PROJ Service Migration in Cellular Networks
This project has three focus parts: Cellular Networks Testbed, LLMs, and optimization algorithms.
Cellular Networks Testbed
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.
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
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
- Time-Series LLM (TBD)
https://ai-on-openshift.io/generative-ai/llm-serving/
Optimization Algorithms
Prioritize time series forecasting and decision making.