6.1d Bonding and teaming: aggregating multiple interfaces for redundancy and bandwidth¶

🌐 Context Introduction¶

When managing AI infrastructure, network reliability and performance are critical. A single network interface can become a bottleneck or a single point of failure — if that cable breaks or the switch port fails, your AI training job or data transfer stops. Bonding (also called teaming in some contexts) solves this by combining two or more physical network interfaces into one logical interface. This provides redundancy (if one link fails, traffic continues over the others) and increased bandwidth (traffic can be distributed across multiple links).

Think of it like having multiple lanes on a highway: if one lane is closed, traffic still flows on the others, and during normal operation, more cars can travel simultaneously.

⚙️ How Bonding Works¶

• Bonding creates a single virtual network interface (e.g., bond0) that represents the group of physical interfaces (e.g., eth0 and eth1).
• The Linux kernel handles the distribution of traffic across the member interfaces based on a configured mode.
• The bonded interface gets its own IP address, MAC address, and routing rules — the physical interfaces behind it become transparent to the network.
• Bonding is configured using the bonding kernel module and can be managed via configuration files or tools like nmcli (NetworkManager command-line tool).

🛠️ Common Bonding Modes¶

Different modes offer different trade-offs between redundancy, bandwidth, and compatibility with network switches. Here are the most relevant ones for AI infrastructure:

• Mode 1 (Active-backup) is the simplest for redundancy: one interface is active, the other is standby. If the active link fails, the standby takes over.
• Mode 4 (LACP) is the most common for AI workloads requiring both redundancy and maximum bandwidth. It uses the Link Aggregation Control Protocol to negotiate with the switch.
• Mode 0 sends packets in round-robin order across all interfaces, but requires the switch to be configured for the same aggregation group.

📊 Visual Representation: Network Interface Bonding (Link Aggregation)¶

This diagram displays how multiple physical network interfaces are combined into a single logical bond interface for redundancy and higher throughput.

📊 Redundancy vs. Bandwidth — What Matters for AI?¶

• Redundancy is critical for long-running AI training jobs. A single network failure should not interrupt data streaming or checkpoint saving.
• Bandwidth aggregation helps when moving large datasets (e.g., training data, model checkpoints) between storage and compute nodes.
• For AI infrastructure, Mode 4 (LACP) is often preferred because it provides both benefits and is standards-based, ensuring compatibility with enterprise switches.
• Mode 1 is useful for simpler setups where bandwidth is not the bottleneck but uptime is essential.

🕵️ Verifying Bonding Configuration¶

After setting up a bond, you can verify its status using standard Linux tools:

• Check the bond's status by reading the /proc/net/bonding/bond0 file. This shows the active interfaces, link status, and mode.
• Use ip link show to see the bond interface and its member interfaces. The bond will appear as a separate interface with its own state.
• Use ip addr show bond0 to confirm the IP address is assigned to the bond, not to the individual physical interfaces.
• Monitor link failures by checking the carrier status of each member interface. If a physical link drops, the bond should automatically fail over.

🧩 Key Considerations for AI Infrastructure¶

• Switch configuration must match the bonding mode. For Mode 4, the switch ports must be configured in an LACP port channel. For Mode 1, no switch changes are needed.
• Naming consistency is important. Use predictable interface names (e.g., eno1, eno2) or rename them to something meaningful like eth0 and eth1 to avoid confusion.
• Testing failover is essential. After configuring bonding, physically unplug one cable and verify that traffic continues without interruption. Use a continuous ping or iperf test to validate.
• Monitoring should include bond status. Set up alerts for when a member interface goes down, so you know you are running on reduced redundancy.
• Performance tuning may be needed. Some bonding modes distribute traffic based on MAC addresses or IP hashes, which can cause uneven load. Test with your actual AI traffic patterns.

✅ Summary¶

Bonding and teaming are fundamental techniques for building resilient, high-performance network connections in AI infrastructure. By aggregating multiple physical interfaces, you protect against hardware failures and increase available bandwidth for data-intensive workloads. Choose the mode that matches your switch capabilities and operational needs — Mode 1 for simple redundancy, Mode 4 for full aggregation with LACP. Always verify the configuration with real failover tests before putting the system into production.