The BGP Link Bandwidth feature is used to advertise the bandwidth of an autonomous system exit link as an extended community. This feature is configured for links between directly connected external BGP neighbors. This feature is used with BGP multipath features to configure load balancing over links with unequal bandwidth.
Prerequisites
•BGP load balancing or multipath must be configured
•BGP extended community must be configured to be sent between iBGP neighbors
•CEF must be enabled
Restrictions
•Can only be configured under IPv4 and VPNv4 address family
•BGP can originate the link bandwidth community only for directly connected eBGP peers
•iBGP and eBGP load balancing are supported in IPv4 and VPNv4 address families
•eiBGP load balancing is supported only in VPNv4 address-family
Overview
The BGP Link Bandwidth feature is used to enable multipath load balancing for external links with unequal bandwidth. When this feature is enabled, paths learned from directly connected eBGP peers are propagated through to the iBGP peers with the bandwidth of the external link. The link bandwidth extended community indicates the preference of an autonomous system exit link.
The attribute is a 4-byte value that is configured for a link that on the interface that connects two single hop eBGP peers. The extended community is used as a traffic sharing value. Two paths are equal for load balancing if the weight, local-pref, as-path, MED and IGP costs are the same.
source: http://www.cisco.com/en/US/docs/ios/12_2t/12_2t2/feature/guide/ftbgplb.html
Update: I made a practical example of the feature. Be sure to have a quick look as it will start to make a bit more sense:
BGP dmzlink-bw Unequal-cost load-balance
