Announcing a route into the Internet Default-Free Zone (DFZ) using the Border Gateway Protocol (BGP) has a cost borne by "everybody else" on the Internet. Because this cost is very large, network operators (Internet Service Providers) generally refuse to carry routes more specific than /24 and have asked the Internet registries to avoid allocating CIDR blocks with a longer prefix than /20. This article will attempt to assess the actual cost given numbers that are accurate circa 2008.
This article assumes you have a basic familiarity with how BGP works on the Internet and how provider aggregatable (PA) address space, provider independent (PI) address space and traffic engineering (TE) interact with the BGP system. If you are not familiar with these terms, you are at risk of misinterpreting the results.
Summary of results: The economic impact of one BGP route announcement is around $8000 per year regardless of the type of announcement or the number of included IP addresses. The cost of operating the BGP system as a whole is around $2B per year.
Analysis:
BGP has two cost components:
#1 has a trivial cost: generally less than a couple-hundred kilobits per second on a typically 100mbps or gigabit link. Because it can have only a minor effect, I exclude it from the cost calculation.
#2 has a significant cost. The price of DFZ-capable routers starts around $40,000 each and rises through $500,000.
[The cost of one route in one such router per year] will equal [the portion of the router's cost attributable to the routing table] divided by [the expected deployment lifetime of the router in its DFZ application] divided by [the total route count in the DFZ].
The deployment lifetime for a DFZ router is about 3 years. After 3 years they are either significantly upgraded, retired to another role or resold on the second-hand market.
The number of routes in the DFZ as of 2/2008 is around 245,000 as measured on any DFZ router.
An entry level DFZ router, such as the Cisco 7606 with the RSP720-3CXL card has a street price around $40,000 but how much of that is attributeable to the size of the routing table?
A Cisco 6506 with a Supervisor2+PFC/MSFC is capable of essentially the same tasks and used to be usable as a DFZ router. It supports fewer packets per second, but still well within bounds for the entry-level DFZ router task. The key reason it's no longer usable as a DFZ router is that it doesn't support a large enough route table. Such 6506 systems are readily available for around $10,000.
Modern stackable 1U switches such as the Cisco 3750G are also capable of performing the layer-3 tasks of a 7606 with the exception of the route table. They support comparable numbers of packets per second but can only support a few thousand routes while a 7606 can support one million. A 3750G setup comparable to a DFZ-tasked 7606 except for the route table size also costs around $10,000.
The cost attributable to the size of the routing table is [the cost of a router which supports the full DFZ task] minus [the cost of a router which supports everything except the large route table]: $40,000 - $10,000 = $30,000.
Note that this number is a lower bound. If all DFZ routers cost $40,000 then this would be the cost of the routing table. Some DFZ routers cost much more, as much as $500,000. The cost attributable to the route table is higher in these devices.
Thus [the lower bound cost of one route in one router per year] = ($30,000 / 3) / 245,000 = $0.04, four cents per route per router per year.
There are, however, a large number of routers in the Internet DFZ. There are 27,000 active autonomous systems (ASes) on the Internet, each of which has at least one DFZ router, most of which have at least two DFZ routers and some of which have hundreds. The exact number of routers is unknown, but per discussions on the ARIN PPML and NANOG mailing lists it is believed to be somewhere between 120,000 and 180,000 with the consensus number near 150,000.
The [economic impact of a BGP route] must then be [the lower bound cost of the route in one router per year] times [the number of routers which must carry that route]: $0.04 * 150,000 = $6,200/year. It costs "everybody else" at least $6200/year to carry each BGP route you announce.
A similar calculation yields an upper bound in the neighborhood of $12,000 per year with a probable cost in the neighborhood of $8,000 per year.
In most router implementations, IPv6 routes consume twice the route table capacity of IPv4 routes. The cost of an IPv6 route should thus be double: $16,000 per year.
Note that $8,000/routeyear * 245,000 routes = two billion dollars per year. Perhaps this explains Cisco and Juniper's annual gross revenues.
Now that you know the cost of a BGP route in the DFZ, can you really afford Provider Independent space? Are you sure you want to disaggreagate your network into /24's?
Food for thought.
William Herrin, 2/2008