INTERNET-DRAFT Matt Crawford
Fermilab
<draft-ietf-ipngwg-esd-analysis-05.txt> Allison Mankin
ISI
Thomas Narten
IBM
John W. Stewart, III
Juniper
Lixia Zhang
UCLA
October, 1999
Separating Identifiers and Locators in Addresses: |
An Analysis of the GSE Proposal for IPv6
<draft-ietf-ipngwg-esd-analysis-05.txt> |
Status of this Memo |
This document is an Internet-Draft and is in full conformance with
all provisions of Section 10 of RFC2026 except that the right to
produce derivative works is not granted.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF), its areas, and its working groups. Note that
other groups may also distribute working documents as Internet-
Drafts.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet- Drafts as reference
material or to cite them other than as "work in progress."
The list of current Internet-Drafts can be accessed at
http://www.ietf.org/ietf/1id-abstracts.txt
The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html.
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-05.txt [Page 1]
INTERNET-DRAFT October, 1999
includes the feature of configuring a node internal to a site with
only the local routing and end-point identification 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.
It has often been said that IPv4 "got it wrong" by treating its |
addresses simultaneously as locators and identifiers. However, there |
has never beeeen a detailed and comprehensive 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
identifiers 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 |
4. The GSE Proposal......................................... 14 |
5. Analysis: The Pros and Cons of Overloading Addresses..... 21 |
6. Conclusion............................................... 39 |
7. Security Considerations.................................. 40 |
8. Acknowledgments.......................................... 40 |
9. References............................................... 41 |
10. Authors' Addresses...................................... 43 |
Appendix A: Increased Reliance on Domain Name System (DNS)... 43 |
draft-ietf-ipngwg-esd-analysis-05.txt [Page 2]
INTERNET-DRAFT October, 1999
Appendix B: Additional Issues Related to Specifically to GSE. 47 |
Appendix C: Ideas Incorporated Into IPv6..................... 48 |
Appendix D: Reverse Mapping of Complete GSE Addresses........ 49 |
1. Introduction
In October of 1996, Mike O'Dell published an Internet-Draft (dubbed
"8+8") that proposed significant changes to the IPv6 unicast
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
Internet-Draft was produced. This version changed the name of the
proposal from "8+8" to "GSE" to identify the three separate
components of a unicast 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 several
improvements to the then existing IPv6 specifications (including
increasing the aggregatability of addresses, having hard boundaries
between routing and non-routing parts of the address, and easing the
DNS aspects of renumbering).
This document focuses primarily on the issue of separating unicast
addresses into distinct portions for identification and location
purposes, a separation that IPv4 does not make but that is
fundamental to GSE. 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 new challenges, 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 3]
INTERNET-DRAFT October, 1999
results on IPv6 addressing.
Finally, this document's focus on unicast issues should not be
interpreted to mean that the impact of separating identifier and
locating functions on non-unicast aspects of routing and addressing
are well understood or trivial to deal with. Specifically,
understanding how multicasting and anycast addressing [ANYCAST,
RFC1884] fits into such a model requires further work.
2. Definitions and Terminology
The following terminology is used throughout this document.
Routing Goop --- A term defined by the GSE document. It refers to
the first six bytes of a sixteen byte 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 on the public Internet the site
housing the address resides.
Site Topology Partition --- A term defined by the GSE document
that refers to the two bytes of a sixteen byte 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.
Routing Stuff --- The part of an address that identifies which
link the address resides on. Within the context of
GSE, the Routing Stuff comprises the Routing Goop
and Site Topology Partition parts of an address
(i.e., the left mots eight bytes).
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 (i.e., the rightmost
eight bytes) 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.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 4]
INTERNET-DRAFT October, 1999
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 count
on. 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/re-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.
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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 5]
INTERNET-DRAFT October, 1999
addresses within transport protocol addresses that identify the end-
point of a connection couples those transport protocols with routing
to a degree. This entanglement is inconsistent with a (too) 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 complicates 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 since the same packet field is used for
both.
Consequently, there has been a train of thought for some time 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.
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 a single 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 by 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 6]
INTERNET-DRAFT October, 1999
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 the central Internet Assigned Numbers
Authority (IANA) to obtain an address block 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
actual 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.
+----------------+
| |------- 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 7]
INTERNET-DRAFT October, 1999
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 many of them 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, even though
it is going to take the same path for all of them; in other words,
there is an opportunity to do data abstraction.
Figure 1 shows network numbers using the older "classful" notation.
Since 1981, the first few bits of an address syntactically identified
which parts of an address identified the "network" and "local"
portions of an address. There were a small number of Class A
addresses (intended for very large sites), a medium number of Class B
addresses (for medium-sized sites) and a very large number of Class C
addresses (for very small sites). In practice, the actual size of
real networks didn't match the original allocation of Class A, B, and
C addresses. Class B addresses were bigger than most sites needed
(and there weren't enough of them), and Class C addresses were too
small (i.e., typical sites would need to get 10 or more C blocks to
cover all of the hosts on their networks). Consequently, classless
addressing was developed [CIDR], which made the boundaries between
the network and local parts of an address more flexible. With
classless addressing, a separate prefix-length (i.e., network mask)
specifies how many of the left-most bits of an address identify the
network part of the address.
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 cost of computing a routing table
and the size of the forwarding table computed from the routing table.
To achieve this goal CIDR aggressively aggregates network addresses.
Aggregating network addresses means "merging" multiple addresses into
a single "bigger" one, that is to use a common prefix to provide
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.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 8]
INTERNET-DRAFT October, 1999
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. 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, and Provider
B's routers need only a single routing table entry to reach all of
Provider A's customers. Note the important difference between the
cases described in Figures 1 and 2; the latter example uses fewer
entries in the routing table to reach the same number of
destinations.
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 a significant decrease in the growth rate of the routing
tables. 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" [RFC2008]. In pre-CIDR
draft-ietf-ipngwg-esd-analysis-05.txt [Page 9]
INTERNET-DRAFT October, 1999
days sites acquired addresses from a central authority independent of
their provider, and a site could assume it "owned" the address block
it was given. Owning addresses meant that once one had been given a
set of network addresses, one could always use them; no matter where
one's site connected to the Internet, the prefix for that network
could be injected into the public routing system. Today, however, it
is simply not possible for all individual sites to have their own
prefixes injected into the DFZ; there would be too many of them.
Consequently, if a site decides to change providers, 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
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 Customer3, 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 address block. From
Provider B's view, the address space underneath 204.1.0.0/16 is no
draft-ietf-ipngwg-esd-analysis-05.txt [Page 10]
INTERNET-DRAFT October, 1999
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 Customer1-5,
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 a small number of
new exceptions could be tolerated, and certain prefixes have been
grandfathered, 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, with the current practice of
provider-based addressing, 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 today do
route filtering based, in part, on prefix length; as a result, a site
which does not renumber may have only partial connectivity to the
Internet. That is, a site may advertise a long prefix into the
routing system, but there is no assurance that all parts of the
Internet will accept the route; some simply ignore it.
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 renumbering 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.) This can lead to
a "tragedy of the commons" situation, where everyone agrees that some
sites should renumber, but they themselves want to be one of those
that do not.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 11]
INTERNET-DRAFT October, 1999
3.4. Multi-Homed Sites 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, the prefix of their own address block. 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 might 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 level, there are three possible ways
to deal with multi-homed sites. The first possibility is to stay
with pre-CIDR approach, allowing each multi-homed site to receive its
address block directly from a registry, independent of its providers.
The problem with this approach is that, because the address block 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 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. 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, if you want shortest-path routing load-spreading.
Consider Figure 4. Provider C has two paths for reaching Customer1.
Provider A advertises 204.1/16, an aggregate which includes
Customer1. 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 perspective, the
draft-ietf-ipngwg-esd-analysis-05.txt [Page 12]
INTERNET-DRAFT October, 1999
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 Customer1 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 | |
/ +------------+ +------------+ |
/ +----------/ |
/ / |
Customer1 --- / 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 and not advertise the prefix obtained from
one provider to any of its other 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 faces 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. While this may be a problem for the (limited) IPv4
draft-ietf-ipngwg-esd-analysis-05.txt [Page 13]
INTERNET-DRAFT October, 1999
address space, it is not a significant issue in IPv6. Second,
when the connection to ISP-A goes down, addresses numbered out
of ISP-A's space become unreachable. 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 network applications 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 support multi-homing with IPv4/CIDR faces 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 concern 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 when they
change providers. 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 asserts that IPv4 with CIDR
has not achieved the aggressive aggregation required for the route
draft-ietf-ipngwg-esd-analysis-05.txt [Page 14]
INTERNET-DRAFT October, 1999
computation functions of the DFZ of the Internet to scale for IPv4
and that the much larger address space of IPv6 simply exacerbates the
problem.
The GSE proposal does not propose to eliminate the need for
renumbering. Indeed, it asserts that end sites will have to renumber
more frequently in order to continue scaling the Internet. However,
GSE proposes to make the cost of renumbering small enough that sites
can be renumbered at essentially any time with little or no
disruption to its network connectivity, and in particular with no
impact on communications that are strictly within the site.
Finally, GSE attempts to address the problem of sites that have
multiple Internet connections. In CIDR, the pressure for better
multi-homing support can create exceptions to route aggregation and
result in poor scaling. That is, the public routing infrastructure
may have to carry multiple distinct routes for some demanding multi-
homed sites, one for each independent path. GSE recognizes the
"special work done by the global Internet infrastructure on behalf of
multi-homed sites" [GSE], and proposes a way for multi-homed sites to
gain certain benefit without impacting global scaling. This includes
a specific mechanism that providers can 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 functions, it splits the address into two elements: the
high-order 8 bytes used for routing purposes (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 is:
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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 15]
INTERNET-DRAFT October, 1999
4.2.1. Routing Stuff (RG and STP)
The Routing Goop (RG) identifies where within the public Internet
topology a site connects and is used to route datagrams to the site.
RG is structured as follows:
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 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 now examine the bits
in the RG. After the 3-bit prefix identifying the address as having
a GSE format, 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.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 16]
INTERNET-DRAFT October, 1999
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.
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 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 even know its RG, as described later) and on receipt
of a packet only the ESD would be used in testing whether the 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 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 an appropriate MAC address or
draft-ietf-ipngwg-esd-analysis-05.txt [Page 17]
INTERNET-DRAFT October, 1999
EUI-64 identifier. A globally unique ESD must then be 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). Although we do not explore them in detail
here, we note that a global coordination structure is required here
to control the allocation of globally unique identifiers.
4.3. Address Rewriting by Border Routers
To obviate the need to renumber devices within sites because of
changing providers, the GSE design hides the global Routing Goop (RG)
from hosts in each site by having 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.
GSE's approach to easing renumbering isn't so much to ease
renumbering as to make it transparent to end users. The RG by which
a site is known is hidden 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 is to insure the changeability of the RG without impacting the
site itself. It is expected that the RG would need to change
relatively frequently (e.g., several times a year) in order to
support sufficient 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 18]
INTERNET-DRAFT October, 1999
accomplished through a dynamic protocol with the upstream provider. |
In addition, the site's DNS records would need updating to properly |
indicate the current RG value.
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 may well 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 shields a site's intra-site traffic from any
instabilities resulting from renumbering.
In addition to rewriting source addresses that leave a site,
destination addresses must be 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 is expected to know the full 16-byte address for
the destination through mechanisms such as a DNS query. The
initiating node does not, however, know its own RG, and 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 19]
INTERNET-DRAFT October, 1999
to stay the same, all of SmallISP1's customers would have to renumber
into address space covered by an aggregate of BigISP2. GSE deals
with this problem by handling the RG in DNS with indirection.
Specifically, a site's DNS server specifies the RG portion of its
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. Note that this mechanism does not depend on the GSE address
format per se and can also be applied to IPv4 addressing.
4.5. Support for Multi-Homed Sites
GSE defines 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 addresses
returned from DNS queries that include RG1 would not be viable
addresses. If PBR1 and PBR2 knew about each other, however, then in
draft-ietf-ipngwg-esd-analysis-05.txt [Page 20]
INTERNET-DRAFT October, 1999
this case PBR1 could tunnel packets destined for RG1-prefixed
addresses to PBR2, thus keeping the communication working. (Note
that IP-in-IP encapsulation is necessary since routers between PBR1
and PBR2 would forward packets destined for addresses with PBR1's
prefix back towards PBR1.)
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 multicast can be made to work under GSE.
3) It is not known how mobility 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 simply 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.
At the network layer, an identifier refers to an interface, at
the transport layer it refers to a process or other endpoint of
a "connection".
2) Identifiers must be mapped into a locator that the network layer
draft-ietf-ipngwg-esd-analysis-05.txt [Page 21]
INTERNET-DRAFT October, 1999
can use to actually deliver a packet to its intended
destination.
3) There must be a suitable way to adequately authenticate the user
of an identifier, so that communicating peers have sufficient
confidence that packets sent to or received from a particular
identifier correspond to the intended recipient.
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.
In IPv4, when applications communicate, transport "identifiers"
consist of addresses and port numbers. For the purposes of this
discussion, we use the term "identifier" to mean 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 a structure to keep
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 identifier
portion of an address appears to provide 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 22]
INTERNET-DRAFT October, 1999
packet to its higher-layer end point. This means that if a packet is
delivered to the correct destination node (that is the identifier
carried in the packet address matches to one interface identifier of
the node), the node will accept the packet, regardless of how the
packet got there. The exact locator used does not matter, within
most Internet circumstances, so long as it gets the packet delivered
to its proper destination.
The most obvious example that could benefit from the separation of
locators and identifiers involves communication with a mobile host.
Transport protocols such as TCP are unable to keep connections open
if either of the two endpoint identifiers for an open connection
changes. 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 [MOBILITY].
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 changing its identification.
As a side note, it is a requirement in GSE that packets be
demultiplexed to higher layer endpoints 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 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 accepted.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 23]
INTERNET-DRAFT October, 1999
Not surprisingly, having separate locator and identifiers in
addresses leads to additional problems as well. 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 packets being demultiplexed
upward in the protocol stack, it becomes much easier for an intruder
to masquerade as someone else.
5.2. Mapping an Identifier to a Locator
The idea of using addresses that cleanly separate location and
identification information is not new. 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 determining 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 difficult if not impossible
with a flat identifier space (e.g., the ESD identifier).
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.
However, in the case of TCP, such a scheme would have the
benefit that applications would probably not need to be
modified. For UDP-based applications, this may not hold, since
most UDP-based protocols are implemented within applications.
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,
draft-ietf-ipngwg-esd-analysis-05.txt [Page 24]
INTERNET-DRAFT October, 1999
especially if long-running connections are required to survive
renumbering and/or deal with mobile nodes.
The GSE proposal uses the last approach. 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. Note that addresses containing invalid
Routing Stuff will result any time when 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.
Whatever the causes, the failures are fundamentally due to dynamic
topological changes at the network layer, yet in GSE such failures
are left to be dealt with at the application level (through DNS),
because neither the transport nor the network level has the ability
to re-map identifiers to corresponding locators. |
To avoid the above problem a network architecture must provide the
ability to map an identifier to a locator. In IPv4, this mapping is
trivial, since the identifier and locator are combined in a single
quantity (i.e., the IPv4 address). GSE does not provide this mapping
functionality directly. Instead, GSE assumes that a node's DNS name
serves as its stable identifier, and uses normal DNS queries to map
the DNS "identifier" into an IPv6 address. The IPv6 address contains
both the ESD identifier together with its Routing Stuff, that is an
initial binding/mapping between the identifier and locator. When
this binding breaks (for example due to dynamic topological changes),
the ESD identifier cannot be mapped into a new locator by itself.
Instead one must resort back to application level, hoping another DNS
query would provide rescue to the broken binding between identifier
to locator that is needed for network delivery.
The use of DNS to provide identifier to locator mapping contributes
to GSE's apparent simplicity. 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 25]
INTERNET-DRAFT October, 1999
directly on the application, because lower layers (i.e., transport
and network layers) cannot perform the mapping themselves due to
layering violation concerns (i.e., TCP and UDP can't perform a DNS
query). Second, following all RG changes the DNS database must be
promptly updated and all expired information must be flushed out of
all DNS caches. This stringent timing requirement imposed by lower
level operation would represent a departure from the original DNS
design, which provides DNS names to address mappings that only change
slowly over time if at all, and which relies heavily on caching over
relatively long time periods to scale well.
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 Identifiers 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
space appropriately to support scalable 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 any structure.
Imposing a hierarchy on identifiers poses the following difficulties:
- - It increases the size of the identifier. The exact size
necessary to support sufficient hierarchy 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.
- - Due to the requirement of tying an identifier to the
delegation structure the identifier of a node cannot be burned
in during manufacturing. Instead a mechanism is needed to allow
a node to learn its identifier. To be practical, such a
mechanism would need to be automated and avoid the need for
manual configuration.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 26]
INTERNET-DRAFT October, 1999
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 hierarchy 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 hierarchy; 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. The delegation of the two-level hierarchy
(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
significance, would be simpler than the procedures for managing IPv4
addresses, 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].
The topic of mapping full 16-byte GSE addresses to a locator or other
information is discussed in Appendix D.
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, there is an
expectation that repeated and subsequent use of the same identifier
results in continued communication with the same end point. To be
useful then, a valid identifier must either be easily distinguishable
from a fraudulent one, or the system must have a way to prevent
identifiers from being used in an unauthorized manner.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 27]
INTERNET-DRAFT October, 1999
The remainder of this section discusses how identifier authentication
is done in both IPv4 and GSE, and shows how overloading an address
with both an identifier and a locator provides a significant
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 cover it here.
Instead, we focus on the relative strengths in the two schemes.
The following discussion assumes an absence of cryptographic |
authentication to bind an identifier to an end site. Many of the |
concerns described below would become non-issues if an appropriate |
cryptographic infrastructure were available. Section 5.5 discusses |
this issue in more detail. |
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, an |
(uncompromised) 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 28]
INTERNET-DRAFT October, 1999
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.
The previous discussion describes the operation of ARP in the absence |
of intruders or other malicious users. ARP has a number of security |
vulnerabilities that make it trivial for an intruder to intercept |
traffic and selectively process traffic that traverses a link, |
provided the intruder is attached to the link the traffic of interest |
traverses. For example, an intruder could intercept all traffic to an |
address by being the last to return an ARP response, and then |
selectively relay the traffic (after examining and/or modifying it) |
to its intended recipient. This is a classic man-in-the-middle |
attack. |
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.
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. This alone prevents certain types
of spoofing attacks. For example, if a destination receives an
unexpected packet corresponding to a TCP connection that it is
unaware of, it may return a TCP segment resetting the connection. |
Second, routers performing ingress filtering can refuse to forward
traffic claiming to originate from a source whose source 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.
To summarize, the routing subsystem in IPv4 provides a limited (but |
quite significant) defense against arbitrary hijacking of packets to |
an improper destination. We do not claim that this defense is |
sufficient against all types of attacks by a determined intruder. |
However, it does provide some degree of defense against accidental |
draft-ietf-ipngwg-esd-analysis-05.txt [Page 29]
INTERNET-DRAFT October, 1999
misconfigurations (e.g., assigning an improper address to an |
interface) and does erect hurdles that prevent an abritrary node from |
impersonating another node. The more dangerous attack, subverting |
the routing subsystem by injecting unauthorized routes, can be traced |
and detected by appropriate tools. |
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. |
Thus, even the limited protection offered by IPv4 is not immediately |
available. |
An interesting question is whether any such protection is needed. One |
argument is that address-based authentication is so inherently weak |
as to be useless, thus the increased vulnerability of a GSE-like |
scheme is not significant. Where authentication is desired, the use |
of something based on cryptography is necessary (e.g., IPsec |
[RFC2401]). |
There are at least two arguments against this line of thought. |
First, the lack of protection comparable to IPv4 may lead to a new |
set of (poorly understood) security threats; Section 5.5 below |
describes one possible threat. These threats must be dealt with at |
the transport (or lower) layer because the threats are to the |
integrety of the transport layer itself. Attempting to solve them at |
higher-layers (e.g., via IPsec [RFC2401] and IKE [RFC2409]) results |
in a potential layering circularity, where the security mechanisms |
rely on a correctly functioning transport, but the transport relies |
on those same security mechanisms to provide a service. Whether such |
a mechanism can be designed is an area of future work. |
Second, requiring that basic threats to the transport layer be dealt |
with using cryptographic techniques significantly increases the cost |
of formerly simple packet exchanges. Cryptographic security no longer |
becomes a choice an application can make, but quite possibly a |
requirement to protect against certain types of attacks. Thus, the |
cost of deploying effective defenses against a new class of denial of |
service attacks may be quite significant.
5.4. Transport Layer: What Locator Should Be Used?
In the following, we focus on what Routing Stuff to use with TCP; UDP
also depends on the Routing Stuff in similar way. Indeed, we believe
draft-ietf-ipngwg-esd-analysis-05.txt [Page 30]
INTERNET-DRAFT October, 1999
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. UDP-based communications are stateless,
and remember nothing from one packet to the next. Consequently,
changing UDP to remember locator information in addition to the
identifier of the peer may require the introduction of "session"
features, perhaps as part of a common "library". Second, changes to
UDP 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
active open)
- - changes to the Routing Stuff during an open connection.
5.4.1. 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 the
appropriate RG for the remote peer. That is, the initiator of
communication is assumed to provide the correct Routing Stuff when
initiating communication to a specific destination.
5.4.2. 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, which the server
will send the reply to. Whether the Routing Stuff paired with the
received ESD actually matches the Routing Stuff located at the site
where the legitimate owner of the ESD currently resides is not known
and cannot be determined. 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.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 31]
INTERNET-DRAFT October, 1999
5.4.3. 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 that traffic-balancing schemes could result in the use
of two egress routers, with roughly every other packet exiting
through a different egress router.
Because TCP under GSE demultiplexes packets using only ESDs, newly
arrived packets will be delivered to the correct end-point regardless
of whether their source RG have changed. The GSE proposal calls for
return traffic to continue 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. That is, the RG to use for reaching a
peer is bound to a connection when the connection is established and
does not change thereafter. 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 scenario, we consider ways of allowing
the RG change to be made to existing established connections.
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 the source RG. Because this
vulnerability is already present in today's Internet (forging the
source address of a packet 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, acceptance of
traffic independent of its source RG does not appear to significantly
worsen existing robustness. Note, however, that ingress filtering as
described in Section 5.3.1, cannot be performed on packets containing
GSE addresses. This does make it more difficult to prevent certain
types of attacks.
We also considered allowing TCP to reply to each segment using the RG
of the most recently-received segment. Although this allows TCP
connections to survive certain important events (e.g., renumbering),
it also makes it trivial for anyone 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 32]
INTERNET-DRAFT October, 1999
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 the RG used for sending traffic must be properly
authenticated (e.g., using cryptographic means). In the GSE
proposal, the is no apparent way 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, and accept the loss of connectivity whenever the RG
changes.
5.4.4. The Impact of Corrupted Routing Goop
Another interesting issue that arises is what impact corrupted RG |
would have on robustness, given that there is no IPv6 header checksum |
that could help detect a corrupted source address field. Because the |
RG is not covered by the TCP checksum (the sender doesn't know what |
source RG will be inserted), no TCP mechanism can detect such |
corruption at the receiver. Moreover, once a specific RG is in use, |
it does not change for the duration of a connection. One 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 a 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 destination 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 the two RGs worked correctly. We conclude that
the only way TCP can determine that a particular RG is correct is by
receiving an ACK for a specific sequence number in which all
draft-ietf-ipngwg-esd-analysis-05.txt [Page 33]
INTERNET-DRAFT October, 1999
transmissions of that sequence number used the same RG. This would
involve non-trivial changes to TCP implementations.
At best, an RG selection algorithm for TCP would require new logic in
implementations of TCP's opening handshake --- a significant
transition and 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.
5.5. On The Uniqueness Of ESDs
Although ESDs are expected to be globally unique, their uniqueness
property may be violated either due to mistakes in allocation or by
malicious attacks. The exact uniqueness requirements for ESDs
depends on what purpose they serve and how they are used. If the
correctness of some applications relies on the global uniqueness of
ESDs, then active checking and enforcement will be necessary. On the
other hand if ESDs are used only to uniquely identify individual
endpoints within a session, then one may consider global uniqueness
as unnecessary.
5.5.1. Impact of Duplicate ESDs
Consider what happens when two nodes using the same ESD attempt to |
communicate with each other. In the GSE proposal, a node queries the |
DNS to obtain an IPv6 address. The returned address includes the |
Routing Stuff of an address (the RG+STP portions). At this point, |
the sender might notice the destination ESD is the same as its own |
ESD and indicate an error. If it doesn't check, however, it may well |
forward the packet to a router that delivers the packet to its |
correct destination (using the information in the Routing Stuff). On |
receipt of the packet, again, the destination node could examine the |
ESD portion of the source address and determine that it is the same |
as its own and indicate an error. Alternatively, it could just |
process the packet without detecting the duplication and |
communication would proceed as normal (unless there are port number |
conflicts due to the sender and receiver allocating port numbers from |
draft-ietf-ipngwg-esd-analysis-05.txt [Page 34]
INTERNET-DRAFT October, 1999
the same name space).
A more problematic case occurs if two nodes having 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 all carry 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) Following the GSE specification, the transport end-point would
accept the packet, without regard to the Routing Stuff of the
source address. This may lead to a number of robustness
problems (and 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.
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).
5.5.2. 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 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 in the absence of
any ESD enforcement (i.e. ensuring each host use only the assigned
ESD), intruders will actively impersonate other sites for the sole
draft-ietf-ipngwg-esd-analysis-05.txt [Page 35]
INTERNET-DRAFT October, 1999
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 impersonating 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. |
Note that this attack is worse than the common syn-flood attack |
because it not only ties up resources on the target machine, it |
blocks out legitimate access to the target machine by a specific |
third party. |
Another potential attack involves an intruder assuming the ESD of a |
target site (e.g., mit.edu), then opening TCP connections using |
mit.edu's ESD to a targer server (e.g., big-server.com). Because the |
RG would point back to the attacker, the attacker could create a |
number of TCP connections in an OPEN state without needing to guess |
the sequence numbers needed to complete a 3-way handshake. Once those |
connections are open, it would be difficult to (automatically) |
distinguish between connections that are part of a denial-of-service |
attack from those (idle) connections that are part of a legitimate |
activity. |
The previous discussion indicates that separating identifiers and |
locators opens up new potential denial-of-service attack policies |
that would need to be carefully studied. One way of addressing them |
would be to have a way to authenticate the RG associated with an |
identifier, as the attacks take advantage of the distinction between |
identifiers and locators.
5.6. 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 value the RG portion of the |
address will have when it reaches the receiver. This issue is |
specific to GSE and other approaches in which the end node knows |
its own RG would not automatically have this problem. |
draft-ietf-ipngwg-esd-analysis-05.txt [Page 36]
INTERNET-DRAFT October, 1999
2) It can't be easily done in GSE using just the ESD because there |
is no mechanism at 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) It is conceivable that one could send a "who are you" type |
message to a peer asking it to return a more suitable identifier |
that can be used to jump start the authentication process. This |
additional information would include information needed to |
obtain keys, certificates, etc. from an appropriate source that |
can be used to verify proper use of an ESD by a particular node. |
Note, however, that the "who are you" makes use of the full |
address, not just the ESD portion.
4) Any scheme that uses the full IPv6 address to do the |
authentication can be used with today's 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. For all other cases, use
the first RG received with a given ESD for all sending.
3) Simultaneously, we understand that, with the above rules, there
are still open issues with regard to invalid RGs, either through
corruption or through a active hostile attacks.
One difficulty With the above recommendation is that 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 quandary. 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.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 37]
INTERNET-DRAFT October, 1999
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 authenticating |
the RG associated with an ESD. How to authentic the RG associated |
with an ESD in GSE does not appear to have a trivial solution is an |
open problem. |
5.7. The Need For Strong Authentication |
The problems described earlier stem from an inability to verify |
whether a particular RG is legitimately associated with an ESD. One |
approach that would address this problem is to use cryptographic |
techniques to verify the binding between RG and an ESD. There are two |
cases to consider. |
First, for an existing connection, switching from one RG to another |
risks the possibility of an intruder hijacking a connection. |
Addressing this risk involves having one endpoint verify |
(cryptographically) with its peer that proposed new RG is acceptable. |
This requires only an ability to communicate with the peer using the |
older (i.e., current) RG and using the older RG to verify the new RG. |
For example, a node could send its peer a message requesting |
cryptographic verification for a new RG prior to actually switching |
to it. Such verification would not require a public key |
infrastrucutre, as the purpose is not to verify that the legitimate |
owner of the ESD approves use of the RG, but that the peer with which |
one is currently communicating with (and who is using a particular |
ESD -- possibly illegally) approves switching to a different RG. |
A more problematic case involves the wedging of connections as |
described in Section 5.5.2. Here, an intruder improperly uses an |
identifier legitimately belonging to someone else, denying the |
legitimate owner service. Addressing this problem is more difficult. |
One approach is to verify the RG associated with an identifier the |
first time it is used. This would appear to require a global PKI |
infrastructure (not available today) in which every potential node is |
registered so that in the case of conflicts, it becomes possible to |
determine the legitimate owner of an identifier. |
Another interesting question concerns at what layer such |
cryptographic mechanisms would be needed. Ideally, the denial of |
service threats must be dealt with at the transport (or lower) layer |
because the threats are to the integrety of the transport layer |
itself. Attempting to solve them at higher-layers (e.g., via IPsec |
and IKE) results in a potential layering circularity, where the |
draft-ietf-ipngwg-esd-analysis-05.txt [Page 38]
INTERNET-DRAFT October, 1999
security mechanisms rely on a correctly functioning transport, but |
the transport relies on those same security mechanisms to provide a |
service. Further work is needed to determine whether such a mechanism |
can be designed using IPsec. |
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's CIDR-style addressing to
better understand the pros and cons 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 |
three 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 hierarchy 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 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.
Second, when communicating with a remote site, if the RG changes |
there begins to be uncertainty as to whether a reliable TCP handshake
is possible (because of the need for passively opened TCP to use the
RG's it obtains from the packets). Because the reliability of TCP's
byte stream is critically dependent on its three-way handshake, this
is a significant issue.
Finally, 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 39]
INTERNET-DRAFT October, 1999
to receive packets from a forged source, but the potential for
mischief between communicating peers is significantly limited because
return traffic will not generally 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 lose the built-in
protection available in classical IP and thus face great difficulty
assuring that traffic from a source identified only by an identifier
actually comes from the correct source. Short of using cryptographic
techniques (e.g. IPsec), there is no known mechanism that can use an
identifier alone to perform this remote entity authentication. 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. Indeed, the main conclusion of
this paper is that a GSE-like addressing structure introduces new
security vulnerabilities that are not present in IP, and that those
problems are serious enough to question the benefits of an
architecture that separates locaters and identifiers in addresses.
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 interim
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
attendees of the interim meeting whose high caliber discussions
helped motivate and shape this document.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 40]
INTERNET-DRAFT October, 1999
9. References
[ANYCAST] "Host Anycasting Service", C. Partridge, T. Mendez, & W.
Milliken, RFC 1546.
[BATES] Scalable support for multi-homed multi-provider
connectivity, Tony Bates & Yakov Rekhter, RFC 2260,
January, 1998.
[Bellovin 89] "Security Problems in the TCP/IP Protocol Suite",
Bellovin, Steve, Computer Communications Review, Vol. 19,
No. 2, pp32-48, April 1989.
[CIDR] "Classless Inter-Domain Routing (CIDR): an Address
Assignment and Aggregation Strategy". V. Fuller, T. Li, J.
Yu, & K. Varadhan, RFC 1519, September 1993.
[DHCP-DDNS] Interaction between DHCP and DNS, Internet Draft, Yakov
Rekhter, (Work in Progress.)
[DDNS] "Dynamic Updates in the Domain Name System (DNS UPDATE)",
Paul Vixie (Editor), RFC 2136, April, 1997.
[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, (Work in progress).
[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."
[IPv6-ADDRESS] "An IPv6 Aggregatable Global Unicast Address
Format", R. Hinden, M. O'Dell, S. Deering, RFC 2374, July,
1998.
[MOBILITY] "IP Mobility Support", C. Perkins, RFC 2002, October,
1996.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 41]
INTERNET-DRAFT October, 1999
[NAT] "IP Network Address Translator (NAT) Terminology and |
Considerations", P. Srisuresh, M. Holdrege, RFC 2663, |
August, 1999. |
[RFC1752] "The Recommendation for the IP Next Generation Protocol",
S. Bradner, A. Mankin, RFC 1752, January, 1995.
[RFC1788] "ICMP Domain Name Messages", W. Simpson, RFC 1788, April,
1995.
[RFC1884] "IP Version 6 Addressing Architecture", R. Hinden & S.
Deering, Editors, RFC 1884.
[RFC1958] "Architectural Principles of the Internet", B. Carpenter,
RFC 1958, June, 1996.
[RFC1971] "IPv6 Stateless Address Autoconfiguration", S. Thomson,
T. Narten, RFC 1971, August, 1996.
[RFC2008] "Implications of Various Address Allocation Policies for
Internet Routing", Y. Rekhter, T. Li, RFC 2008, October
1996.
[RFC2073] An IPv6 Provider-Based Unicast Address Format. Y.
Rekhter, P. Lothberg, R. Hinden, S. Deering, J. Postel. RFC
2073, January, 1997.
[RFC2267] Network Ingress Filtering: Defeating Denial of Service
Attacks which employ IP Source Address Spoofing, P.
Ferguson, D. Senie, RFC 2267, January, 1998. |
[RFC2401] Security Architecture for the Internet Protocol. S. Kent, |
R. Atkinson, RFC 2401, November 1998. |
[RFC2409] The Internet Key Exchange (IKE). D. Harkins, D. Carrel, |
RFC 2267 November 1998.
[ROUTER-RENUM] "Router Renumbering for IPv6", M. Crawford, draft-
ietf-ipngwg-router-renum-06.txt. |
[SITE-PREFIXES] "Site prefixes in Neighbor Discovery", E. Nordmark, |
draft-ietf-ipngwg-site-prefixes-03.txt. |
draft-ietf-ipngwg-esd-analysis-05.txt [Page 42]
INTERNET-DRAFT October, 1999
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
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: 703-812-3706
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
Appendix A: Increased Reliance on Domain Name System (DNS)
As we've discussed in previous sections, the motivation for
separating identifiers from locators in IP address is to allow the
locator portion to change more easily. However because GSE does not
provide a mapping from an ESD to its locator, whenever the locator
changes, GSE falls back on DNS to provide such mapping.
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.
A.1: Renumbering and DNS: 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 to achieve a high degree of route
aggregation. On closer examination, frequent (e.g., daily)
draft-ietf-ipngwg-esd-analysis-05.txt [Page 43]
INTERNET-DRAFT October, 1999
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
adequate overlap period after the RG changes, during which both the
old Routing Stuff and the new Routing Stuff can be used
simultaneously. During the overlap period, DNS glue records will
need to be updated to use the new addresses (including Routing Stuff)
and DNS RR's needs to be updated. Only after all relevant DNS
servers have been updated and all previously cached RRs containing
the old addresses have timed out can the old RG 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.
A.2: Efficient DNS support for Site Renumbering
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.
With GSE, 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. One DNS modification that
could reduce the cost of updating the DNS when a site is renumbered
is to store addresses in two distinct RR's: one for the Routing Goop
that reflects where a node attaches to the Internet and the other for
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. That way, renumbering a site would only
draft-ietf-ipngwg-esd-analysis-05.txt [Page 44]
INTERNET-DRAFT October, 1999
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
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.
A.2.1: 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
draft-ietf-ipngwg-esd-analysis-05.txt [Page 45]
INTERNET-DRAFT October, 1999
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).
A.2.2: 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.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 46]
INTERNET-DRAFT October, 1999
Appendix B: Additional Issues Related to Specifically to GSE |
This paper focused primarily on the issues of separating identifiers
and locators in unicast addresses. It is worth noting that a number |
of GSE-specific additional issues were identified during the IPng |
interim meeting. These stem from a GSE end node not knowing its own |
RG and the need for border routers to translate the RG of addresses. |
These issues would need to be considered before an architecture such |
as GSE could be deployed. Specifically:
- - it is not known how multicast would work under GSE. One
identified issue is that a site with multiple egress routers
would (by default) inject multicast traffic through each egress |
routers, each would then replace the source Routing Goop with a
differing value. This would lead to multiple copies of the same
packet each carrying a different IPv6 address, thus being
considered as from different sources.
- - It would be more difficult to create tunnels. Any tunnel that
crosses a site boundary (i.e., the entry and exit points are in
differing sites) would in effect require that both tunnel
endpoints be border routers to insure that the addresses in the
inner headers were rewritten correctly.
- - In order for the DNS to hide a site's Routing Goop from
internal nodes yet make it visible to external nodes requires a
two-faced DNS. The current DNS model assumes a single global
database in which all queries are answered the same way,
irregardless of who issued the query. It is unclear how to make
the DNS answer queries in a context-sensitive manner without |
also negatively impacting (i.e., crippling) its caching model. |
- - Applications that send addresses in payloads (e.g., FTP PORT |
command) may run into difficulties with GSE. Because the sender |
does not know its own RG, the addresses it sends in payloads |
will contain only the site-local prefix in the RG portion of the |
address. In order for the receiver to open a connection back to |
that address, it needs the proper RG. This problem is analagous |
to that of NATs, where addresses in payloads need to be |
rewritten (e.g., via an ALG) when crossing the boundary between |
different addressing realms [NAT]. |
- - Border routers need to rewrite the source address of outgoing |
packets. Additional parsing of packet headers is also required, |
to find and rewrite any other addresses containing the site- |
local prefix. For example, the source routing header may contain |
additional addresses.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 47]
INTERNET-DRAFT October, 1999
Appendix C: 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.
The unicast address format was changed to improve the aggregability
of unicast addresses. Instead of a topologically insignificant
Registry ID immediately following the Format Prefix [RFC2073], there
is now a Top-Level Aggregation Identifier [IPv6-ADDRESS]. This field
identifies a large routable aggregate to which an address belongs
rather than an administrative unit that assigned the address. 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) into IPv6 addresses
[IPv6-ADDRESS] also came from GSE. The previous "provider-based"
addressing architecture for IPv6 [RFC2073] 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 part indicating an interface on a specific link required
specifying an Interface Identifier that would be suitable 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.
Another change was the redefinition of the interface identifier to be
a 64-bit quantity. In the common case where a node has at least one
IEEE interface, the interface identifier is constructed from an IEEE
identifier (i.e., a MAC address) in such a way that there is a very
high probability that the identifier will be globally unique. In the
case where a globally unique identifier can't easily be constructed
automatically, a bit in the identifier indicates that the address is
not globally unique. At present, there are no plans for transport
protocols such as TCP to exploit interface identifiers, but the door
has been left open for a future protocol (e.g., TCPng) to take
advantage of the ESD concept.
Another change to come out of the GSE discussions relates to reducing
draft-ietf-ipngwg-esd-analysis-05.txt [Page 48]
INTERNET-DRAFT October, 1999
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
the ip6.int domain or some new mechanism which involves participation
by the routers or the end-nodes themselves.
Another example of follow-on work is site prefixes [SITE-PREFIXES], |
whose aim is to have communicating parties prefer site-local |
addresses for internal communication. Applications using site-local |
addresses are generally immune to renumbering issues that effect only |
global-scope addresses. |
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 [ROUTER-RENUM], and some inter-provider
cooperative method of providing multihoming to mutual customers with
minimal impact on routing tables in distant parts of the network.
Appendix D: Reverse Mapping of Complete GSE Addresses
The ability to map an IP address into its corresponding DNS name is
used in several contexts:
1) Network packet tracing utilities (e.g., tcpdump) display the
contents of packets. Printing out the DNS names appearing in
those packets (rather than dotted IP addresses) requires access
to an address-to-name mapping mechanism.
2) Some applications perform a "poor-man's" authentication by using
the DNS to map the source address of a peer into a DNS name.
The client then queries the DNS a second time, this time asking
for the address(es) corresponding to the peer's DNS name. Only
if one of the addresses returned by the DNS matches the peer
address of the TCP connection is the source of the TCP
connection accepted as being from the indicated DNS name.
It is important to note that although two DNS queries are made
during the above operation, it is the second one --- mapping the
draft-ietf-ipngwg-esd-analysis-05.txt [Page 49]
INTERNET-DRAFT October, 1999
peer's DNS name back into an IP address --- that provides the
authentication property. The first transaction simply obtains
the peer's DNS name, but no assumption is made that the returned
DNS name is correct. Thus, the first DNS query could be
replaced by an alternate mechanism without weakening the already
weak authentication check described above. One possible
alternate mechanism, an ICMP "Who Are You" message, is described |
below.
3) Applications that log all incoming network connections (e.g.,
anonymous FTP servers) may prefer logging recognizable DNS names
to addresses.
4) Network administrators examining logs or other trace data
containing addresses may wish to determine the DNS name of some
addresses. Note that this may occur sometime after those
addresses were actually used.
The following subsections 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.
D.1: DNS-Like Reverse Mapping of Full GSE Addresses
Although it seems infeasible to have a global scale, reverse mapping
of ESDs, within a site, it may be feasible to maintain a database
keyed on unstructured 8-byte ESDs. However, it is an open question
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, since the Routing Stuff
is organized hierarchically, if an ESD is always used in conjunction
with Routing Stuff (i.e., a full 16-byte address), it becomes
feasible to maintain a DNS-like tree that maps full GSE addresses
into DNS names, in a fashion analagous to what is done with IPv4 PTR |
draft-ietf-ipngwg-esd-analysis-05.txt [Page 50]
INTERNET-DRAFT October, 1999
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 hold valid indefinitely.
As a consequence, a packet trace recorded in the past might not
contain 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, together with
accurate timing information, 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 incidence in IPv4 where a block of returned
addresses was reassigned and reused within 24 hours of the
renumbering event.
D.2: The ICMP Who-Are-You Message
There is widespread agreement on the utility of being able to
determine the DNS name one is communicating with from the address
being used. In addition to the fact that DNS names are more
meaningful to human users and more stable than addresses, many users
use this reverse mapping as part of a poor-man's authentication for
the remote peer; if one can map the obtained DNS name back to the
same address, one has an increased confidence of the peer being a
legitimate one.
In practice, however, 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.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 51]
INTERNET-DRAFT October, 1999
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 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 time. 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.
draft-ietf-ipngwg-esd-analysis-05.txt [Page 52]
