMF control plane updates) | Critical Service Continuity: Directly implements the necessary 5G Core control procedures (PSA Relocation/UL-CL) at all anchor points to shift traffic seamlessly |
| Policy Enforcement Agent (PEA) | Central Cloud (Azure) | Global Policy Management: Distributes high-level, long-term policies, cost objectives, and optimization models across all PLA instances | Long-term historical data, Global business objectives, Failure tolerance settings, Regulatory/Tenancy policies. | Global Governance: Provides the top-level goals and learning feedback to the decentralized PLA instances, ensuring consistency and alignment with global business objectives. |
TODO Optimization Algorithms
Prioritize time series forecasting and decision making (TBD).