Ports and uplinks

A switch's front panel is a set of roles, not just a set of holes: downlinks that face devices, uplinks that face the rest of the network, stacking ports that face sibling switches, management ports that face the operator. Each role has a typical speed, medium, module and failure mode. This page sorts the ports by role, explains the confusing overlap of "access", "trunk" and "uplink", quantifies oversubscription, and covers uplink modules, link aggregation on uplinks and how uplink optics are chosen.

Port roles

Role	Faces	Typical medium	Speeds	Count on a 48-port access switch
Downlink / access	end devices, servers	RJ45 (PoE), SFP/SFP+/SFP28	1G–25G (mGig 2.5/5/10G on copper)	24–48
Uplink	distribution/spine	SFP+/SFP28/QSFP+/QSFP28/QSFP-DD	10G–400G	2–8
Stacking	sibling switches in a stack	proprietary cable or QSFP+/QSFP28 with stacking firmware	40–480 Gb/s per ring	2
Management (OOB)	management network	RJ45 (sometimes SFP)	100M/1G	1
Console	operator	RJ45 (RS-232) and/or USB-C/mini-USB	—	1–2
USB	file storage, ZTP images	USB-A	—	1
Combo	either RJ45 or SFP, one active	—	1G	2–4 on small switches

On data-centre leaf/spine switches every front port is the same (32 × QSFP-DD); "uplink" is a configuration role, not a hardware class (Switch roles & topologies).

Access, trunk, uplink — three different vocabularies

Term	Layer	Meaning	Confusion
Access port (L2)	2	carries one untagged VLAN	vs. "access switch" (a role in the hierarchy)
Trunk port (L2)	2	carries several VLANs with 802.1Q tags	vs. Juniper/HP "trunk" = LAG in some CLIs
Uplink	physical/role	the link toward the core; usually an L2 trunk or an L3 routed port	an uplink can be an access port (single VLAN) or a routed interface
Downlink	physical/role	toward devices or lower tier	—
Routed port	3	has an IP address, no VLAN tagging	leaf–spine uplinks are routed

Forwarding-layer detail: Forwarding basics.

Oversubscription

Design	Downlink capacity	Uplink capacity	Ratio	Typical use
48 × 1G + 4 × 10G	48 Gb/s	40 Gb/s	1.2 : 1	campus access
48 × 1G + 2 × 10G	48	20	2.4 : 1	campus access, budget
48 × 10G + 6 × 40G	480	240	2 : 1	older DC leaf
48 × 25G + 8 × 100G	1 200	800	1.5 : 1	DC leaf, general compute
48 × 25G + 6 × 100G	1 200	600	2 : 1	DC leaf, budget
32 × 100G split 16/16	1 600	1 600	1 : 1	storage, AI
64 × 400G split 32/32	12 800	12 800	1 : 1	AI back-end

Ratio = sum of downlink speeds ÷ sum of uplink speeds. 3 : 1 is tolerable for office traffic, 1 : 1 required where any-to-any traffic runs at line rate (GPU fabrics). Uplinks are usually aggregated (LAG or ECMP), so their optics fail as a group of independent links rather than one big pipe.

Uplink modules and slots

Campus switches often take a removable uplink module: 4 × SFP+, 2 × SFP28 + 2 × QSFP28, 4 × 25G, 2 × 40G/100G. Points to check:

the module has its own supported-optics list and sometimes its own power budget;
ports on the module may not support breakout or 1G;
some modules are hot-swappable, others need a reload;
stacking may consume uplink ports or a dedicated slot (Stacking & MLAG).

Aggregating uplinks

Method	What it gives	Optics consequence
LAG / port-channel (LACP)	N links to one neighbour act as one	all members same speed and, ideally, same module type; one member with CRC errors degrades hashed flows
MLAG / vPC / VLT	LAG across two neighbours	members land on two switches — count optics for both
L3 ECMP	N routed uplinks, equal cost	leaf–spine default; each uplink independent, BFD detects failures
Stacking ring	switch-to-switch backplane	dedicated cables/optics, not counted as uplinks

A single 100G uplink versus 4 × 25G LAG: same capacity on paper, but a single flow is capped at one member's speed in a LAG, while the 100G carries it whole (Breakout & MPO cabling).

Breakout as an uplink strategy

A 100G QSFP28 uplink split into 4 × 25G reaches four leaves from one spine port; a 400G port split into 4 × 100G quadruples radix. Which ports allow it, how children are named and what the module must advertise: Port types & cages, Port naming & LEDs.

Choosing uplink optics

Question	Answer decides
Distance and fibre type	SR/SR4 on MMF ≤ 100–400 m; DR/FR/LR on SMF for 500 m–10 km; ER/ZR/CWDM/DWDM beyond (Reach tables)
Fibre count available	parallel (MPO) SR4/DR4 need 8 fibres; duplex LC FR4/LR4/CWDM4 need 2
Both ends' cages	100G QSFP28 to 25G SFP28 → breakout; QSFP-DD to QSFP28 → QSFP28 module in the DD cage
FEC and speed policy of both NOSes	must match (Port configuration recipes)
Module policy of both vendors	OEM, coded or third-party (How each NOS validates a module)
Power budget of the uplink cage	40 km ER4 / ZR modules may exceed it (Power & consumption)
Redundancy	two uplinks on two different modules/paths; spares on site

Downlink specifics

PoE on copper access ports: 15.4 / 30 / 60 / 90 W per port (802.3af/at/bt); PoE budget shares the PSU with the optics — a fully loaded PoE switch may derate.
Multigig RJ45 (2.5/5/10GBASE-T) uses USXGMII PHYs (SGMII & serial gigabit).
Server downlinks at 25G/100G: DAC ≤ 3 m, AOC beyond, optics rarely (DAC vs AOC).
Copper SFPs in SFP downlinks: SGMII mode needed for 10/100 clients.

Typical port faults by role

Role	Frequent problem	Where to look
Uplink	FEC/speed mismatch between two vendors; wrong fibre polarity on MPO	Speed & rate, No link
Downlink (server)	DAC without AN, 25G port group speed conflict	Port types & cages
Downlink (copper SFP)	10/100 client will not link	SGMII & serial gigabit
Stacking	non-stacking cable/module in a stack port	Stacking & MLAG
Combo	SFP inserted but RJ45 side active	media-type setting
Management	in-band and OOB confused, wrong VRF	Management & monitoring

In CodingBox

Whatever the role, each fibre port reads its module's memory first. Confirming a module's speed codes, power class and identity on the bench before install — especially for uplinks between different vendors — avoids a port that silently stays down (Check transceiver).

Choosing the module for the fibre behind the port — matching table by fibre type and distance, budget from the module's side, connector and polish rules, common mismatches: Transceivers in the link.