Top-of-rack vs end-of-row switching
Where the access switch sits decides what runs through the cable trays, how many ports sit idle, and how much of the estate a single switch failure touches.
The options
- Top-of-rack (ToR — switch in the cabinet) — One or two access switches per cabinet, with short copper inside it.
- End-of-row (EoR / MoR — shared switch per row) — Shared access switches in a cabinet at the end or middle of a row.
Both names describe where the access layer physically lives. In a top-of-rack arrangement each cabinet contains its own access switching, servers connect over short copper inside the cabinet, and the cabinet leaves with a small number of uplinks. In an end-of-row arrangement, several cabinets share switching housed in a network cabinet at the end — or the middle, which is the same idea and usually called middle-of-row — and every server connection runs horizontally to reach it.
The names are also slightly misleading. “Top-of-rack” is a topology, not a mounting position: the switch is frequently installed in the middle of the cabinet, where the cable runs to the servers above and below are shortest. What matters is that the access port is in the cabinet, not that it is at the top of it.
Side by side
| Top-of-rack | End-of-row | |
|---|---|---|
| Where the access switch lives | In each cabinet, one or two per cabinet. | In a shared network cabinet serving several cabinets in the row. |
| Server-to-switch cabling | Short copper runs that stay inside the cabinet. | Horizontal copper runs from every cabinet to the row position. |
| What leaves the cabinet | A small number of fibre uplinks. | Every server connection, as structured copper. |
| Switch count | Higher — scales with cabinets. | Lower — scales with rows. |
| Port utilization | Ports can strand in partly filled cabinets. | Ports pool across cabinets, so utilization is usually higher. |
| Failure domain of one switch | One cabinet. | A group of cabinets in the row. |
| Cable pathway load | Light — mostly fibre in the trays. | Heavy — dense copper trunks along the row. |
| Adding a cabinet later | Add the cabinet with its own switching and uplinks. | Pull new horizontal cabling, and hope the row switch still has ports. |
| Where it commonly fits | Dense, uniform cabinets and leaf-spine fabrics. | Mixed or lightly filled cabinets and structured-cabling estates. |
Port stranding is the usual argument against top-of-rack
A switch with forty-eight access ports in a cabinet holding twelve servers leaves thirty-six ports doing nothing, and they cannot be lent to the cabinet next door. Multiply that across a hall of partly filled cabinets and the waste is real — in capital, in the rack units the switches occupy, and in the power and heat they add to the cabinet.
End-of-row pools those ports across several cabinets, so the same physical ports serve whichever cabinet actually has servers in it. The price is paid in cabling: every one of those connections is a horizontal run that has to be installed, dressed, documented and maintained, and it has to travel through the same pathways as everything else.
Cabling is not only a cost — it is a physical load
Dense copper trunks take up tray and under-floor space, add weight, and can obstruct airflow where they pass. In a raised-floor room, congested under-floor pathways compete directly with the air the floor is supposed to deliver. That is a thermal consequence of a network decision, and it is easy to miss because the two live in different drawings.
Top-of-rack shifts most of that volume out of the shared pathways: the copper is short and stays in the cabinet, and what travels the room is a handful of fibre per cabinet. The cabinet itself gets busier instead, which is a trade worth checking against the cabinet’s own airflow and cable management.
Failure domains follow the switch, not the label
A single access switch failure takes down whatever connects to it. Under top-of-rack that is one cabinet; under end-of-row it is a group of them. Neither is automatically better — what matters is whether the workload can tolerate the resulting domain, and whether the design has dual access paths at all.
Dual switching changes the picture in both arrangements, and it raises the same question the Academy asks about A/B power feeds: are the two paths genuinely independent, or do they meet again upstream in a shared uplink, a shared chassis, a shared power feed or a shared cable route? Two switches in one cabinet on one power strip are not two paths.
The choice is entangled with the fabric above it
Top-of-rack fits naturally with leaf-spine fabrics, where each cabinet switch is a leaf with uplinks to several spines and east-west traffic has predictable hop counts. End-of-row grew up alongside older three-tier designs, where aggregation sat between access and core.
That history explains part of the preference pattern, but it is not a rule. The useful question is not which name is modern; it is what the traffic actually does, how uniform the cabinets are, how full they will be on day one versus year three, and what the cable pathways can carry.
Where each one tends to fit
These are conditions commonly associated with each option, not a decision procedure. A real choice is made against site data, the adopted standards and the constraints of the specific building.
Conditions that tend to point toward top-of-rack
- Cabinets are dense, uniform and filled to a predictable level.
- The fabric is leaf-spine and east-west traffic dominates.
- Cable pathways are constrained and copper volume has to stay out of them.
- Cabinets are deployed as repeatable units rather than filled gradually.
Conditions that tend to point toward end-of-row or middle-of-row
- Cabinets are mixed, lightly filled, or fill slowly over time.
- The estate is built around structured cabling and a cabling standard.
- Fewer managed devices is worth more than shorter cable runs.
- Equipment in a cabinet varies enough that per-cabinet port planning is wasteful.
Where this comparison stops
This page compares placement, not products. Switch selection, oversubscription ratios, uplink sizing, media choice, power and cooling budgets inside the cabinet and the cabling standard the site follows are all separate engineering decisions with their own constraints.
Real estates are rarely pure. Mixed arrangements — top-of-rack for dense compute and end-of-row for storage or legacy equipment — are common and can be the right answer. The comparison is useful for reasoning about a row, not for declaring a site-wide policy.
Frequently asked questions
What is a top of rack switch?
An access switch installed inside a server cabinet, so servers in that cabinet connect over short copper runs and the cabinet leaves with a small number of fibre uplinks. The name describes the topology rather than the mounting height — the switch is often physically in the middle of the cabinet.
Is top of rack better than end of row?
Neither is generally better. Top-of-rack keeps copper out of the shared pathways and bounds a switch failure to one cabinet, at the cost of more switches and stranded ports. End-of-row pools ports across cabinets and uses fewer switches, at the cost of dense horizontal cabling and a larger failure domain.
Does a top-of-rack switch have to be at the top of the rack?
No. It is frequently mounted in the middle of the cabinet so the cable runs to servers above and below are shortest. What defines the arrangement is that the access port is inside the cabinet.
What is middle-of-row switching?
The same idea as end-of-row, with the shared network cabinet placed in the middle of the row instead of at one end, which roughly halves the longest horizontal cable run.
How many top-of-rack switches per cabinet?
One if there is a single access path, two where the design calls for independent A and B paths. Two switches only provide independence if their uplinks, power feeds and cable routes are also independent.
Does top-of-rack work with leaf-spine?
Yes — it is the common pairing. Each cabinet switch acts as a leaf with uplinks to multiple spines, which gives east-west traffic a predictable number of hops.
Sources
These sources support the summary above. Use the current adopted edition, verified site information and qualified professional review for real work.