NCP-NET Networking Professional Study Strategy¶

Study Approach¶

Phase 1: Foundation (1-2 weeks)¶

1. InfiniBand Core
2. Protocol stack and speed generations
3. Subnet management and routing
4. Partitioning and QoS

5. 📖 NVIDIA Networking Docs

6. RDMA and RoCE

7. RDMA operations and programming model
8. RoCE v1 vs v2
9. Lossless Ethernet (PFC, ECN, DCQCN)
10. GPUDirect RDMA
11. 📖 GPUDirect Docs

Phase 2: Advanced (2-3 weeks)¶

1. Spectrum Ethernet - Switch family, Cumulus Linux, DOCA
2. Topologies - Fat-tree, rail-optimized, dragonfly, adaptive routing
3. UFM - Monitoring, telemetry, diagnostics, fabric management

Phase 3: Exam Prep (1-2 weeks)¶

1. Practice scenarios and troubleshooting
2. Review specifications and commands
3. Focus on weak areas

Recommended Resources¶

• NVIDIA Networking Documentation - Complete reference
• UFM Documentation - Fabric management
• DGX SuperPOD Guide - Reference architecture
• NVIDIA DLI Courses - Official training
• NVIDIA Developer Blog - Technical articles

Exam Tactics¶

Keywords¶

• "Lossless" - PFC + ECN for RoCE
• "Isolation" - InfiniBand partitions (P_Keys)
• "Congestion" - Adaptive routing, ECN, DCQCN
• "Monitor" or "manage" - UFM
• "GPU communication" - GPUDirect RDMA, NVLink, NCCL
• "Large scale" - Dragonfly topology
• "DGX cluster" - Rail-optimized topology

Common Pitfalls¶

• RoCE v2 is routable (UDP/IP), RoCE v1 is L2 only
• PFC is per-priority, not per-port
• Adaptive routing should be per-flow for NCCL, not per-packet
• GPUDirect RDMA requires nvidia-peermem module
• InfiniBand uses P_Keys for isolation, not VLANs