I understand (can someone confirm this?) that the CPU and memory
burden on each DFZ router scales roughly in proportion to the number
of prefixes multiplied by the number of peer routers, with a
significant contributing factor being the number of changed
advertisements. This last factor is driven by outages, permanent
changes in network structure and in particular by a small subset of
operators chopping and changing their advertised prefixes between
different routers on an hourly or daily basis to achieve Traffic
Engineering of incoming traffic.