Network Types and Topologies

This is what topology decisions actually cost. The engineer who wired that building in 2014 chose the cheapest option — one switch per floor, connected in a chain. No redundancy. No alternative paths. Nobody asked "what happens if this one switch dies?" That question should have been answered on day one.

This module gives you two frameworks for classifying and designing networks. The first is network type — how geographically large is this network, and who owns it? The second is topology — what shape does the physical wiring take, and how does that shape determine resilience, cost, and scale? These are not academic classifications. They are the vocabulary you need to design networks that survive reality.

The most fundamental way to classify a network is by geographic scale. Engineers use these terms constantly — in job descriptions, architecture diagrams, vendor docs, and incident reports. Understanding what each acronym actually means (not just the expansion) saves you from confusion in every technical conversation.

The Lines Are Blurring — Cloud Changed Everything

These categories were crisp in 1990. Today, they blur in fascinating ways. When your startup team connects via WireGuard VPN to a private subnet on AWS, is that a LAN or a WAN? Technically, it is a WAN traversal (traffic crosses the internet) that behaves exactly like a LAN — same private IP subnet, no NAT, sub-10ms latency. AWS calls this a VPC (Virtual Private Cloud), and it deliberately gives you LAN semantics over WAN infrastructure.

Similarly, when you use SD-WAN (Software-Defined WAN), you are abstracting the physical WAN links (MPLS, broadband, LTE) into a single logical network managed by software. A branch office in Denver and a headquarters in Chicago appear to be on the same "flat" network — but the physical connection might be switching between MPLS and LTE depending on cost and latency. The topology is virtual, not physical.

When to Use What

Physical topology describes how devices are physically wired to each other — the actual shape of the cables. There are six fundamental topologies. Every real-world network is built from some combination of them.

In the early 1980s, a coaxial cable ran through an office building and every computer was physically clamped to it with a vampire tap — a connector that pierced the cable shielding to make contact. This was 10BASE5 (Thick Ethernet), and it was revolutionary. One cable, unlimited devices, no dedicated hardware beyond the cable itself. By 1985, it evolved to 10BASE2 (Thin Ethernet, BNC connectors) which you could wire yourself with cheap coaxial cable.

The fundamental problem was collision domains. Since all devices shared the same wire, only one device could transmit at a time. If two devices transmitted simultaneously, their signals collided and destroyed each other — both had to wait a random interval and retry. This is called CSMA/CD (Carrier Sense Multiple Access with Collision Detection). As devices were added, collisions multiplied, performance dropped. A 10 Mbps bus with 20 devices might only deliver 2-3 Mbps effective throughput.

1. Listen: Is anyone transmitting? (Carrier Sense)
2. If yes: wait until channel is free
3. If no: start transmitting your data
4. While transmitting: keep listening (Collision Detection)
5. If collision detected:
   a. Stop transmitting immediately
   b. Send jam signal (tell everyone there was a collision)
   c. Wait a random "backoff" time (exponential backoff)
   d. Go back to step 1
6. After 16 failed attempts: give up, report error

Efficiency drops sharply with more devices:
 5 devices  → ~85% channel efficiency
20 devices  → ~40% channel efficiency
50 devices  → ~10% channel efficiency

The single biggest failure mode of bus topology was physical: one break, everyone goes down. A cable cut, a bad connector, even a loose BNC terminator at either end — and every device on the bus lost connectivity instantly. There was no alternative path. Network troubleshooting in the bus era involved walking the entire cable run with an ohmmeter looking for a break. Engineers who grew up in this era still flinch at coaxial cable.

Today, virtually every network — from your home WiFi (all devices connect to one router) to Google's data centers (all servers connect to a Top-of-Rack switch) — is a star at its core. The key innovation was replacing the hub with a switch.

Hub vs Switch — A Critical Distinction

Both create a star topology, but they behave completely differently:

When data arrives at a switch port, the switch reads the destination MAC address from the Ethernet frame header and looks it up in its MAC address table (also called the CAM table — Content Addressable Memory). If it knows which port that MAC is on, it forwards the frame only to that port. If it does not know yet (no entry in the table), it floods the frame to all ports except the one it came from — and then learns from the response.

Switch CAM table (builds up over time):
MAC Address       Port   VLAN   Age
A4:C3:F0:11:22:33   1     10    50s   ← Device A on port 1
B8:E8:56:44:55:66   2     10    30s   ← Device B on port 2
DC:A6:32:77:88:99   3     20    10s   ← Device C on port 3

When a frame arrives (src: A4:C3:F0:11:22:33 → dst: B8:E8:56:44:55:66):
1. Learn: A4:C3:F0:11:22:33 is on port 1 (already known)
2. Look up: B8:E8:56:44:55:66 → port 2
3. Forward: send frame ONLY to port 2

When dst MAC is unknown:
1. Look up: not in table → FLOOD to all ports (except source port)
2. When B replies, learn B's MAC → add to table
3. Future frames to B go only to port 2

Spanning Tree Protocol — Preventing Loops

A star topology with redundant switches creates a problem: loops. If Switch A is connected to Switch B, which is connected to Switch C, which is connected back to Switch A, a broadcast frame would loop forever — a broadcast storm that consumes 100% of bandwidth and crashes the network within seconds.

STP (Spanning Tree Protocol — IEEE 802.1D) prevents this by logically blocking redundant paths. STP elects a root bridge, calculates the shortest path from every switch to the root, and puts redundant paths in blocking state. If the active path fails, STP unblocks an alternative path — but this takes 30–50 seconds with classic STP. Modern networks use RSTP (Rapid Spanning Tree — 802.1w) which converges in 1–2 seconds, or port channel/LAG aggregation which bonds multiple links without needing STP at all.

Ring topology was IBM's answer to Ethernet's collision problem in 1984 — Token Ring (IEEE 802.5). Instead of random collision detection, a special packet called a token circulated around the ring. Only the device holding the token was allowed to transmit. Deterministic access: no collisions, equal fairness, predictable performance. But Token Ring was slow to expand (adding a device required breaking the ring), expensive (IBM's MAU — Multistation Access Unit — cost a fortune), and fundamentally lost the market to switched Ethernet by 1998.

Token Ring died. Ring topology did not. The world's telephone networks and long-haul internet backbone still use ring topology in the form of SONET/SDH (Synchronous Optical Networking / Synchronous Digital Hierarchy) — the standard physical layer for carrier networks since the 1990s.

Dual-Ring Self-Healing — How SONET Achieves 99.999% Uptime

The weakness of a simple ring — one break = total failure — is solved by running two rings in opposite directions simultaneously. This is the SONET/SDH BLSR (Bidirectional Line-Switched Ring) architecture.

Normal operation: traffic flows clockwise on working ring
        Node A ──→──→──→──→ Node B
       ↑                           ↓
       Node D ←──←──←──←── Node C
       (protection ring flows counter-clockwise, idle)

Cable cut between B and C:
        Node A ──→──→──→──→ Node B
       ↑                           ↓ ← traffic "wraps" here
       Node D ←──←──←──←── Node C
       ↑ ← traffic wraps here      ↓
       ← now flowing counter-clockwise on protection ring →

Result: all traffic continues in under 50 milliseconds.
This is why fiber cuts rarely cause noticeable telephone outages.

The 50ms protection switching time of SONET/SDH is a standard specification — the ITU requires it for carrier-grade networks. This is tight enough that voice calls don't experience audible disruption. For comparison, a human typically does not notice audio interruptions under 200ms. Modern networks use OTN (Optical Transport Network) which extends SONET concepts with more capacity and better overhead management.

The internet is the world's largest mesh network — specifically, a partial mesh of ~80,000 Autonomous Systems (ASes) connected by BGP peering relationships. No single AS connects to every other AS (that would be a full mesh requiring billions of physical links). Instead, each AS connects to a few well-chosen peers, and BGP propagates reachability information so every AS can eventually reach every other.

Full Mesh vs Partial Mesh

Full mesh links = n × (n - 1) / 2

 4 nodes  →   6 links   (feasible)
10 nodes  →  45 links   (expensive)
20 nodes  → 190 links   (very expensive)
50 nodes  → 1,225 links (impractical)

Partial mesh: each node connects to 2-4 strategic peers
→ Reduces links while maintaining redundancy
→ Used by: the internet, corporate WANs, data center spines

In data centers, the dominant topology today is spine-leaf — a two-tier partial mesh. Every leaf switch (ToR — Top of Rack) connects to every spine switch. No leaf connects directly to another leaf. This creates equal-cost multi-path routing (ECMP) — traffic from any server to any other server takes exactly 2 hops through the fabric, and there are always multiple equal-cost paths to load-balance across.

       Spine 1      Spine 2      Spine 3      Spine 4
         │ │ │ │      │ │ │ │      │ │ │ │      │ │ │ │
         └─┼─┼─┼──────┼─┼─┼─┘      └─┼─┼─┼──────┼─┼─┘
           │ │ │      │ │ │            │ │ │      │ │
        Leaf 1       Leaf 2         Leaf 3       Leaf 4
        │ │ │ │      │ │ │ │       │ │ │ │      │ │ │ │
       Servers      Servers        Servers      Servers

Every leaf connects to every spine (partial mesh between tiers)
Traffic from Server A to Server B: A → Leaf 1 → any Spine → Leaf 3 → B
Always 2 hops regardless of which server pair. Always ECMP load balanced.
Used by: Google, Facebook, AWS, Azure, Netflix, every hyperscale data center.

When a company needs to connect 5,000 employees across a 20-floor office building, star topology at one level is not enough. You need a hierarchy — a tree of star networks, where each level aggregates the bandwidth from the level below.

The Three-Tier Enterprise Model

The critical insight of the three-tier model is traffic locality. Most office communication stays within a department — emails, file shares, printers. By organizing the tree so that each department is under the same distribution switch, most traffic never needs to go higher than the distribution layer. Only inter-department or internet traffic climbs to the core. This is why enterprise networks can handle 10,000 devices with only a handful of core switches.

Collapsed Core — For Smaller Networks

For networks under ~500 devices (small offices, branch offices), the distribution and core layers merge into a single collapsed core layer. Two high-performance switches handle everything that distribution and core would normally split. This reduces hardware cost while maintaining the essential isolation that access layer star topology provides.

        [Core/Distribution Switch A] ←→ [Core/Distribution Switch B]
         ╱   ╲     ╱   ╲                ╱   ╲     ╱   ╲
        ↓     ↓   ↓     ↓              ↓     ↓   ↓     ↓
      Floor1 Floor2 Floor3 Floor4    Floor5 Floor6 Floor7 Floor8
       (Access switches with end devices)

Both core switches active (HSRP/VRRP for gateway redundancy)
STP or port channels prevent loops between core switches
If one core switch fails, HSRP failover in ~3 seconds

Pure topologies exist in textbooks. Real networks combine multiple topologies at different scales, choosing the right shape for each requirement. Understanding this is what separates someone who can recite topology names from someone who can design real infrastructure.

A Real Enterprise Network Architecture

Personal devices:
  PAN (Bluetooth): employee phone ↔ headset, mouse, keyboard

Building LAN (Star + Tree):
  Access switches (Star) → Distribution switches (Tree) → Core switches

Between buildings on campus (Star + Partial Mesh):
  Each building's core switch connects to 2+ campus core switches
  (partial mesh for redundancy, star within each building)

Between cities (Ring + Partial Mesh WAN):
  SONET/SDH dual rings carry voice and critical traffic
  MPLS partial mesh carries data between city offices

To the internet (Partial Mesh):
  Dual ISP connections for redundancy
  BGP peering with multiple upstream providers

To cloud (Hybrid WAN + VPC):
  AWS Direct Connect: private fiber to AWS (bypasses internet)
  VPN tunnels: encrypted overlay on top of internet WAN
  Inside AWS VPC: virtual star/tree topology (subnets = segments)

Storage (SAN):
  Fibre Channel SAN fabric in data center (dual fabric for HA)
  All database servers connect to both SAN fabrics simultaneously

Notice how each layer uses the appropriate topology: star for end-device connections (simple, manageable), tree for building-scale aggregation (hierarchical, scalable), partial mesh for building-to-building (redundant), ring for long-haul backbone (self-healing), and mesh peering for internet connectivity (maximum resilience).

SD-WAN — Topology as Software

SD-WAN (Software-Defined Wide Area Network) abstracts the physical WAN topology from the logical network. Instead of manually configuring MPLS circuits between offices, SD-WAN creates a virtual full mesh — any branch can talk directly to any other branch over whatever physical links are available (MPLS, broadband, 4G/5G, satellite). The SD-WAN controller decides in real-time which physical path each traffic flow should take based on latency, packet loss, and cost.

A branch office in Tulsa might send Zoom video directly to a branch in Charlotte via broadband (low cost, acceptable latency), while SAP traffic goes via MPLS (guaranteed QoS, predictable latency), and backup data goes via LTE (expensive per MB, but available when MPLS is down). The topology is virtual, dynamic, and policy-driven. Cisco Viptela, VMware VeloCloud, and Fortinet Secure SD-WAN are the major vendors.

Leaf-Spine Goes Mainstream

The three-tier enterprise model (core/distribution/access) was designed for north-south traffic — client to server, up and down the hierarchy. Modern workloads are increasingly east-west — server to server, microservice to microservice, container to container. The three-tier model is inefficient for east-west: traffic from any server to any other server has to climb to the core and back down.

Leaf-spine (spine-leaf) topology solves this: every server is always exactly two hops from every other server, regardless of which leaf they are on. As workloads shift from client-server to microservices, leaf-spine is replacing three-tier even in enterprise data centers. If you are designing a new data center today, leaf-spine is the answer.

Intent-Based Networking (IBN)

Traditional networks are configured device by device, command by command. Intent-Based Networking lets you declare what you want ("all HR computers should not be able to talk to engineering servers") and software translates that into thousands of ACL rules, VLAN assignments, and routing policies automatically. Cisco DNA Center and Juniper Apstra implement this concept. The topology still matters — but you are increasingly describing intent rather than configuring individual devices.

The Five Questions to Ask Before Any Topology Decision

Every topology has a characteristic failure mode. Knowing these is essential for both design and troubleshooting.

The pattern is clear: resilience requires redundant paths, and redundant paths require mesh characteristics — either a partial mesh between critical nodes, or a redundant star where two switches both connect to all devices. The most cost-effective approach for most enterprise networks is redundant star topology — two switches in every critical position, connected to each other, with devices dual-homed where necessary.