Network Working Group T. Narten
Internet-Draft IBM
Intended status: Informational February 17, 2010
Expires: August 21, 2010
On the Scalability of Internet Routing
draft-narten-radir-problem-statement-05.txt
Abstract
There has been much discussion over the last years about the overall
scalability of the Internet routing system. Some have argued that
the resources required to maintain routing tables in the core of the
Internet are growing faster than available technology will be able to
keep up. Others disagree with that assessment. This document
attempts to describe the factors that are placing pressure on the
routing system and the growth trends behind those factors.
Status of this Memo
This Internet-Draft is submitted to IETF in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF), its areas, and its working groups. Note that
other groups may also distribute working documents as Internet-
Drafts.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
The list of current Internet-Drafts can be accessed at
http://www.ietf.org/ietf/1id-abstracts.txt.
The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html.
This Internet-Draft will expire on August 21, 2010.
Copyright Notice
Copyright (c) 2010 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Narten Expires August 21, 2010 [Page 1]
Internet-Draft On the Scalability of Internet Routing February 2010
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include Simplified BSD License text as described in Section 4.e of
the Trust Legal Provisions and are provided without warranty as
described in the BSD License.
This document may contain material from IETF Documents or IETF
Contributions published or made publicly available before November
10, 2008. The person(s) controlling the copyright in some of this
material may not have granted the IETF Trust the right to allow
modifications of such material outside the IETF Standards Process.
Without obtaining an adequate license from the person(s) controlling
the copyright in such materials, this document may not be modified
outside the IETF Standards Process, and derivative works of it may
not be created outside the IETF Standards Process, except to format
it for publication as an RFC or to translate it into languages other
than English.
Narten Expires August 21, 2010 [Page 2]
Internet-Draft On the Scalability of Internet Routing February 2010
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 3
2. Terms and Definitions . . . . . . . . . . . . . . . . . . . . 4
3. Background . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1. Technical Aspects . . . . . . . . . . . . . . . . . . . . 7
3.2. Business Considerations . . . . . . . . . . . . . . . . . 7
3.3. Alignment of Incentives . . . . . . . . . . . . . . . . . 8
3.4. Table Growth Targets . . . . . . . . . . . . . . . . . . . 9
4. Pressures on Routing Table Size . . . . . . . . . . . . . . . 10
4.1. Traffic Engineering . . . . . . . . . . . . . . . . . . . 10
4.2. Multihoming . . . . . . . . . . . . . . . . . . . . . . . 11
4.3. End Site Renumbering . . . . . . . . . . . . . . . . . . . 12
4.4. Acquisitions and Mergers . . . . . . . . . . . . . . . . . 12
4.5. RIR Address Allocation Policies . . . . . . . . . . . . . 12
4.6. Dual Stack Pressure on the Routing Table . . . . . . . . . 13
4.7. Internal Customer Routes . . . . . . . . . . . . . . . . . 14
4.8. IPv4 Address Exhaustion . . . . . . . . . . . . . . . . . 14
5. Pressures on Control Plane Load . . . . . . . . . . . . . . . 15
5.1. Interconnection Richness . . . . . . . . . . . . . . . . . 15
5.2. Multihoming . . . . . . . . . . . . . . . . . . . . . . . 15
5.3. Traffic Engineering . . . . . . . . . . . . . . . . . . . 15
5.4. Questionable Operational Practices? . . . . . . . . . . . 16
5.4.1. Rapid shuffling of prefixes . . . . . . . . . . . . . 16
5.4.2. Anti-Route Hijacking . . . . . . . . . . . . . . . . . 16
5.4.3. Operational Ignorance . . . . . . . . . . . . . . . . 16
5.5. RIR Policy . . . . . . . . . . . . . . . . . . . . . . . . 17
6. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
7. Security Considerations . . . . . . . . . . . . . . . . . . . 20
8. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 21
9. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 22
10. Informative References . . . . . . . . . . . . . . . . . . . . 23
Author's Address . . . . . . . . . . . . . . . . . . . . . . . . . 24
Narten Expires August 21, 2010 [Page 3]
Internet-Draft On the Scalability of Internet Routing February 2010
1. Introduction
Prompted in part by the October, 2006 IAB workshop on Routing &
Addressing [RFC4984], there has been a renewed focus on the topic of
routing scalability within the Internet. The issue itself is not
new, with discussions dating back at least 10-15 years [GSE, ROAD].
This document attempts to describe the "pain points" being placed on
the routing system, with the aim of describing the essential aspects
so that the community has a way of evaluating whether proposed
changes to the routing system actually address or impact existing
pain points in a significant manner.
Narten Expires August 21, 2010 [Page 4]
Internet-Draft On the Scalability of Internet Routing February 2010
2. Terms and Definitions
Control Plane: The routing setup protocols, their associated state
and the activity needed to create and maintain the data structures
used to forward packets from one network to another. The term is
defined broadly to include all protocols and activities needed to
construct and maintain the forwarding tables used to forward
packets.
Control Plane Load: The actual load associated with operating the
Control Plane. The higher the control plane load, the higher the
cost of operating the control plane (in terms of hardware,
bandwidth, power, etc.). The terms "routing load" and "control
plane load" are used interchangeably throughout this document.
Control Plane Cost: The overall cost associated with operating the
Control Plane. The cost consists of capital costs (for hardware),
bandwidth costs (for the control plane signalling) and any other
ongoing operational cost associated with operating and maintaining
the control plane.
Default Free Zone (DFZ): That part of the Internet where routers
maintain full routing tables. Many routers maintain only partial
tables, having explicit routes for "local" destinations (i.e.,
prefixes) plus a "default" for everything else. For such routers,
building and maintaining routing tables is relatively simple
because the amount of information learned and maintained can be
small. In contrast, routers in the DFZ maintain complete
information about all reachable destinations, which at the time of
this writing number in the hundreds of thousands of entries.
Routing Information Base (RIB): The data structures a router
maintains that hold the information about destinations and paths
to those destinations. The amount of state information maintained
is dependent on a number of factors, including the number of
individual prefixes learned from peers, the number of BGP peers,
the number of distinct paths interconnecting destinations, etc.
In addition to maintaining information about active paths used for
forwarding, the RIB may also include information about unused
("backup") paths.
Forwarding Information Base (FIB): The actual table consulted while
making forwarding decisions for individual packets. The FIB is a
compact, optimized subset of the RIB, containing only the
information needed to actually forward individual packets, i.e.,
mapping a packet's destination address to an outgoing interface
and next-hop. The FIB only stores information about paths
actually used for forwarding; it typically does not store
Narten Expires August 21, 2010 [Page 5]
Internet-Draft On the Scalability of Internet Routing February 2010
information about backup paths. The FIB is typically constructed
from specialized hardware components, which have different (and
higher) cost properties than the hardware typically used to
maintain the RIB.
Traffic Engineering (TE): In this document, "traffic engineering"
refers to the current practice of inbound, inter-AS traffic
engineering. TE is accomplished by injecting additional, more-
specific routes into the routing system and/or increasing the
frequency of routing updates in order to arrange for inbound
traffic at the boundary of an Autonomous system (AS) to travel
over a different path than it otherwise would.
Provider Aggregatable (PA) address space: Address space that an end
site obtains from an upstream ISP's address block. The main
benefit of PA address space is that reachability to all of a
provider's customers can be achieved by advertising a single
"provider aggregate" address prefix into the DFZ, rather than
needing to announce individual prefixes for each customer. An
important disadvantage is that when a customer changes providers,
the customer must renumber their site into addresses belonging to
the new provider and return the previously used addresses to the
former provider.
Provider Independent (PI) address space: Address space that an end
site obtains directly from a Regional Internet Registry (RIR) for
addressing its devices. The main advantage (for the end site) is
that it does not need to renumber its site when changing
providers, since it continues to use its PI block. However, PI
address blocks are not aggregatable and thus each individual PI
assignment results in an individual prefix being injected into the
DFZ.
Site: Any topologically and administratively distinct entity that
connects to the Internet. A site can range from a large
enterprise or ISP to a small home site.
Narten Expires August 21, 2010 [Page 6]
Internet-Draft On the Scalability of Internet Routing February 2010
3. Background
Within the DFZ, both the size of the RIB and FIB and the overall
update rate have historically increased at a greater than linear
rate. Specifically:
o The number of individual prefixes that are being propagated into
the DFZ over time has been and continues to increase at a faster-
than-linear rate. The term "super-linear" has been used to
characterize the growth. The exact nature of the growth is much
debated (e.g., quadratic, polynomial, etc.), but growth is clearly
faster than linear. The reasons behind the rate increase are
varied and discussed below. Because each individual prefix
requires resources to process, any increase in the number of
prefixes produces a corresponding increase in control plane load
of the routing system. Each individual prefix that appears in
routing updates requires state in the RIB (and possibly the FIB)
and consumes processing and other resources when updates related
to the prefix are received.
o The overall rate of routing updates is increasing [1], requiring
routers to process updates at an increased rate or converge more
slowly if they cannot. The rate increase of the control plane
load is driven by a number of factors (discussed below). Further
study is needed to better understand the factors behind the
increasing update rate. For example, it appears that a
disproportionate increase in observed updates originates from a
small percentage of the total number of advertised prefixes.
The super-linear growth in the routing load presents a scalability
challenge for current and/or future routers. While there appears to
be general agreement that we will be able to build routers (i.e.,
hardware & software) actually capable of handling the control plane
load, both today and going forward, there is considerable debate
about the cost. In particular, will it be possible for ISPs that
currently (or would like to) maintain routes as part of the DFZ be
able to afford to do so, or will only the largest (and a shrinking
number) of top tier ISPs be able to afford the investment and cost of
operating the control plane while being a part of the DFZ?
Finally, the scalability challenge is aggravated by the lack of any
firm limiting architectural upper-bound on the growth rate of the
routing load and a weakening of social constraints that historically
have helped restrain the growth rate so far. Going forward, there is
considerable uncertainty (some would say doubt) whether future growth
rates will continue to be sufficiently constrained so that router
development can keep up at an acceptable price point.
Narten Expires August 21, 2010 [Page 7]
Internet-Draft On the Scalability of Internet Routing February 2010
3.1. Technical Aspects
The technical challenge of building routers relates to the resources
needed to process a larger and increasingly dynamic amount of routing
information. More specifically, routers must maintain an increasing
amount of associated state information in the RIB, they must be
capable of populating a growing FIB, they must perform forwarding
lookups at line rates (while accessing the FIB) and they must be able
to initialize the RIB and FIB after system restart. All of these
activities must take place within acceptable time frames (i.e., paths
for individual destinations must converge and stabilize within an
acceptable time period). Finally, the hardware needed to achieve
this cannot have unreasonable power consumption or cooling demands.
3.2. Business Considerations
While the IETF does not (and cannot) concern itself with business
models or the profitability of the ISP community, the cost of running
the routing subsystem as a whole is directly influenced by the
routing architecture of the Internet, which clearly is the IETF's
business. Thus, it is useful to consider the overall business
environment that underlies operation of the DFZ routing
infrastructure. The DFZ is run entirely by the private sector with
no overall governmental oversight or regulatory framework to oversee
or even influence what routes are propagated where, who must carry
them, etc. ISPs decide (on their own) which routing updates to
accept and how (if at all) to process them. Thus, there is no
overall authority that can limit the number of prefixes that are
injected into the DFZ or that insure that any particular prefixes are
accepted at all. Today, the system functions because the set of
entities that comprise the DFZ are (generally) able to accept the
prefixes that are being advertised and some loose best practices have
emerged that are generally followed (e.g., minimum prefix sizes that
are routed coupled with RIR policies that place limitations on who
may obtain PI prefixes).
In general the Internet would benefit if the cost of the (routing)
infrastructure did not grow too rapidly as the Internet grows, since
a lower infrastructure cost makes it possible to provide Internet
service at a lower cost to a larger number of users. That said, some
types of Internet growth tie directly to revenue opportunities or
cost savings for an ISP (e.g., adding more users/customers,
increasing bandwidth, technological advances, providing new or
additional services, etc.). Upgrading or changing infrastructure is
most feasible (and expected) when supported by a workable cost
recovery model. Hence limiting the cost of self-induced scaling is a
nice-to-have benefit, but not a requirement.
Narten Expires August 21, 2010 [Page 8]
Internet-Draft On the Scalability of Internet Routing February 2010
On the other hand, it is problematic when the infrastructure cost for
an ISP grows (rapidly) due to factors outside of its own control,
e.g., resulting from overall Internet growth external to the ISP. If
an ISP that does not add new customers, upgrade the bandwidth for
their customers, or provide new services needs to upgrade or replace
their infrastructure in unexpected ways, then they have no natural
cost recovery mechanisms. This is in essence what is happening with
the scaling of the global routing table. An ISP that is part of the
DFZ may need to upgrade its routers to handle an increased routing
load just to maintain the same level of service with respect to their
current customers and services.
Even if it is technically possible to build routers capable of
meeting the technical and operational requirements, it is also
necessary that the overall cost to build, maintain and deploy such
equipment meet reasonable business expectations. ISPs, after all,
are run as businesses. As such, they must be able to plan, develop
and construct viable business plans that provide an acceptable return
on investment (i.e., one acceptable to investors).
3.3. Alignment of Incentives
Today's growth pattern is influenced by the scaling properties of the
current routing system. If the routing system had better scaling
properties, we would be able support and enable more widespread usage
of such services as multihoming and traffic engineering. The current
system simply would not be able to handle to the routing load if
everyone were to choose to multihome. There are millions of
potential end sites that would benefit from being able to multihome.
This compares with a low few hundred thousand prefixes being carried
today. Broader availability of multihoming is limited by barriers
imposed by operational practices that try to strike a balance between
the amount of multihoming and preservation of routing slots. It is
desirable that the routing and addressing system exert the least
possible back pressure on end user applications and deployment
scenarios, to enable the broadest possible use of the Internet.
One aspect of the current architecture is a misalignment of cost and
benefit. Injecting individual prefixes into the DFZ creates a small
amount of "pain" for those routers that are part of the DFZ. Each
individual prefix adds only a small cost to the routing load, but the
aggregate sum of all prefixes is significant, and leads to the key
issue at hand. Those that inject prefixes into the DFZ do not
generally pay the cost associated with the individual prefix -- it is
carried by the routers in the DFZ. But the originator of the prefix
receives the benefit. Hence, there is misalignment of incentives
between those receiving the benefit and those bearing the cost of
providing the benefit. Consequently, incentives are not aligned
Narten Expires August 21, 2010 [Page 9]
Internet-Draft On the Scalability of Internet Routing February 2010
properly to produce a natural feedback loop to balance the cost and
benefit of maintaining routing tables.
3.4. Table Growth Targets
A precise target for the rate of table size or routing update
increase that should reasonably be supported going forward is
difficult to state in quantitative terms. One target might simply be
to keep the growth at a stable, but manageable growth rate so that
the increased router functionality can roughly be covered by
improvements in technology (e.g., increased processor speeds,
reductions in component costs, etc.).
However, it is highly desirable to significantly bring down (or even
reverse) the growth rate in order to meet user expectations for
specific services. As discussed below, there are numerous pressures
to deaggregate routes. These pressures come from users seeking
specific, tangible service improvements that provide "business-
critical" value. Today, some of those services simply cannot be
supported to the degree that future demand can reasonably be expected
because of the negative implications on DFZ table growth. Hence,
valuable services are available to some, but not all potential
customers. As the need for such services becomes increasingly
important, it will be difficult to deny such services to large
numbers of users, especially when some "lucky" sites are able to use
the service and others are not.
Narten Expires August 21, 2010 [Page 10]
Internet-Draft On the Scalability of Internet Routing February 2010
4. Pressures on Routing Table Size
There are a number of factors behind the increase in the quantity of
prefixes appearing in the DFZ. From a theoretical perspective, the
number of prefixes in the DFZ can be minimized through aggressive
aggregation [RFC4632]. In practice, strict adherence to the CIDR
principles is difficult.
4.1. Traffic Engineering
Traffic engineering (TE) is the act of arranging for certain Internet
traffic to use or avoid certain network paths (that is, TE attempts
to place traffic where capacity exists, or where some set of
parameters of the path is more favorable to the traffic being placed
there).
Outbound TE is typically accomplished by using internal interial
gateway protocol (IGP) metrics to choose the shortest exit for two
equally good BGP paths. Adjustment of IGP metrics controls how much
traffic flows over different internal paths to specific exit points
for two equally good BGP paths. Additional traffic can be moved by
applying some policy to depreference or filter certain routes from
specific BGP peers. Because outbound TE is achieved via a site's own
IGP, outbound TE does not impact routing outside of a site.
Inbound TE is performed by announcing a more-specific route along the
preferred path that "catches" the desired traffic and channels it
away from the path it would take otherwise (i.e., via a larger
aggregate). At the BGP level, if the address range requiring TE is a
portion of a larger address aggregate, network operators implementing
TE are forced to de-aggregate otherwise aggregatable prefixes in
order to steer the traffic of the particular address range to
specific paths.
TE is performed by both ISPs and customer networks, for three primary
reasons:
o to match traffic with network capacity, or to spread the traffic
load across multiple links (frequently referred to as "load
balancing")
o to reduce costs by shifting traffic to lower cost paths or by
balancing the incoming and outgoing traffic volume to maintain
appropriate peering relations
o to enforce certain forms of policy (e.g., to prevent government
traffic from transiting through other countries)
Narten Expires August 21, 2010 [Page 11]
Internet-Draft On the Scalability of Internet Routing February 2010
TE impacts route scaling in two ways. First, inbound TE can result
in additional prefixes being advertised into the DFZ. Second,
Network operators usually achieve traffic engineering by "tweaking"
the processing of routing protocols to achieve desired results, e.g.,
by sending updates at an increased rate. In addition, some devices
attempt to automatically find better paths and then advertise those
preferences through BGP, though the extent to which such tools are in
use and contributing to the control plane load is unknown.
In today's highly competitive environment, providers require TE to
maintain good performance and low cost in their networks.
4.2. Multihoming
Multihoming refers generically to the case in which a site is served
by more than one ISP [RFC4116]. Multihoming is used to provide
backup paths (i.e., to remove single points of failure), to achieve
load-sharing, and to achieve policy or performance objectives (e.g.,
to use lower latency or higher bandwidth paths). Multihoming may
also be a requirement due to contract or law.
Multihoming can be accomplished using either PI or PA address space.
A multihomed site advertises its site prefix into the routing system
of each of its providers. For PI space, the site's PI space is used,
and the prefix is propagated throughout the DFZ. For PA space, the
PA site prefix may (or may not) be propagated throughout the DFZ,
with the details depending on what type of multihoming is sought.
If the site uses PA space, the PA site prefix allocated from one of
its providers (whom we'll call the Primary Provider) is used. The PA
site prefix will be aggregatable by the Primary Provider but not the
others. To achieve multihoming with comparable properties to that
when PI addresses are used as described above, the PA site prefix
will need to be injected into the routing system of all of its ISPs,
and throughout the DFZ. In addition, because of the longest-match
forwarding rule, the Primary Provider must advertise both its
aggregate and the individual PA site prefix; otherwise, the path via
the primary provider (as advertised via the aggregate) will never be
selected due to the longest match rule. For the type of multihoming
described here, where the PA site prefix is propagated throughout the
DFZ, the use of PI vs. PA space has no impact on the control plane
load. The increased load is due entirely to the need to propagate
the site's individual prefix throughout the DFZ.
The demand for multihoming is increasing [2]. The increase in
multihoming demand is due to the increased reliance on the Internet
for mission and business-critical applications (where businesses
require 7x24 availability for their services) and the general
Narten Expires August 21, 2010 [Page 12]
Internet-Draft On the Scalability of Internet Routing February 2010
decrease in cost of Internet connectivity.
4.3. End Site Renumbering
It is generally considered painful and costly to renumber a site,
with the cost proportional to the size and complexity of the network
and most importantly, to the degree that addresses are stored in
places that are difficult in practice to update. When using PA
space, a site must renumber when changing providers. Larger sites
object to this cost and view the requirement to renumber akin to
being held "hostage" to the provider from which PA space was
obtained. Consequently, many sites desire PI space. Having PI space
provides independence from any one provider and makes it easier to
switch providers (for whatever reason). However, each individual PI
prefix must be propagated throughout the DFZ and adds to the control
plane load.
It should be noted that while larger sites may also want to
multihome, the cost of renumbering drives some sites to seek PI
space, even though they do not multihome.
4.4. Acquisitions and Mergers
Acquisitions and mergers take place for business reasons, which
usually have little to do with the network topologies of the impacted
organizations. When a business sells off part of itself, the assets
may include networks, attached devices, etc. A company that
purchases or merges with other organizations may quickly find that
its network assets are numbered out of many different and
unaggragatable address blocks. Consequently, an individual
organization may find itself unable to announce a single prefix for
all of their networks without renumbering a significant portion of
its network.
Likewise, selling off part of a business may involve selling part of
a network as well, resulting in the fragmentation of one address
block into two (or more) smaller blocks. Because the resultant
blocks belong to different organizations, they can no longer be
advertised by a single aggregate and the resultant fragments may need
to be advertised individually into the DFZ.
4.5. RIR Address Allocation Policies
ISPs and multihoming end sites obtain address space from RIRs. As an
entity grows, it needs additional address space and requests more
from its RIR. In order to be able to obtain additional address space
that can be aggregated with the previously-allocated address space,
the RIR must keep a reserve of space that the requester can grow into
Narten Expires August 21, 2010 [Page 13]
Internet-Draft On the Scalability of Internet Routing February 2010
in the future. But any reserved address space cannot be used for any
other purpose (i.e., assigned to another organization). Hence, there
is an inherent conflict between holding address space in reserve to
allow for the future growth of an existing allocation holder and
using address space efficiently. In IPv4, there has been a heavy
emphasis on conserving address space and obtaining efficient
utilization. Consequently, insufficient space has been held in
reserve to allow for the growth of all sites and some allocations
have had to be made from discontiguous address blocks. That is, some
sites have received discontiguous address blocks because their growth
needs exceeded the amount of space held in reserve for them.
In IPv6, its vast address space allows for a much a greater emphasis
to be placed placed on preserving future aggregation than was
possible in IPv4.
4.6. Dual Stack Pressure on the Routing Table
The recommended IPv6 deployment model is dual-stack, where IPv4 and
IPv6 are run in parallel across the same links. This has two
implications for routing. First, although alternative scenarios are
possible, it seems likely that many routers will be supporting both
IPv4 and IPv6 simultaneously and will thus be managing both IPv4 and
IPv6 routing tables within a single router. Second, for sites
connected via both IPv4 and IPv6, both IPv4 and IPv6 prefixes will
need to be propagated into the routing system. Consequently, dual-
stack routers will maintain both an IPv4 and IPv6 route to reach the
same destination.
It is possible to make some simple estimates on the approximate size
of the IPv6 tables that would be needed if all sites reachable via
IPv4 today were also reachable via IPv6. In theory, each autonomous
system (AS) needs only a single aggregate route. This provides a
lower bound on the size of the fully-realized IPv6 routing table.
(As of Feb 2010, [3] states there are 33,548 active ASes in the
routing system.)
A single IPv6 aggregate will not allow for inbound traffic
engineering. End sites will need to advertise a number of smaller
prefixes into the DFZ if they desire to gain finer grained control
over their IPv6 inbound traffic. This will increase the size of the
IPv6 routing table beyond the lower bound discussed above. There is
reason to expect the IPv6 routing table will be smaller than the
current IPv4 table, however, because the larger initial assignments
to end sites will minimize the de-aggregation that occurs when a site
must go back to its upstream address provider or RIR and receive a
second, non-contiguous assignment.
Narten Expires August 21, 2010 [Page 14]
Internet-Draft On the Scalability of Internet Routing February 2010
It is possible to extrapolate what the size of the IPv6 Internet
routing table would be if widespread IPv6 adoption occurred, from the
current IPv4 Internet routing table. Each active AS (33,548) would
require at least one aggregate. In addition, the IPv6 Internet table
would also carry more-specific prefixes for traffic engineering.
Assume that the IPv6 Internet table will carry the same number of
more specifics as the IPv4 Internet table. In this case one can take
the number of IPv4 Internet routes and subtract the number of CIDR
aggregates that they could easily be aggregated down to. As of Feb
2010, the 313,626 routes can be easily aggregated down to 193,844
CIDR aggregates [3]. That difference yields 119,782 extra more-
specific prefixes. Thus if each active AS (33,548) required one
aggregate, and an additional 119,782 more specifics were required,
then the IPv6 Internet table would be 153,330 prefixes.
4.7. Internal Customer Routes
In addition to the Internet routing table, networks must also support
their internal routing table. Internal routes are defined as more-
specific routes that are not advertised to the DFZ. This primarily
consists of prefixes that are a more-specific of a provider aggregate
(PA) and are assigned to a single-homed customer. The DFZ need only
carry the PA aggregate in order to deliver traffic to the provider.
However, the provider's routers require the more-specific route to
deliver traffic to the end site.
Internal routes could also come from more-specific prefixes
advertised by multihomed customers with the "no-export" BGP
community. This is useful when the fine grained control of traffic
to be influenced can be contained to the neighboring network.
For a large ISP, the internal IPv4 table can be between 50,000 and
150,000 routes. During the dot com boom some ISPs had more internal
prefixes than there were in the Internet table. Thus the size of the
internal routing table can have significant impact on the scalability
and should not be discounted.
4.8. IPv4 Address Exhaustion
The IANA and RIR free pool of IPv4 addresses will be exhausted within
a few years. As the free pool shrinks, the size of the remaining
unused blocks will also shrink and unused blocks previously held in
reserve for expansion of existing allocations or otherwise not used
due to their smaller size will be allocated for use. Consequently,
as the community looks to use every piece of available address space
(no matter how small) there will be an increasing pressure to
advertise additional prefixes in the DFZ.
Narten Expires August 21, 2010 [Page 15]
Internet-Draft On the Scalability of Internet Routing February 2010
5. Pressures on Control Plane Load
This section describes a number of trends and pressures that are
contributing to the overall routing load. The previous section
described pressures that are increasing the size of the routing
table. Even if the size could be bounded, the amount of work needed
to maintain paths for a given set of prefixes appears to be
increasing.
5.1. Interconnection Richness
The degree of interconnectedness between ASes has increased in recent
years. That is, the Internet as a whole is becoming "flatter" with
an increasing number of possible paths interconnecting sites [4]. As
the number of possible paths increase, the amount of computation
needed to find a best path also increases. This computation comes
into effect whenever a change in path characteristics occurs, whether
from a new path becoming available, an existing path failing, or a
change in the attributes associated with a potential path. Thus,
even if the total number of prefixes were to stay constant, an
increase in the interconnection richness implies an increase in the
resources needed to maintain routing tables.
5.2. Multihoming
Multihoming places pressure on the routing system in two ways.
First, an individual prefix for a multihomed site (whether PI or PA)
must be propagated into the routing system, so that other sites can
find a good path to the site. Even if the site's prefix comes out of
a PA block, an individual prefix for the site needs to be advertised
so that the most desirable path to the site can be chosen when the
path through the aggregate is sub-optimal. Second, a multihomed site
will be connected to the Internet in more than one place, increasing
the overall level of interconnection richness. If an outage occurs
on any of the circuits connecting the site to the Internet, those
changes will be propagated into the routing system. In contrast, a
singly-homed site numbered out of a Provider Aggregate places no
additional control plane load in the DFZ as the details of the
connectivity status to the site are kept internal to the provider to
which it connects.
5.3. Traffic Engineering
The mechanisms used to achieve multihoming and inbound Traffic
Engineering are the same. In both cases, a specific prefix is
advertised into the routing system to "catch" traffic and route it
over a different path than it would otherwise be carried. When
multihoming, the specific prefix is one that differs from that of its
Narten Expires August 21, 2010 [Page 16]
Internet-Draft On the Scalability of Internet Routing February 2010
ISP or is a more-specific of the ISP's PA. Traffic Engineering is
achieved by taking one prefix and dividing it into a number of
smaller and more-specific ones, and advertising them in order to gain
finer-grained control over the paths used to carry traffic covered by
those prefixes.
Traffic Engineering increases the number of prefixes carried in the
routing system. In addition, when a circuit fails (or the routing
attributes associated with the circuit change), additional load is
placed on the routing system by having multiple prefixes potentially
impacted by the change, as opposed to just one.
5.4. Questionable Operational Practices?
Some operators are believed to engage in operational practices that
increase the load on the routing system.
5.4.1. Rapid shuffling of prefixes
Some networks try to assert fine-grained control of inbound traffic
by modifying route announcements frequently in order to migrate
traffic to less loaded links quickly. The goal of this is to achieve
higher utilization of multiple links. In addition, some route
selection devices actively measure link or path utilization and
attempt to optimize inbound traffic by withholding or depreferencing
certain prefixes in their advertisements. In short, any system that
actively measures load and modifies route advertisements in real time
increases the load on the routing system, as any change in what is
advertised must ripple through the entire routing system.
5.4.2. Anti-Route Hijacking
In order to reduce the threat of accidental (or intentional)
hijacking of its address space by an unauthorized third party, some
sites advertise their space as a set of smaller prefixes rather than
as one aggregate. That way, if someone else advertised a path for
the larger aggregate (or a small piece of the aggregate), it will be
ignored in favor of the more-specific announcements. This increases
both the number of prefixes advertised, and the number of updates.
5.4.3. Operational Ignorance
It is believed that some undesirable practices result from operator
ignorance, where the operator is unaware of what they are doing and
the impact that has on the DFZ.
The default behavior of most BGP configurations is to automatically
propagate all learned routes. That is, one must take explicit
Narten Expires August 21, 2010 [Page 17]
Internet-Draft On the Scalability of Internet Routing February 2010
configuration steps to prevent the automatic propagation of learned
routes. In addition, it is often significant work to figure out how
to (safely) aggregate routes (and which ones to aggregate) in order
to reduce the number of advertisements propagated elsewhere. While
vendors could provide additional configuration "knobs" to reduce
leakage, the implementation of additional features increases
complexity and some operators may fear that the new configuration
will break their existing routing setup. Finally, leaking routes
unnecessarily does not generally harm those responsible for the
misconfiguration, hence, there may be little incentive to change such
behavior.
5.5. RIR Policy
RIR address policy has direct impact on the control plane load
because address policy determines who is eligible for a PI assignment
(which impacts how many are given out in practice) and the size of
the assignment (which impacts how much address space can be
aggregated within a single assignment). If PI assignments for end
sites did not exist, then those end sites would not advertise their
own prefix directly into the global routing system; instead their
address block would be covered by their provider's aggregate. That
said, RIRs have adopted PI policies in response to community demand,
for reasons described elsewhere (e.g., to support multihoming and to
avoid the need to renumber). In short, RIR policy can be seen as a
symptom rather than a root cause.
Narten Expires August 21, 2010 [Page 18]
Internet-Draft On the Scalability of Internet Routing February 2010
6. Summary
As discussed in previous sections, in the current operating
environment, an ISP may experience an overall increase in the routing
load due entirely to external factors outside of its control. These
external pressures can make it increasingly difficult for ISPs to
recover control plane related costs associated with the growth of the
Internet. Moreover, real business and user needs are creating
increasing pressure to use techniques that increase the control plane
load for ISPs operating within the DFZ. While the system largely
works today, there is a real risk that the current cost and incentive
structures will be unable to keep control plane costs manageable
(within the context of then-available routing hardware) over the next
decades. The Internet would strongly benefit from a routing and
addressing model designed with this in mind. Thus, in the absence of
a business model that better supports such cost recovery, there is a
need for an approach to routing and addressing that fulfils the
following criteria:
1. Provides sufficient benefits to the party bearing the costs of
deploying and maintaining the technology to recover the cost for
doing so.
2. Reduces the growth rate of the DFZ control plane load. In the
current architecture, this is dominated by the routing, which is
dependent on:
A. The number of individual prefixes in the DFZ
B. The update rate associated with those prefixes.
Any change to the control plane architecture must result in a
reduction in the overall control plane load, and shouldn't simply
shift the load from one place in the system to another, without
reducing the overall load as a whole.
3. Allows any end site wishing to multihome to do so
4. Supports ISP and enterprise TE needs
5. Allows end sites to switch providers while minimizing
configuration changes to internal end site devices.
6. Provides end-to-end convergence/restoration of service at least
comparable to that provided by the current architecture
This document has purposefully been scoped to focus on the growth of
the routing control plane load of operating the DFZ. Other problems
Narten Expires August 21, 2010 [Page 19]
Internet-Draft On the Scalability of Internet Routing February 2010
that may seem related, but do not directly impact on route scaling
are not considered to be "in scope" at this time. For example,
Mobile IP [RFC3344] [RFC3775] and NEMO [RFC3963] place no pressures
on the routing system. They are layered on top of existing IP, using
tunneling to forward packets via a care-of addresses. Hence,
"improving" these technologies (e.g., by having them leverage a
solution to the multihoming problem), while a laudable goal, is not
considered a necessary goal.
Narten Expires August 21, 2010 [Page 20]
Internet-Draft On the Scalability of Internet Routing February 2010
7. Security Considerations
This document does not introduce any security considerations.
Narten Expires August 21, 2010 [Page 21]
Internet-Draft On the Scalability of Internet Routing February 2010
8. IANA Considerations
This document contains no IANA actions.
Narten Expires August 21, 2010 [Page 22]
Internet-Draft On the Scalability of Internet Routing February 2010
9. Acknowledgments
The initial version of this document was produced by the Routing and
Addressing Directorate (http://www.ietf.org/IESG/content/radir.html).
The membership of the directorate at that time included Marla
Azinger, Vince Fuller, Vijay Gill, Thomas Narten, Erik Nordmark,
Jason Schiller, Peter Schoenmaker, and John Scudder.
Comments should be sent to rrg@iab.org or to radir@ietf.org.
Narten Expires August 21, 2010 [Page 23]
Internet-Draft On the Scalability of Internet Routing February 2010
10. Informative References
[RFC3344] Perkins, C., "IP Mobility Support for IPv4", RFC 3344,
August 2002.
[RFC3775] Johnson, D., Perkins, C., and J. Arkko, "Mobility Support
in IPv6", RFC 3775, June 2004.
[RFC3963] Devarapalli, V., Wakikawa, R., Petrescu, A., and P.
Thubert, "Network Mobility (NEMO) Basic Support Protocol",
RFC 3963, January 2005.
[RFC4116] Abley, J., Lindqvist, K., Davies, E., Black, B., and V.
Gill, "IPv4 Multihoming Practices and Limitations",
RFC 4116, July 2005.
[RFC4632] Fuller, V. and T. Li, "Classless Inter-domain Routing
(CIDR): The Internet Address Assignment and Aggregation
Plan", BCP 122, RFC 4632, August 2006.
[RFC4984] Meyer, D., Zhang, L., and K. Fall, "Report from the IAB
Workshop on Routing and Addressing", RFC 4984,
September 2007.
[1] <http://www3.ietf.org/proceedings/06mar/slides/grow-3.pdf>
[2] <http://www.cidr-report.org/as2.0/, http://www.cidr-report.org/
cgi-bin/
plota?file=%2fvar%2fdata%2fbgp%2fas2.0%2fbgp%2das%2dcount%2etxt&
descr=Unique%20ASes&ylabel=Unique%20ASes&with=step,http://
www.potaroo.net/tools/asn32/>
[3] <http://www.cidr-report.org/as2.0/>
[4] <http://www.potaroo.net/bgprpts/bgp-average-aspath-length.png>
Narten Expires August 21, 2010 [Page 24]
Internet-Draft On the Scalability of Internet Routing February 2010
Author's Address
Thomas Narten
IBM
Email: narten@us.ibm.com
Narten Expires August 21, 2010 [Page 25]
