TRACK 01 / 06 • HYPERSCALE CORE ARCHITECTURE

🌐 Network for AI & Hyperscale Infrastructure Architect Learning Path¶

🚀 Elite Infrastructure Masterclass: Design, build, and operate non-blocking AI training fabrics (RoCEv2, PFC, ECN), 5-stage BGP Clos fabrics (RFC 7938), Segment Routing Ti-LFA backbones, and EVPN-VXLAN ESI multihomed clusters on real Arista cEOS containers.

📊 Learning Path Overview¶

Metric	Target Specification
Estimated Completion Time	40 – 50 Hours (Self-paced, hands-on lab driven)
Milestone Stages	6 Progressive Stages (Underlay → BGP Core → SR-MPLS → AI/EVPN Fabrics → Telemetry → Capstone)
Lab Framework	Containerlab + Arista cEOS (Runs 100% locally on macOS OrbStack or Linux Docker)
Target Roles	Architect - Network for AI, Hyperscale Infrastructure Architect, Principal Network Engineer, Core Backbone Architect
Target Employers	Hyperscalers (Google, Meta, AWS, Microsoft), AI Supercomputing Labs (NVIDIA, OpenAI, Anthropic), OEM Titans (HPE/Aruba, Arista, Cisco), and Tier-1 Service Providers

🎯 Industry Alignment: The AI & Hyperscale Revolution¶

Modern AI training clusters (LLM pre-training, mixture-of-experts, distributed GPU compute) have fundamentally shifted networking requirements from traditional enterprise designs to lossless, high-radix, ultra-low-latency fabrics.

This learning path directly mirrors the production competencies required by leading architecture positions (such as the HPE Architect - Network for AI, Routing and Automation and Meta Production Network Architect roles):

⚡ AI Fabric & Lossless Transport
- Zero Packet Drop: Priority Flow Control (PFC 802.1Qbb) to prevent buffer overflows
- Congestion Avoidance: ECN (RFC 3168) & WRED marking before pause storms
- Ultra-Low Latency: RDMA over Converged Ethernet (RoCEv2) for GPU memory access
- Buffer Sizing: Incast mitigation and dynamic headroom partition sizing
🌐 Hyperscale Clos Datacenter Fabrics
- Non-Blocking Scale: 5-Stage Clos topology math supporting 16,384+ GPUs
- Overlay Routing: EVPN-VXLAN (RFC 8365 / RFC 7432) with Symmetric IRB
- Open Multihoming: ESI All-Active multihoming replacing proprietary MLAG/vPC
- Multi-Pod Interconnect: VXLAN DCI and seamless inter-fabric data movement
🛣️ High-Radix Routing & Backbone
- Datacenter BGP: RFC 7938 leaf-spine BGP design with per-tier private ASNs
- BGP Unnumbered: RFC 5549 IPv4 peering over IPv6 Link-Local interfaces
- Segment Routing: SR-MPLS with SRGB 16000–23999 and Node/Prefix SIDs
- Sub-50ms Protection: Topology-Independent LFA (Ti-LFA) Fast Reroute
🤖 NetDevOps & Telemetry Automation
- Push Observability: Sub-second gNMI streaming telemetry over gRPC HTTP/2
- Standardized Schemas: Multi-vendor OpenConfig YANG telemetry models
- Time-Series Monitoring: Prometheus metrics scraping and Grafana dashboards
- Automated Verification: Cisco PyATS/Genie pre- and post-maintenance test suites

🗺️ 6-Stage Progressive Milestone Roadmap¶

01

Stage 1 · Underlay Routing & High-Performance Fabrics

Foundation

Establish deterministic ECMP load-balancing, sub-second convergence, and point-to-point link-state fabrics without DR/BDR election overhead.

OSPFv2/v3 Area 0 IS-IS Wide Metrics RFC 5305 ECMP Hashing

02

Stage 2 · Enterprise Edge & Hyperscale BGP-4 Core

Control Plane

Scale the control plane across massive leaf-spine fabrics with blast-radius containment, 10-step path selection, and iBGP route reflectors.

RFC 7938 BGP Route Reflectors 10-Step Decision BGP Communities Dual-Homed DIA

03

Stage 3 · Backbone Transport & Segment Routing

Transport Core

Eliminate LDP and RSVP-TE state bloat using Source Routing, establish multi-tenant MP-BGP VPNv4, and guarantee sub-50ms failover.

SR-MPLS SRGB 16000-23999 Prefix SIDs Ti-LFA Sub-50ms MP-BGP VPNv4

04

Stage 4 · AI & Datacenter Fabrics (EVPN-VXLAN + Lossless)

AI Flagship

Deploy high-radix leaf-spine fabrics with distributed Symmetric IRB routing, vendor-neutral ESI multihoming, and lossless RoCEv2/PFC tuning for GPU clusters.

RoCEv2 Lossless PFC 802.1Qbb ECN RFC 3168 EVPN-VXLAN ESI Multihoming

05

Stage 5 · NetDevOps & Real-Time Streaming Telemetry

Observability

Replace 5-minute SNMP polling with sub-second gRPC push streams, standardized OpenConfig YANG models, Prometheus alerting, and PyATS verification.

gNMI gRPC Protobuf OpenConfig YANG Prometheus Grafana PyATS Assertions

06

Stage 6 · Capstone System Design & Failure Triage Drills

Architecture Mastery

Tackle real-world hyperscale scaling calculations for 16,384+ GPU fabrics, debug silent packet drop and PFC deadlock, and defend designs in staff-level interview drills.

5-Stage Clos Sizing PFC Deadlock Mitigation CoPP Defense BGP Unnumbered RFC 5549

🚀 Interactive Lesson Directory (Click Any Lesson to Start)¶

Milestone Stage	Key Protocol Focus	Clickable Lessons & Hands-on Labs	Runnable Lab	Action
Stage 1 `Underlay Routing`	OSPFv2/v3, IS-IS Wide Metrics, Point-to-Point Adjacencies, ECMP	• 01 · Link-State Routing Foundations • 02 · OSPF Multi-Area Core Architecture • 03 · IS-IS Backbone Engineering • 05 · IGPs at Hyper-Scale	`labs/igp-lab`	Start Stage 1 →
Stage 2 `BGP-4 Core & Edge`	RFC 7938 BGP Clos, 10-Step Decision, Route Reflectors, Multi-Homing	• Lab 01 · eBGP, iBGP & next-hop-self • Lab 02 · iBGP over IS-IS Underlay • Lab 03 · Scalable Route Reflectors • Lab 04 · Multihomed BGP Edge • DIA Lab 01 · Multi-Provider Transit	`labs/bgp-lab`	Start Stage 2 →
Stage 3 `Backbone & SR-MPLS`	MP-BGP VPNv4, SRGB 16000–23999, Prefix SIDs, Sub-50ms Ti-LFA	• MPLS Lab 01 · MPLS + LDP Underlay • MPLS Lab 02 · Single-AS L3VPN & VRF • SR Lab 01 · SR-MPLS Node & Prefix SIDs • SR Lab 02 · Ti-LFA Sub-50ms FRR • SR Lab 03 · BGP Color Traffic Steering	`labs/segment-routing-lab`	Start Stage 3 →
Stage 4 `AI & EVPN Fabrics`	Lossless RoCEv2, PFC 802.1Qbb, ECN, Symmetric IRB, ESI Multihoming	• EVPN Lab 01 · Pure Layer-2 VNI • EVPN Lab 02 · Symmetric IRB Routing • EVPN Lab 03 · ESI All-Active Multihoming • EVPN Lab 04 · EVPN-VPWS & E-LAN • EVPN Lab 05 · EVPN DCI Multi-Site	`labs/evpn-datacenter-lab`	Start Stage 4 →
Stage 5 `NetDevOps & Telemetry`	gNMI Streaming Protobuf, OpenConfig YANG, Prometheus, PyATS Assertions	• NetDevOps Lab 01 · Jinja2/YAML Modeling • NetDevOps Lab 02 · PyATS Assertions • Telemetry Lab 01 · gNMI & OpenConfig • Telemetry Lab 02 · pygnmi Python Streams • Telemetry Lab 04 · Real-Time Grafana	`labs/telemetry-lab`	Start Stage 5 →
Stage 6 `Capstone System Design`	5-Stage Clos Sizing, Buffer Exhaustion, CoPP Defense, BGP Unnumbered	• System Design Masterclass & Scenario Drills • Security Lab 01 · CoPP CPU Protection • IPv6 Lab 02 · BGP Unnumbered (RFC 5549)	`labs/security-lab`	Start Stage 6 →

🧪 Detailed Milestone Curricula & Verification Gates¶

📍 Stage 1: Underlay Routing & High-Performance Fabrics¶

Core Focus: Deterministic equal-cost multi-pathing (ECMP), sub-second convergence, and carrier-grade link-state protocols.
Protocol Mechanics: OSPFv2 LSA types ½/⅗, Point-to-Point network types (bypassing DR/BDR election latency), IS-IS Level-1/Level-2 hierarchy, and TLV-based wide metric extensions (RFC 5305).
Interactive Labs:
Local Runner:
```
cd labs/igp-lab
./run.sh --guided
```
Milestone Gate: Full bidirectional reachability across all loopbacks with ECMP load balancing verified by automated tests.

📍 Stage 2: Enterprise Edge & Hyperscale BGP-4 Core¶

Core Focus: Autonomous System boundaries, multi-homed transit edge, and massive-scale datacenter routing based on RFC 7938.
Protocol Mechanics: 10-step BGP decision algorithm (Weight → Local Pref → AS-PATH → Origin → MED → eBGP over iBGP), iBGP full-mesh avoidance via Route Reflectors (cluster-id, originator-id), BGP communities for traffic engineering, and BGP Unnumbered over IPv6 Link-Local (RFC 5549).
Interactive Labs:

Local Runners:

cd labs/bgp-lab && ./run.sh --guided
cd labs/bgp-dia-lab && ./run.sh --guided

Milestone Gate: Route reflectors successfully reflect routes without routing loops; primary/backup egress traffic shifts deterministically during simulated provider transit failure.

📍 Stage 3: Backbone Transport & Segment Routing (SR-MPLS)¶

Core Focus: Removing control plane state (eliminating LDP and RSVP-TE) using Source Routing, establishing multi-tenant VRF isolation, and guaranteeing sub-50ms failover.
Protocol Mechanics: MPLS label stacks, Penultimate Hop Popping (PHP, Implicit Null label 3), MP-BGP VPNv4 with Route Distinguishers (RD) and Route Targets (RT), Segment Routing Global Block (SRGB 16000–23999), Node SIDs, Adjacency SIDs, and Topology-Independent Loop-Free Alternate (Ti-LFA).
Interactive Labs:

Local Runners:

cd labs/mpls-l3vpn-lab && ./run.sh --guided
cd labs/segment-routing-lab && ./run.sh --guided

Milestone Gate: Zero packet drop beyond 50ms during core link cutover; automated test verifies backup repair path pre-programmed in forwarding plane.

📍 Stage 4: AI & Datacenter Fabrics (EVPN-VXLAN + Lossless Ethernet)¶

Core Focus: High-radix leaf-spine Clos fabrics, multi-tenant overlay routing, and lossless transport required for high-throughput AI GPU training (RoCEv2).
Protocol Mechanics:
- EVPN-VXLAN: Symmetric Integrated Routing & Bridging (IRB), Anycast Virtual Gateway, Ethernet Segment Identifier (ESI) Type-0/Type-1 All-Active Multihoming (eliminating proprietary MLAG/vPC), EVPN Route Types 2 (MAC/IP), 3 (Inclusive Multicast), 4 (Ethernet Segment), and 5 (IP Prefix).
- Lossless AI Transport: Priority Flow Control (PFC, IEEE 802.1Qbb) to prevent packet drop on GPU ingress buffers; Explicit Congestion Notification (ECN, RFC 3168) with Random Early Detection (WRED) to signal bottleneck congestion before pause frames trigger PFC deadlocks.
Interactive Labs:

Local Runner:

cd labs/evpn-datacenter-lab
./run.sh --guided

Milestone Gate: Dual-homed servers actively hash traffic across independent leaf switches via ESI without loops; zero packet loss on simulated failovers.

📍 Stage 5: NetDevOps & Real-Time Streaming Telemetry¶

Core Focus: Modernizing operations from human CLI typing and 5-minute SNMP polling to Infrastructure-as-Code and sub-second push telemetry.
Protocol Mechanics: gNMI (gRPC Network Management Interface) streaming over HTTP/2 with Protocol Buffers; OpenConfig standardized YANG schemas; Prometheus time-series scraping; Grafana real-time telemetry dashboards; and Cisco PyATS/Genie automated state assertions.
Interactive Labs:

Local Runners:

cd labs/netdevops-lab && ./run.sh --guided
cd labs/telemetry-lab && ./run.sh --guided

Milestone Gate: Micro-burst interface traffic spikes visible in Grafana within 250ms of generation; automated PyATS test suite validates network health pre- and post-deployment.

📍 Stage 6: Capstone System Design & Failure Triage Drills¶

Core Focus: Synthesis, capacity planning, and high-pressure incident mitigation.
Topics:
- Calculating oversubscription ratios for a 16,384 GPU cluster using 64-port 800G switches (5-stage Clos scaling math).
- Designing blast-radius containment boundaries using eBGP Private AS numbering schemes.
- Diagnosing silent packet drops caused by PFC deadlock and microburst buffer exhaustion.
- System design interview drills covering real trade-offs (e.g. RoCEv2 vs. InfiniBand vs. Ultra Ethernet Consortium).
Curriculum References:
- Google & Hyperscale System Design Masterclass
- NetForge Master Curriculum Architecture Roadmap

🛠️ Executable Local Lab Environment¶

NetForge Labs uses an automated step runner architecture. You never need to manually copy-paste hundreds of lines of syntax unless you choose to practice CLI typing.

# 1. Navigate to any lab directory
cd labs/evpn-datacenter-lab

# 2. Launch the guided interactive runner
./run.sh --guided

# Or deploy the complete verified topology in one command
./run.sh --all

Every runner provides: 1. Config Previews: Inspect exact Arista cEOS commands before execution. 2. Manual Practice Guidance: Exact syntax if you prefer manual configuration via clab exec. 3. Automated Verifiers: Instant health checks validating operational state tables, routes, and ping matrixes.

🎓 Career Defense: 3 Portfolio Projects You Can Present¶

Upon completing this learning path, you will possess concrete, reproducible projects you can demonstrate and defend in staff-level technical interviews:

Non-Blocking 5-Stage Clos Datacenter Fabric:
Defend your BGP ASN scheme (RFC 7938), BGP Unnumbered design, and ECMP hash symmetry.
Explain how your ESI All-Active multihoming design completely eliminated proprietary vendor lock-in (MLAG/vPC).
Lossless RoCEv2 AI Transport Architecture:
Walk interviewers through the exact buffer threshold configurations, headroom sizing math, and PFC/ECN tuning required to prevent packet loss under distributed model all-reduce operations.
Autonomous Sub-Second Fast Reroute Backbone:
Demonstrate how you migrated from legacy RSVP-TE to Segment Routing (SR-MPLS) with Ti-LFA, achieving deterministic sub-50ms failover without keeping per-flow state in the core.