Link aggregation is the technique of combining several physical Ethernet links into one logical link between two devices, so that they behave as a single interface with more capacity and built-in redundancy. In networking, and specifically in Ethernet and switching, it is standardized as IEEE 802.1AX, originally IEEE 802.3ad, and is known under vendor names such as EtherChannel, port channel, LAG, bond or team.
An aggregated link with four 10 Gb/s members offers 40 Gb/s of total capacity, and if one cable or port fails, traffic continues over the remaining members. The key subtlety is that a single traffic flow normally does not exceed the speed of one member link, because of the way traffic is distributed.
How link aggregation distributes traffic
Each device in a link aggregation group (LAG) chooses a member link for every outgoing frame with a hash function. The hash takes header fields, such as MAC addresses, IP addresses and TCP or UDP ports, and maps the result to one member. All frames of the same flow produce the same hash, so they travel over the same physical link and arrive in order.
This per-flow approach avoids the packet reordering that per-packet round-robin distribution would cause, which would hurt TCP performance. The price is that balance depends on traffic diversity: many flows spread well, while a few large flows can overload one member while others stay idle. The table below compares common hashing policies.
| Hash input | Typical name | Balance quality | Best suited to |
|---|---|---|---|
| Source and destination MAC | Layer 2 | Poor when traffic passes one router | Pure layer 2 segments with many hosts |
| Source and destination IP | Layer 3 | Good with many host pairs | Routed traffic between many endpoints |
| IP addresses and ports | Layer 3+4 | Best for many sessions | Servers with many client connections |
| Tunnel inner headers | Encapsulation-aware | Good for overlay traffic | VXLAN and similar overlays |
Each side of a LAG hashes its own outgoing traffic independently, so the path of a request and its reply can use different member links. This is normal and does not affect correctness.
LACP and static link aggregation
A LAG can be configured statically or negotiated with the Link Aggregation Control Protocol (LACP). LACP exchanges small control frames, called LACPDUs, so that both ends agree which ports belong to the group and detect misconfiguration before traffic is affected.
| Aspect | Static LAG | LACP |
|---|---|---|
| Negotiation | None; ports are forced into the group | Partners exchange LACPDUs |
| Detection of miswiring | No; can create loops or black holes | Yes; mismatched ports stay out |
| Failure detection | Physical link state only | Link state plus protocol timeout |
| Timers | Not applicable | Fast: 1 s, 3 s timeout; slow: 30 s, 90 s timeout |
| Recommended use | Devices without LACP support | Default choice wherever supported |
Each LACP port runs in active or passive mode. At least one side must be active, and using active on both ends is the most robust configuration. Fast timers detect a failed partner within about three seconds, which matters when a link stays physically up but stops forwarding.
Requirements and configuration parameters
Member links of one LAG must be consistent in speed, duplex and layer 2 settings. Mismatched VLAN lists, MTU or native VLAN settings usually keep a port suspended. The table summarizes the parameters administrators set most often.
| Parameter | Typical values | Effect |
|---|---|---|
| Number of members | 2 to 8 active, more on some platforms | Total capacity and redundancy |
| Member speed | 1, 10, 25, 40 or 100 Gb/s, all equal | Maximum speed of a single flow |
| Minimum links | 1 up to the member count | Takes the LAG down below a capacity threshold |
| LACP rate | Fast or slow | Speed of failure detection |
| Hash policy | Layer 2, layer 3 or layer 3+4 | Evenness of load distribution |
Using a power-of-two number of members, such as two, four or eight, gives the most even distribution on many switches, because hash results are often split into a fixed number of buckets.
Link aggregation compared with related technologies
Standard link aggregation connects exactly two devices. When a server must connect to two switches for redundancy, vendors offer multi-chassis link aggregation (MLAG or MC-LAG), in which two switches present themselves as one LACP partner. Stacking achieves a similar result by making several switches act as one managed switch.
| Method | Layer | Uses all links | Typical use |
|---|---|---|---|
| Link aggregation | Layer 2 | Yes | Switch to switch, server to switch |
| MLAG or switch stacking | Layer 2 | Yes, across two switches | Dual-homed servers and access switches |
| Spanning tree | Layer 2 | No; blocks redundant links | Loop prevention in switched networks |
| ECMP routing | Layer 3 | Yes, per flow | Routed fabrics and data center spines |
| NIC failover bonding | Host | No; standby link only | Simple server redundancy |
Spanning tree treats a LAG as one logical port, so redundant members are not blocked. This is one of the main reasons link aggregation is preferred over parallel independent links between the same pair of switches.
Common mistakes and limitations
Most link aggregation problems come from expectations about speed or from inconsistent configuration on the two ends.
- Expecting a single file copy to double in speed is the most common misconception; one TCP flow stays on one member.
- Configuring a static LAG on one side and LACP on the other leaves ports suspended or causes loops.
- Choosing a layer 2 hash for routed traffic puts nearly all traffic on one member, because every frame carries the router’s MAC address.
- Connecting LAG members to two independent switches without MLAG support breaks the aggregation.
- Using different cable types or optics on members can produce subtle errors on only one link, visible in interface counters.
Storage protocols such as iSCSI often achieve better results with multipath I/O over separate links instead of a LAG, because multipathing balances individual sessions explicitly.
In simple terms
Link aggregation is like turning several single-lane roads into one multi-lane highway between two buildings. More cars can travel at the same time, and if one lane is closed, traffic keeps moving on the others.
Each car still stays in its own lane for the whole trip, though. That is why one large download does not become faster, while many downloads together can use all the lanes at once.
A practical example
A small company runs a virtualization host that serves about 60 employees and is connected to its core switch by a single 10 Gb/s link. During backups the link reaches full use, and a cable fault once took every virtual machine offline for an hour.
The administrator adds a second 10 Gb/s port, creates an LACP group with active mode on both ends, fast timers and a layer 3+4 hash policy. Daytime traffic from many users now spreads across both links, the nightly backup no longer slows other services, and a later test of pulling one cable causes no interruption for users. The backup job itself still peaks at about 10 Gb/s, as expected for a single large flow.