Switched Ethernet (Switches, Forwarding, LAN Design)
Context: FIT1047_MOC · the fix for every shared-Ethernet problem from Physical and Data Link Layers (Signals, Ethernet, CSMA-CD) — true logical star, full duplex, no collisions · plus how real LANs are designed
Quick Revision
- 🎯 Objective: a switch is a layer-2 device that reads frame MAC addresses and forwards ONLY to the destination socket — learning its forwarding table by watching sender addresses.
- 📦 Core Components: learn-by-observing ➔ flood-when-unknown ➔ store-and-forward buffering ➔ switch vs router ➔ design levers (standards, segmentation).
- ⚡ Key Constraint: an EMPTY forwarding table makes a switch behave exactly like a hub (flood to all except source) — the learning story is the exam story.
📝 Core
1. How a Switch Works
- Logical star ➔ point-to-point circuits per device; looks like a hub (4–24 UTP sockets) but the circuit is NOT shared.
- Learning ➔ every arriving frame’s sender MAC gets recorded (MAC → socket); destination in table ⟹ deliver to that socket only; unknown ⟹ flood to all sockets except the source.
- Full duplex ➔ separate send/receive paths per device ⟹ collisions impossible; CSMA/CD becomes unnecessary.
- Store and forward ➔ two frames for the same destination at once? The switch buffers the second and forwards after the first — memory replaces collisions.
- Consequence ➔ switched Ethernet uses almost ALL available bandwidth simultaneously (A→B while C→D).
- Multi-switch LANs ➔ tables map MACs to the socket toward them — possibly another switch’s uplink.
2. Switch vs Router (lecture table)
| Switch | Router | |
|---|---|---|
| connects | devices within one LAN | different LANs |
| address used | MAC | IP |
| delivery | direct to destination device | to destination device or next network |
| table | learned automatically from traffic | needs configuration / routing protocols |
3. LAN Design (best practice)
- Standard kit ➔ switched 1000BASE-T + Cat 6/6a (10 Gbps upgrade path); optic fibre where traffic is extreme or future-proofing pays.
- Throughput thinking ➔ find the bottleneck (congestion point ⟹ dropped packets); fix by (1) upgrading the standard (new NICs+switches, e.g. 16×100BASE-T + 2×1000BASE-T uplink ports for servers) or (2) segmentation — second server NIC, split into two LANs, router between them.
⚠️ Common Mistakes
- 💡 Learning uses SENDER addresses ➔ the table fills from who talks, not who is talked to; silent devices stay unknown and keep being flooded.
- 💡 Switch ≠ hub despite looks ➔ hub repeats blindly (one collision domain); switch forwards selectively (no collisions) — the W7↔W8 contrast question.
- 💡 No collisions ≠ no queuing ➔ two frames to one destination still serialise; store-and-forward hides it, latency shows it.
🧠 Active Recall
A frame from A (socket 0) to new device D arrives at a fresh switch. Trace table growth over A→D, D→A, A→D again.
Answer
- Short answer: A→D: learn A=0, flood; D→A (say socket 3): learn D=3, deliver direct to 0; A→D: both known — direct to 3.
- Why: Symmetric traffic teaches fastest ➔ each direction’s first frame registers its sender; two exchanges complete the pair.
Your file server saturates its link on a 100BASE-T switched LAN. Give both remedies from the design lecture with their costs.
Answer
- Short answer: Upgrade the path (gigabit uplink ports/NIC/cabling to the server) or segment (second NIC / split LANs joined by a router) to parallelise paths.
- Why: Bottleneck relief, not average speed ➔ throughput rises by widening or multiplying the congested link only.