ethernet channel bonding

George Bonser george at captech.com
Wed Nov 3 12:51:57 MST 1999


> 
> Most of the ethernet drivers detect this. In an HA situation you don't rely
> on that anyway. If that interface isnt hearing multicasted heartbeats or
> OSPF routing frames assume its broken

One of the things that is interesting to me is to add some reliability at
the hardware level. For example, if I have a NIC that will properly
support channel bonding, I can connect the channels to different physical
switches that will also do bonding. One channel goes to one switch and one
channel goes to the other switch. The two switches are on the same vlan.
In normal operation, the switches will load balance the traffic over the
two connections. If one switch fails, traffic still keeps coming.

The topology currently used a great deal here in the Silicon Valley at
many ha colocation operations are two ethernet channels coming from the
network provider that go to different switches but are bonded together.
Internet traffic is balanced over the connections to the customer's
space.

Each of these goes to a different switch in the customer's rack or cage.
These are usually layer 2/3 switches that will do things like packet
filtering, NAT, routing, etc. You then have two load balancers and a
server farm. Each device has two NICs that are bonded together but
connect to each switche. This provides redundancy at the physical
level that software can not provide. If a NIC goes bad, someone
disconnects a cable, or a switch dies, traffic still flows.

Once you have the physical redundancy, software can take care of failover
between the load balancers. The use of bonding saves IP addresses and
reduces complexity.

Bottom line is that software only provides part of the solution in HA.





More information about the Linux-HA mailing list