14.2a How a network switch works: MAC address tables and frame forwarding

📦 AI Data Center Networking 📖 Enterprise Networking Baseline

🧭 Context Introduction

In an AI data center, network switches are the silent workhorses that move data between servers, storage systems, and GPUs. Unlike a simple hub that blindly repeats every signal, a switch is intelligent — it learns which devices are connected to which ports and forwards frames only where they need to go. This reduces congestion and keeps AI training jobs running smoothly.

For a new engineer, understanding how a switch builds its MAC address table and uses it for frame forwarding is the first step to troubleshooting network issues and designing efficient AI infrastructure.

⚙️ What is a MAC Address Table?

A MAC address table (also called a Content Addressable Memory or CAM table) is a database stored inside a switch. It maps:

  • MAC address — the unique hardware address of a device's network interface
  • Port number — the physical port on the switch where that device is connected
  • VLAN ID — the virtual LAN segment the device belongs to (covered in a later section)

How the Table is Built (Learning Process)

  1. Initial state — The table is empty.
  2. First frame arrives — A device sends a frame. The switch reads the source MAC address from the frame header.
  3. Learning — The switch records: "MAC address AA:BB:CC:DD:EE:FF is on port 1".
  4. Aging — Entries have a timer (typically 300 seconds). If no frames are received from that MAC, the entry is removed to keep the table current.

💡 Key point: The switch learns MAC addresses passively — it never asks devices for their MAC. It simply listens to traffic.

🕵️ How Frame Forwarding Works

When a frame arrives at a switch port, the switch performs a simple three-step decision:

Step 1: Learn

  • Record the source MAC address and port into the MAC table (if not already present).

Step 2: Lookup

  • Check the destination MAC address against the MAC table.

Step 3: Forward or Flood

  • Known unicast — If the destination MAC is in the table, forward the frame only to that specific port.
  • Unknown unicast — If the destination MAC is not in the table, flood the frame out all ports except the one it arrived on.
  • Broadcast/multicast — Always flood (except the incoming port).

📊 Visual Representation: Layer 2 Switch MAC Address Learning

This diagram displays how switches learn source MAC addresses and populate forwarding tables to direct packets to target ports.

flowchart LR Incoming["Incoming Packet (Src MAC: A)"] --> Port["Switch Physical Port 1"] Port --> MACTable{"Store (MAC: A -> Port 1)"} MACTable --> Lookup{"Lookup Dest MAC"} Lookup -->|Found| Forward["Forward to associated port"] Lookup -->|Not Found| Flood["Flood to all other ports"] classDef cpu fill:#eafaf1,stroke:#76b900,stroke-width:2px,rx:6px,ry:6px; classDef memory fill:#f0f7ff,stroke:#3498db,stroke-width:1.5px,rx:4px,ry:4px; classDef system fill:#f1f5f9,stroke:#64748b,stroke-width:1.5px; class MACTable,Lookup cpu; class Incoming,Forward memory; class Flood system;

📊 Comparison: Switch vs. Hub vs. Router

Feature Network Switch Hub Router
Intelligence Learns MAC addresses None (repeats everything) Uses IP addresses
Forwarding decision Based on MAC table No decision (flood all) Based on routing table
Collision domains One per port Single shared domain One per interface
Used in AI DC for Server-to-server, GPU-to-storage Rarely used Connecting to WAN or subnets

🛠️ Practical Example: Frame Flow in an AI Cluster

Imagine a small AI training cluster with:

  • Server A (MAC: 00:11:22:33:44:AA) on switch port 1
  • Server B (MAC: 00:11:22:33:44:BB) on switch port 2
  • Storage array (MAC: 00:11:22:33:44:CC) on switch port 3

Scenario: Server A sends data to Storage

  1. Switch receives frame from port 1.
  2. Switch learns: 00:11:22:33:44:AA → port 1.
  3. Switch looks up destination MAC (00:11:22:33:44:CC).
  4. If found in table → forward only to port 3.
  5. If not found → flood to ports 2 and 3.

Scenario: Storage replies to Server A

  1. Switch receives frame from port 3.
  2. Switch learns: 00:11:22:33:44:CC → port 3.
  3. Destination MAC (00:11:22:33:44:AA) is already in table.
  4. Forward only to port 1.

✅ Result: Only the necessary ports carry traffic. Port 2 (Server B) is not disturbed.

🔍 Why This Matters for AI Infrastructure

In AI data centers, latency is critical. A poorly configured switch can cause:

  • Flooding — If MAC tables are too small, the switch floods frames, wasting bandwidth.
  • MAC flapping — If a device moves between ports (e.g., VM migration), the switch must re-learn quickly.
  • Table exhaustion — AI clusters can have thousands of MAC addresses. Enterprise switches support 32K–128K entries.

Common Troubleshooting Commands (for reference)

For reference, here are typical commands used to inspect MAC tables on managed switches:

show mac address-table
show mac address-table aging-time
show mac address-table count

📤 Output: A list of MAC addresses, VLANs, and port assignments.

✅ Key Takeaways for New Engineers

  • A switch learns MAC addresses by reading source addresses in incoming frames.
  • It forwards frames intelligently — only to the port where the destination device lives.
  • Unknown destinations cause flooding, which wastes bandwidth in AI clusters.
  • MAC tables have a finite size and aging timer — both are important for performance tuning.
  • In AI data centers, minimizing flooding and ensuring fast table lookups directly impacts training job completion times.

🧠 Remember: Every time you see a slow AI training job, check if the network switch is flooding traffic unnecessarily — it might be a MAC table issue, not a GPU problem.