32.2d Network provisioning: automated switch configuration and fabric setup via BCM¶

📘 Context Introduction¶

In an AI cluster, the network is the backbone that connects compute nodes, storage systems, and management services. Without proper network provisioning, even the most powerful GPUs will sit idle waiting for data. NVIDIA Base Command Manager (BCM) simplifies this by automating switch configuration and fabric setup — turning what used to be a manual, error-prone process into a repeatable, reliable operation.

For new engineers, think of network provisioning as the "plumbing" of your AI cluster. BCM handles the configuration of network switches (the valves and pipes) and the fabric (the entire network topology) so that all nodes can communicate efficiently.

⚙️ What is Network Provisioning in BCM?¶

Network provisioning in BCM refers to the automated process of: - Configuring network switches (both Ethernet and InfiniBand) - Setting up the network fabric (the interconnection between switches and nodes) - Applying consistent network policies across the cluster

BCM uses a declarative model — you define the desired network state, and BCM makes it happen.

🛠️ Key Components of Automated Switch Configuration¶

🔄 How Automated Switch Configuration Works¶

The process follows a simple workflow:

1. Discovery — BCM detects all switches connected to the cluster management network
2. Validation — BCM checks if switches match expected models and firmware versions
3. Configuration — BCM applies the appropriate switch profile (VLANs, port settings, routing)
4. Verification — BCM confirms the switch is operating as expected
5. Monitoring — BCM continuously checks switch health and alerts on changes

📊 Visual Representation: BCM Multi-Network interface provisioning¶

This diagram displays BCM network architecture, showing dedicated management and high-speed data networks.

🌐 Fabric Setup via BCM¶

Fabric setup is about connecting all switches into a unified network. BCM handles this by:

• Topology Mapping — BCM builds a map of all switch-to-switch and switch-to-node connections
• Path Optimization — BCM configures routing protocols (like ECMP or adaptive routing) for best performance
• Redundancy Configuration — BCM sets up failover paths so the network survives a switch failure
• QoS Policies — BCM applies traffic prioritization for AI workloads (e.g., NCCL traffic gets priority)

📊 Comparison: Manual vs. BCM-Automated Provisioning¶

🕵️ Common Use Cases for New Engineers¶

As a new engineer, you'll encounter these scenarios:

• Adding a new compute node — BCM automatically configures the switch port for that node
• Replacing a failed switch — BCM detects the new switch and applies the correct configuration
• Upgrading fabric firmware — BCM coordinates firmware updates across all switches with minimal downtime
• Troubleshooting connectivity — BCM provides a unified view of the entire fabric state

✅ Best Practices for Network Provisioning with BCM¶

• Always use switch profiles — Never manually configure individual switches; always create and apply profiles
• Test in a staging fabric first — Validate new profiles on a small test network before production
• Enable automatic discovery — Let BCM find new switches rather than adding them manually
• Monitor fabric health daily — Check BCM dashboards for port errors, link flaps, or bandwidth saturation
• Document topology changes — Update the BCM topology database whenever physical connections change

🎯 Summary¶

Network provisioning via BCM transforms complex switch and fabric management into a simple, automated process. For new engineers, the key takeaway is: define your desired network state once, and let BCM handle the rest. This approach reduces human error, speeds up deployments, and ensures your AI cluster's network performs at its best.

As you gain experience, you'll learn to customize switch profiles and optimize fabric settings for specific AI workloads — but the foundation is always the same: automated, repeatable, and validated network provisioning.