LACP

Why it exists

Parallel Ethernet links between two switches are a loop unless something collapses them into one logical port. Spanning tree’s answer is to block; aggregation’s answer is to forward on all of them as one.

Static bundling forwards on a member as soon as the interface is up, with no confirmation the far end agrees. LACP exchanges identity and state first, and refuses members that do not match — which is how it tells a one-way failure or a crossed cable from a working extra link. IEEE 802.1AX-2020 §5

Two cables become one port

Active LACP on both ends exchanges LACPDUs, matches system ID and key, and moves both members into one aggregator. Spanning tree then sees a single logical link.

How it works

Each end advertises an Actor System identifier, an operational key, port priority, port number, and state flags. It also mirrors what it believes about the partner. Matching Actor and Partner information is what moves a member into the aggregator. IEEE 802.1AX-2020 §5

Activity is a flag: active transmits LACPDUs unsolicited; passive waits to be asked. Two passive ends wait forever. Prefer active on both.

Hashing is a local transmit decision. Each end independently chooses which member to send a flow on. A bundle distributes flows, not packets — one conversation is capped at one member’s speed.

On the wire

Constructed examples, encoded from the field table below them — not captured traffic.

Configure it

When it breaks

Symptom first, because that is what you have when it happens.

1. A bundle is configured on both ends but only one member is passing traffic, or the bundle never forms.

   Narrow it down

   1. Check the mode on each end: active, passive, or on.
   2. Confirm both ends are not passive — two passive ends never start.
   3. Look for one end configured statically while the other speaks LACP.

   Cause

   Passive waits to be asked. Two passive ends wait forever. A static bundle facing an LACP bundle is worse: the static end forwards immediately on every member while the other end has not agreed, which is a loop rather than a bundle.

   Fix

   Use active on at least one end, and prefer active on both. Never configure a static bundle facing a protocol one.

   IEEE 802.1AX-2020 §5

2. Traffic is unevenly distributed across members — one link saturated, the rest nearly idle.

   Narrow it down

   1. Read the hashing policy in use on the transmitting device.
   2. Characterise the traffic: how many distinct flows, and do they vary in the fields being hashed?
   3. Check whether the heavy traffic is a single flow.

   Cause

   A bundle distributes flows, not packets. A hash over source and destination MAC address puts everything between two routers on one member, and a single large transfer is one flow no matter how the hash is configured.

   Fix

   Hash on fields that actually vary — usually including layer 4 ports. Accept that one flow cannot exceed one member; a bundle adds capacity for many conversations, not for one.

3. A member fails but traffic on the bundle keeps being blackholed.

   Narrow it down

   1. Check whether the failure was detected — a dark fibre with light on one strand keeps the interface up.
   2. Look at whether LACP is still receiving on that member.
   3. Check the configured min-links.

   Cause

   The interface never went down, so the member stayed in the bundle. LACP would have noticed the missing protocol packets, but a static bundle has nothing to notice with.

   Fix

   Run LACP rather than static bundling, and consider micro-BFD on the members where sub-second detection matters.

   IEEE 802.1AX-2020 §5

4. Members stay individual or show partner information that does not match the neighbour you expect.

   Narrow it down

   1. Compare Actor System and Actor Key on each end of every member.
   2. Trace cables — a crossed member lands on the wrong partner System.
   3. Confirm channel-group numbers and keys are consistent for the intended aggregator.

   Cause

   LACP only bundles when Actor and Partner information agree. A crossed cable or a key mismatch leaves the member out of the aggregator by design.

   Fix

   Fix the cabling or the key. Do not switch the member to static to “make it come up” — that removes the check that caught the mistake.

   IEEE 802.1AX-2020 §5

5. show etherchannel lists a member as standby and operators treat it as a fault.

   Narrow it down

   1. Check max-bundle and port-priority configuration.
   2. Confirm the member is exchanging LACPDUs and shows a matched partner.
   3. Distinguish hot-standby from independent (mismatch) state.

   Cause

   Under a max-bundle limit, lower-priority members are intentionally standby. That is not the same as a member that never matched Partner Information.

   Fix

   If standby is unintended, raise the member’s port priority or raise max-bundle. If the member is independent, fix the Actor/Partner mismatch instead.

   IEEE 802.1AX-2020 §5

Design notes

A static bundle facing an LACP bundle is a loop risk: the static end forwards immediately while the other end has not agreed. The overhead being avoided is one small frame per second.

PAgP is Cisco-proprietary and is not this protocol. Mixed PAgP/LACP ends do not form a standard bundle. Cisco IOS XE — Configuring EtherChannels §LACP vs PAgP

Misconceptions

“A four-member bundle gives one host four times the throughput.”

It gives the bundle four times the aggregate capacity across many flows. Any single flow is hashed to one member and is capped at that member’s speed — which is why a bundle is a poor answer to "this one backup job is too slow".

“Static bundling is the same thing with less overhead.”

A static bundle forwards on a member as soon as the interface is up, with no confirmation the far end agrees. Miscabling that LACP would refuse to bundle becomes a forwarding loop. The overhead being avoided is one small frame per second.

“Both ends must use the same hashing policy.”

Hashing is a local transmit decision. Each end independently chooses which member to send on, and the two need not match — which also means an imbalance can exist in one direction only.

More walkthroughs

Terms

vPC peer-link

The Cisco MLAG link between two peers. It is not a member of the downstream vPC. If it fails while keepalive is alive, the secondary suspends member ports.

Check yourself

• Both ends of a bundle are configured passive. What happens?

  • The bundle forms normally
  • Neither end initiates, so the bundle never forms
  • One end is elected active automatically
  • The bundle forms with one member
• A single large file transfer across a four-member bundle runs at one member’s speed. Why?

  • The hash policy is misconfigured
  • A bundle distributes flows, and one transfer is one flow
  • The members negotiated different speeds
  • LACP is limiting it
• One end is LACP active and the other is static (`mode on`). What is the risk?

  • The bundle forms more slowly
  • The static end forwards before the LACP end agrees — a loop risk, not a bundle
  • Spanning tree blocks both members
  • Nothing — static and LACP interoperate
• Must both ends of a bundle use the same hash algorithm?

  • Yes — otherwise the bundle will not form
  • No — hashing is a local transmit decision and need not match
  • Yes — LACP negotiates one algorithm for both ends
  • Only if the members have different speeds
• A member’s Partner System does not match the neighbour you cabled toward. What should LACP do?

  • Force the member into the aggregator anyway
  • Keep the member individual until Actor and Partner information agree
  • Fall back to static bundling
  • Elect a new System ID
• A third member shows as standby while two others are bundling. When is that expected?

  • Never — standby always means a fault
  • When max-bundle or port priority intentionally keeps it hot-standby
  • Only when the member is passive
  • Only on Cisco platforms
• Why does LACP catch a unidirectional fibre failure that a static bundle misses?

  • It measures optical power
  • Missing LACPDUs take the member out of the aggregator even if the interface stays up
  • It relies on spanning tree
  • It disables the interface immediately
• What does the Actor System field identify?

  • The port channel number
  • This end’s System identifier — usually a MAC — used when matching Partner Information
  • The hashing algorithm
  • The spanning-tree bridge priority

Related

• vPC

  Two Nexus peers present one LACP system downstream; peer-keepalive is the split-brain detector when the peer-link fails.

• STP

  Keeps a switched network free of loops by disabling just enough ports to leave one active path between any two bridges.