One document matched: draft-ietf-ipngwg-esd-analysis-02.txt
Differences from draft-ietf-ipngwg-esd-analysis-01.txt
INTERNET-DRAFT Matt Crawford
Fermilab
<draft-ietf-ipngwg-esd-analysis-02.txt> Allison Mankin
ISI
Thomas Narten
IBM
John W. Stewart, III
Juniper
Lixia Zhang
UCLA
March 13, 1998
Separating Identifiers and Locators in Addresses:
An Analysis of the GSE Proposal for IPv6
<draft-ietf-ipngwg-esd-analysis-02.txt>
Status of this Memo
This document is an Internet-Draft. 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."
To learn the current status of any Internet-Draft, please check the
"1id-abstracts.txt" listing contained in the Internet-Drafts Shadow
Directories on ds.internic.net (US East Coast), nic.nordu.net
(Europe), ftp.isi.edu (US West Coast), or munnari.oz.au (Pacific
Rim).
Distribution of this memo is unlimited.
This Internet-Draft expires May 7, 1998.
Abstract
On February 27-28, 1997, the IPng Working Group held an interim
meeting in Palo Alto, California to consider adopting Mike O'Dell's
"GSE - An Alternate Addressing Architecture for IPv6" proposal [GSE].
In GSE, 16-byte IPv6 addresses are split into distinct portions for
global routing, local routing and end-point identification. GSE
draft-ietf-ipngwg-esd-analysis-02.txt [Page 1]
INTERNET-DRAFT March 13, 1998
includes the feature of configuring a node internal to a site with
only the local routing and end-point identfication portions of the
address, thus hiding the full address from the node. When such a node
generates a packet, only the low-order bytes of the source address
are specified; the high-order bytes of the address are filled in by a
border router when the packet leaves the site.
There is a long history of a vague assertion in certain circles that
IPv4 "got it wrong" by treating its addresses simultaneously as
locators and identifiers. Despite these claims, however, there was
never a complete proposal for a scaleable network protocol which
separated the functions. As a result, it wasn't possible to do a
serious analysis comparing and contrasting a "separated" architecture
and an "overloaded" architecture. The GSE proposal serves as a
vehicle for just such an analysis, and that is the purpose of this
paper.
We conclude that an architecture that clearly separates locators and
indentifiers in addresses introduces new issues and problems that do
not have an easy or clear solution. Indeed, the alleged disadvantages
of overloading addresses turn out to provide some significant
benefits over the non-overloaded approach.
Contents
Status of this Memo.......................................... 1
1. Introduction............................................. 3
2. Definitions and Terminology.............................. 4
3. Addressing and Routing in IPv4........................... 5
3.1. The Need for Aggregation............................ 7
3.2. The Pre-CIDR Internet............................... 7
3.3. CIDR and Provider-Based Addressing.................. 8
3.4. Multi-Homing and Aggregation........................ 12
4. The GSE Proposal......................................... 14
4.1. Motivation For GSE.................................. 14
4.2. GSE Address Format.................................. 15
4.2.1. Routing Stuff (RG and STP)..................... 15
4.2.2. End-System Designator.......................... 17
4.3. Address Rewriting by Border Routers................. 18
4.4. Renumbering and Rehoming Mid-Level ISPs............. 19
4.5. Support for Multi-Homed Sites....................... 20
4.6. Explicit Non-Goals for GSE.......................... 21
5. Analysis: The Pros and Cons of Overloading Addresses..... 21
draft-ietf-ipngwg-esd-analysis-02.txt [Page 2]
INTERNET-DRAFT March 13, 1998
5.1. Purpose of an Identifier............................ 22
5.2. Mapping an Identifier to a Locator.................. 24
5.2.1. Scalable Mapping of Identifers to Locators..... 25
5.2.2. Insufficient Hierarchy Space in ESDs........... 26
5.2.3. Reverse Mapping of Complete GSE Addresses...... 27
5.2.4. DNS-Like Reverse Mapping of Full GSE Addresses. 27
5.2.5. The ICMP Who-Are-You Message................... 28
5.3. Authentication of Identifiers....................... 29
5.3.1. Identifier Authentication in IPv4.............. 30
5.3.2. Identifier Authentication in GSE............... 31
5.3.3. Transport Layer: What Locator Should Be Used?.. 31
5.3.4. RG Selection On An Active Open................. 32
5.3.5. RG Selection On An Passive Open................ 32
5.3.6. Mid-Connection RG Changes...................... 32
5.3.7. The Impact of Corrupt Routing Goop............. 33
5.3.8. On The Uniqueness Of ESDs...................... 35
5.3.9. New Denial of Service Attacks.................. 36
5.3.10. Summary of Identifier Authentication Issues... 36
5.4. Miscellaneous....................................... 38
5.4.1. Renumbering and Domain Name System (DNS) Issues 38
5.4.2. How Frequently Can We Renumber?................ 38
5.4.3. Efficient DNS support for Site Renumbering..... 39
5.4.4. Two-Faced DNS.................................. 40
5.4.5. Bootstrapping Issues........................... 41
6. Conclusion............................................... 41
7. Security Considerations.................................. 42
8. Acknowledgments.......................................... 42
9. References............................................... 43
10. Authors' Addresses...................................... 44
1. Introduction
In October of 1996, Mike O'Dell published an Internet-Draft (dubbed
"8+8") that proposed significant changes to the IPv6 addressing
architecture. The 8+8 proposal was the topic of considerable
discussion at the December 1996 IETF meeting in San Jose. Because the
proposal offered both potential benefits (e.g., enhanced routing
scalability) and risks (e.g., changes to the basic IPv6
architecture), the IPng Working Group held an interim meeting on
February 27-28, 1997 to consider adopting the 8+8 proposal.
Shortly before the interim meeting, an updated version of the
draft-ietf-ipngwg-esd-analysis-02.txt [Page 3]
INTERNET-DRAFT March 13, 1998
Internet-Draft was produced. This version changed the name of the
proposal from "8+8" to "GSE" to identify the three separate
components of the address: Global, Site and End-System Designator.
The well-attended meeting generated high caliber, focused technical
discussions on the issues involved, with participation by almost all
of the attendees. By the middle of the second day there was unanimous
agreement that the GSE proposal as written presented too many risks
and should not be adopted as the basis for IPv6. The proposal did,
however, challenge the group to make improvements to the then
existing IPv6 specifications (e.g., increasing the aggregatability of
addresses, having hard boundaries in addresses between routing parts
and non-routing parts and easing the DNS aspects of renumbering).
This document focuses primarily on the issue of separating addresses
into distinct portions for identification and location: a separation
that GSE has but IPv4 does not. We start with a discussion of the
current architecture of IPv4 addressing and its impact on route
scalability, identification, multi-homing, etc. Next, the details of
the GSE proposal are described. Finally, the fundamental issue of
decomposing addresses into multiple separate functional parts is
analyzed in the context of the GSE proposal. Here we detail some of
the practical reasons why separating addresses into locators and
identifier poses a number of challenging problems, making it clear
that having such a separation is no panacea. An appendix contains a
summary of the IPng Working Group's deliberations of GSE and the
results on IPv6 addressing.
2. Definitions and Terminology
The following terminology is used throughout this document.
Routing Goop --- A term defined by the GSE document that refers to
first six bytes of an IPv6 GSE address. The Routing
Goop portion of an address identifies where a site
connects to the public Internet. More generally,
the term refers to the portion of an address's
routing prefix that identifies where a site at which
an address resides connects to the public Internet.
Site Topology Partition --- A term defined by the GSE document
that refers to the two bytes of an IPv6 GSE address
immediately to the right of the Routing Goop. The
Site Topology Partition part of an address
identifies which link within a site an address
resides on.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 4]
INTERNET-DRAFT March 13, 1998
Routing Stuff --- The part of an address that identifies which
link the address resides on. Within the context of
GSE, the Routing Goop and Site Topology Partition
parts of an address comprise the Routing Stuff.
identifier --- a value that indicates the sender of a packet, or
the intended recipient of a packet. Within the
context of GSE, the ESD portion of the address is an
identifier.
locator --- a field in a packet header that is used by the routing
subsystem to deliver a packet to the link on which a
destination resides. The terms locator and Routing
Stuff are similar, we use Routing Stuff when
referring to the specific locator in GSE.
3. Addressing and Routing in IPv4
Before dealing with details of GSE, we present some background about
how routing and addressing works in "classical IP" (i.e., IPv4). We
present this background because the GSE proposal proposes a fairly
major change to the base model. In order to properly evaluate GSE,
one must understand what problems in IPv4 it alleges to improve or
fix.
The structure and semantics of a network layer protocol's addresses
are absolutely core to that protocol. Addressing substantially
impacts the way packets are routed, the ability of a protocol to
scale and the kinds of functionality higher layer protocols can
provide. Indeed, addressing is intertwined with both routing and
transport layer issues; a change in any one of these can impact
another. Issues of administration and operation (e.g., address
allocation and required renumbering), while not part of the pure
exercise of engineering a network layer protocol, turn out to be
critical to the scalability of that protocol in a global and
commercial network. The interaction between addressing, routing and
especially aggregation is particularly relevant to this document, so
some time will be spent describing it.
Addresses in IPv4 serve two purposes:
1) Unique identification of an interface. A sending host tells the
network the identity of the intended recipient by placing an IP
address into the destination address field. In addition, the
receiving host checks the destination address field of received
packets to ensure that the packet is, in fact, for it.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 5]
INTERNET-DRAFT March 13, 1998
2) Location information of that interface. Routers use the packet's
destination address in deciding where to forward the packet to
get it closer to its ultimate destination. That is, addresses
identify "where" the intended recipient is located within the
Internet topology.
For scalability, the location information contained in addresses
must be aggregatable. In practice, this means that nodes
topologically close to each other (e.g., connected to the same
link, residing at the same site, or customers of the same ISP)
must use addresses that share a common prefix.
What is important to note is that these identification and location
requirements have been met through the use of the same value, namely
the IP address. As will be noted repeatedly in this document, the
"overloading" of IPv4 addresses with multiple semantics has some
undesirable implications. For example, the embedding of IPv4
addresses within transport protocol addresses that identify the end-
point of a connection couples those transport protocols with routing.
This entanglement is inconsistent with a strictly layered model in
which routing would be a completely independent function of the
network layer and not directly impact the transport layer.
Combining locator and identifier functions also has the practical
impact of complicating the support for mobility. In a mobile
environment, the location of an end-station may change even though
its identity stays the same; ideally, transport connections should be
able to survive such changes. In IPv4, however, one cannot change the
locator without also changing the identifier.
Consequently, there has been a train of thought for some time has
been that having separate values for location and identification
could be of significant benefit. The GSE proposal, among other
things, attempts to make such a separation.
This document frequently uses mobility as an example to demonstrate
the pros and cons of separating the identifier from the locator.
However, the reader should note the fundamental equivalence between
the problems faced by mobile hosts and the problem faced by sites
that change providers yet don't want to renumber their network. When
a site changes providers, it moves topologically in much the same way
a mobile node does when it moves from one place to another.
Consequently, techniques that help or hinder mobility are often
relevant to the issue of site renumbering.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 6]
INTERNET-DRAFT March 13, 1998
3.1. The Need for Aggregation
IPv4 has seen a number of different addressing schemes. Since the
original specification, the two major additions have been subnetting
and classless routing. The motivation for adding subnetting was to
allow a collection of networks located at one site to be viewed from
afar as a single IP network (i.e., to aggregate all of the individual
networks into one bigger network). The practical benefit of
subnetting was that all of a site's hosts, even if scattered among
tens or hundreds of LANs, could be represented with a single routing
table entry in routers located far from the site. In contrast, prior
to subnetting, a site with ten LANs would advertise ten separate
network entries, and all routers would have to maintain ten separate
entries, even though they contained essentially redundant
information.
The benefits of aggregation should be clear. The amount of work
involved in constructing forwarding tables (i.e., selecting best
routes and installing them into the switching subsystem) is dependent
in part on the number of network routes (i.e., destinations) to which
best paths are computed. If each site has 10 internal networks, and
each of those networks is individually advertised to the global
routing system, the complexity of computing forwarding tables can
easily be an order of magnitude greater than if each site advertised
a single entry that covered all of the addresses used within the
site.
3.2. The Pre-CIDR Internet
In the early days of the Internet, its topology and addressing were
orthogonal. Specifically, when a site wanted to connect to the
Internet, it approached a centralized address allocation authority to
obtain an address and then approached a provider about procuring
connectivity. This procedure for address allocation resulted in a
system where the addresses used by customers of the same provider
bore little relation to the addresses used by other customers of that
same provider. In other words, though the topology of the Internet
was mostly hierarchical, the addressing was not. An example of such a
topology and addressing scheme is shown in Figure 1.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 7]
INTERNET-DRAFT March 13, 1998
+----------------+
| |------- Customer1 (192.2.2.0)
| |------- Customer2 (128.128.0.0)
| Provider A |------- Customer3 (18.0.0.0)
| |------- Customer4 (193.3.3.0)
| |------- Customer5 (194.4.4.0)
+----------------+
|
|
|
|
+----------------+
| Provider B |
+----------------+
Figure 1
Figure 1 shows Provider A having 5 customers, each with their own
independently obtained network address. Providers A and B connect to
each other. In order for Provider B to be able to send traffic to
Customers1-5, Provider A must announce a separate route to Provider B
for each of the 5 networks. That is, the routers within Provider B
must have explicit routing entries for each of Provider A's customers
-- 5 separate routes.
Experience has shown that this approach scales very poorly. In the
Default-Free Zone (DFZ) of the Public Internet, where routers must
maintain routing entries for all reachable destinations, the cost of
computing forwarding tables quickly becomes unacceptably large. A
large part of the cost is related to the seemingly redundant
computations that must be made for each individual network, even
though the reality is that many reside in the same topological
location (e.g., under the same provider). Looking at Figure 1, the
problem is that provider B performs 5 separate calculations to
construct the forwarding table needed to reach each of A's customers.
Said another way, from Provider B's perspective, it doesn't matter
where Provider A's customers connect to Provider A because Provider B
is going to take the same path for all of them; in other words, there
is an opportunity to do data abstraction.
3.3. CIDR and Provider-Based Addressing
One of the reasons CIDR (Classless Inter-Domain Routing) and its
associated provider-assigned address allocation policy were
introduced was to help reduce the size of a routing table and the
complexity of computing a forwarding table from that routing table.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 8]
INTERNET-DRAFT March 13, 1998
CIDR does this by aggressively aggregating network addresses.
Aggregating network addresses means "merging" multiple addresses into
a single "bigger" one. In CIDR, this means that addresses share a
common prefix. The common prefix provides location information for
all addresses sharing that same prefix.
With CIDR, sites that want to connect to the Internet approach a
provider to procure both connectivity and a network address.
Individual providers have a block of address space covered by one
prefix and assign pieces of that space to customers. Consequently,
customers of the same provider have addresses that share the same
prefix. Note that CIDR started to use the term "prefix" to refer to a
classless network. The combination of CIDR and provider-based
addressing results in the ability of a provider to address many
hundreds of sites while introducing just one network address into the
global routing system. An example of such a topology and addressing
scheme is shown in Figure 2.
+----------------+
| |------- Customer1 (204.1.0.0/19)
| |------- Customer2 (204.1.32.0/23)
| Provider A |------- Customer3 (204.1.34.0/24)
| |------- Customer4 (204.1.35.0/24)
| |------- Customer5 (204.1.36.0/23)
+----------------+
|
| A announces
| 204.1/16 to B
|
+----------------+
| Provider B |
+----------------+
Figure 2
In Figure 2, Provider A has been assigned the classless block, or
"aggregate," 204.1.0.0/16 (i.e., a prefix with the high-order 16 bits
denoting a single network). Provider A has 5 customers, each of which
has been assigned a prefix subordinate to the aggregate. In order
for Provider B to be able to reach Customers1-5, Provider A only
needs to announce the single prefix 204.1.0.0/16. The benefit for
Provider B is that its routers need only a single routing table entry
to reach all of Provider A's customers. Note the difference between
the cases described in Figures 1 and 2. The important difference in
the two Figures is that the latter example uses fewer entries in the
routing table to reach the same number of destinations.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 9]
INTERNET-DRAFT March 13, 1998
CIDR was a critical step for the Internet: in the early 1990s the
size of default-free routing tables required to support the classful
Internet was almost more than the commercially-available hardware and
software of the day could handle. The introduction of BGP4's
classless routing and provider-based address allocation policies
resulted in an immediate relief. At the same time, however, CIDR
introduced some new weaknesses. First, the Internet addressing model
had to shift from one of "address owning" to "address lending." In
pre-CIDR days sites acquired addresses from a central authority
independent of their provider, and a site could assume it "owned" the
address it was given. Owning addresses meant that once one had been
given a set of network addresses, one could always use them and
assume that no matter where a site connected to the Internet, the
prefix for that network could be injected into the public routing
system. Today, however, it is simply no longer possible for each
individual site to have its own private prefix injected into the DFZ;
there would simply be too many of them. Consequently, if a site
decides to change providers, then it needs to renumber all of its
nodes using address space given to it by the new provider. The "old"
addresses it had used are returned back to its previous provider. To
understand this, consider if, from Figure 2, Customer3 changes its
provider from Provider A to Provider C, but does not renumber. The
picture would be as follows:
+----------------+
| |---- Customer1 (204.1.0.0/19)
| |---- Customer2 (204.1.32.0/23)
| Provider A |
+---------------| |---- Customer4 (204.1.35.0/24)
| A announces | |---- Customer5 (204.1.36.0/23)
| 204.1/16 to B +----------------+
| |
| |
| |
+----------------+ |
| Provider B | |
+----------------+ |
| |
| |
| |
| C announces |
| 204.1.34/24 |
| to B +----------------+
+---------------| Provider C |---- Customer3 (204.1.34.0/24)
+----------------+
Figure 3
draft-ietf-ipngwg-esd-analysis-02.txt [Page 10]
INTERNET-DRAFT March 13, 1998
In Figure 3, Providers A, B and C are all directly connected to each
other. In order for Provider B to reach Customers 1, 2, 4 and 5,
Provider A still only announces the 204.1.0.0/16 aggregate. However,
in order for Provider B to reach Customer 3, Provider C must announce
the prefix 204.1.34.0/24. Prefix 204.1.34.0/24 is called a "more-
specific" of 204.1.0.0/16; another term used is that Customer3 and
Provider C have "punched a hole in" Provider A's block. The result
of this is that from Provider B's view, the address space underneath
204.1.0.0/16 is no longer cleanly aggregated into a single prefix and
instead the aggregation has been broken because the addressing is
inconsistent with the topology; in order to maintain reachability to
Customer3, Provider B must carry two prefixes where it used to have
to carry only one.
The example in Figure 3 explains why sites must renumber if existing
levels of aggregation are to be maintained. While it is certainly
clear that a small number of exceptions can be tolerated, the reality
in today's Internet is that there are thousands of providers, many
with thousands of individual customers. It is generally accepted that
renumbering of sites is essential for maintaining sufficient
aggregation.
The empirical cost of renumbering a site in order to maintain
aggregation has been the subject of much discussion. The practical
reality, however, is that forcing all sites to renumber is difficult
given the size and wealth of companies that now depend on the
Internet for running their business. Thus, although the technical
community came to consensus that address lending was necessary in
order for the Internet to continue to operate and grow, the reality
has been that some of CIDR's benefits have been lost because not all
sites renumber. It is worth noting that a number of providers do
route filtering based, in part, on prefix length; as a result, a site
which does not renumber may have, at best, partial connectivity to
the Internet.
One unfortunate characteristic of CIDR at an architectural level is
that the pieces of the infrastructure that benefit from the
aggregation (i.e., the providers which make up the DFZ) are not the
pieces that incur the cost (i.e., the end site). The logical
corollary of this statement is that the pieces of the infrastructure
that do incur cost to achieve aggregation (e.g., sites which renumber
when they change providers) don't directly see the benefit. (The word
"directly" is used here because the continued operation of the
Internet is a benefit, though it requires selflessness on the part of
the site to recognize.)
draft-ietf-ipngwg-esd-analysis-02.txt [Page 11]
INTERNET-DRAFT March 13, 1998
3.4. Multi-Homing and Aggregation
As sites become more dependent on the Internet, they have begun to
install additional connections to the Internet to improve robustness
and performance. Such sites are called "multi-homed." Unfortunately,
when a site connects to the Internet at multiple places, the impact
on routing can be much like a site that switches providers but
refuses to renumber.
In the pre-CIDR days, multi-homed sites were typically known by only
one network prefix. When that site's providers announced the site's
network into the global routing system, a "shortest path" type of
routing would occur so that pieces of the Internet closest to the
first provider would use the first provider while other pieces of the
Internet would use the second provider. This allowed sites to use the
routing system itself to load balance traffic across their multiple
connections. This type of multi-homing assumes that a site's prefix
can be propagated throughout the DFZ, an assumption that is no longer
universally true.
With CIDR, issues of addressing and aggregation complicate matters
significantly. At the highest levels, there are three possible ways
to deal with multi-homed sites. The first approach is for multi-
homed sites to receive address space directly from a registry,
independent of its providers. The problem with this approach is
that, because the address space is obtained independent of either
provider, it is not aggregatable and therefore has a negative impact
on the scaling of global routing.
The second approach is for a multi-homed site to receive an
allocation from one of its providers and just use that single prefix.
The site would advertise its prefix to all of the providers to which
it connects. There are two problems with this is approach. First,
although the prefix is aggregatable by the provider which made the
allocation, it is not aggregatable by the other providers. To the
other providers, the site's prefix poses the same problem that a
provider-independent address would. This has a negative impact on
the scaling of global routing. Second, due to CIDR's rule for
longest-match routing, it turns out that the site's prefix is not
always aggregatable in practice even by the provider that made the
allocation. Consider Figure 4. Provider C has two paths for reaching
Customer 1. Provider A advertises 204.1/16, an aggregate which
includes Customer 1. But Provider C will also receive an
advertisement for prefix 204.1.0/19 from Provider B, and because the
prefix match through B is longer, C will choose that path. In order
for Provider C to be able to choose between the two paths, Provider A
would also have to advertise the longer prefix for 204.1.0/19 in
addition to the shorter 204.1/16. At this point, from the routing
draft-ietf-ipngwg-esd-analysis-02.txt [Page 12]
INTERNET-DRAFT March 13, 1998
perspective, the situation is very similar to the general problem
posed by the use of provider-independent addresses.
It should be noted that the above example simplifies a very complex
issue. For example, consider the example in Figure 4 again. Provider
A could choose not to propagate a route entry for the longer
204.1.0/19 prefix, advertising only the shorter 204.1/16. In such
cases, provider C would always select Provider B. Internally,
Provider A would continue to route traffic from its other customers
to Customer 1 directly. If Provider A had a large enough customer
base, effective load sharing might be achieved.
A advertises
+------------+ 204.1/16 to C +------------+
___| Provider A |-----------------| Provider C |
/ +------------+ +------------+
/ +----------/
/ /
Customer 1 --- / B advertises 204.1.0/19 to C
204.1.0.0/19 | /
| +------------+
----- | Provider B |
+------------+
Figure 4
The third approach is for a multi-homed site to receive an allocation
from each of its providers. This approach has advantages from the
perspective of route scaling because both allocations are
aggregatable. Unfortunately, the approach doesn't necessarily meet
the demands of the multi-homed site. A site that has a prefix from
each of its providers has a number of choices about how to use that
address space. Possibilities include:
1) The site can number a distinct set of hosts out of each of the
prefixes. Consider a configuration where a site is connected to
ISP-A and ISP-B. If the link to ISP-A goes down, then unless the
ISP-A prefix is announced to ISP-B (which breaks aggregation),
the hosts numbered out of the ISP-A prefix would be unreachable.
2) The site could assign each host multiple addresses (i.e., one
address for each ISP connection). There are two problems with
this. First, it accelerates the consumption of the address
space. Second, when the connection to ISP-A goes down,
addresses numbered out of ISP-A's space become unreachable.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 13]
INTERNET-DRAFT March 13, 1998
Remote peers would have to have sufficient intelligence to use
the second address. For example, when initiating a connection to
a host, the DNS would return multiple candidate addresses.
Clients would need to try them all before concluding that a
destination is unreachable (something not all hosts currently
do). In addition, a site's hosts would need a significant amount
of intelligence for choosing the source addresses they use. A
host shouldn't choose a source address corresponding to a link
that is down. At present, hosts do not have such sophistication.
In summary, how best to achieve multi-homing with IPv4 in the face of
CIDR is an unsolved problem. There is a delicate balance between the
scalability of routing versus the site's requirements of robustness
and load-sharing. At this point in time, no solution has been
discovered that satisfies the competing requirements of route scaling
and robustness/performance. It is worth noting, however, that some
people are beginning to study the issue more closely and propose
novel ideas [BATES].
4. The GSE Proposal
This section provides a description of GSE with the intent of making
this document stand-alone with respect to the GSE "specification." We
begin by reviewing the motivation for GSE. Next we review the salient
technical details, and we conclude by listing the explicit non-goals
of the GSE proposal.
4.1. Motivation For GSE
The primary motivation for GSE was the fact that the chief initial
IPv6 global unicast address structure, provider-based [RFC 2073], was
fundamentally the same as IPv4 with CIDR and provider-based
aggregation. Provider-based addressing requires that sites renumber
when they switch providers, so that sites are always aggregated
within their provider's prefix. In practice, the cost of renumbering
(which can only grow as a site grows in size and becomes more
dependent on the Internet for day-to-day business) is high enough
that an increasing number of sites refuse to renumber. This cost is
particularly relevant in cases where end-users are asked to renumber
because an upstream provider has changed its transit provider (i.e.,
the end site is asked to renumber for reasons outside of its control
and for which it sees no direct benefit). Consequently, the GSE
draft asserted that IPv4 with CIDR has not achieved the aggressive
aggregation required for the route computation functions of the DFZ
of the Internet to scale for IPv4 and that the larger addresses of
IPv6 simply exacerbate the problem.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 14]
INTERNET-DRAFT March 13, 1998
The GSE proposal did not propose to eliminate the need for
renumbering. Indeed, it asserted that end sites will have to be
renumbered more frequently in order to continue scaling the Internet.
However, GSE proposed to make the cost of such a renumbering so small
that sites could be renumbered at essentially any time with little or
no disruption.
Finally, GSE dealt significantly with sites that have multiple
Internet connections. In some addressing schemes (e.g., CIDR), this
"multi-homing" can create exceptions to the aggregation and result in
poor scaling. That is, the public routing infrastructure needs to
carry multiple distinct routes for the multi-homed site, one for each
independent path. GSE recognized the "special work done by the global
Internet infrastructure on behalf of multi-homed sites," [GSE] and
proposed a way for multi-homed sites to gain some benefit without
impacting global scaling. This included a specific mechanism that
providers could use to support multi-homed sites, presumably at a
cost that the site would consider when deciding whether or not to
become multi-homed.
4.2. GSE Address Format
The key departure of GSE from classical IP addressing (both v4 and
v6) was that rather than over-loading addresses with both locator and
identifier purposes, it split the address into two elements: the
high-order 8 bytes for routing (called "Routing Stuff" throughout the
rest of this document) and the low-order 8 bytes for unique
identification of an end-point. The structure of GSE addresses was:
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
| Routing Goop | STP| End System Designator |
+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
6+ bytes ~2 bytes 8 bytes
Figure 5
4.2.1. Routing Stuff (RG and STP)
The Routing Goop (RG) identifies the place in the Internet topology
where a site connects and is used to route datagrams to the site. RG
is structured as follows:
draft-ietf-ipngwg-esd-analysis-02.txt [Page 15]
INTERNET-DRAFT March 13, 1998
1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| xxx | 13 Bits of LSID | Upper 16 bits of Goop |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
3 4
2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| Bottom 18 bits of Routing Goop |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Figure 6
The RG describes the location of a site's connection by identifying
smaller and smaller regions of topology until finally it identifies
the link which connects the site. Before interpreting the bits in the
RG, it is important to understand that routing with GSE depends on
decomposing the Internet's topology into a specific graph. At the
highest level, the topology is broken into Large Structures (LSs). An
LS is basically a region that can aggregate significant amounts of
topology. Examples of potential LSs are large providers and exchange
points. Within an LS the topology is further divided into another
graph of structures, with each LS dividing itself however it sees
fit. This division of the topology into smaller and smaller
structures can recurse for a number of levels, where the trade-off is
"between the flat-routing complexity within a region and minimizing
total depth of the substructure." [ESD]
Having described the decomposition process, we can now examine the
bits in the RG. After the 3-bit prefix identifying the address as
GSE, the next 13 bits identify the LS. By limiting the field to 13
bits, a ceiling is defined on the complexity of the top-most routing
level (i.e., what we currently call the DFZ). In the next 34 bits, a
series of subordinate structure(s) are identified until finally the
leaf subordinate structure is identified, at which point the
remaining bits identify the individual link within that leaf
structure.
The remaining 14 bits of the Routing Stuff (i.e., the low-order 14
bits of the high-order 8 bytes) comprise the STP and are used for
routing structure within a site, similar to subnetting with IPv4.
These bits are not part of the Routing Goop per se. The distinction
between Routing Stuff and Routing Goop is that RG controls routing in
the Public Internet, while Routing Stuff includes the RG plus the
Site Topology Partition (STP). The STP is used for routing structure
within a site. [Note that the term "Routing Stuff" was a creation of
draft-ietf-ipngwg-esd-analysis-02.txt [Page 16]
INTERNET-DRAFT March 13, 1998
the author's of this analysis document and was not used in the GSE
document.]
The GSE proposal formalized the ideas of sites and of public versus
private topology. In the first case, a site is a set of hosts,
routers and media under the same administrative control which have
zero or more connections to the Internet. A site can have an
arbitrarily complicated topology, but all of that complexity is
hidden from everyone outside of the site. A site only carries
packets which originated from, or are destined to, that site; in
other words, a site cannot be a transit network. A site is private
topology, while the transit networks form the public topology.
A datagram is routed through public topology using just the RG, but
within the destination site, routing is based on the Site Topology
Partition (STP).
4.2.2. End-System Designator
The End-System Designator (ESD) is an unstructured 8-byte field that
uniquely identifies an interface from all others. The most important
feature of the ESD is that it alone identifies an interface; the
Routing Stuff portion of an address, although used to help deliver a
packet to its destination, is not used to actually identify an end
point. End-points of communication care about the ESD; as examples,
TCP peers could be identified by the source and destination ESDs
alone (together with port numbers), checksums would exclude the RG
(the sender doesn't know its RG, as described later) and on receipt
of a datagram only the ESD would be used in testing whether a packet
is intended for local delivery.
The leading contender for the role of a 64-bit globally unique ESD is
the recently defined "EUI-64" identifier. [EUI64] These identifiers
consist of a 24-bit "company_id" concatenated with a 40-bit
"extension." (Company_id is just a new name for the
"Organizationally Unique Identifier" that forms the first half of an
802 MAC address.) Manufacturers are expected to assign locally unique
values to the extension field, guaranteeing global uniqueness for the
complete 64-bit identifier.
A range of the EUI-64 space is reserved to cover pre-existing 48-bit
MAC addresses, and a defined mapping insures that an ESD derived from
a MAC address will not duplicate the ESD of a device that has a
built-in EUI-64.
In some cases, interfaces may not have access to an appropriate MAC
address or EUI-64 identifier. A globally unique ESD must then be
draft-ietf-ipngwg-esd-analysis-02.txt [Page 17]
INTERNET-DRAFT March 13, 1998
obtained through some alternate mechanism. Several possible
mechanisms can be imagined (e.g., the IANA could hand out addresses
from the company_id it has been allocated), but we do not explore
them in detail here.
4.3. Address Rewriting by Border Routers
GSE site border routers rewrite addresses of the packets they forward
across the boundary between the site and public topology. Within a
site, nodes need not know the RG associated with their addresses.
They simply use a designated "Site-Local RG" value for internal
addresses. When a packet is forwarded to the public topology, the
border router replaces the Site-Local RG portion of the packet's
source address with an appropriate value. Likewise, when a packet
from the public topology is forwarded into a site, the border router
replaces the RG part of the destination address with the designated
Site-Local RG.
To simplify discussion, the following text uses the singular term RG
as if a site could have only one RG value (i.e., one connection to
the Internet). In fact, a site could have multiple Internet
connections and consequently multiple RGs.
Having border routers rewrite addresses obviates the need to renumber
devices within sites because of changing providers --- GSE's approach
wasn't so much to ease renumbering as to make it transparent. To
achieve transparency, the RG by which a site is known is hidden
(i.e., kept secret) from nodes within that site. Instead, the RG for
the site would be known only by the exit router, either through
static configuration or through a dynamic protocol with an upstream
provider.
Because end hosts don't know their RG, they don't know their entire
16-byte address, so they can't specify the full address in the source
fields of packets they originate. Consequently, when a datagram
leaves a site, the egress border router fills in the high-order
portion of the source address with the appropriate RG.
The point of keeping the RG hidden from nodes within the core of a
site was to insure the changeability of the RG without impacting the
site itself. It was expected that the RG would need to change
relatively frequently (e.g., several times a year) in order to
support scalable aggregation as the topology of the Internet changes.
A change to a site's RG would only require a change at the site's
egress point, and it's well possible that this change could be
accomplished through a dynamic protocol with the upstream provider.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 18]
INTERNET-DRAFT March 13, 1998
Hiding a site's RG from its internal nodes does not, however, mean
that changes to RG have no impact on end sites. Since the full 16-
byte address of a node isn't a stable value (the RG portion can
change), a stored address may contain invalid RG and be unusable if
it isn't "refreshed" through some other means. For example, opening a
TCP connection, writing the address of the peer to a file and then
later trying to reestablish a connection to that peer is likely to
fail. For intra-site communication, however, it is expected that
only the Site-Local RG would be used (and stored) which would
continue to work for intra-site communication regardless of changes
to the site's external RG. This has the benefit of shielding a site's
intra-site traffic from any instabilities resulting from renumbering.
In addition to rewriting source addresses upon leaving a site,
destination addresses are rewritten upon entering a site. To
understand the motivation behind this, consider a site with
connections to three Internet providers. Because each of those
connections has its own RG, each destination within the site would be
known by three different 16-byte addresses. As a result, intra-site
routers would have to carry a routing table three times larger than
expected. To work around this, GSE proposed replacing the RG in
inbound packets with the special "Site-Local RG" value to reduce
intra-site routing tables to the minimum necessary.
In summary, when a node initiates a flow to a node at another site,
the initiating node knows the full 16-byte address for the
destination through some mechanism like a DNS query. The initiating
node does not, however, know its RG, so uses the Site-Local RG values
in the RG part of the source address. When the datagram reaches the
exit border router, the router replaces the RG of the packet's source
address. When the datagram arrives at the entry router at the
destination site, the router replaces the RG portion of the
destination address with the distinguished "Site-Local RG" value.
When the destination host needs to send return traffic, that host
knows the full 16-byte address for the other host because it appeared
in the source address field of the arriving packet.
4.4. Renumbering and Rehoming Mid-Level ISPs
One of the most difficult-to-solve components of the renumbering
problem with CIDR is that of renumbering mid-level service providers.
Specifically, if SmallISP1 changes its transit provider from BigISP1
to BigISP2, then in order for the overall size of the routing tables
to stay the same, all of SmallISP1's customers would have to renumber
into address space covered by an aggregate of BigISP2. GSE dealt
with this problem by handling the RG in DNS with indirection.
Specifically, a site's DNS server specifies the RG portion of its
draft-ietf-ipngwg-esd-analysis-02.txt [Page 19]
INTERNET-DRAFT March 13, 1998
addresses by referencing the "name" of its immediate provider, which
is a resolvable DNS name (this implies a new Resource Record type).
That provider may define some of the low-order bits of the RG and
then reference its immediate provider. This chain of reference allows
mid-level service providers to change transit providers, and the
customers of that mid-level will simply "inherit" the change in RG.
4.5. Support for Multi-Homed Sites
GSE defined a specific mechanism for providers to use to support
multi-homed customers that gives those customers more reliability
than singly-homed sites, but without a negative impact on the scaling
of global routing. This mechanism is not specific to GSE and could be
applied to any multi-homing scenario where a site is known by
multiple prefixes (including provider-based addressing). Assume the
following topology:
Provider1 Provider2
+------+ +------+
| | | |
| PBR1 | | PBR2 |
+----x-+ +-x----+
| |
RG1 | | RG2
| |
+--x-----------x--+
| SBR1 SBR2 |
| |
+-----------------+
Site
Figure 7
PBR1 is Provider1's border router while PBR2 is Provider2's border
router. SBR1 is the site's border router that connects to Provider1
while SBR2 is the site's border router that connects to Provider2.
Imagine, for example, that the line between Provider1 and the site
goes down. Any already existing flows that use a destination address
including RG1 would stop working. In addition, any DNS queries that
return addresses including RG1 would not be viable addresses. If PBR1
and PBR2 knew about each other, however, then in this case PBR1 could
tunnel packets destined for RG1-prefixed addresses to PBR2, thus
keeping the communication working. (Note that true tunneling, i.e.,
re-encapsulation, is necessary since routers between PBR1 and PBR2
would forward RG1 addresses towards PBR1.)
draft-ietf-ipngwg-esd-analysis-02.txt [Page 20]
INTERNET-DRAFT March 13, 1998
4.6. Explicit Non-Goals for GSE
It is worth noting explicitly that GSE did not attempt to address the
following issues:
1) Survival of TCP connections through renumbering events. If a
site is renumbered, TCP connections using a previous address
will continue to work only as long as the previous address still
works (i.e., while it is still "valid" using RFC 1971
terminology). No attempt is made to have existing connections
switch to the new address.
2) It is not known how mobility can be made to work under GSE.
3) It is not known how multicast can be made to work under GSE.
4) The performance impact of having routers rewrite portions of the
source and destination address in packet headers requires
further study.
That GSE didn't address the above does not mean they cannot be
solved. Rather the issues weren't studied in sufficient depth.
5. Analysis: The Pros and Cons of Overloading Addresses
At this point we have given complete descriptions of two addressing
architectures: IPv4, which uses the overloading technique, and GSE,
which uses the separated technique. We now compare and contrast the
two techniques.
The following discussion is organized around three fundamental
points:
1) Identifiers indicate who the intended recipient of a packet is,
2) Identifiers must be mapped into a locator that the network layer
uses to actually deliver a packet to its intended destination,
and
3) There must be a suitable way to sufficiently authenticate the
user of an identifier, so that peers using identifiers have
sufficient confidence that packets sent to or received from a
particular identifier correspond to the intended recipient.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 21]
INTERNET-DRAFT March 13, 1998
5.1. Purpose of an Identifier
An identifier gives an entity the ability to refer to a communication
end point and to refer to the same endpoint over an extended period
of time. In terms of semantics, two or more packets sent to the same
identifier should be delivered to the same end point. Likewise, one
expects multiple packets received from the same identifier to have
been originated by the same sending entity. That is, a source
identifier indicates who the packet is from and a destination
identifier indicates who the packet is intended for.
When applications communicate, "identifiers" consist of addresses and
port numbers. For the purposes of this discussion, the term
"identifier" means the identifier of an interface. It is assumed that
port numbers will be present when higher layer entities communicate;
the exact port numbers used are not relevant to this discussion.
In small networks, flat routing can be used to deliver packets to
their destination based only on the destination identifier carried in
a packet header (i.e., the identifier is the locator and is not
required to have any structure). However, in such systems, a distinct
route entry is required for every destination, an approach that does
not scale. In larger networks, packet addresses include a locator
that helps the network layer deliver a packet to its destination.
Such a locator typically has structure (i.e., is an aggregate for
many destinations) that keeps routing tables small relative to the
total number of reachable destinations. In IPv4, the identifier and
locator are combined in a single address; it is not possible to
separate the locator portion of an address from the identifier
portion. In contrast, the ESD portion of a GSE address (which can
easily be extracted from the address) serves as an identifier, while
the Routing Stuff plays the role of a locator.
Having a clear separation between the locator and the identifer
portion of an address appears to give protocols some additional
flexibility. Once a packet has been delivered to its intended
destination interface (i.e., node), for example, the locator has
served its purpose and is no longer needed to further demultiplex a
packet to its higher-layer end point. This means that if a packet is
delivered to the correct destination node, the node will accept the
packet, regardless of how the packet got there. The exact locator
used does not matter, so long as it corresponds to one that delivers
a packet to its proper destination.
The most obvious example that could benefit from the separation of
locators and indentifiers involves communication with a mobile host.
Transport protocols such as TCP are unable to keep connections open
if either of the endpoint identifiers for an open connection changes.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 22]
INTERNET-DRAFT March 13, 1998
Fundamentally, the endpoint identifiers indicate the two endpoint
entities that are communicating. If a node were to receive a packet
from a node with which it had been communicating previously, but the
identifier used by the sending node has changed, the recipient would
be unable to distinguish this case from that of a packet received
from a completely different node.
In the specific case of TCP and IPv4, connections are identified
uniquely by the tuple: (srcIPaddr, dstIPaddr, srcport, dstport).
Because IPv4 addresses contain a combined locator/identifier, it is
not possible to have a node's location change without also having its
identifier change. Consequently, when a mobile node moves, its
existing connections no longer work, in the absence of special
protocols such as Mobile IP [RFC2002].
In contrast, connections in GSE are identified by the ESDs rather
than full IPv6 addresses. That is, connections are identified
uniquely by the tuple: (srcESD, dstESD, srcport, dstport).
Consequently, when demultiplexing incoming packets to their proper
end point, TCP would ignore the Routing Stuff portions of addresses.
Because the Routing Stuff portion of an address is ignored during
demultiplexing operations, a mobile node is free to move -- and
change its Routing Stuff -- without consequences for the
demultiplexing operation.
As a side note, it is a requirement in GSE that packets be
demultiplexed on ESDs alone independent of the Routing Stuff. If a
site is multi-homed, the packets it sends may exit the site at
different egress border routers during the lifetime of a connection.
Because each border router will place its own RG into the source
addresses of outgoing packets, the receiving TCP must ignore (at
least) the RG portion of addresses when demultiplexing received
packets. The alternative would be to make TCP unable to cope with
common routing changes, i.e., if the path changed, packets delivered
correctly would be discarded by the receiving TCP rather than
processed.
Not surprisingly, having separate locator and identifiers in
addresses leads to some additional problems. First, an identifier by
itself provides only limited value. In order to actually deliver
packets to a destination identifier, a corresponding locator must be
known. The general problem of mapping identifiers into locators is
non-trivial to solve, and is the topic of the next Section. Second,
because the Routing Stuff is ignored when demultiplexing packets
upward in the protocol stack, it becomes much easier for an intruder
to masquerade as someone else.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 23]
INTERNET-DRAFT March 13, 1998
5.2. Mapping an Identifier to a Locator
The idea of using addresses that cleanly separate location and
identification information is not new [references XXX]. However,
there are several different flavors. In its pure form, a sender need
only know the identifier of an end-point in order to send packets to
it. When presented with a datagram to send, network software would be
responsible for finding the locator associated with an identifier so
that the packet can be delivered. A key question is: "who is
responsible for finding the Routing Stuff associated with a given
identifier"? There are a number of possibilities, each with a
different set of implications:
1) The network layer could be responsible for doing the mapping.
The advantage of such a system is that an ESD could be stored
essentially forever (e.g., in configuration files), but whenever
it is actually used, network layer software would automatically
perform the mapping to determine the appropriate Routing Stuff
for the destination. Likewise, should an existing mapping become
invalid, network layer software could dynamically determine the
updated value. Unfortunately, building such a mapping mechanism
that scales is a hard problem.
2) The transport layer could be responsible for doing the mapping.
It could perform the mapping when a connection is first opened,
periodically refreshing the binding for long-running
connections. Implementing such a scheme would change the
existing transport layer protocols TCP and UDP significantly.
3) Higher-layer software (e.g., the application itself) could be
responsible for performing the mapping. This potentially
increases the burden on application programmers significantly,
especially if long-running connections are required to survive
renumbering and/or deal with mobile nodes.
It should be noted that the GSE proposal does not embrace the general
model, it uses the last. The network and transport layers are always
presented with both the Routing Stuff (RG + STP) and the ESD together
in one IPv6 address. It is neither of these layers' jobs to determine
the Routing Stuff given only the ESD or to validate that the Routing
Stuff is correct. When an application has data to send, it queries
the DNS to obtain the IPv6 AAAA record for a destination. The
returned AAAA record contains both the Routing Stuff and the ESD of
the specified destination. While such an approach eliminates the need
for the lower layers to be able to map ESDs into corresponding
Routing Stuff, it also means that when presented with an address
containing an incorrect (i.e., no longer valid) Routing Stuff, the
network is unable to deliver the packet to its correct destination.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 24]
INTERNET-DRAFT March 13, 1998
It is up to applications themselves to deal with such failures. Note
that addresses containing invalid Routing Stuff will result any time
cached addresses are used after the Routing Stuff of the address
becomes invalid. This may happen if addresses are stored in
configuration files, a mobile node moves to a new location, long-
running applications (clients and servers) cache the result of DNS
queries, a long-running connection attempts to continue operating
during a site renumbering event, etc.
A network architecture must provide the ability to map an identifier
to a locator. In IPv4, this mapping is trivial (the identity
function), since the identifier and locator are combined in a single
quantity (i.e., the IPv4 address). GSE does not provide mapping
functionality directly. Indeed, GSE uses two different identifiers.
At the highest level, a node's DNS name serves as its identifer, with
normal DNS queries used to map the DNS "identifier" into a locator
(i.e., the first 8 bytes of the IPv6 address). At a lower layer, the
IPv6 address contains the ESD identifier together with its Routing
Stuff (i.e., locator). Note that the DNS name is expected to be the
stable identifier that can be mapped into an appropriate locator at
any time. In contrast, the ESD identifier, cannot be mapped into a
locator by itself.
The use of two identifiers contributes to making GSE appear simple.
However, there are two fundamental problems with this approach, if
the intention is to make it transparently easy to change locators
over time. First, the burden of performing the mapping from
identifier to locator is placed directly on the application,
requiring active participation from the application. Second, The
lower layers (i.e., transport and network layers) cannot make use of
this mapping themselves due to layering violation concerns (i.e., TCP
and UDP can't depend on the DNS to perform a query).
The following subsections discuss a number of issues related to
keeping track of or determining the locator associated with an
identifier.
5.2.1. Scalable Mapping of Identifers to Locators
It is not difficult to construct a mapping from an identifier (such
as an ESD) to a locator (as well as other information such as a name,
cryptographic keys, etc.) provided one can structure the identifier
appropriately to support such lookups. In particular, identifiers
must have sufficient structure to support the delegating mechanism of
a distributed database such as DNS. On the other hand, no scalable
mechanism is known for performing such a mapping on arbitrary
identifiers taken from a flat space lacking structure.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 25]
INTERNET-DRAFT March 13, 1998
Imposing a heirarchy on identifiers poses the following difficulties:
- it increases the size of the identifier. The exact size
necessary to support sufficient heirarchy is unclear, though it
is likely to be roughly the same as that used for the routing
hierarchy. Analysis done during the original IPng debates
[RFC1752] suggests that close to 48-bits of hierarchy are needed
to identify all the possible sites 30-40 years from now.
- the assignment of identifiers must be tied to the delegation
structure. That is, the site that "owns" an identifier is the
one responsible for maintaining the identifier-to-locator
mapping information about it.
- a mechanism would be needed to make it possible for a node to
determine what its identifier is. To be practical, such a
mechanism would need to be automated and avoid the need for
manual configuration.
5.2.2. Insufficient Hierarchy Space in ESDs
In the case of GSE's 8-byte ESD, the size of the identifier is not
large enough to contain sufficient heirarchy to both create DNS-like
delegation points and support stateless address autoconfiguration.
Stateless address autoconfiguration [RFC1971] already assumes that an
interface's 6-byte link-layer (i.e., MAC) address can be appended to
a link's routing prefix to produce a globally unique IPv6 address.
With GSE, only two bytes would be available for hierarchy and
delegation.
It is also the case that the sorts of built-in identifiers now found
in computing hardware, such as "EUI-48" and "EUI-64" addresses
[IEEE802, IEEE1212], do not have the structure required for this
delegation. Such identifiers have only two-levels of heirarchy; the
top-level typically identifies a manufacturer, with the remaining
part of the address being the equivalent of the serial number unique
to the manufacturer. In addition, the delegation of the two-level
heirarchy (i.e., equipment manufacturer) does not correspond to the
administrator under which the end-user operates. Hence, stateless
autoconfiguration [RFC1971] cannot create addresses with the
necessary hierarchical property in the ESD portion of an address.
Finally, imposing a required hierarchical structure on identifiers
such as an ESD would also introduce a new administrative burden and a
new or expanded registry system to manage ESD space (i.e., to insure
that ESDs are globally unique). While the procedures for assigning
ESDs, which need only organizational and not topological
draft-ietf-ipngwg-esd-analysis-02.txt [Page 26]
INTERNET-DRAFT March 13, 1998
significance, would be simpler than the procedures for managing IPv4
addresses (or DNS names), it is hard to imagine such a process being
universally well-received or without controversy; it seems a laudable
goal to avoid the problem altogether if possible. In addition, it
would likely increase the complexity for connecting new nodes to the
Internet, a goal inconsistent with Stateless Address
autoconfiguration [RFC1971].
5.2.3. Reverse Mapping of Complete GSE Addresses
The following two sections describe techniques for mapping a full
IPv6 address back into some quantity (e.g., a DNS name or locator).
We include these descriptions for completeness even though they do
not address the fundamental problem of how to perform the mapping on
an identifier alone. It should also be noted that because both
techniques operate on complete IPv6 addresses, they are both directly
applicable to provider-based addressing schemes and are not specific
to GSE.
5.2.4. DNS-Like Reverse Mapping of Full GSE Addresses
Although it seems infeasible to have a global scale, reverse mapping
of ESDs, within a site, one could imagine maintaining a database
keyed on unstructured 8-byte ESDs. However, it is a matter of debate
whether such a database can be kept up-to-date at reasonable cost,
without making unreasonable assumptions as to how large sites are
going to grow, and how frequently ESD registrations will be made or
updated. Note that the issue isn't just the physical database itself,
but the operational issues involved in keeping it up-to-date. For the
rest of this section, however, let us assume that such a database can
be built.
A mechanism supporting a lookup keyed on a flat-space ESD from an
arbitrary site requires having sufficient structure to identify the
site that needs to be queried. In practice, an ESD will almost always
be used in conjunction with Routing Stuff (i.e., a full 16-byte
address). Since the Routing Stuff is organized hierarchically, it
becomes feasible to maintain a DNS-like tree that maps full GSE
addresses into DNS names, in a fashion analogous to what is done with
IPv4 PTR records today.
It should be noted that a GSE address lookup will work only if the
Routing Stuff portion of the address is correctly entered in the DNS
tree. Because the Routing Stuff portion of an address is expected to
change over time, this assumption will not be valid indefinitely. As
a consequence, a packet trace recorded in the past might not contain
draft-ietf-ipngwg-esd-analysis-02.txt [Page 27]
INTERNET-DRAFT March 13, 1998
enough information to identify the off-Site sources of the packets in
the present. This problem can be addressed by requiring that the
database of RG delegations be maintained for some period of time
after the RG is no longer usable for routing packets.
Finally, it should be noted that the problem where an address's RG
"expires" with the implication that the mapping of "expired"
addresses into DNS names may no longer hold is not a problem specific
to the GSE proposal. With provider-based addressing, the same issue
arises when a site renumbers into a new provider prefix and releases
the allocation from a previous block. The authors are aware of one
such renumbering in IPv4 where a block of returned addresses was
reassigned and reused within 24 hours of the renumbering event.
5.2.5. The ICMP Who-Are-You Message
Although there is widespread agreement on the utility of being able
to determine the DNS name one is communicating with, there is also
widespread concern that repeating the experience of the "IN-
ADDR.ARPA" domain is undesirable. In practice, the IN-ADDR.ARPA
domain is not fully populated and poorly maintained. Consequently,
an old proposal to define an ICMP Who-Are-You message was resurrected
[RFC1788]. A client would send such a message to a peer, and that
peer would return an ICMP message containing its DNS name.
Asking a remote host to supply its own name in no way implies that
the returned information is accurate. However, having a remote peer
provide a piece of information that a client can use as input to a
separate authentication procedure provides a starting point for
performing strong authentication. The actual strength of the
authentication depends on the authentication procedure invoked,
rather than the untrustable piece of information provided by a remote
peer.
Reconsidering the "cheap" authentication procedure described earlier,
the ICMP Who-Are-You replaces the DNS PTR query used to obtain the
DNS name of a remote peer. The second DNS query, to map the DNS name
back into a set of addresses, would be performed as before. Because
the latter DNS query provides the strength of the authentication,
the use of an ICMP Who-Are-You message does not in any way weaken the
strength of the authentication method. Indeed, it can only make it
more useful in practice, because virtually all hosts can be expected
to implement the Who-Are-You message.
The Who-Are-You message is robust against renumbering, since it
follows the paths of valid routable prefixes. Essentially, it uses
the Internet routing system in place of the DNS delegation scheme. It
draft-ietf-ipngwg-esd-analysis-02.txt [Page 28]
INTERNET-DRAFT March 13, 1998
is attractive in the context of GSE-style renumbering, since no host
or DNS server needs to be updated after a renumbering event for Who-
Are-You-based lookups to work. It has advantages outside the context
of GSE as well, including a more decentralized, and hence more
scalable, administration and easier upkeep than a DNS reverse-lookup
zone. It also has drawbacks: it requires the target node to be up and
reachable at the time of the query and to know its fully qualified
domain name. It is also not possible to resolve addresses once those
addresses become unroutable. In contrast, the DNS PTR mirrors, but is
independent of, the routing hierarchy. The DNS can maintain mappings
long after the routing subsystem stops delivering packets to certain
addresses.
The requirement that the target node be up and reachable at the time
of the query makes it very uncertain that one would be able to take
addresses from a packet log and translate them to correct domain
names at a later date. One can argue that this is a design flaw in
the logging system, as it violates the architectural principle,
"Avoid any design that requires addresses to be ... stored on non-
volatile storage." [RFC1958] A better-designed system would look up
domain names promptly from logged addresses. Indeed, one of the
authors has been doing that for some years.
5.3. Authentication of Identifiers
The true value of a globally unique identifier lies not on its
uniqueness but on an ability to use the same identifier repeatedly
and have it refer to the same end point. That is, when an identifier
is used, there is an expectation that repeated and subsequent use of
the identifier results in continued communication with the same end
point. To be useful then, a valid identifier must either be easily
distinguishable from a fraudulant one, or the system must have a way
to prevent identifiers from being used in an unauthorized manner.
The remainder of this section discusses how identifer authentication
is done in both IPv4 and GSE, and shows how overloading an address
with both an identifier and a locator provides automatic identifier
authentication. In contrast, there is essentially no identifier
authentication in GSE. It should be noted that the actual strength
of authentication that would be considered sufficient is a topic in
its own right, and we do not spent much time on it. Instead, we focus
on the relative strengths in the two schemes.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 29]
INTERNET-DRAFT March 13, 1998
5.3.1. Identifier Authentication in IPv4
As described earlier, an IPv4 address simultaneously plays two roles:
a unique identifier and a locator. Using an overloaded address as an
identifier has the side-effect of insuring that (for all practical
purposes) the identifier is globally unique. Furthermore, because
the same number is used both to identify an interface and to deliver
data to that interface, it is impossible for some interface A to use
the identification of another interface B in an attempt to receive
data destined to B without being detected, unless the routing system
is compromised.
When both interfaces A and B claim the same unicast address, the
routing subsystem generally delivers packets to only one of them. The
other node will quickly realize that something is wrong (since
communication using the duplicate address fails) and take corrective
actions, either correcting a misconfiguration or otherwise detecting
and thwarting the intruder. To understand how the routing subsystem
prevents the same address from being used in multiple locations,
there are two cases to consider, depending on whether the two
interfaces using duplicate addresses are attached to the same or to
different links.
When two interfaces on the same link use the same address, a node
(host or router) sending traffic to the duplicate address will in
practice send all packets to one of the nodes. On Ethernets, for
example, the sender will use ARP (or Neighbor Discovery in IPv6) to
determine the link-layer address corresponding to the destination
address. When multiple ARP replies for the target IP address are
received, the most recently received response replaces whatever is
already in the cache. Consequently, the destinations a node using a
duplicate IP address can communicate with depends on what its
neighboring nodes have in their ARP caches. In most cases, such
communication failures become apparent relatively quickly, since it
is unlikely that communication can proceed correctly on both nodes.
It is also the case that a number of ARP implementations (e.g., BSD-
derived implementations) log warning messages when an ARP request is
received from a node using the same address as the machine receiving
the ARP request.
When two interfaces on different links use the same address, the
routing subsystem generally delivers packets to only one of the nodes
because only one of the links has the right subnet corresponding to
the IP address. Consequently, the node using the address on the
"wrong" link will generally never receive any packets sent to it and
will be unable to communicate with anyone. For obvious reasons, this
condition is usually detected quickly.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 30]
INTERNET-DRAFT March 13, 1998
It should be noted that although an address containing a combined
identifier and locator can be forged, the routing subsystem
significantly limits communication using the forged address. First,
return traffic will be sent to the correct destination and not the
originator of the forged address. Second, routers performing ingress
filtering can refuse to forward traffic claiming to originate from a
source whose claimed address does not match the expected addresses
(from a topology perspective) for sources located within a particular
region [RFC 2267]. To effectively masquerade as someone else
requires subverting the intermediate routing subsystem.
5.3.2. Identifier Authentication in GSE
In GSE, it is not possible for the routing subsystem to provide any
enforcement on the authenticity of identifiers with respect to their
corresponding Routing Stuff, since the Routing Stuff and ESD portions
of an address are by definition completely orthogonal quantities.
This fundamental problem is compounded by the fact that GSE provides
no way (at the transport or network layer) to map an ESD into its
corresponding Routing Stuff. Thus, when looking at the source address
of a received packet, there is no way to ascertain whether the
Routing Stuff portion of the address corresponds to legitimate
Routing Stuff with respect to the corresponding ESD. Consequently, it
becomes trivial in many cases for one node to masquerade as another.
5.3.3. Transport Layer: What Locator Should Be Used?
In the following, we focus on what Routing Stuff to use with TCP.
UDP-based protocols also depend on the Routing Stuff in similar way.
Indeed, we believe that TCP is the "easier" case to deal with, for
two reasons. First, TCP is a stateful protocol in which both ends of
the connection can negotiate with each other. Some UDP-based
protocols are stateless, and remember nothing from one packet to the
next. Consequently, changing UDP-based protocols may require the
introduction of "session" features, perhaps as part of a common
"library", for use by applications whose transport protocol is
relatively stateless. Second, changes to UDP-based protocols in
practice mean changing individual applications themselves, raising
deployability questions.
There are three cases of interest from TCP's perspective:
- the sending side of an active open
- the sending side of a passive open (i.e., how to respond to an
draft-ietf-ipngwg-esd-analysis-02.txt [Page 31]
INTERNET-DRAFT March 13, 1998
active open)
- changes to the Routing Stuff during an open connection.
5.3.4. RG Selection On An Active Open
If the host is performing a TCP "active open", the application first
queries the DNS to obtain the destination address, which contains
appropriate RG. That is, the initiator of communication is assumed to
provide the correct Routing Stuff when initiating communication to a
specific destination.
5.3.5. RG Selection On An Passive Open
When a server passively accepts connections from arbitrary clients,
it has no choice but to assume that the Routing Stuff in the source
address of a received packet that initiated the communication is
correct, because it has no way to authenticate its validity. Note
that the Routing Stuff is "correct" only in the sense that it
corresponds to the site originating the connection. Whether the
Routing Stuff paired with the received ESD is actually located at
that site where the legitimate owner of the ESD currently resides is
not known. Because the ESD alone cannot be mapped into a locator (or
some other quantity that can provide input to an authentication
procedure), there is no way to determine whether the received Routing
Stuff corresponds to that legitimately associated with the source
identifier of the received packet. The issue of spoofing is
discussed in more detail later.
5.3.6. Mid-Connection RG Changes
While packets are flowing as part of an open connection, the RG
appearing on subsequent packets is susceptible to change through
renumbering events, or as a result of site-internal routing changes
that cause the egress point for off-site traffic to change. It is
even possible (in the worst case) that traffic-balancing schemes
could result in the use of two egress routers, with roughly every
other packet exiting through a different egress router. In GSE, the
RG does not change once a connection has been opened.
Because TCP under GSE demultiplexes packets using only ESDs, packets
will be delivered to the correct end-point regardless of what source
RG is used. However, in GSE return traffic continues to be sent via
the "old" RG, even though it may have been deprecated or become less
optimal because the peer's border router has changed. It would seem
draft-ietf-ipngwg-esd-analysis-02.txt [Page 32]
INTERNET-DRAFT March 13, 1998
highly desirable for TCP connections to be able to survive such
events. However, the completion of renumbering events (so that an
earlier RG is now invalid) and certain topology changes would require
TCP to switch sending to a new RG mid-connection. To explore the
whole space, we considered ways of allowing this mid-connection RG
change to happen.
If TCP connection identifiers are based on ESDs rather than full
addresses, traffic from the same ESD would be viewed as coming from
the same peer, regardless of its source RG. Because this
vulnerability is already present in today's Internet (forging full
source addresses is trivial), the mere delivery of incoming datagrams
with the same ESD but a different RG does not introduce new
vulnerability to TCP. In today's Internet, any node can already
originate FINs/RSTs from an arbitrary source address and potentially
or definitely disrupt the connection. Therefore, changing RG for
acceptance, or acceptance of traffic independent of its source RG,
does not appear to significantly worsen existing robustness. (See the
comment on ingress filtering in Section 5.3.1, however.)
We also considered allowing TCP to reply to each segment using the RG
of the most recently-received segment. Although this allows TCP to
survive some important events (e.g., renumbering), it also makes it
trivial to hijack connections, unacceptably weakening robustness
compared with today's Internet. A sender simply needs to guess the
sequence numbers in use by a given TCP connection [Bellovin 89] and
send traffic with a bogus RG to hijack a connection to an intruder at
an arbitrary location.
Providing protection from hijacking implies that the RG used to send
packets must be bound to a connection end-point (e.g., it is part of
the connection state). Although it may be reasonable to accept
incoming traffic independent of the source RG, the choice of sending
RG requires more careful consideration. Indeed, any subsequent change
in what RG is used for sending traffic must be properly authenticated
(e.g., using cryptographic means). In the GSE proposal, it is not
clear how to authenticate such a change, since the remote peer
doesn't even know its own RG. Consequently, the only reasonable
approach in GSE is to send to the peer using the first RG used for
the entire life of a connection. That is, always use the first RG
seen.
5.3.7. The Impact of Corrupt Routing Goop
Another interesting issue that arises is what impact corrupted RG
would have on robustness. Because the RG is not covered by the TCP
checksum (the sender doesn't know what source RG will be inserted),
draft-ietf-ipngwg-esd-analysis-02.txt [Page 33]
INTERNET-DRAFT March 13, 1998
it would be difficult to detect such corruption at the receiver.
Moreover, once a specific RG is in use, it does not change for the
duration of a connection. The interesting case occurs on the passive
side of a TCP connection, where a server accepts incoming connections
from remote clients. If the initial SYN from the client includes
corrupted RG, the server TCP will create a TCP connection (in the
SYN-RECEIVED state) and cache the corrupted RG with the connection.
The second packet of the 3-way handshake, the SYN-ACK packet, would
be sent to the wrong RG and consequently not reach the correct
destination. Later, when the client retransmits the unacknowledged
SYN, the server will continue to send the SYN-ACK using the bad RG.
Eventually the client times out, and the attempt to open a TCP
connection fails.
We next consider relaxing the restriction on switching RGs in an
attempt to avoid the previous failure scenario. The situation is
complicated by the fact that the RG on received packets may change
for legitimate reasons (e.g., a multi-homed site load-shares traffic
across multiple border routers). The key question is how one can
determine which RG is valid and which is not. That is, for each of
the RGs a sender attempts to use, how can it determine which RG
worked and which did not? Solving this problem is more difficult than
first appears, since one must cover the cases of delayed segments,
lost segments, simultaneous opens, etc. If a SYN-ACK is retransmitted
using different RGs, it is not possible to determine which of those
RGs worked correctly. We conclude that the only way TCP could
determine that a particular RG is correct is by receiving an ACK for
a specific sequence number in which all transmissions of that
sequence number used the same RG (a non-trivial addition to TCP).
At best, an RG selection algorithm for TCP would be relatively
straightforward but would require new logic in implementations of
TCP's opening handshake --- a significant transition/deployment
issue. We are not certain that a valid algorithm is attainable,
however. RG changes would have to be handled in all cases handled by
the opening handshake: delayed segments, lost segments, undetected
bit errors in RG, simultaneous opens, old segments, etc.
In the end, we conclude that although the corrupted SYN case
introduces potential problems, the changes that would need to be made
to TCP to robustly deal with such corruption would be significant, if
tractable at all. This would result in a transition to GSE also
having a significant TCPng component, a significant drawback.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 34]
INTERNET-DRAFT March 13, 1998
5.3.8. On The Uniqueness Of ESDs
The uniqueness requirements for ESDs depends on what purpose they
serve and how they are used. In GSE, ESDs identify interfaces,
requiring that they be globally unique. It does not make sense for
two different interfaces to use the same ESD; every interface must
have its own ESD to distinguish it from others.
If ESDs are only used to identify session endpoints, the situation
becomes more complex. At first glance it might appear that two nodes
using the same ESD cannot communicate. However, this is not
necessarily the case. In the GSE proposal, for example, a node
queries the DNS to obtain an IPv6 address. The returned address
includes the Routing Stuff of an address (the RG+STP portions). Since
the sending host transmits packets based on the entire destination
IPv6 address, the sender may well forward the packet to a router that
delivers the packet to its correct destination (using the information
in the Routing Stuff). It is only on receipt of a packet that a node
would extract the ESD portion of a datagram's destination address and
ask "is this for me?" That is, a sender may not notice the
destination ESD is the same as the sending ESD because of the Routing
Stuff part of the address.
A more problematic case occurs if two nodes using the same ESD
communicate with a third party. To the third party, packets received
from either machine might appear to be coming from the same machine
since they are both using the same ESD. Consequently, at the
transport level, if both machines choose the same source and
destination port numbers (one of the ports --- a server's well-known
port number --- will likely be the same), packets belonging to two
distinct transport connections will be demultiplexed to a single
transport end-point.
When packets from different sources using the same source ESD are
delivered to the same transport end-point, a number of possibilities
come to mind:
1) The transport end-point could accept the packet, without regard
to the Routing Stuff of the source address. This may lead to a
number of robustness problems, if data from two different
sources mistakenly using the same ESD are delivered to the same
transport or application end-point (which at best will confuse
the application).
2) The transport end-point could verify that the Routing Stuff of
the source address matches one of a set of expected values
before processing the packet further. If the Routing Stuff
doesn't match any expected value, the packet could be dropped.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 35]
INTERNET-DRAFT March 13, 1998
This would result in a connection from one host operating
correctly, while a connection from another host (using the same
ESD) would fail.
3) When a packet is received with an unexpected Routing Stuff the
receiver could invoke special-purpose code to deal with this
case. Possible actions include attempting to verify whether the
Routing Stuff is indeed correct (the saved values may have
expired) or attempting to verify whether duplicate ESDs are in
use (e.g., by inventing a protocol that sends packets using both
Routing Stuff and verifies that they are delivered to the same
end-point).
5.3.9. New Denial of Service Attacks.
It is clear that there are potential problems if identifiers are not
globally unique. How common such problems would actually occur in
practice depends on how many duplicates there are actually are. Thus,
one might be tempted to make the argument that a scheme for assigning
identifiers could be made to be "unique enough" in practice. This
would be a dangerous and naive assumption, because intruders will
actively impersonate other sites for the sole purpose of invalidating
the uniqueness assumption. For example, one could deny service to
host foo.bar.com by querying the DNS for its corresponding ESD, and
then impersonaiting that ESD.
As a specific example, one GSE-specific denial-of-service attack
would be for an intruder to masquerade as another host and "wedge"
connections in a SYN-RECEIVED state by sending SYN segments
containing an invalid RG in the source IP address for a specific ESD.
Subsequent connection attempts to the wedged host from the legitimate
owner of the ESD (if they used the same TCP port numbers) would then
not complete, since return traffic would be sent to the wrong place.
5.3.10. Summary of Identifier Authentication Issues
In summary, changing the RG dynamically in a safe way for a
connection requires that an originator of traffic be able to
authenticate a proposed change in the RG before sending to a
particular ESD via that RG. This is difficult for several reasons:
1) It can't be done on an end-to-end basis in GSE (e.g., via IPSec)
because the sender doesn't know what the RG portion of the
address will be when it reaches the sender.
2) It can't easily be done in GSE because there is no mechanism at
draft-ietf-ipngwg-esd-analysis-02.txt [Page 36]
INTERNET-DRAFT March 13, 1998
or below the transport layer to map ESDs into a quantity that
can be used as a key to jump start the authentication process
(using the DNS would be problematic due to layering circularity
considerations).
3) Any scheme that uses the full IPv6 address to do the
authentication can be used with standard provider-based
addressing, raising the question of what benefit is retained
from having separate identifiers and locators.
Our final conclusion is that with the GSE approach, transport
protocol end-points must make an early, single choice of the RG to
use when sending to a peer and stick with that choice for the
duration of the connection. Specifically:
1) The demultiplexing of arriving packets to their transport end
points should use only the ESD, and not the Routing Stuff.
2) If the application chooses an RG for the remote peer (i.e., an
active open), use the provided RG for all traffic sent to that
peer, even if alternative RGs are received on subsequent
incoming datagrams from the same ESD.
3) For all other cases, use the first RG received with a given ESD
for all sending. Simultaneously, we understand that with this
rule, there are still open issues with regard to invalid RGs,
either through corruption or through a active hostile attacks.
With the above recommendation, there does not appear to be a
straightforward way to use ESDs in conjunction with mobility or site
renumbering (in which existing connections survive the renumbering).
This presents a quandry. The main benefit of separating identifiers
and locators is the ability to have communication (e.g., a TCP
connection) continue transparently, even when the Routing Stuff
associated with a particular ESD changes. However, switching to a new
Routing Stuff without properly authenticating it makes it trivial to
hijack connections.
We cannot emphasize enough that the use of an ESD independent of an
associated RG can be very dangerous. That is, communicating with a
peer implies that one is always talking to the same peer for the
duration of the communication. But as has been described in previous
sections, such assurance can only come from properly authenticated
RG.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 37]
INTERNET-DRAFT March 13, 1998
5.4. Miscellaneous
5.4.1. Renumbering and Domain Name System (DNS) Issues
Because any mapping scheme is complicated by renumbering, and because
recent IPv4 experience has shown a requirement for renumbering at
some frequency, it is worthwhile to explore the general renumbering
issue.
5.4.2. How Frequently Can We Renumber?
One premise of the GSE proposal [GSE] is that an ISP can renumber the
Routing Goop portion of a site's addresses transparently to the site
(i.e., without coordinating the change with the site). This would
make it possible for backbone providers to aggressively renumber the
Routing Goop part of addresses and achieve a high degree of route
aggregation. On closer examination, frequent (e.g., daily)
renumbering turns out to be difficult in practice because of a
circular dependency between the DNS and routing. Specifically, if a
site's Routing Stuff changes, nodes communicating with the site need
to obtain the new Routing Stuff. In the GSE proposal, one queries the
DNS to obtain this information. However, in order to reach a site's
DNS servers, the pointers controlling the downward delegation of
authoritative DNS servers (i.e., DNS "glue records") must use
addresses with Routing Stuff that are reachable. That is, in order to
find the address for the web server "www.foo.bar.com", DNS queries
might need to be sent to a root DNS server, as well as DNS servers
for "bar.com" and "foo.bar.com". Each of these servers must be
reachable from the querying client. Consequently, there must be an
overlap period during which both the old Routing Stuff and the new
Routing Stuff can be used simultaneously. During the overlap period,
DNS glue records would need to be updated to use the new addresses
(including Routing Stuff). Only after all relevant DNS servers have
been updated and older cached RRs containing the old addresses have
timed out can the old address be deleted.
An important observation is that the above issue is not specific to
GSE: the same requirement exists with today's provider-based
addressing architecture. When a site is renumbered (e.g., it switches
ISPs and obtains a new set of addresses from its new provider), the
DNS must be updated in a similar fashion.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 38]
INTERNET-DRAFT March 13, 1998
5.4.3. Efficient DNS support for Site Renumbering
When a site renumbers to satisfy its ISP, only the site's routing
prefix needs to change. That is, the prefix reflects where within the
Internet the site resides.
In the current Internet, when a site is renumbered, the addresses of
all the site's internal nodes change. This requires a potentially
large update to the RR database for that site. Although Dynamic DNS
[DDNS] could potentially be used, the cost is likely to be large due
to the large number of individual records that would need to be
updated. In addition, when DHCP and DDNS are used together [DHCP-
DDNS], it may be the case that individual hosts "own" their own A or
AAAA records, further complicating the question of who is able to
update the contents of DNS RRs.
One change that could reduce the cost of updating the DNS when a site
is renumbered is to split addresses into two distinct portions: a
Routing Goop that reflects where a node attaches to the Internet and
a STP-plus-ESD that is the site-specific part of an address. During a
renumbering, the Routing Goop would change, but the "site internal
part" would remain fixed. Furthermore, the two parts of the address
could be stored in the DNS as separate RRs. That way, renumbering a
site would only require that the Routing Goop RR of a site be
updated; the "site-internal part" of individual addresses would not
change.
To obtain the address of a node from the DNS, a DNS query for the
name would return two quantities: the "site internal part" and the
DNS name of the Routing Stuff for the site. An additional DNS query
would then obtain the specific RR of the site, and the complete
address would be synthesized by concatenating the two pieces of
information.
Implementing these DNS changes increases the practicality of using
Dynamic DNS to update a site's DNS records as it is renumbered. Only
the site's Routing Goop RRs would need updating.
Finally, it may be useful to divide a node's AAAA RR into the three
logical parts of the GSE proposal, namely RG, STP and ESD. Whether or
not it is useful to have separate RRs for the STP and ESD portions of
an address or a single RR combining both is an issue that requires
further study.
If AAAA records are comprised of multiple distinct RRs, then one
question is who should be responsible for synthesizing the AAAA from
its components: the resolver running on the querying client's machine
or the queried name server? To minimize the impact on client hosts
draft-ietf-ipngwg-esd-analysis-02.txt [Page 39]
INTERNET-DRAFT March 13, 1998
and make it easier to deploy future changes, it is recommended that
the synthesis of AAAA records from its constituent parts be done on
name servers rather than in client resolvers.
5.4.4. Two-Faced DNS
The GSE proposal attempts to hide the RG part of addresses from nodes
within a site. If the nodes do not know their own RG, then they can't
store or use them in ways that cause problems should the site be
renumbered and its RG change (i.e., the cached RG become invalid). A
site's DNS servers, however, will need to have more information about
the RG its site uses. Moreover, the responses it returns will depend
on who queries the server. A query from a node within the site should
return an address with a Site Local RG, whereas a query for the same
name from a client located at a different site should return the
global scope RG. This facilitates intra-site communication to be
more resilient to failures outside of the site. Such context-
dependent DNS servers are commonly referred as "two-faced" DNS
servers.
Some issues that must be considered in this context:
1) A DNS server may recursively attempt to resolve a query on
behalf of a requesting client. Consequently, a DNS query might
be received from a proxy rather than from the client that
actually seeks the information. Because the proxy may not be
located at the same site as the originating client, a DNS server
cannot reliably determine whether a DNS request is coming from
the same site or a remote site. One solution would be to
disallow recursive queries for off-site requesters, though this
raises additional questions.
2) Since cached responses are, in general, context sensitive, a
name server may be unable to correctly answer a query from its
cache, since the information it has is incomplete. That is, it
may have loaded the information via a query from a local client,
and the information has a site-local prefix. If a subsequent
request comes in from an off-site requester, the DNS server
cannot return a correct response (i.e., one containing the
correct RG).
draft-ietf-ipngwg-esd-analysis-02.txt [Page 40]
INTERNET-DRAFT March 13, 1998
5.4.5. Bootstrapping Issues
If Routing Stuff information is distributed via the DNS, key DNS
servers must always be reachable. In particular, the addresses
(including Routing Stuff) of all root DNS servers are, for all
practical purposes, well-known and assumed to never change. It is not
uncommon for the addresses of root servers to be hard-coded into
software distributions. Consequently, the Routing Stuff associated
with such addresses must always be usable for reaching root servers.
If it becomes necessary or desirable to change the Routing Stuff of
an address at which a root DNS server resides, the routing subsystem
will likely need to continue carrying "exceptions" for those
addresses. Because the total number of root DNS servers is relatively
small, the routing subsystem is expected to be able to handle this
requirement.
All other DNS server addresses can be changed, since their addresses
are typically learned from an upper-level DNS server that has
delegated a part of the name space to them. So long as the delegating
server is configured with the new address, the addresses of other
servers can change.
6. Conclusion
The GSE proposal provides a concrete example of a network protocol
design that separates identifiers from locators in addresses. In
this paper we compared GSE with IPv4 to better understand the pro's
and con's of the respective design approaches.
Functionally speaking, identifiers and locators each have a logically
different role to play. Thus overloading both in one field causes
problems whenever the location of a node changes but its identity
does not. However, our analysis shows that overloading also presents
two critically important benefits.
First, for network entity A to send data to network entity B, A must
not only know B's end identifier but also B's locator. No scalable
way is known at this time to provide this mapping at the network
layer, other than overloading the two quantities into an address as
is done in IPv4. Fundamentally, a scalable mapping algorithm strongly
suggests that the identifier space be structured hierarchically, yet
identifiers in GSE are not sufficiently large to both contain
sufficient heirarchy and support stateless address autoconfiguration.
Instead, GSE forces applications to supply up-to-date locators.
However, relying on the locator provided at the time communication is
established as GSE does is inadequate when the remote locator can
draft-ietf-ipngwg-esd-analysis-02.txt [Page 41]
INTERNET-DRAFT March 13, 1998
change dynamically, precisely the scenario that is supposed to
benefit from the separation. That is, the benefits of separating the
identifier from the locator are largely lost, if the changes in the
identifier to locator binding are not tracked quickly.
Secondly, when communicating with a remote site, a receiver must be
able to insure (with reasonable certainty) that received data does
indeed come from the expected remote entity. In IPv4, it is possible
to receive packets from a forged source, but the potential for
mischief between communicating peers is significantly limited because
return traffic will not reach the source of the forged traffic. That
is, communication involving packets sent in both directions will not
succeed. In contrast, architectures like GSE that decouple the
identifier and locator functions have great difficulty assuring that
traffic from a source identified only by an identifer actually comes
from the correct source. Short of using cryptographic techniques in
which both end points share a private secret (e.g., using IPSec),
there is no known mechanism that can use an identifier alone to
perform this remote entity authentication in a scalable way. That
is, using an identifier alone for authentication of received packets
is dangerously unsafe.
In summary, although overloading the address field with a combined
identifier and locator leads to difficulties in retaining the
identity of a node whenever its address changes, analysis in this
paper suggests that the benefit of the overloading actually out-
weighs its cost. Completely separating an identifier from its
locator renders the identifier untrustworthy, thus useless, in the
absence of an accompanying authentication system.
7. Security Considerations
The primary security consideration with GSE or, more generally, a
network layer with addresses split into locator and identifier parts,
is that of one node impersonating another by copying the
identification without the location.
8. Acknowledgments
Thanks go to Steve Deering and Bob Hinden (the Chairs of the IPng
Working Group) as well as Sun Microsystems (the host for the PAL1
meeting) for the planning and execution of the interim meeting.
Thanks also go to Mike O'Dell for writing the 8+8 and GSE drafts; by
publishing these documents and speaking on their behalf, Mike was the
catalyst for some valuable discussions, both for IPv6 addressing and
for addressing architectures in general. Special thanks to the
draft-ietf-ipngwg-esd-analysis-02.txt [Page 42]
INTERNET-DRAFT March 13, 1998
attendees of the PAL1 meeting whose high caliber discussions helped
motivate and shape this document.
9. References
[BATES] Scalable support for multi-homed multi-provider
connectivity, Internet Draft, Tony Bates & Yakov Rekhter,
draft-bates-multihoming-01.txt.
[Bellovin 89] "Security Problems in the TCP/IP Protocol Suite",
Bellovin, Steve, Computer Communications Review, Vol. 19,
No. 2, pp32-48, April 1989.
[CERT] CERT(sm) Advisory CA-96.21
(ftp://info.cert.org/pub/cert_advisories)
[DANVERS] Minutes of the IPNG working Group, April 1995.
ftp://ftp.ietf.cnri.reston.va.us/ietf-online-proceedings/
95apr/area.and.wg.reports/ipng/ipngwg/ ipngwg-minutes-
95apr.txt.
[DHCP-DDNS] Interaction between DHCP and DNS, Internet Draft, Yakov
Rekhter, draft-ietf-dhc-dhcp-dns-04.txt.
[DDNS] "Dynamic Updates in the Domain Name System (DNS UPDATE)",
Paul Vixie (Editor), draft-ietf-dnsind-dynDNS-11.txt,
November, 1996.
[EUI64] 64-Bit Global Identifier Format Tutorial.
http://standards.ieee.org/db/oui/tutorials/EUI64.html.
Note: "EUI-64" is claimed as a trademark by an organization
which also forbids reference to itself in association with
that term in a standards document which is not their own,
unless they have approved that reference. However, since
this document is not standards-track, it seems safe to name
that organization: the IEEE.
[GSE] "GSE - An Alternate Addressing Architecture for IPv6", Mike
O'Dell, draft-ietf-ipngwg-gseaddr-00.txt.
[IEEE802] IEEE Std 802-1990, Local and Metropolitan Area Networks:
IEEE Standard Overview and Architecture.
[IEEE1212] IEEE Std 1212-1994, Information technology--
Microprocessor systems: Control and Status Registers (CSR)
Architecture for microcomputer buses.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 43]
INTERNET-DRAFT March 13, 1998
[RFC1122] "Requirements for Internet hosts - communication layers",
R. Braden, 10/01/1989.
[RFC1715] The H Ratio for Address Assignment Efficiency. C.
Huitema.
[RFC1726] Technical Criteria for Choosing IP:The Next Generation
(IPng). F. Kastenholz, C. Partridge.
[RFC1752] "The Recommendation for the IP Next Generation Protocol,"
S. Bradner, A. Mankin, 01/18/1995.
[RFC1788] "ICMP Domain Name Messages", W. Simpson, 04/14/1995
[RFC1958] Architectural Principles of the Internet. B. Carpenter.
[RFC1971] IPv6 Stateless Address Autoconfiguration. S. Thomson, T.
Narten.
[RFC2002] "IP Mobility Support", C. Perkins, RFC 2002, October,
1996.
[RFC2008] "Implications of Various Address Allocation Policies for
Internet Routing", Y. Rekhter, T. Li.
[RFC2065] Domain Name System Security Extensions. D. Eastlake, C.
Kaufman.
[RFC2073] An IPv6 Provider-Based Unicast Address Format. Y.
Rekhter, P. Lothberg, R. Hinden, S. Deering, J. Postel
[RFC2267] Network Ingress Filtering: Defeating Denial of Service
Attacks which employ IP Source Address Spoofing, P.
Ferguson, D. Senie, January 1988.
10. Authors' Addresses
Matt Crawford John Stewart
Fermilab MS 368 Juniper Networks, Inc.
PO Box 500 385 Ravendale Drive
Batavia, IL 60510 USA Mountain View, CA 94043
Phone: 630-840-3461 Phone: +1 650 526 8000
EMail: crawdad@fnal.gov EMail: jstewart@juniper.net
Allison Mankin Lixia Zhang
USC/ISI UCLA Computer Science Department
4350 North Fairfax Drive 4531G Boelter Hall
draft-ietf-ipngwg-esd-analysis-02.txt [Page 44]
INTERNET-DRAFT March 13, 1998
Suite 620 Los Angeles, CA 90095-1596 USA
Arlington, VA 22203 USA Phone: 310-825-2695
EMail: mankin@isi.edu EMail: lixia@cs.ucla.edu
Phone: 703-807-0132
Thomas Narten
IBM Corporation
3039 Cornwallis Ave.
PO Box 12195 - F11/502
Research Triangle Park, NC 27709-2195
Phone: 919-254-7798
EMail: narten@raleigh.ibm.com
draft-ietf-ipngwg-esd-analysis-02.txt [Page 45]
INTERNET-DRAFT March 13, 1998
Appendix B -- Ideas Incorporated Into IPv6
This section summarizes changes made to IPv6 specifications which
originated in the GSE proposal or in the discussions arising from it.
First and most visible was the change to the unicast address format.
Instead of an topologically insignificant Registry ID immediately
following the Format Prefix, there is now a Top-Level Aggregation
Identifier. This field will identify a large routable aggregate to
which an address belongs rather than an administrative unit by which
an address was assigned. The TLA corresponds to the "Large
Structure" of GSE. The IPv6 Next-Level Aggregation Identifier (NLA)
is roughly the rest of the GSE "Routing Goop" and the Site-Level
Aggregation Identifier (SLA) is a slightly expanded GSE Site Topology
Partition.
The decision to put fixed boundaries between parts of the unicast
address (TLA, NLA, SLA, Interface Identifier) also came from GSE.
The previous "provider-based" addressing architecture for IPv6 had
fluid boundaries between Registry ID, Provider ID, Subscriber ID and
the Intra-Subscriber part, as well as undefined divisions within the
Provider-ID and Intra-Subscriber part. (On subnetworks with a MAC-
layer address, the latter boundary was generally placed to
accommodate use of that address as an Interface ID.) The new
addressing architecture still expects divisions within the NLA
portion of the address, placed to reflect topological aggregation
points.
Defining a fixed boundary between the routable portion of the address
and the node-on-link identifier required the specification of an
Interface Identifier which would be as suitable as possible for all
subnetwork technologies. The IEEE "EUI-64" identifier was selected,
having the advantages of an easy mapping from 48 bit MAC addresses
and a defined escape flag into locally-administered values.
The second change to come out of the GSE discussions relates to
reducing the number of DNS record changes required in the event of
site renumbering. This work is not finalized as of this writing, but
the result may be that individual IPv6 addresses are stored (and
signed, in the case of Secure DNS) as a partial address and an
indirect pointer which leads to the high-order part of the address.
There may be multiple levels of indirection and a changed record at
any one level would suffice to update the DNS's record of the IPv6
addresses of every node in a given branch of the addressing
hierarchy.
A change in the method of doing DNS address-to-name lookups is also
in the works. This may be a change in the form and/or operation of
draft-ietf-ipngwg-esd-analysis-02.txt [Page 46]
INTERNET-DRAFT March 13, 1998
the ip6.int domain or some new mechanism which involves participation
by the routers or the end-nodes themselves.
Two other changes arising from GSE will not affect the IPv6 base
specifications themselves, but do direct additional work. Those are
the injection of global prefix information into a site from a
provider or exchange, and some inter-provider cooperative method of
providing multihoming to mutual customers with minimal impact on
routing tables in distant parts of the network.
draft-ietf-ipngwg-esd-analysis-02.txt [Page 47]
| PAFTECH AB 2003-2026 | 2026-04-23 06:11:18 |