The tier structure of a campus network is a decision that must be made before choosing devices. Changing it later tears up cabling, the IP plan and the routing design together. In a single-building office, a three-tier architecture brings unnecessary cost and complexity; in a growing multi-building site, a collapsed core turns into a bottleneck within a few years. To determine which structure a campus network should be built with, we compare the core–distribution–access model with the collapsed core and routed access alternatives, and show with an oversubscription calculation and a decision tree which structure suits when.
The purpose of the hierarchical campus model
The hierarchical model divides network tasks into specific layers and aims to solve three problems:
- Limiting the failure domain: A loop, broadcast storm or misconfiguration on an access switch shouldn’t spread to the whole campus.
- Predictable growth: When a new floor or building is added, the same block should be repeatable without changing the design.
- Defined policy points: Where routing, route summarization, QoS and access control will be applied should be clear from the start.
The number of tiers is a means of reaching these goals, not the goal itself. The number of physical devices and the logical tier aren’t the same thing either: in a collapsed core, the core and distribution roles merge on the same device pair, but the roles don’t disappear. They just sit in the same box.
Three-tier architecture: roles and responsibilities
| Layer | Role | Typical characteristics |
|---|---|---|
| Access | End device connectivity, PoE, port security, the QoS trust boundary | High port density, 1G/mGig ports |
| Distribution | Aggregating access blocks, the L2/L3 boundary, route summarization, policy | Redundant pair, 10G/25G downlinks |
| Core | Fast transport between distribution blocks, the data center and the WAN edge | L3 only, no policy, high-capacity uplinks |
The reason the core exists is the number of links. If you want to connect four distribution blocks in a full mesh without a core, you need n(n−1)/2 = 6 link bundles between block pairs; when the number of blocks rises to eight, this becomes 28. With a core layer, each new block connects only to the core pair. No policy (ACL, NAT, deep inspection) is applied at the core; that work is done at the distribution layer or on the firewall. The core should stay simple. One of the mistakes that can prove costly is loading ACLs and NAT onto the core because “it’s the most powerful device anyway,” after which every change puts the whole campus at risk.
Collapsed core: when is it enough?
In a collapsed core, the core and distribution roles merge onto a redundant switch pair (stack, MLAG or vPC). All access switches, as well as servers, firewalls and the WAN edge, connect to this pair. It’s usually enough under these conditions:
- A single building, or a few nearby buildings with short fiber distances.
- A layout where all floor cabinets (IDFs) can be connected directly to the core pair with fiber.
- The core pair’s port capacity covering the current and planned number of access switches.
- An environment where the data center function is small or gathered on a separate server switch pair.
The limits of a collapsed core are port capacity and the failure domain. A software bug or a wrong change on the core pair affects the whole campus. That’s why the redundancy architecture of the pair is the most sensitive decision in a collapsed core design. For the control plane difference between a stack and MLAG/vPC and downtime behavior during upgrades, see the switch stack, MLAG and vPC comparison article. A stack is a single control plane; a single software bug can take both members down together.
The L2 and L3 boundary: the routed access approach
In the classic design, the access layer is L2 and the distribution layer is L3. VLANs are carried from the access switch to distribution over a trunk, and the default gateway is the SVI at distribution. This model has two variants and one alternative:
| Model | L2/L3 boundary | Advantage | Limitation |
|---|---|---|---|
| L2 access, loop-free | Distribution | VLANs specific to each access switch; an L3 link between the distribution pair, or no STP blocking thanks to MLAG/stack | Depends on the redundancy architecture of the distribution pair |
| L2 access, VLANs spread | Distribution | The same VLAN on many access switches | Large failure domain, dependence on STP |
| Routed access | Access | No STP on uplinks, both uplinks active with ECMP, small failure domain | VLANs don’t extend beyond the access switch; requires an L3 license and more subnets |
In routed access, each access switch is the default gateway for its own user subnets and connects to distribution with routed ports. The two uplinks are used at the same time as equal-cost paths (ECMP); on an uplink failure, convergence depends on the speed of the routing protocol. STP is removed from the uplink entirely. For dynamic routing design within the campus, the OSPF design in a campus network article complements this approach.
Where routed access most often gets stuck is the need for a VLAN that must span more than one access switch: L2 roaming on controllerless Wi-Fi, old building automation devices expecting the same subnet, printers gathered into a single VLAN. With controller-based Wi-Fi, client traffic is usually tunneled to the controller, so this problem often disappears on its own. In other cases, question at the design stage whether the need is really mandatory; “it used to be like this” isn’t a justification. If a VLAN spanning many access switches is unavoidable, STP design must be done carefully; details are in the STP, RSTP and MSTP article.
Oversubscription ratio: a worked calculation
The oversubscription ratio is the ratio of total port capacity at one tier to the uplink capacity going to the tier above. Because users don’t all generate traffic at full speed at the same time, a 1:1 ratio isn’t needed on a campus. A practical rule often cited in Cisco’s older campus design guides is to take 20:1 from access to distribution and 4:1 from distribution to core as upper limits. These aren’t rules, they’re starting points. If video, backup and large file traffic are heavy, a lower ratio is targeted.
Worked example 1: 48 × 1G access ports and 2×10G uplinks. Total access capacity is 48 Gbps.
- If both uplinks are active (with MLAG/stack or routed access): 48 ÷ 20 = 2.4:1.
- If STP blocks one uplink: 48 ÷ 10 = 4.8:1. With the same hardware, the ratio doubles purely because of the architecture.
Worked example 2: A two-switch stack with 24 × 2.5G mGig ports for Wi-Fi 6/6E access points and 48 × 1G ports; uplinks 2×10G. Total access capacity is 24×2.5 + 48×1 = 108 Gbps; the ratio is 108 ÷ 20 = 5.4:1. If the uplinks are increased to 2×25G, the ratio becomes 108 ÷ 50 ≈ 2.2:1.
Worked example 3 (distribution–core): 10 access switches connected to the distribution pair, each with 2×10G; 200 Gbps of total downlink. If there are 4×40G uplinks from distribution to core, 200 ÷ 160 = 1.25:1; with 4×10G it would be 5:1.
The calculation shows the ratio, not the real need. Ratio design done without measuring the peak utilization and output drops of existing uplinks leads either to unnecessary investment or to a bottleneck. Don’t design without measuring. For the method, see the uplink utilization and 95th percentile (P95) reporting article.
Failure domains and growth scenarios
Choose the architecture not for today’s need but for the three-to-five-year growth expectation. Scenarios that should be put on the table:
- A new building or floor: In a collapsed core, new cabinets require fiber to be pulled to the core pair. If distance or fiber capacity is insufficient, a distribution block needs to be added for the new building and the architecture extended to three tiers. For fiber routes between buildings, the campus backbone fiber design article is useful.
- Data center growth: As the number of servers grows, separating server switches from the user core prevents campus changes from touching the data center.
- Uplink speed increase: mGig access points and 10G desktops push access uplinks to 25G or 40G/100G; distribution and core port types should be chosen accordingly.
- Shrinking the failure domain: An L2 problem in one building shouldn’t pass to other buildings. Making links between buildings L3 is the simplest and most effective boundary. A broadcast storm can’t cross a router.
Decision tree and sample topologies
Answer the questions below in order; for most campuses, the architecture becomes clear on its own:
- Can all access cabinets be connected to a single point with fiber, and is the port capacity of the core pair enough? If yes, it’s a collapsed core candidate; if no, a distribution layer is needed.
- Is there more than one building or independent distribution block? For two blocks, connecting the distribution pairs directly may be enough; with three or more blocks, a separate core should be considered.
- Is there a VLAN that must span more than one access switch? If not, routed access or loop-free L2 access can be chosen; if so, L2 access and careful STP design are needed.
- Do the access switches support L3 routing and the required protocols (such as OSPF) with their license? If not, routed access turns into a cost item.
- Which model can the team operate? Routed access requires more subnet and routing knowledge. Operating capacity is also part of the design; when a fault occurs at night, the person connecting to the console needs to understand the model.
Three typical topologies come out of these answers:
- A single building, a few floors: A collapsed core pair (stack or MLAG), access switches each connected to the pair with two uplinks, VLANs specific to each access switch.
- Two or three nearby buildings: A collapsed core in the main building, small distribution pairs connected with L3 in the other buildings; summarized routes between buildings.
- A multi-building site: A distribution pair per building, a separate core pair, the data center and WAN edge connected to the core; L2 or routed access at the access layer as needed.
Evaluating the existing topology against these criteria and determining the target architecture can be done within enterprise network consulting; redesigning the L2/L3 boundary and the VLAN layout is the subject of the VLAN, routing and network segmentation service. As a general design reference, Cisco’s Design Zone campus guides can be reviewed.
Frequently asked questions
After how many users should you move to a three-tier architecture?
The number of users alone isn’t decisive. What decides is the number of buildings and cabinets, whether all cabinets can be connected to a single point with fiber, the port capacity of the core pair and the need to limit the failure domain. When there are three or more independent distribution blocks, a separate core usually starts to make sense.
Is routed access better for every campus?
No. Routed access removes STP from the uplinks, uses both uplinks at the same time and shrinks the failure domain. But it prevents VLANs from spanning more than one access switch, and requires an L3 license on access switches and managing more subnets. If you don’t have those needs, it’s a good option.
Is the 20:1 oversubscription ratio still valid?
This ratio is a starting value cited as an upper limit in Cisco’s older campus design guides. In environments with growing numbers of high-bandwidth end devices such as video, backup and Wi-Fi 6/6E, a lower ratio is targeted. Make the final decision by measuring the uplinks’ peak utilization and output drops.
Does moving from collapsed core to three tiers require downtime?
If links between buildings were designed as L3 from the start and the IP plan is split into per-building blocks, the migration can be done in a controlled way by adding the new core pair and moving the blocks one at a time. Each block move requires a short maintenance window; if there are VLANs spread at L2, the migration needs a more extensive redesign.