Why bond server network ports

Port bonding (or, at the switch level, link aggregation) solves two problems: it increases channel throughput beyond the speed of a single port and adds fault tolerance — if one cable or switch port goes down, traffic keeps flowing through the remaining interfaces without dropping connections.

Most servers have 2-4 network ports at 1 or 10 Gbps. Individually each one is limited to its own speed, while bonding two 10 Gbps ports in LACP mode theoretically gives 20 Gbps of total throughput, although a single TCP connection is still limited to the speed of one physical link.

Bonding modes: which one to choose

In Linux, the bonding driver supports several modes, but in practice three are used:

active-backup (mode 1) — one port is active, the other is standby. Requires no switch configuration but does not increase speed.
balance-rr (mode 0) — packets go out round-robin across ports. Gains speed for UDP but often breaks TCP packet order.
802.3ad, also known as LACP (mode 4) — the standard aggregation protocol, negotiated with the switch. It distributes connections by a hash function of MAC/IP/port and requires LACP configuration on both sides.

Production servers in a data center use LACP: it is the only one of the three officially negotiated with the switch side and it correctly handles a link failure.

Configuring LACP on the server and switch

On Linux with NetworkManager, a bond is created like this:

nmcli connection add type bond ifname bond0 mode 802.3ad
nmcli connection modify bond0 +bond.options "lacp_rate=fast,miimon=100"
nmcli connection add type ethernet ifname eth0 master bond0
nmcli connection add type ethernet ifname eth1 master bond0
nmcli connection up bond0

With netplan the configuration looks like this:

network:
  version: 2
  bonds:
    bond0:
      interfaces: [eth0, eth1]
      parameters:
        mode: 802.3ad
        lacp-rate: fast
        mii-monitor-interval: 100
      addresses: [203.0.113.10/24]

On the switch side, the ports the server is connected to are grouped into a port-channel with LACP active enabled. If the switch is set to static mode (without LACP), bonding in 802.3ad mode will not work — both ends must negotiate the protocol.

How to check that bonding is working

Interface status and LACP negotiation are checked like this:

cat /proc/net/bonding/bond0
ip -s link show bond0

In the output, the important line is MII Status: up for each slave interface, and the Aggregator ID — it must match on both ports if they are actually aggregated. After configuration, throughput should be measured with iperf3 on both sides at once, running several parallel streams (-P 4) to engage hashing across different connections.

Common mistakes: why LACP "does not come up"

The problem is most often a mismatch between server and switch settings. Common causes:

Symptom	Cause
Bond is up, but speed is like a single port	Switch ports are not in the same port-channel
Aggregator ID differs between ports	Different VLANs or speed on switch ports
Link keeps flapping	miimon is too small or the cable/SFP is faulty
A single TCP connection does not speed up	The LACP hash algorithm distributes by flow, not by splitting one flow

Another common mistake is enabling LACP on the server but leaving the switch ports in separate VLANs without a port-channel. Then bond0 comes up locally, but there is no real aggregation: traffic simply goes through one physical port.

Checklist before going into production

LACP mode (802.3ad) is configured the same way on the server and the switch, and the ports are grouped into one port-channel.
In /proc/net/bonding/bond0, all slave ports show MII Status: up and the same Aggregator ID.
Throughput has been verified with a load test using several parallel streams.
Failover has been tested manually: unplugging one cable must not drop active connections.
If the server needs several public addresses on top of bond0, IP aliases or additional subnets are configured on the bond0 interface itself, not on eth0/eth1.

Network checks should be part of the general list during server acceptance — a failed link or a misconfigured LACP setup is easier to catch in the first hour of operation than a month later under load.

Bonding and LACP: Combining Server Network Ports