Since a couple of years I have been using bonding to achieve load balancing, higher bandwidth and redundancy on Linux (and BSD, Windows, Solaris, ...). At home this can be rather challenging since lots of consumer switches don't understand all the bonding methods and Realtek adapters have the unpleasant habit of keeping the spoofed MAC address after a reboot which messes things up.

For this set up I will use a bunch of cheap Realtek and Intel NICs to connect 2 PC's using 4 NICs on every side to benchmark the actual speed.

I already mentioned that my Realtek adapters tend to remember the spoofed MAC address during reboot which causes UDEV to give the adapters a new name and rendering your config worthless. To prevent this I added some lines inside /etc/network/interfaces that start and stop the bond and load and unload the bonding module which causes the NIC to have it's factory MAC address back before rebooting.

This is my basic config using bonding mode 0 (round-robin). I only changed the mode= or in case of mode=2 and mode=4 with L3+L4 hashing I uncommented the xmit_hash_policy variable. All benchmarked modes give fault tolerance and load balancing (albeit not always visible in the results)

# The primary network interface
auto bond0
iface bond0 inet static
pre-up modprobe bonding mode=0 miimon=100 #xmit_hash_policy=layer3+4
pre-up ip link set enp4s0 master bond0
pre-up ip link set enp6s0 master bond0
pre-up ip link set enp8s0 master bond0
pre-up ip link set enp9s0 master bond0
mtu 9000
up /bin/true
down /bin/true
post-down ip link set enp4s0 nomaster
post-down ip link set enp6s0 nomaster
post-down ip link set enp8s0 nomaster
post-down ip link set enp9s0 nomaster
post-down ip link set dev bond0 down
post-down rmmod bonding

Mode=0 is only good for connecting 2 systems with multiple NICs directly to each other to achieve a single connection with a higher bandwidth. In switched environments or even worse in a bridge this method will really mess up your connections. In a switched connection you will see random packet loss and out of order packets. In a bridge you are lucky if you even get some packages over.

To get the most out of this I set the MTU to 9000 (jumbo frames) and connected 2 systems directly to each other. My NICs are all auto sensing so I didn't have to use crossed cables.

I used these scripts to run multiple instances of iperf (network benchmark tool) in parallel

method mode 0 (Round Robin) mode 2 (XOR) mode 2 (XOR L3+L4) mode 4 (LACP L2) mode 4 (LACP L3+L4)
single connection speed 3.14 Gbits/sec 719 Mbits/sec 661 Mbits/sec 730 Mbits/sec 725 Mbits/sec
total speed of 4 simultaneous connections 3.157 Gbits/sec 722 Mbits/sec 2.107 Gbit/s 735 Mbits/sec 1.484 Gbits/sec
ethtool advertised speed 4000Mb/s 4000Mb/s 4000Mb/s 4000Mb/s 4000Mb/s
Where to use inter system connection only

inter system connection OR on a switch that only supports static LAGs

inter system connection OR on a L3 switch than only supports static LAGs (...)

inter system connection OR on a L2 managed switch that supports LACP inter system connection OR on a L3 switch that supports LACP

As you can see mode=0 wins the benchmark but as I already said that comes with a price. Netgear for instance will recommend mode=2 for unmanaged switches or switches that only can handle static LAGs and mode=4 for switches that support LACP.

But in real life you should always use LACP since it is the most robust method out there and if you really need higher single connection speed you will have to invest in 10, 25, 40 or 100Gbit connections.

LACP can combine any form of hashing. You can combine hashing the MAC address (L2), the IP address (L3) or the UDP/TCP session (L4). This isn't only the case for Linux. Solaris and derivates like SmartOS just as well give you the option to combine any of these 3 hashing methods in your LACP aggregate. In Solaris L4 hashing is the default one, in Linux it is L2.

You can see in my result that running multiple iperf sessions on different ports really does make a difference if you use L4 hashing since we have 4 different TCP sessions. In the real world LACP will rarely be used for single connection configurations. If you use LACP underneath a bridge for your hypervisor for instance, since all your VMs and containers will have their own MAC address, IP address and use their own UDP/TCP sessions all your physical connections will be actually used and you will get a higher total bandwidth (albeit hard to benchmark) but you will never get more than 1 Gbit per session.

To finish, one mode that doesn't have anything to do with load balancing but can be nice is mode=1. I used this in the past to set up a fiber connection as primary link with a WiFi link as backup. In case the fiber stops working, traffic would be send over the WiFi link. Of course this kind of behavior can be achieved by using STP as well if you have a managed switch on both ends.