EtherChannel and LACP Explained: Many Cables, One Link
An EtherChannel ties several physical cables between the same two switches into one logical link called a port-channel: more capacity, a spare path if one cable dies, and one link as far as spanning tree is concerned. This guide explains the three ways to form one — LACP, PAgP, and static on — which mode pairs actually work, the identical-settings rule behind the %EC-5-CANNOT_BUNDLE2 log message, and the per-conversation hashing that means one file transfer still only ever uses one cable.
The short answer
What is an EtherChannel? It is several physical cables between the same two switches, tied together so both switches treat them as one single link — one logical interface, called a port-channel. The bundle carries more traffic than one cable, keeps working (slower) if one cable fails, and counts as one link for spanning tree, the safety system that would otherwise block all but one of the parallel cables as a loop risk. LACP is the standard protocol that negotiates the bundle automatically; PAgP is Cisco's older own version; and static mode, called on, forces the bundle with no negotiation at all.
Think of a supermarket adding checkout lanes. One lane moves one queue of shoppers at a fixed speed. Four lanes move roughly four queues at once, and if one till breaks the shop stays open on three. But one shopper with one very large trolley still goes through a single lane at single-lane speed — extra lanes help the crowd, not the individual. An EtherChannel behaves in exactly that way, and the trolley rule is the one people forget.
What a bundle buys — and what it does not
Three things, and they are worth separating because they are often sold as one. Capacity: four 10-gigabit members give the pair of switches a 40-gigabit logical link for traffic as a whole. Survival: pull one cable and the port-channel stays up on the remaining members; nothing has to reconverge, because as far as every protocol above it is concerned nothing changed except available capacity. Simplicity: spanning tree, routing, and every feature configured on the port-channel interface see one link, so four parallel cables do not look like a loop and do not need three of them blocked.
What it does not buy is single-conversation speed. A bundle is not one fat pipe; it is several pipes with a dispatcher. The next sections explain the dispatcher, but hold the trolley in mind: one backup job, one replication stream, one very large transfer between the same two endpoints is one conversation, and it will ride one member at that member's speed no matter how wide the bundle is.
A suspended member does not take the bundle down. It quietly shrinks it — which is why the log message below matters.
Three ways to form a channel: LACP, PAgP, and on
The two ends have to agree to bundle, and there are three ways to reach that agreement. In two of them the ports exchange small negotiation messages and check each other before bundling. In the third, each side simply assumes. Modes are where almost every failed bundle is born, so read the table by pairs: what matters is the combination, not either side alone.
| Method | Modes | Combinations that form a channel | Combinations that never form one |
|---|---|---|---|
| LACP — open standard, every vendor | active starts the negotiation; passive only answers | active + active, active + passive | passive + passive — both sides wait for the other to speak first, forever |
| PAgP — Cisco only | desirable starts; auto only answers | desirable + desirable, desirable + auto | auto + auto — the same polite deadlock as passive + passive |
| Static — no protocol | on forces the port into the bundle | on + on | on mixed with anything else — there is no negotiation to meet the other side halfway |
Two rules sit across the whole table. First, the methods do not mix: Cisco's own configuration guide states that ports configured for PAgP cannot form EtherChannels with ports configured for LACP, and neither protocol interoperates with ports configured manually with on. Second, at least one side of a negotiated bundle must be the talker — active or desirable. In new designs the boring default is also the right one: LACP active on both ends. It works between any vendors, it starts promptly from either side, and because the two ends keep exchanging LACP messages it notices a member that stops behaving like part of the bundle. Static on has one legitimate use — a partner device that speaks no aggregation protocol at all — and one real danger: with no negotiation there is no check, so if the far end is cabled or configured differently the bundle can black-hole traffic or open a loop that a negotiated bundle would have refused to form.
One size note, stated carefully because it is platform territory: on many Cisco platforms LACP lets you assign up to 16 ports to one channel, of which up to 8 are active at once and the rest wait as hot standbys, while PAgP and static channels are commonly limited to 8. Your platform's configuration guide is the authority for its own numbers — check it before designing to a count.
The identical-settings rule, and %EC-5-CANNOT_BUNDLE2
Negotiation agreeing is only half the test. The switch also requires every member of a bundle to be configured identically, because the members are interchangeable: any conversation may land on any of them, so a member that treats traffic differently would corrupt the illusion that this is one link. Cisco's troubleshooting guidance lists the settings that must match, and they are the checklist to work through whenever a bundle misbehaves:
| Setting | Rule |
|---|---|
| Speed and duplex | Identical on every member — a 1-gigabit port cannot join a bundle of 10-gigabit ports. |
| Switchport mode | All access, or all trunk — never a mixture. |
| Access VLAN | The same VLAN on every member, if the ports are access ports. |
| Native VLAN and allowed VLANs | Identical on every member, if the ports are trunks. |
| Trunking behaviour such as DTP | The same on every member — one port still negotiating trunking dynamically while its partners are fixed is a mismatch. |
A port that breaks the rule is suspended: taken out of the bundle, carrying nothing, while the bundle carries on without it. And the switch tells you, precisely, in the log. The decoder entry for %EC-5-CANNOT_BUNDLE2 walks the message word by word; its sample reads: Gi1/0/25 is not compatible with Gi1/0/26 and will be suspended (dtp mode of Gi1/0/25 is on, Gi1/0/26 is off). Everything you need is in that sentence — which port, which partner it was compared with, and the exact setting that differs, in brackets at the end. The operational trap is not the failure, it is the silence afterwards: the link is up, the bundle is up, users are unaffected, and the uplink has been running at reduced capacity since a change three weeks ago. Bundles deserve a member count check after every change, not just an up/down check.
A working configuration, and how to verify it
On Cisco IOS, a two-member LACP trunk looks like this. The habit that prevents most mismatches is in it: configure the members as a range so they cannot drift apart, and put the trunk settings on the port-channel interface itself, where they apply to the bundle as a whole.
interface range GigabitEthernet1/0/25 - 26 switchport mode trunk channel-group 1 mode activeinterface Port-channel1 switchport mode trunk switchport trunk allowed vlan 10,20,30
Verify with show etherchannel summary, and read the letters, not just the link state. The port-channel line should show the bundle in use as a Layer 2 channel — marked (SU), S for Layer 2 and U for in use — and every member should carry a (P), bundled in the port-channel. Two letters are the fault-finders: (s) means suspended — go straight to the log for the %EC-5-CANNOT_BUNDLE2 line naming the differing setting, then compare that port's configuration against a working member line by line with show running-config interface — and (I) means stand-alone, a port acting as an ordinary individual link because no channel formed at all, which points back at the mode pairs in the table above: wrong protocol, or two listeners waiting for each other. The fix for a suspended port is never a reload; it is making that one port's configuration identical to its partners, after which it renegotiates and joins.
The dispatcher: load balancing is a hash, per conversation
When a frame leaves on a port-channel, the switch must pick one member for it. It does not spray packets round-robin — spraying would deliver packets out of order and break more than it fixed. Instead it hashes chosen fields of the conversation — depending on platform and configuration, source and destination MAC addresses, IP addresses, or TCP/UDP port numbers — and the hash result selects the member. Every packet of that same conversation hashes the same way, so it stays on the same member, in order.
Two consequences follow, and both show up in real troubleshooting. First, the trolley rule: one conversation is capped at one member's speed, always. A 4 x 10G bundle does not give any single flow 40G. Second, balance is statistical, not guaranteed: a handful of heavy conversations can hash onto the same member and saturate it while its partners idle — the same polarization problem that haunts equal-cost routing in fabrics, in miniature. If a bundle looks lopsided, the lever is the hash inputs: including Layer 4 port numbers spreads conversations that share the same pair of IP addresses (the classic case is many flows between the same two servers) across members far better than MAC addresses alone. Cisco IOS lets you inspect the method with show etherchannel load-balance and predict a given conversation's member with test etherchannel load-balance — check where a flow will land before blaming the cable it landed on.
Four mistakes that waste the evidence
1. Two polite ends. Passive + passive, or auto + auto, is a bundle that will never exist and never error — both sides are waiting to be asked. If a new channel simply does nothing, check who was supposed to start before checking cables.
2. Mixing protocols. LACP on one end and PAgP on the other speak different languages and form nothing; on mixed with either forms nothing safely. Standardise on LACP unless you have a specific reason not to.
3. Editing one member by hand. A quick VLAN change on a single member port is exactly how a port gets suspended at 2 a.m. three weeks later. Change the port-channel interface, or change every member identically — and afterwards confirm every member still shows (P).
4. Expecting one flow to use the whole bundle. Testing a new 4-member uplink with a single-stream speed test, seeing one member's worth of throughput, and declaring the bundle broken is the trolley mistake in uniform. Test with many parallel streams, or check the hash placement first.
For the switching foundations a bundle builds on — trunking, native VLANs, and why spanning tree blocks parallel links in the first place — read VLANs, trunks, and spanning tree; for what happens when the far side is a pair of switches pretending to be one, read vPC vs VSS vs StackWise explained.
Try it yourself
Paste a two-switch configuration pair with a port-channel into the free Network Config Verifier and look for the member mismatches it flags, then map the same pair with the free Config-to-Topology Mapper to see the bundle drawn as the single logical link both switches believe it is. And the next time a log shows %EC-5-CANNOT_BUNDLE2, open the decoder entry for %EC-5-CANNOT_BUNDLE2 — the setting named in brackets at the end of the message is the fix.
Fact-check sources: EtherChannel modes and their meanings (on forces a channel with no negotiation; auto and passive respond but do not initiate; desirable and active initiate), PAgP and LACP packets exchanged only between ports in their respective mode pairs, and the rule that PAgP, LACP, and manually configured ports do not interoperate with each other — Cisco's EtherChannel configuration guide for Catalyst switches (Understanding PAgP / IEEE 802.3ad LACP EtherChannel Configuration). LACP as the IEEE 802.3ad standard, now in IEEE 802.1AX — same Cisco guide and the IEEE 802.1AX standard family. The identical-settings requirements (same speed and duplex; same VLAN assignment or trunking; same native VLAN; the same allowed VLAN range on trunks, failing which ports do not form a channel even in PAgP auto or desirable) — Cisco's EtherChannel configuration and troubleshooting guidance for Catalyst switches. The %EC-5-CANNOT_BUNDLE2 sample message and its suspended-port meaning are quoted from the site's own decoder entry, which is sourced from Cisco EtherChannel documentation. Member counts (LACP up to 16 assigned, up to 8 active, on many platforms) are stated as platform-dependent because Cisco states them per platform — the Catalyst 3750 configuration guide is the example checked. Load-balancing behaviour (a per-conversation hash over address and port fields, configured and tested with the etherchannel load-balance commands) — Cisco IOS EtherChannel load-balancing documentation.
Related guides
- An EtherChannel (a link aggregation group on other vendors) bundles several physical links between the same pair of devices into one logical port-channel interface. Spanning tree treats the bundle as a single link, traffic is shared across the members, and if one member fails the others keep carrying traffic — the bundle gets slower, it does not go down.
- There are three ways to form a channel and the two ends must speak the same one. LACP (IEEE 802.3ad, now in IEEE 802.1AX) uses active and passive: active/active and active/passive form a channel, passive/passive never does because neither side starts. PAgP is Cisco-only and uses desirable and auto with the same pattern. Static mode on uses no protocol at all and only pairs with on. LACP and PAgP do not interoperate with each other, and neither interoperates with on.
- Every member port must be configured identically — same speed and duplex, same access or trunk mode, same access VLAN or, on trunks, the same native VLAN and allowed VLAN list. A port that differs is suspended and the switch logs %EC-5-CANNOT_BUNDLE2 naming the exact setting that differs, while the rest of the bundle keeps working at reduced capacity.
- Load balancing is per conversation, not per packet: the switch hashes fields such as source and destination addresses and puts each conversation on one member, so many conversations together can use the whole bundle but a single conversation can never go faster than one member link.