Section

Switches

A switch is the device that decides, frame by frame, which physical port gets the next packet — and almost everything that matters about buying, configuring or troubleshooting one comes down to three things: whether it can move packets at the speed its ports advertise (non-blocking), whether its forwarding table is large enough for your network, and whether it does anything beyond forwarding (VLANs, PoE, link aggregation, management). Everything below is the reasoning behind those three, in the order you actually have to decide them.

5 articles

Switches
01

All in Security

Newest first
Pulling switches out of a mechanical keyboard safely
Switches

Pulling switches out of a mechanical keyboard safely

Push the switch straight down until the two retention clips on its top and bottom faces clear the plate, then lift it out — pulling at an angle or yanking the keycap still attached is what breaks…

14 SepRead →
What makes a keyboard mechanical
Switches

What makes a keyboard mechanical

A mechanical keyboard is one where each key sits on its own independent physical switch, and that switch closes the circuit itself when you press it.

8 SepRead →
Mechanical keyboard switches and how they feel
Switches

Mechanical keyboard switches and how they feel

The feel of a mechanical keyboard switch comes down to four measurable things: the force curve it presents as you press, where along the travel that force peaks and drops, whether it makes a noise…

7 SepRead →
What exactly is a mechanical keyboard switch
Switches

What exactly is a mechanical keyboard switch

What is the difference between linear, tactile, and clicky switches? Simply put, the distinction lies in their actuation profile: linear switches glide smoothly from top to bottom with no bump…

1 SepRead →
A guide to different mechanical keyboard switches
Switches

A guide to different mechanical keyboard switches

The best mechanical keyboard switch for you depends on whether your priority is the audible click, the tactile bump, or silent speed.

30 AugRead →

A switch is the device that decides, frame by frame, which physical port gets the next packet — and almost everything that matters about buying, configuring or troubleshooting one comes down to three things: whether it can move packets at the speed its ports advertise (non-blocking), whether its forwarding table is large enough for your network, and whether it does anything beyond forwarding (VLANs, PoE, link aggregation, management). Everything below is the reasoning behind those three, in the order you actually have to decide them.

The real constraint is the switching fabric, not the port count

A switch is non-blocking when its internal switching capacity is at least the sum of all its ports in both directions at once. A 24-port gigabit switch has 24 ports × 1 Gbps × 2 (full duplex) = 48 Gbps of aggregate traffic to absorb, and a unit rated below that will drop frames under a full-mesh load even though every link negotiates at a gigabit. Vendors quote this as "switching capacity" or "switching bandwidth"; the number to look for on the datasheet is one at or above the port arithmetic. If the spec sheet lists a figure below it, the switch is oversubscribed: fine for a desk cluster where only a few ports ever saturate, wrong for a storage or uplink-heavy rack.

The other half of the same problem is the uplink. Twenty-four gigabit access ports sharing one gigabit uplink to the core is a 24:1 contention ratio; the access ports negotiate fine and the congestion only shows up as latency when several hosts burst at once. The fix is either a bigger uplink (10G SFP+ or higher) or link aggregation — bundling multiple physical links into one logical link using LACP. Aggregation adds bandwidth and redundancy, but it costs switch ports and requires configuration on both ends; a misconfigured LAG is one of the most common causes of a "random" packet loss that turns out to be a single-link hash collision.

Where this breaks down: for a small office with a handful of clients and no local servers, a non-blocking fabric and a fat uplink are money spent on a problem you do not have. The uplink is the binding constraint far more often than the fabric.

Forwarding is a table lookup, and the table's size is a hard limit

Every switch keeps a MAC address table (also called the forwarding database or CAM table). When a frame arrives, the switch looks up the destination MAC; if it finds an entry, it forwards out that one port; if it does not, it floods the frame out every port in the same VLAN. The table has a fixed size — typically quoted in the low thousands of entries on unmanaged and small managed units, and much higher on enterprise gear. When the table fills, the switch ages out entries and starts flooding, and the symptom is a network that works but leaks broadcast-like traffic everywhere and slows under load.

For a home or small office this number is irrelevant. It becomes relevant the moment you put a virtualization host on the switch: a single hypervisor with many virtual machines and bridged networking can present hundreds of MAC addresses on one physical port, and a few such hosts can exhaust a small table. If you run containers or VMs on the same switch as your laptops, check the MAC table size before you buy, not after the network starts flooding.

The forwarding decision itself is done in hardware (an ASIC) on any real switch, which is why store-and-forward switches add latency proportional to frame size — the switch buffers the whole frame and runs a CRC check before forwarding, so a large frame takes longer to cross than a small one. Cut-through switches start forwarding after reading the destination address, cutting latency at the cost of forwarding frames that later turn out to be corrupt. For most office and home use store-and-forward is the right default; cut-through earns its keep in latency-sensitive trading or HPC fabrics, and the difference is measured in microseconds, not milliseconds.

Managed, smart, or unmanaged: decide by whether you need VLANs

If you need VLANs, link aggregation, port mirroring, or any per-port policy, you need a managed switch. If you need none of those, an unmanaged switch is cheaper, has no configuration to break, and is the correct choice for a desk cluster or a media cabinet. The middle category — variously sold as "smart," "web-managed," or "easy smart" — gives you VLANs and a web interface without a full CLI, and it is where most small networks should land.

The trade-off is not just price. Managed switches carry firmware, and firmware means an attack surface and a patching obligation. An unmanaged switch has no IP address to compromise; a managed one on a flat network is a device you must keep updated and must not leave on the default password. The practical sequence is: buy unmanaged where the only job is "more ports," buy smart-managed where you want to separate a guest network or a lab from the main LAN, and buy fully managed where you need SNMP monitoring, redundant power, or stacking. Paying for a managed switch and then never configuring it is the most common waste in small-network purchasing.

VLANs themselves are the reason most people move up a tier. A VLAN segments broadcast domains on one physical switch; traffic between VLANs must be routed, either by a router (router-on-a-stick, where one link carries all VLANs as tagged sub-interfaces) or by a Layer 3 switch that routes between them in hardware. Router-on-a-stick is cheap and fine until inter-VLAN traffic grows, at which point the single link becomes the bottleneck and an L3 switch pays for itself.

Power over Ethernet changes the power budget, and the budget is the spec

PoE lets one cable carry both data and power to access points, cameras, and phones. The decision is not "PoE or not" but which PoE standard and how many watts the switch can supply in total. The switch's total power budget — often quoted separately from its per-port wattage — is the number that bites: a switch rated for a certain per-port output can still be unable to power every port at that output simultaneously, because the total budget is smaller than ports × per-port watts. Add up the actual draw of every powered device, leave headroom, and compare against the total budget, not the per-port figure.

The cost of getting this wrong is a switch that powers most of your cameras and then reboots one of them when a new device is plugged in, or refuses to power the last port. The other cost is heat and noise: a PoE switch running near its budget runs hot and its fans run loud, which matters if it lives in a closet next to a desk. Non-PoE devices on a PoE switch are unaffected — the switch negotiates before applying power — so buying PoE "for the future" costs only the price difference and the heat, not compatibility.

Wireless is not a switch, and the numbers show why the two get confused

An access point connects to a switch through one wired port, and the wireless side is governed by a completely different set of limits. IEEE 802.11ax (Wi-Fi 6) supports a maximum throughput of 20 gigabits per second at the MAC data service access point, but that is an aggregate figure for the standard, not what any one client sees. The realistic per-client numbers are far smaller: 2.4 Gbps theoretical maximum for 802.11ax with 2 spatial streams and a 160 MHz channel, 1.2 Gbps for 2×2 at 80 MHz, 287 Mbps for 2×2 at 20 MHz, and 143 Mbps for a single-stream 1×1 client at 20 MHz.

Those figures explain the most common misdiagnosis in small networks: someone upgrades to a Wi-Fi 6 access point, wires it to a gigabit switch port, and wonders why throughput did not improve. The wired port is the ceiling. A single gigabit uplink caps the access point regardless of what the radio can do, and even a 2.5G or 10G uplink only helps if the switch port and the cabling support it. The access point's radio capability and the switch's port speed have to be matched, and the switch is usually the cheaper half to upgrade.

It also explains why 20 MHz and 80 MHz channels behave so differently in dense environments: narrower channels carry less per client but interfere less, and the 143 Mbps and 287 Mbps figures for 20 MHz are the honest expectation for a crowded apartment building where 160 MHz channels are unusable. The wide-channel numbers are real but only in clean spectrum.

Buying sequence, and what each step costs

Work in this order, because each decision constrains the next.

The honest trade-off across all of it: the cheapest switch that meets your port count will usually forward fine, and the money you save buys nothing you will miss until you add VLANs, a virtualization host, or a PoE camera. The failure mode of overbuying is a managed switch you never configure and firmware you never patch. The failure mode of underbuying is a flooding MAC table or a saturated uplink that looks like an intermittent fault for weeks. Buy the fabric and the table size you can justify, buy management only if you will use it, and treat the uplink and the PoE budget as the two specifications most likely to be the real limit.

Adding a switch to a live network without a maintenance window

Plug the new switch in with no uplink, let it boot, then connect the uplink last. A switch with an empty MAC address table floods every unknown unicast frame out of every port in the VLAN, and if you introduce it between two already-running segments it will briefly forward traffic in loops until STP settles — the symptom is a broadcast storm that saturates links for the duration of convergence, not a clean cutover. Bringing it up in isolation means the table is empty, STP has already elected a root on the new device's own ports, and the only flooding after you connect the uplink is the normal learning burst.

Check spanning tree before you connect anything. On a managed switch, confirm the bridge priority on your existing core is lower than any default (most units ship at the same default priority, so the root election is decided by MAC address — effectively random), and confirm the new switch's ports are set to the edge/portfast equivalent for access ports so a workstation plugging in does not trigger a topology change. A port without that setting will run the full listening-and-learning sequence on every link event, which on a busy desk cluster shows up as a few seconds of dropped traffic each time someone docks a laptop.

The cost of doing it this way is a second trip to the rack and a few minutes with a console cable. The cost of skipping it is a storm that takes down the segment you were trying to extend, at whatever hour you chose to avoid disruption.

Where a switch sits relative to a router, and why the distinction still matters

A switch forwards frames inside a subnet by MAC address; a router forwards packets between subnets by IP address. A Layer 3 switch does both, but the reason to keep them distinct in your head is that inter-VLAN traffic on an L3 switch never leaves the box, while the same traffic routed through a separate router crosses a physical link twice. That is the mechanism behind the router-on-a-stick bottleneck: every packet between VLANs traverses the single trunk in and back out, so the trunk's speed caps all inter-VLAN throughput combined.

An L3 switch avoids that by routing in the ASIC, but it costs more and it moves your default gateway off the router — which means your firewall rules, NAT, and any DHCP reservations tied to the router now sit on the wrong device unless you re-plan them. For a small network with one guest VLAN and light cross-subnet traffic, router-on-a-stick is cheaper and simpler and the trunk is not the limit. The moment you add a NAS or a virtualization host that clients in another VLAN pull from heavily, the single trunk becomes the constraint and the L3 switch earns its price.

The limit worth knowing: an L3 switch routes between VLANs, but it does not replace the firewall. ACLs on a switch filter by address and port, not by application, and putting your only inter-VLAN policy there means the guest network's access to the management VLAN is governed by a rule set most people configure once and never audit.