[Search] [txt|pdf|bibtex] [Tracker] [WG] [Email] [Diff1] [Diff2] [Nits]

Versions: 00 01 02 03 04 05                                             
INTERNET-DRAFT                                             Matt Crawford
                                                                Fermilab
<draft-ietf-ipngwg-esd-analysis-02.txt>                   Allison Mankin
                                                                     ISI
                                                           Thomas Narten
                                                                     IBM
                                                    John W. Stewart, III
                                                                 Juniper
                                                             Lixia Zhang
                                                                    UCLA
                                                          March 13, 1998

             Separating Identifiers and Locators in Addresses:
                 An Analysis of the GSE Proposal for IPv6

                  <draft-ietf-ipngwg-esd-analysis-02.txt>


Status of this Memo

   This document is an Internet-Draft. Internet-Drafts are working
   documents of the Internet Engineering Task Force (IETF), its areas,
   and its working groups. Note that other groups may also distribute
   working documents as Internet-Drafts.

   Internet-Drafts are draft documents valid for a maximum of six months
   and may be updated, replaced, or obsoleted by other documents at any
   time. It is inappropriate to use Internet-Drafts as reference
   material or to cite them other than as "work in progress."

   To learn the current status of any Internet-Draft, please check the
   "1id-abstracts.txt" listing contained in the Internet-Drafts Shadow
   Directories on ds.internic.net (US East Coast), nic.nordu.net
   (Europe), ftp.isi.edu (US West Coast), or munnari.oz.au (Pacific
   Rim).

   Distribution of this memo is unlimited.

   This Internet-Draft expires May 7, 1998.


Abstract

   On February 27-28, 1997, the IPng Working Group held an interim
   meeting in Palo Alto, California to consider adopting Mike O'Dell's
   "GSE - An Alternate Addressing Architecture for IPv6" proposal [GSE].
   In GSE, 16-byte IPv6 addresses are split into distinct portions for
   global routing, local routing and end-point identification. GSE



draft-ietf-ipngwg-esd-analysis-02.txt                           [Page 1]


INTERNET-DRAFT                                            March 13, 1998


   includes the feature of configuring a node internal to a site with
   only the local routing and end-point identfication portions of the
   address, thus hiding the full address from the node. When such a node
   generates a packet, only the low-order bytes of the source address
   are specified; the high-order bytes of the address are filled in by a
   border router when the packet leaves the site.

   There is a long history of a vague assertion in certain circles that
   IPv4 "got it wrong" by treating its addresses simultaneously as
   locators and identifiers. Despite these claims, however, there was
   never a complete proposal for a scaleable network protocol which
   separated the functions. As a result, it wasn't possible to do a
   serious analysis comparing and contrasting a "separated" architecture
   and an "overloaded" architecture. The GSE proposal serves as a
   vehicle for just such an analysis, and that is the purpose of this
   paper.

   We conclude that an architecture that clearly separates locators and
   indentifiers in addresses introduces new issues and problems that do
   not have an easy or clear solution. Indeed, the alleged disadvantages
   of overloading addresses turn out to provide some significant
   benefits over the non-overloaded approach.

   Contents

   Status of this Memo..........................................    1

   1.  Introduction.............................................    3

   2.  Definitions and Terminology..............................    4

   3.  Addressing and Routing in IPv4...........................    5
      3.1.  The Need for Aggregation............................    7
      3.2.  The Pre-CIDR Internet...............................    7
      3.3.  CIDR and Provider-Based Addressing..................    8
      3.4.  Multi-Homing and Aggregation........................   12

   4.  The GSE Proposal.........................................   14
      4.1.  Motivation For GSE..................................   14
      4.2.  GSE Address Format..................................   15
         4.2.1.  Routing Stuff (RG and STP).....................   15
         4.2.2.  End-System Designator..........................   17
      4.3.  Address Rewriting by Border Routers.................   18
      4.4.  Renumbering and Rehoming Mid-Level ISPs.............   19
      4.5.  Support for Multi-Homed Sites.......................   20
      4.6.  Explicit Non-Goals for GSE..........................   21

   5.  Analysis: The Pros and Cons of Overloading Addresses.....   21



draft-ietf-ipngwg-esd-analysis-02.txt                           [Page 2]


INTERNET-DRAFT                                            March 13, 1998


      5.1.  Purpose of an Identifier............................   22
      5.2.  Mapping an Identifier to a Locator..................   24
         5.2.1.  Scalable Mapping of Identifers to Locators.....   25
         5.2.2.  Insufficient Hierarchy Space in ESDs...........   26
         5.2.3.  Reverse Mapping of Complete GSE Addresses......   27
         5.2.4.  DNS-Like Reverse Mapping of Full GSE Addresses.   27
         5.2.5.  The ICMP Who-Are-You Message...................   28
      5.3.  Authentication of Identifiers.......................   29
         5.3.1.  Identifier Authentication in IPv4..............   30
         5.3.2.  Identifier Authentication in GSE...............   31
         5.3.3.  Transport Layer: What Locator Should Be Used?..   31
         5.3.4.  RG Selection On An Active Open.................   32
         5.3.5.  RG Selection On An Passive Open................   32
         5.3.6.  Mid-Connection RG Changes......................   32
         5.3.7.  The Impact of Corrupt Routing Goop.............   33
         5.3.8.  On The Uniqueness Of ESDs......................   35
         5.3.9.  New Denial of Service Attacks..................   36
         5.3.10.  Summary of Identifier Authentication Issues...   36
      5.4.  Miscellaneous.......................................   38
         5.4.1.  Renumbering and Domain Name System (DNS) Issues   38
         5.4.2.  How Frequently Can We Renumber?................   38
         5.4.3.  Efficient DNS support for Site Renumbering.....   39
         5.4.4.  Two-Faced DNS..................................   40
         5.4.5.  Bootstrapping Issues...........................   41

   6.  Conclusion...............................................   41

   7.  Security Considerations..................................   42

   8.  Acknowledgments..........................................   42

   9.  References...............................................   43

   10.  Authors' Addresses......................................   44


1.  Introduction

   In October of 1996, Mike O'Dell published an Internet-Draft (dubbed
   "8+8") that proposed significant changes to the IPv6 addressing
   architecture. The 8+8 proposal was the topic of considerable
   discussion at the December 1996 IETF meeting in San Jose. Because the
   proposal offered both potential benefits (e.g., enhanced routing
   scalability) and risks (e.g., changes to the basic IPv6
   architecture), the IPng Working Group held an interim meeting on
   February 27-28, 1997 to consider adopting the 8+8 proposal.

   Shortly before the interim meeting, an updated version of the



draft-ietf-ipngwg-esd-analysis-02.txt                           [Page 3]


INTERNET-DRAFT                                            March 13, 1998


   Internet-Draft was produced. This version changed the name of the
   proposal from "8+8" to "GSE" to identify the three separate
   components of the address: Global, Site and End-System Designator.

   The well-attended meeting generated high caliber, focused technical
   discussions on the issues involved, with participation by almost all
   of the attendees. By the middle of the second day there was unanimous
   agreement that the GSE proposal as written presented too many risks
   and should not be adopted as the basis for IPv6. The proposal did,
   however, challenge the group to make improvements to the then
   existing IPv6 specifications (e.g., increasing the aggregatability of
   addresses, having hard boundaries in addresses between routing parts
   and non-routing parts and easing the DNS aspects of renumbering).

   This document focuses primarily on the issue of separating addresses
   into distinct portions for identification and location: a separation
   that GSE has but IPv4 does not.  We start with a discussion of the
   current architecture of IPv4 addressing and its impact on route
   scalability, identification, multi-homing, etc. Next, the details of
   the GSE proposal are described.  Finally, the fundamental issue of
   decomposing addresses into multiple separate functional parts is
   analyzed in the context of the GSE proposal. Here we detail some of
   the practical reasons why separating addresses into locators and
   identifier poses a number of challenging problems, making it clear
   that having such a separation is no panacea.  An appendix contains a
   summary of the IPng Working Group's deliberations of GSE and the
   results on IPv6 addressing.


2.  Definitions and Terminology

   The following terminology is used throughout this document.

      Routing Goop --- A term defined by the GSE document that refers to
                    first six bytes of an IPv6 GSE address. The Routing
                    Goop portion of an address identifies where a site
                    connects to the public Internet.  More generally,
                    the term refers to the portion of an address's
                    routing prefix that identifies where a site at which
                    an address resides connects to the public Internet.

      Site Topology Partition --- A term defined by the GSE document
                    that refers to the two bytes of an IPv6 GSE address
                    immediately to the right of the Routing Goop. The
                    Site Topology Partition part of an address
                    identifies which link within a site an address
                    resides on.




draft-ietf-ipngwg-esd-analysis-02.txt                           [Page 4]


INTERNET-DRAFT                                            March 13, 1998


      Routing Stuff --- The part of an address that identifies which
                    link the address resides on. Within the context of
                    GSE, the Routing Goop and Site Topology Partition
                    parts of an address comprise the Routing Stuff.

      identifier --- a value that indicates the sender of a packet, or
                    the intended recipient of a packet. Within the
                    context of GSE, the ESD portion of the address is an
                    identifier.

      locator --- a field in a packet header that is used by the routing
                    subsystem to deliver a packet to the link on which a
                    destination resides. The terms locator and Routing
                    Stuff are similar, we use Routing Stuff when
                    referring to the specific locator in GSE.


3.  Addressing and Routing in IPv4

   Before dealing with details of GSE, we present some background about
   how routing and addressing works in "classical IP" (i.e., IPv4). We
   present this background because the GSE proposal proposes a fairly
   major change to the base model. In order to properly evaluate GSE,
   one must understand what problems in IPv4 it alleges to improve or
   fix.

   The structure and semantics of a network layer protocol's addresses
   are absolutely core to that protocol. Addressing substantially
   impacts the way packets are routed, the ability of a protocol to
   scale and the kinds of functionality higher layer protocols can
   provide. Indeed, addressing is intertwined with both routing and
   transport layer issues; a change in any one of these can impact
   another. Issues of administration and operation (e.g., address
   allocation and required renumbering), while not part of the pure
   exercise of engineering a network layer protocol, turn out to be
   critical to the scalability of that protocol in a global and
   commercial network. The interaction between addressing, routing and
   especially aggregation is particularly relevant to this document, so
   some time will be spent describing it.

   Addresses in IPv4 serve two purposes:

     1) Unique identification of an interface. A sending host tells the
        network the identity of the intended recipient by placing an IP
        address into the destination address field.  In addition, the
        receiving host checks the destination address field of received
        packets to ensure that the packet is, in fact, for it.




draft-ietf-ipngwg-esd-analysis-02.txt                           [Page 5]


INTERNET-DRAFT                                            March 13, 1998


     2) Location information of that interface. Routers use the packet's
        destination address in deciding where to forward the packet to
        get it closer to its ultimate destination. That is, addresses
        identify "where" the intended recipient is located within the
        Internet topology.

        For scalability, the location information contained in addresses
        must be aggregatable. In practice, this means that nodes
        topologically close to each other (e.g., connected to the same
        link, residing at the same site, or customers of the same ISP)
        must use addresses that share a common prefix.

   What is important to note is that these identification and location
   requirements have been met through the use of the same value, namely
   the IP address. As will be noted repeatedly in this document, the
   "overloading" of IPv4 addresses with multiple semantics has some
   undesirable implications. For example, the embedding of IPv4
   addresses within transport protocol addresses that identify the end-
   point of a connection couples those transport protocols with routing.
   This entanglement is inconsistent with a strictly layered model in
   which routing would be a completely independent function of the
   network layer and not directly impact the transport layer.

   Combining locator and identifier functions also has the practical
   impact of complicating the support for mobility. In a mobile
   environment, the location of an end-station may change even though
   its identity stays the same; ideally, transport connections should be
   able to survive such changes. In IPv4, however, one cannot change the
   locator without also changing the identifier.

   Consequently, there has been a train of thought for some time has
   been that having separate values for location and identification
   could be of significant benefit. The GSE proposal, among other
   things, attempts to make such a separation.

   This document frequently uses mobility as an example to demonstrate
   the pros and cons of separating the identifier from the locator.
   However, the reader should note the fundamental equivalence between
   the problems faced by mobile hosts and the problem faced by sites
   that change providers yet don't want to renumber their network. When
   a site changes providers, it moves topologically in much the same way
   a mobile node does when it moves from one place to another.
   Consequently, techniques that help or hinder mobility are often
   relevant to the issue of site renumbering.







draft-ietf-ipngwg-esd-analysis-02.txt                           [Page 6]


INTERNET-DRAFT                                            March 13, 1998


3.1.  The Need for Aggregation

   IPv4 has seen a number of different addressing schemes. Since the
   original specification, the two major additions have been subnetting
   and classless routing. The motivation for adding subnetting was to
   allow a collection of networks located at one site to be viewed from
   afar as a single IP network (i.e., to aggregate all of the individual
   networks into one bigger network). The practical benefit of
   subnetting was that all of a site's hosts, even if scattered among
   tens or hundreds of LANs, could be represented with a single routing
   table entry in routers located far from the site. In contrast, prior
   to subnetting, a site with ten LANs would advertise ten separate
   network entries, and all routers would have to maintain ten separate
   entries, even though they contained essentially redundant
   information.

   The benefits of aggregation should be clear. The amount of work
   involved in constructing forwarding tables (i.e., selecting best
   routes and installing them into the switching subsystem) is dependent
   in part on the number of network routes (i.e., destinations) to which
   best paths are computed. If each site has 10 internal networks, and
   each of those networks is individually advertised to the global
   routing system, the complexity of computing forwarding tables can
   easily be an order of magnitude greater than if each site advertised
   a single entry that covered all of the addresses used within the
   site.


3.2.  The Pre-CIDR Internet

   In the early days of the Internet, its topology and addressing were
   orthogonal. Specifically, when a site wanted to connect to the
   Internet, it approached a centralized address allocation authority to
   obtain an address and then approached a provider about procuring
   connectivity. This procedure for address allocation resulted in a
   system where the addresses used by customers of the same provider
   bore little relation to the addresses used by other customers of that
   same provider. In other words, though the topology of the Internet
   was mostly hierarchical, the addressing was not. An example of such a
   topology and addressing scheme is shown in Figure 1.











draft-ietf-ipngwg-esd-analysis-02.txt                           [Page 7]


INTERNET-DRAFT                                            March 13, 1998



                +----------------+
                |                |------- Customer1 (192.2.2.0)
                |                |------- Customer2 (128.128.0.0)
                |   Provider A   |------- Customer3 (18.0.0.0)
                |                |------- Customer4 (193.3.3.0)
                |                |------- Customer5 (194.4.4.0)
                +----------------+
                        |
                        |
                        |
                        |
                +----------------+
                |   Provider B   |
                +----------------+

                                 Figure 1

   Figure 1 shows Provider A having 5 customers, each with their own
   independently obtained network address. Providers A and B connect to
   each other. In order for Provider B to be able to send traffic to
   Customers1-5, Provider A must announce a separate route to Provider B
   for each of the 5 networks.  That is, the routers within Provider B
   must have explicit routing entries for each of Provider A's customers
   -- 5 separate routes.

   Experience has shown that this approach scales very poorly. In the
   Default-Free Zone (DFZ) of the Public Internet, where routers must
   maintain routing entries for all reachable destinations, the cost of
   computing forwarding tables quickly becomes unacceptably large. A
   large part of the cost is related to the seemingly redundant
   computations that must be made for each individual network, even
   though the reality is that many reside in the same topological
   location (e.g., under the same provider). Looking at Figure 1, the
   problem is that provider B performs 5 separate calculations to
   construct the forwarding table needed to reach each of A's customers.
   Said another way, from Provider B's perspective, it doesn't matter
   where Provider A's customers connect to Provider A because Provider B
   is going to take the same path for all of them; in other words, there
   is an opportunity to do data abstraction.


3.3.  CIDR and Provider-Based Addressing

   One of the reasons CIDR (Classless Inter-Domain Routing) and its
   associated provider-assigned address allocation policy were
   introduced was to help reduce the size of a routing table and the
   complexity of computing a forwarding table from that routing table.



draft-ietf-ipngwg-esd-analysis-02.txt                           [Page 8]


INTERNET-DRAFT                                            March 13, 1998


   CIDR does this by aggressively aggregating network addresses.
   Aggregating network addresses means "merging" multiple addresses into
   a single "bigger" one. In CIDR, this means that addresses share a
   common prefix. The common prefix provides location information for
   all addresses sharing that same prefix.

   With CIDR, sites that want to connect to the Internet approach a
   provider to procure both connectivity and a network address.
   Individual providers have a block of address space covered by one
   prefix and assign pieces of that space to customers. Consequently,
   customers of the same provider have addresses that share the same
   prefix. Note that CIDR started to use the term "prefix" to refer to a
   classless network. The combination of CIDR and provider-based
   addressing results in the ability of a provider to address many
   hundreds of sites while introducing just one network address into the
   global routing system. An example of such a topology and addressing
   scheme is shown in Figure 2.


                +----------------+
                |                |------- Customer1 (204.1.0.0/19)
                |                |------- Customer2 (204.1.32.0/23)
                |   Provider A   |------- Customer3 (204.1.34.0/24)
                |                |------- Customer4 (204.1.35.0/24)
                |                |------- Customer5 (204.1.36.0/23)
                +----------------+
                        |
                        |  A announces
                        |  204.1/16 to B
                        |
                +----------------+
                |   Provider B   |
                +----------------+

                                  Figure 2

   In Figure 2, Provider A has been assigned the classless block, or
   "aggregate," 204.1.0.0/16 (i.e., a prefix with the high-order 16 bits
   denoting a single network). Provider A has 5 customers, each of which
   has been assigned a prefix subordinate to the aggregate.  In order
   for Provider B to be able to reach Customers1-5, Provider A only
   needs to announce the single prefix 204.1.0.0/16. The benefit for
   Provider B is that its routers need only a single routing table entry
   to reach all of Provider A's customers. Note the difference between
   the cases described in Figures 1 and 2. The important difference in
   the two Figures is that the latter example uses fewer entries in the
   routing table to reach the same number of destinations.




draft-ietf-ipngwg-esd-analysis-02.txt                           [Page 9]


INTERNET-DRAFT                                            March 13, 1998


   CIDR was a critical step for the Internet: in the early 1990s the
   size of default-free routing tables required to support the classful
   Internet was almost more than the commercially-available hardware and
   software of the day could handle.  The introduction of BGP4's
   classless routing and provider-based address allocation policies
   resulted in an immediate relief. At the same time, however, CIDR
   introduced some new weaknesses. First, the Internet addressing model
   had to shift from one of "address owning" to "address lending." In
   pre-CIDR days sites acquired addresses from a central authority
   independent of their provider, and a site could assume it "owned" the
   address it was given. Owning addresses meant that once one had been
   given a set of network addresses, one could always use them and
   assume that no matter where a site connected to the Internet, the
   prefix for that network could be injected into the public routing
   system. Today, however, it is simply no longer possible for each
   individual site to have its own private prefix injected into the DFZ;
   there would simply be too many of them. Consequently, if a site
   decides to change providers, then it needs to renumber all of its
   nodes using address space given to it by the new provider. The "old"
   addresses it had used are returned back to its previous provider.  To
   understand this, consider if, from Figure 2, Customer3 changes its
   provider from Provider A to Provider C, but does not renumber. The
   picture would be as follows:


                        +----------------+
                        |                |---- Customer1 (204.1.0.0/19)
                        |                |---- Customer2 (204.1.32.0/23)
                        |   Provider A   |
        +---------------|                |---- Customer4 (204.1.35.0/24)
        | A announces   |                |---- Customer5 (204.1.36.0/23)
        | 204.1/16 to B +----------------+
        |                     |
        |                     |
        |                     |
      +----------------+      |
      |   Provider B   |      |
      +----------------+      |
        |                     |
        |                     |
        |                     |
        | C announces         |
        | 204.1.34/24         |
        | to B          +----------------+
        +---------------|   Provider C   |---- Customer3 (204.1.34.0/24)
                        +----------------+

                                  Figure 3



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 10]


INTERNET-DRAFT                                            March 13, 1998


   In Figure 3, Providers A, B and C are all directly connected to each
   other. In order for Provider B to reach Customers 1, 2, 4 and 5,
   Provider A still only announces the 204.1.0.0/16 aggregate.  However,
   in order for Provider B to reach Customer 3, Provider C must announce
   the prefix 204.1.34.0/24. Prefix 204.1.34.0/24 is called a "more-
   specific" of 204.1.0.0/16; another term used is that Customer3 and
   Provider C have "punched a hole in" Provider A's block.  The result
   of this is that from Provider B's view, the address space underneath
   204.1.0.0/16 is no longer cleanly aggregated into a single prefix and
   instead the aggregation has been broken because the addressing is
   inconsistent with the topology; in order to maintain reachability to
   Customer3, Provider B must carry two prefixes where it used to have
   to carry only one.

   The example in Figure 3 explains why sites must renumber if existing
   levels of aggregation are to be maintained. While it is certainly
   clear that a small number of exceptions can be tolerated, the reality
   in today's Internet is that there are thousands of providers, many
   with thousands of individual customers. It is generally accepted that
   renumbering of sites is essential for maintaining sufficient
   aggregation.

   The empirical cost of renumbering a site in order to maintain
   aggregation has been the subject of much discussion. The practical
   reality, however, is that forcing all sites to renumber is difficult
   given the size and wealth of companies that now depend on the
   Internet for running their business. Thus, although the technical
   community came to consensus that address lending was necessary in
   order for the Internet to continue to operate and grow, the reality
   has been that some of CIDR's benefits have been lost because not all
   sites renumber.  It is worth noting that a number of providers do
   route filtering based, in part, on prefix length; as a result, a site
   which does not renumber may have, at best, partial connectivity to
   the Internet.

   One unfortunate characteristic of CIDR at an architectural level is
   that the pieces of the infrastructure that benefit from the
   aggregation (i.e., the providers which make up the DFZ) are not the
   pieces that incur the cost (i.e., the end site). The logical
   corollary of this statement is that the pieces of the infrastructure
   that do incur cost to achieve aggregation (e.g., sites which renumber
   when they change providers) don't directly see the benefit. (The word
   "directly" is used here because the continued operation of the
   Internet is a benefit, though it requires selflessness on the part of
   the site to recognize.)






draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 11]


INTERNET-DRAFT                                            March 13, 1998


3.4.  Multi-Homing and Aggregation

   As sites become more dependent on the Internet, they have begun to
   install additional connections to the Internet to improve robustness
   and performance. Such sites are called "multi-homed."  Unfortunately,
   when a site connects to the Internet at multiple places, the impact
   on routing can be much like a site that switches providers but
   refuses to renumber.

   In the pre-CIDR days, multi-homed sites were typically known by only
   one network prefix. When that site's providers announced the site's
   network into the global routing system, a "shortest path" type of
   routing would occur so that pieces of the Internet closest to the
   first provider would use the first provider while other pieces of the
   Internet would use the second provider. This allowed sites to use the
   routing system itself to load balance traffic across their multiple
   connections. This type of multi-homing assumes that a site's prefix
   can be propagated throughout the DFZ, an assumption that is no longer
   universally true.

   With CIDR, issues of addressing and aggregation complicate matters
   significantly.  At the highest levels, there are three possible ways
   to deal with multi-homed sites.  The first approach is for multi-
   homed sites to receive address space directly from a registry,
   independent of its providers.  The problem with this approach is
   that, because the address space is obtained independent of either
   provider, it is not aggregatable and therefore has a negative impact
   on the scaling of global routing.

   The second approach is for a multi-homed site to receive an
   allocation from one of its providers and just use that single prefix.
   The site would advertise its prefix to all of the providers to which
   it connects.  There are two problems with this is approach. First,
   although the prefix is aggregatable by the provider which made the
   allocation, it is not aggregatable by the other providers. To the
   other providers, the site's prefix poses the same problem that a
   provider-independent address would.  This has a negative impact on
   the scaling of global routing.  Second, due to CIDR's rule for
   longest-match routing, it turns out that the site's prefix is not
   always aggregatable in practice even by the provider that made the
   allocation. Consider Figure 4. Provider C has two paths for reaching
   Customer 1. Provider A advertises 204.1/16, an aggregate which
   includes Customer 1. But Provider C will also receive an
   advertisement for prefix 204.1.0/19 from Provider B, and because the
   prefix match through B is longer, C will choose that path. In order
   for Provider C to be able to choose between the two paths, Provider A
   would also have to advertise the longer prefix for 204.1.0/19 in
   addition to the shorter 204.1/16.  At this point, from the routing



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 12]


INTERNET-DRAFT                                            March 13, 1998


   perspective, the situation is very similar to the general problem
   posed by the use of provider-independent addresses.

   It should be noted that the above example simplifies a very complex
   issue. For example, consider the example in Figure 4 again. Provider
   A could choose not to propagate a route entry for the longer
   204.1.0/19 prefix, advertising only the shorter 204.1/16. In such
   cases, provider C would always select Provider B. Internally,
   Provider A would continue to route traffic from its other customers
   to Customer 1 directly. If Provider A had a large enough customer
   base, effective load sharing might be achieved.


                                      A advertises
                     +------------+  204.1/16 to C  +------------+
                  ___| Provider A |-----------------| Provider C |
                 /   +------------+                 +------------+
                /                       +----------/
               /                       /
   Customer 1 ---                     / B advertises 204.1.0/19 to C
   204.1.0.0/19  |                   /
                 |      +------------+
                  ----- | Provider B |
                        +------------+


                                Figure 4


   The third approach is for a multi-homed site to receive an allocation
   from each of its providers.  This approach has advantages from the
   perspective of route scaling because both allocations are
   aggregatable. Unfortunately, the approach doesn't necessarily meet
   the demands of the multi-homed site.  A site that has a prefix from
   each of its providers has a number of choices about how to use that
   address space. Possibilities include:

      1) The site can number a distinct set of hosts out of each of the
        prefixes.  Consider a configuration where a site is connected to
        ISP-A and ISP-B. If the link to ISP-A goes down, then unless the
        ISP-A prefix is announced to ISP-B (which breaks aggregation),
        the hosts numbered out of the ISP-A prefix would be unreachable.

      2) The site could assign each host multiple addresses (i.e., one
        address for each ISP connection).  There are two problems with
        this.  First, it accelerates the consumption of the address
        space.  Second, when the connection to ISP-A goes down,
        addresses numbered out of ISP-A's space become unreachable.



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 13]


INTERNET-DRAFT                                            March 13, 1998


        Remote peers would have to have sufficient intelligence to use
        the second address. For example, when initiating a connection to
        a host, the DNS would return multiple candidate addresses.
        Clients would need to try them all before concluding that a
        destination is unreachable (something not all hosts currently
        do). In addition, a site's hosts would need a significant amount
        of intelligence for choosing the source addresses they use. A
        host shouldn't choose a source address corresponding to a link
        that is down. At present, hosts do not have such sophistication.

   In summary, how best to achieve multi-homing with IPv4 in the face of
   CIDR is an unsolved problem.  There is a delicate balance between the
   scalability of routing versus the site's requirements of robustness
   and load-sharing.  At this point in time, no solution has been
   discovered that satisfies the competing requirements of route scaling
   and robustness/performance.  It is worth noting, however, that some
   people are beginning to study the issue more closely and propose
   novel ideas [BATES].


4.  The GSE Proposal

   This section provides a description of GSE with the intent of making
   this document stand-alone with respect to the GSE "specification." We
   begin by reviewing the motivation for GSE. Next we review the salient
   technical details, and we conclude by listing the explicit non-goals
   of the GSE proposal.


4.1.  Motivation For GSE

   The primary motivation for GSE was the fact that the chief initial
   IPv6 global unicast address structure, provider-based [RFC 2073], was
   fundamentally the same as IPv4 with CIDR and provider-based
   aggregation. Provider-based addressing requires that sites renumber
   when they switch providers, so that sites are always aggregated
   within their provider's prefix. In practice, the cost of renumbering
   (which can only grow as a site grows in size and becomes more
   dependent on the Internet for day-to-day business) is high enough
   that an increasing number of sites refuse to renumber.  This cost is
   particularly relevant in cases where end-users are asked to renumber
   because an upstream provider has changed its transit provider (i.e.,
   the end site is asked to renumber for reasons outside of its control
   and for which it sees no direct benefit).  Consequently, the GSE
   draft asserted that IPv4 with CIDR has not achieved the aggressive
   aggregation required for the route computation functions of the DFZ
   of the Internet to scale for IPv4 and that the larger addresses of
   IPv6 simply exacerbate the problem.



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 14]


INTERNET-DRAFT                                            March 13, 1998


   The GSE proposal did not propose to eliminate the need for
   renumbering. Indeed, it asserted that end sites will have to be
   renumbered more frequently in order to continue scaling the Internet.
   However, GSE proposed to make the cost of such a renumbering so small
   that sites could be renumbered at essentially any time with little or
   no disruption.

   Finally, GSE dealt significantly with sites that have multiple
   Internet connections. In some addressing schemes (e.g., CIDR), this
   "multi-homing" can create exceptions to the aggregation and result in
   poor scaling. That is, the public routing infrastructure needs to
   carry multiple distinct routes for the multi-homed site, one for each
   independent path. GSE recognized the "special work done by the global
   Internet infrastructure on behalf of multi-homed sites," [GSE] and
   proposed a way for multi-homed sites to gain some benefit without
   impacting global scaling. This included a specific mechanism that
   providers could use to support multi-homed sites, presumably at a
   cost that the site would consider when deciding whether or not to
   become multi-homed.


4.2.  GSE Address Format

   The key departure of GSE from classical IP addressing (both v4 and
   v6) was that rather than over-loading addresses with both locator and
   identifier purposes, it split the address into two elements: the
   high-order 8 bytes for routing (called "Routing Stuff" throughout the
   rest of this document) and the low-order 8 bytes for unique
   identification of an end-point. The structure of GSE addresses was:

                0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15
              +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
              |  Routing Goop    | STP| End System Designator |
              +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
                     6+ bytes   ~2 bytes       8 bytes

                                 Figure 5



4.2.1.  Routing Stuff (RG and STP)

   The Routing Goop (RG) identifies the place in the Internet topology
   where a site connects and is used to route datagrams to the site. RG
   is structured as follows:






draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 15]


INTERNET-DRAFT                                            March 13, 1998



                           1                   2                   3
       0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      | xxx | 13 Bits of LSID         |      Upper 16 bits of Goop    |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

       3               4
       2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      | Bottom 18 bits of Routing Goop    |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

                                 Figure 6

   The RG describes the location of a site's connection by identifying
   smaller and smaller regions of topology until finally it identifies
   the link which connects the site. Before interpreting the bits in the
   RG, it is important to understand that routing with GSE depends on
   decomposing the Internet's topology into a specific graph. At the
   highest level, the topology is broken into Large Structures (LSs). An
   LS is basically a region that can aggregate significant amounts of
   topology. Examples of potential LSs are large providers and exchange
   points. Within an LS the topology is further divided into another
   graph of structures, with each LS dividing itself however it sees
   fit. This division of the topology into smaller and smaller
   structures can recurse for a number of levels, where the trade-off is
   "between the flat-routing complexity within a region and minimizing
   total depth of the substructure." [ESD]

   Having described the decomposition process, we can now examine the
   bits in the RG. After the 3-bit prefix identifying the address as
   GSE, the next 13 bits identify the LS. By limiting the field to 13
   bits, a ceiling is defined on the complexity of the top-most routing
   level (i.e., what we currently call the DFZ). In the next 34 bits, a
   series of subordinate structure(s) are identified until finally the
   leaf subordinate structure is identified, at which point the
   remaining bits identify the individual link within that leaf
   structure.

   The remaining 14 bits of the Routing Stuff (i.e., the low-order 14
   bits of the high-order 8 bytes) comprise the STP and are used for
   routing structure within a site, similar to subnetting with IPv4.
   These bits are not part of the Routing Goop per se. The distinction
   between Routing Stuff and Routing Goop is that RG controls routing in
   the Public Internet, while Routing Stuff includes the RG plus the
   Site Topology Partition (STP). The STP is used for routing structure
   within a site.  [Note that the term "Routing Stuff" was a creation of



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 16]


INTERNET-DRAFT                                            March 13, 1998


   the author's of this analysis document and was not used in the GSE
   document.]

   The GSE proposal formalized the ideas of sites and of public versus
   private topology. In the first case, a site is a set of hosts,
   routers and media under the same administrative control which have
   zero or more connections to the Internet.  A site can have an
   arbitrarily complicated topology, but all of that complexity is
   hidden from everyone outside of the site.  A site only carries
   packets which originated from, or are destined to, that site; in
   other words, a site cannot be a transit network. A site is private
   topology, while the transit networks form the public topology.

   A datagram is routed through public topology using just the RG, but
   within the destination site, routing is based on the Site Topology
   Partition (STP).


4.2.2.  End-System Designator

   The End-System Designator (ESD) is an unstructured 8-byte field that
   uniquely identifies an interface from all others.  The most important
   feature of the ESD is that it alone identifies an interface; the
   Routing Stuff portion of an address, although used to help deliver a
   packet to its destination, is not used to actually identify an end
   point.  End-points of communication care about the ESD; as examples,
   TCP peers could be identified by the source and destination ESDs
   alone (together with port numbers), checksums would exclude the RG
   (the sender doesn't know its RG, as described later) and on receipt
   of a datagram only the ESD would be used in testing whether a packet
   is intended for local delivery.

   The leading contender for the role of a 64-bit globally unique ESD is
   the recently defined "EUI-64" identifier. [EUI64]  These identifiers
   consist of a 24-bit "company_id" concatenated with a 40-bit
   "extension."  (Company_id is just a new name for the
   "Organizationally Unique Identifier" that forms the first half of an
   802 MAC address.) Manufacturers are expected to assign locally unique
   values to the extension field, guaranteeing global uniqueness for the
   complete 64-bit identifier.

   A range of the EUI-64 space is reserved to cover pre-existing 48-bit
   MAC addresses, and a defined mapping insures that an ESD derived from
   a MAC address will not duplicate the ESD of a device that has a
   built-in EUI-64.

   In some cases, interfaces may not have access to an appropriate MAC
   address or EUI-64 identifier. A globally unique ESD must then be



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 17]


INTERNET-DRAFT                                            March 13, 1998


   obtained through some alternate mechanism. Several possible
   mechanisms can be imagined (e.g., the IANA could hand out addresses
   from the company_id it has been allocated), but we do not explore
   them in detail here.



4.3.  Address Rewriting by Border Routers

   GSE site border routers rewrite addresses of the packets they forward
   across the boundary between the site and public topology. Within a
   site, nodes need not know the RG associated with their addresses.
   They simply use a designated "Site-Local RG" value for internal
   addresses. When a packet is forwarded to the public topology, the
   border router replaces the Site-Local RG portion of the packet's
   source address with an appropriate value. Likewise, when a packet
   from the public topology is forwarded into a site, the border router
   replaces the RG part of the destination address with the designated
   Site-Local RG.

   To simplify discussion, the following text uses the singular term RG
   as if a site could have only one RG value (i.e., one connection to
   the Internet). In fact, a site could have multiple Internet
   connections and consequently multiple RGs.

   Having border routers rewrite addresses obviates the need to renumber
   devices within sites because of changing providers --- GSE's approach
   wasn't so much to ease renumbering as to make it transparent. To
   achieve transparency, the RG by which a site is known is hidden
   (i.e., kept secret) from nodes within that site. Instead, the RG for
   the site would be known only by the exit router, either through
   static configuration or through a dynamic protocol with an upstream
   provider.

   Because end hosts don't know their RG, they don't know their entire
   16-byte address, so they can't specify the full address in the source
   fields of packets they originate. Consequently, when a datagram
   leaves a site, the egress border router fills in the high-order
   portion of the source address with the appropriate RG.

   The point of keeping the RG hidden from nodes within the core of a
   site was to insure the changeability of the RG without impacting the
   site itself. It was expected that the RG would need to change
   relatively frequently (e.g., several times a year) in order to
   support scalable aggregation as the topology of the Internet changes.
   A change to a site's RG would only require a change at the site's
   egress point, and it's well possible that this change could be
   accomplished through a dynamic protocol with the upstream provider.



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 18]


INTERNET-DRAFT                                            March 13, 1998


   Hiding a site's RG from its internal nodes does not, however, mean
   that changes to RG have no impact on end sites. Since the full 16-
   byte address of a node isn't a stable value (the RG portion can
   change), a stored address may contain invalid RG and be unusable if
   it isn't "refreshed" through some other means. For example, opening a
   TCP connection, writing the address of the peer to a file and then
   later trying to reestablish a connection to that peer is likely to
   fail.  For intra-site communication, however, it is expected that
   only the Site-Local RG would be used (and stored) which would
   continue to work for intra-site communication regardless of changes
   to the site's external RG. This has the benefit of shielding a site's
   intra-site traffic from any instabilities resulting from renumbering.

   In addition to rewriting source addresses upon leaving a site,
   destination addresses are rewritten upon entering a site. To
   understand the motivation behind this, consider a site with
   connections to three Internet providers. Because each of those
   connections has its own RG, each destination within the site would be
   known by three different 16-byte addresses. As a result, intra-site
   routers would have to carry a routing table three times larger than
   expected. To work around this, GSE proposed replacing the RG in
   inbound packets with the special "Site-Local RG" value to reduce
   intra-site routing tables to the minimum necessary.

   In summary, when a node initiates a flow to a node at another site,
   the initiating node knows the full 16-byte address for the
   destination through some mechanism like a DNS query. The initiating
   node does not, however, know its RG, so uses the Site-Local RG values
   in the RG part of the source address. When the datagram reaches the
   exit border router, the router replaces the RG of the packet's source
   address.  When the datagram arrives at the entry router at the
   destination site, the router replaces the RG portion of the
   destination address with the distinguished "Site-Local RG" value.
   When the destination host needs to send return traffic, that host
   knows the full 16-byte address for the other host because it appeared
   in the source address field of the arriving packet.


4.4.  Renumbering and Rehoming Mid-Level ISPs

   One of the most difficult-to-solve components of the renumbering
   problem with CIDR is that of renumbering mid-level service providers.
   Specifically, if SmallISP1 changes its transit provider from BigISP1
   to BigISP2, then in order for the overall size of the routing tables
   to stay the same, all of SmallISP1's customers would have to renumber
   into address space covered by an aggregate of BigISP2.  GSE dealt
   with this problem by handling the RG in DNS with indirection.
   Specifically, a site's DNS server specifies the RG portion of its



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 19]


INTERNET-DRAFT                                            March 13, 1998


   addresses by referencing the "name" of its immediate provider, which
   is a resolvable DNS name (this implies a new Resource Record type).
   That provider may define some of the low-order bits of the RG and
   then reference its immediate provider. This chain of reference allows
   mid-level service providers to change transit providers, and the
   customers of that mid-level will simply "inherit" the change in RG.



4.5.  Support for Multi-Homed Sites

   GSE defined a specific mechanism for providers to use to support
   multi-homed customers that gives those customers more reliability
   than singly-homed sites, but without a negative impact on the scaling
   of global routing. This mechanism is not specific to GSE and could be
   applied to any multi-homing scenario where a site is known by
   multiple prefixes (including provider-based addressing). Assume the
   following topology:

                             Provider1     Provider2
                             +------+       +------+
                             |      |       |      |
                             | PBR1 |       | PBR2 |
                             +----x-+       +-x----+
                                  |           |
                              RG1 |           | RG2
                                  |           |
                               +--x-----------x--+
                               | SBR1       SBR2 |
                               |                 |
                               +-----------------+
                                      Site

                                    Figure 7

   PBR1 is Provider1's border router while PBR2 is Provider2's border
   router.  SBR1 is the site's border router that connects to Provider1
   while SBR2 is the site's border router that connects to Provider2.
   Imagine, for example, that the line between Provider1 and the site
   goes down. Any already existing flows that use a destination address
   including RG1 would stop working. In addition, any DNS queries that
   return addresses including RG1 would not be viable addresses. If PBR1
   and PBR2 knew about each other, however, then in this case PBR1 could
   tunnel packets destined for RG1-prefixed addresses to PBR2, thus
   keeping the communication working.  (Note that true tunneling, i.e.,
   re-encapsulation, is necessary since routers between PBR1 and PBR2
   would forward RG1 addresses towards PBR1.)




draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 20]


INTERNET-DRAFT                                            March 13, 1998


4.6.  Explicit Non-Goals for GSE

   It is worth noting explicitly that GSE did not attempt to address the
   following issues:

     1) Survival of TCP connections through renumbering events. If a
        site is renumbered, TCP connections using a previous address
        will continue to work only as long as the previous address still
        works (i.e., while it is still "valid" using RFC 1971
        terminology). No attempt is made to have existing connections
        switch to the new address.

     2) It is not known how mobility can be made to work under GSE.

     3) It is not known how multicast can be made to work under GSE.

     4) The performance impact of having routers rewrite portions of the
        source and destination address in packet headers requires
        further study.

   That GSE didn't address the above does not mean they cannot be
   solved. Rather the issues weren't studied in sufficient depth.


5.  Analysis: The Pros and Cons of Overloading Addresses

   At this point we have given complete descriptions of two addressing
   architectures:  IPv4, which uses the overloading technique, and GSE,
   which uses the separated technique.  We now compare and contrast the
   two techniques.

   The following discussion is organized around three fundamental
   points:

     1) Identifiers indicate who the intended recipient of a packet is,

     2) Identifiers must be mapped into a locator that the network layer
        uses to actually deliver a packet to its intended destination,
        and

     3) There must be a suitable way to sufficiently authenticate the
        user of an identifier, so that peers using identifiers have
        sufficient confidence that packets sent to or received from a
        particular identifier correspond to the intended recipient.







draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 21]


INTERNET-DRAFT                                            March 13, 1998


5.1.  Purpose of an Identifier

   An identifier gives an entity the ability to refer to a communication
   end point and to refer to the same endpoint over an extended period
   of time.  In terms of semantics, two or more packets sent to the same
   identifier should be delivered to the same end point. Likewise, one
   expects multiple packets received from the same identifier to have
   been originated by the same sending entity.  That is, a source
   identifier indicates who the packet is from and a destination
   identifier indicates who the packet is intended for.

   When applications communicate, "identifiers" consist of addresses and
   port numbers. For the purposes of this discussion, the term
   "identifier" means the identifier of an interface. It is assumed that
   port numbers will be present when higher layer entities communicate;
   the exact port numbers used are not relevant to this discussion.

   In small networks, flat routing can be used to deliver packets to
   their destination based only on the destination identifier carried in
   a packet header (i.e., the identifier is the locator and is not
   required to have any structure). However, in such systems, a distinct
   route entry is required for every destination, an approach that does
   not scale. In larger networks, packet addresses include a locator
   that helps the network layer deliver a packet to its destination.
   Such a locator typically has structure (i.e., is an aggregate for
   many destinations) that keeps routing tables small relative to the
   total number of reachable destinations.  In IPv4, the identifier and
   locator are combined in a single address; it is not possible to
   separate the locator portion of an address from the identifier
   portion.  In contrast, the ESD portion of a GSE address (which can
   easily be extracted from the address) serves as an identifier, while
   the Routing Stuff plays the role of a locator.

   Having a clear separation between the locator and the identifer
   portion of an address appears to give protocols some additional
   flexibility. Once a packet has been delivered to its intended
   destination interface (i.e., node), for example, the locator has
   served its purpose and is no longer needed to further demultiplex a
   packet to its higher-layer end point.  This means that if a packet is
   delivered to the correct destination node, the node will accept the
   packet, regardless of how the packet got there. The exact locator
   used does not matter, so long as it corresponds to one that delivers
   a packet to its proper destination.

   The most obvious example that could benefit from the separation of
   locators and indentifiers involves communication with a mobile host.
   Transport protocols such as TCP are unable to keep connections open
   if either of the endpoint identifiers for an open connection changes.



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 22]


INTERNET-DRAFT                                            March 13, 1998


   Fundamentally, the endpoint identifiers indicate the two endpoint
   entities that are communicating. If a node were to receive a packet
   from a node with which it had been communicating previously, but the
   identifier used by the sending node has changed, the recipient would
   be unable to distinguish this case from that of a packet received
   from a completely different node.

   In the specific case of TCP and IPv4, connections are identified
   uniquely by the tuple: (srcIPaddr, dstIPaddr, srcport, dstport).
   Because IPv4 addresses contain a combined locator/identifier, it is
   not possible to have a node's location change without also having its
   identifier change. Consequently, when a mobile node moves, its
   existing connections no longer work, in the absence of special
   protocols such as Mobile IP [RFC2002].

   In contrast, connections in GSE are identified by the ESDs rather
   than full IPv6 addresses. That is, connections are identified
   uniquely by the tuple: (srcESD, dstESD, srcport, dstport).
   Consequently, when demultiplexing incoming packets to their proper
   end point, TCP would ignore the Routing Stuff portions of addresses.
   Because the Routing Stuff portion of an address is ignored during
   demultiplexing operations, a mobile node is free to move -- and
   change its Routing Stuff -- without consequences for the
   demultiplexing operation.

   As a side note, it is a requirement in GSE that packets be
   demultiplexed on ESDs alone independent of the Routing Stuff. If a
   site is multi-homed, the packets it sends may exit the site at
   different egress border routers during the lifetime of a connection.
   Because each border router will place its own RG into the source
   addresses of outgoing packets, the receiving TCP must ignore (at
   least) the RG portion of addresses when demultiplexing received
   packets. The alternative would be to make TCP unable to cope with
   common routing changes, i.e., if the path changed, packets delivered
   correctly would be discarded by the receiving TCP rather than
   processed.

   Not surprisingly, having separate locator and identifiers in
   addresses leads to some additional problems.  First, an identifier by
   itself provides only limited value. In order to actually deliver
   packets to a destination identifier, a corresponding locator must be
   known. The general problem of mapping identifiers into locators is
   non-trivial to solve, and is the topic of the next Section. Second,
   because the Routing Stuff is ignored when demultiplexing packets
   upward in the protocol stack, it becomes much easier for an intruder
   to masquerade as someone else.





draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 23]


INTERNET-DRAFT                                            March 13, 1998


5.2.  Mapping an Identifier to a Locator

   The idea of using addresses that cleanly separate location and
   identification information is not new [references XXX]. However,
   there are several different flavors. In its pure form, a sender need
   only know the identifier of an end-point in order to send packets to
   it. When presented with a datagram to send, network software would be
   responsible for finding the locator associated with an identifier so
   that the packet can be delivered. A key question is: "who is
   responsible for finding the Routing Stuff associated with a given
   identifier"? There are a number of possibilities, each with a
   different set of implications:

     1) The network layer could be responsible for doing the mapping.
        The advantage of such a system is that an ESD could be stored
        essentially forever (e.g., in configuration files), but whenever
        it is actually used, network layer software would automatically
        perform the mapping to determine the appropriate Routing Stuff
        for the destination. Likewise, should an existing mapping become
        invalid, network layer software could dynamically determine the
        updated value. Unfortunately, building such a mapping mechanism
        that scales is a hard problem.

     2) The transport layer could be responsible for doing the mapping.
        It could perform the mapping when a connection is first opened,
        periodically refreshing the binding for long-running
        connections. Implementing such a scheme would change the
        existing transport layer protocols TCP and UDP significantly.

     3) Higher-layer software (e.g., the application itself) could be
        responsible for performing the mapping. This potentially
        increases the burden on application programmers significantly,
        especially if long-running connections are required to survive
        renumbering and/or deal with mobile nodes.

   It should be noted that the GSE proposal does not embrace the general
   model, it uses the last. The network and transport layers are always
   presented with both the Routing Stuff (RG + STP) and the ESD together
   in one IPv6 address. It is neither of these layers' jobs to determine
   the Routing Stuff given only the ESD or to validate that the Routing
   Stuff is correct.  When an application has data to send, it queries
   the DNS to obtain the IPv6 AAAA record for a destination. The
   returned AAAA record contains both the Routing Stuff and the ESD of
   the specified destination. While such an approach eliminates the need
   for the lower layers to be able to map ESDs into corresponding
   Routing Stuff, it also means that when presented with an address
   containing an incorrect (i.e., no longer valid) Routing Stuff, the
   network is unable to deliver the packet to its correct destination.



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 24]


INTERNET-DRAFT                                            March 13, 1998


   It is up to applications themselves to deal with such failures. Note
   that addresses containing invalid Routing Stuff will result any time
   cached addresses are used after the Routing Stuff of the address
   becomes invalid. This may happen if addresses are stored in
   configuration files, a mobile node moves to a new location, long-
   running applications (clients and servers) cache the result of DNS
   queries, a long-running connection attempts to continue operating
   during a site renumbering event, etc.

   A network architecture must provide the ability to map an identifier
   to a locator. In IPv4, this mapping is trivial (the identity
   function), since the identifier and locator are combined in a single
   quantity (i.e., the IPv4 address). GSE does not provide mapping
   functionality directly. Indeed, GSE uses two different identifiers.
   At the highest level, a node's DNS name serves as its identifer, with
   normal DNS queries used to map the DNS "identifier" into a locator
   (i.e., the first 8 bytes of the IPv6 address). At a lower layer, the
   IPv6 address contains the ESD identifier together with its Routing
   Stuff (i.e., locator). Note that the DNS name is expected to be the
   stable identifier that can be mapped into an appropriate locator at
   any time. In contrast, the ESD identifier, cannot be mapped into a
   locator by itself.

   The use of two identifiers contributes to making GSE appear simple.
   However, there are two fundamental problems with this approach, if
   the intention is to make it transparently easy to change locators
   over time. First, the burden of performing the mapping from
   identifier to locator is placed directly on the application,
   requiring active participation from the application. Second, The
   lower layers (i.e., transport and network layers) cannot make use of
   this mapping themselves due to layering violation concerns (i.e., TCP
   and UDP can't depend on the DNS to perform a query).

   The following subsections discuss a number of issues related to
   keeping track of or determining the locator associated with an
   identifier.


5.2.1.  Scalable Mapping of Identifers to Locators

   It is not difficult to construct a mapping from an identifier (such
   as an ESD) to a locator (as well as other information such as a name,
   cryptographic keys, etc.) provided one can structure the identifier
   appropriately to support such lookups. In particular, identifiers
   must have sufficient structure to support the delegating mechanism of
   a distributed database such as DNS. On the other hand, no scalable
   mechanism is known for performing such a mapping on arbitrary
   identifiers taken from a flat space lacking structure.



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 25]


INTERNET-DRAFT                                            March 13, 1998


   Imposing a heirarchy on identifiers poses the following difficulties:

      - it increases the size of the identifier. The exact size
        necessary to support sufficient heirarchy is unclear, though it
        is likely to be roughly the same as that used for the routing
        hierarchy. Analysis done during the original IPng debates
        [RFC1752] suggests that close to 48-bits of hierarchy are needed
        to identify all the possible sites 30-40 years from now.

      - the assignment of identifiers must be tied to the delegation
        structure. That is, the site that "owns" an identifier is the
        one responsible for maintaining the identifier-to-locator
        mapping information about it.

      - a mechanism would be needed to make it possible for a node to
        determine what its identifier is. To be practical, such a
        mechanism would need to be automated and avoid the need for
        manual configuration.


5.2.2.  Insufficient Hierarchy Space in ESDs

   In the case of GSE's 8-byte ESD, the size of the identifier is not
   large enough to contain sufficient heirarchy to both create DNS-like
   delegation points and support stateless address autoconfiguration.
   Stateless address autoconfiguration [RFC1971] already assumes that an
   interface's 6-byte link-layer (i.e., MAC) address can be appended to
   a link's routing prefix to produce a globally unique IPv6 address.
   With GSE, only two bytes would be available for hierarchy and
   delegation.

   It is also the case that the sorts of built-in identifiers now found
   in computing hardware, such as "EUI-48" and "EUI-64" addresses
   [IEEE802, IEEE1212], do not have the structure required for this
   delegation. Such identifiers have only two-levels of heirarchy; the
   top-level typically identifies a manufacturer, with the remaining
   part of the address being the equivalent of the serial number unique
   to the manufacturer.  In addition, the delegation of the two-level
   heirarchy (i.e., equipment manufacturer) does not correspond to the
   administrator under which the end-user operates.  Hence, stateless
   autoconfiguration [RFC1971] cannot create addresses with the
   necessary hierarchical property in the ESD portion of an address.

   Finally, imposing a required hierarchical structure on identifiers
   such as an ESD would also introduce a new administrative burden and a
   new or expanded registry system to manage ESD space (i.e., to insure
   that ESDs are globally unique). While the procedures for assigning
   ESDs, which need only organizational and not topological



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 26]


INTERNET-DRAFT                                            March 13, 1998


   significance, would be simpler than the procedures for managing IPv4
   addresses (or DNS names), it is hard to imagine such a process being
   universally well-received or without controversy; it seems a laudable
   goal to avoid the problem altogether if possible. In addition, it
   would likely increase the complexity for connecting new nodes to the
   Internet, a goal inconsistent with Stateless Address
   autoconfiguration [RFC1971].


5.2.3.  Reverse Mapping of Complete GSE Addresses

   The following two sections describe techniques for mapping a full
   IPv6 address back into some quantity (e.g., a DNS name or locator).
   We include these descriptions for completeness even though they do
   not address the fundamental problem of how to perform the mapping on
   an identifier alone. It should also be noted that because both
   techniques operate on complete IPv6 addresses, they are both directly
   applicable to provider-based addressing schemes and are not specific
   to GSE.


5.2.4.  DNS-Like Reverse Mapping of Full GSE Addresses

   Although it seems infeasible to have a global scale, reverse mapping
   of ESDs, within a site, one could imagine maintaining a database
   keyed on unstructured 8-byte ESDs. However, it is a matter of debate
   whether such a database can be kept up-to-date at reasonable cost,
   without making unreasonable assumptions as to how large sites are
   going to grow, and how frequently ESD registrations will be made or
   updated. Note that the issue isn't just the physical database itself,
   but the operational issues involved in keeping it up-to-date. For the
   rest of this section, however, let us assume that such a database can
   be built.

   A mechanism supporting a lookup keyed on a flat-space ESD from an
   arbitrary site requires having sufficient structure to identify the
   site that needs to be queried. In practice, an ESD will almost always
   be used in conjunction with Routing Stuff (i.e., a full 16-byte
   address). Since the Routing Stuff is organized hierarchically, it
   becomes feasible to maintain a DNS-like tree that maps full GSE
   addresses into DNS names, in a fashion analogous to what is done with
   IPv4 PTR records today.

   It should be noted that a GSE address lookup will work only if the
   Routing Stuff portion of the address is correctly entered in the DNS
   tree. Because the Routing Stuff portion of an address is expected to
   change over time, this assumption will not be valid indefinitely. As
   a consequence, a packet trace recorded in the past might not contain



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 27]


INTERNET-DRAFT                                            March 13, 1998


   enough information to identify the off-Site sources of the packets in
   the present. This problem can be addressed by requiring that the
   database of RG delegations be maintained for some period of time
   after the RG is no longer usable for routing packets.

   Finally, it should be noted that the problem where an address's RG
   "expires" with the implication that the mapping of "expired"
   addresses into DNS names may no longer hold is not a problem specific
   to the GSE proposal. With provider-based addressing, the same issue
   arises when a site renumbers into a new provider prefix and releases
   the allocation from a previous block. The authors are aware of one
   such renumbering in IPv4 where a block of returned addresses was
   reassigned and reused within 24 hours of the renumbering event.


5.2.5.  The ICMP Who-Are-You Message

   Although there is widespread agreement on the utility of being able
   to determine the DNS name one is communicating with, there is also
   widespread concern that repeating the experience of the "IN-
   ADDR.ARPA" domain is undesirable. In practice, the IN-ADDR.ARPA
   domain is not fully populated and poorly maintained.  Consequently,
   an old proposal to define an ICMP Who-Are-You message was resurrected
   [RFC1788]. A client would send such a message to a peer, and that
   peer would return an ICMP message containing its DNS name.

   Asking a remote host to supply its own name in no way implies that
   the returned information is accurate. However, having a remote peer
   provide a piece of information that a client can use as input to a
   separate authentication procedure provides a starting point for
   performing strong authentication. The actual strength of the
   authentication depends on the authentication procedure invoked,
   rather than the untrustable piece of information provided by a remote
   peer.

   Reconsidering the "cheap" authentication procedure described earlier,
   the ICMP Who-Are-You replaces the DNS PTR query used to obtain the
   DNS name of a remote peer. The second DNS query, to map the DNS name
   back into a set of addresses, would be performed as before. Because
   the latter DNS query  provides the strength of the authentication,
   the use of an ICMP Who-Are-You message does not in any way weaken the
   strength of the authentication method. Indeed, it can only make it
   more useful in practice, because virtually all hosts can be expected
   to implement the Who-Are-You message.

   The Who-Are-You  message is robust against renumbering, since it
   follows the paths of valid routable prefixes. Essentially, it uses
   the Internet routing system in place of the DNS delegation scheme. It



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 28]


INTERNET-DRAFT                                            March 13, 1998


   is attractive in the context of GSE-style renumbering, since no host
   or DNS server needs to be updated after a renumbering event for Who-
   Are-You-based lookups to work. It has advantages outside the context
   of GSE as well, including a more decentralized, and hence more
   scalable, administration and easier upkeep than a DNS reverse-lookup
   zone. It also has drawbacks: it requires the target node to be up and
   reachable at the time of the query and to know its fully qualified
   domain name. It is also not possible to resolve addresses once those
   addresses become unroutable. In contrast, the DNS PTR mirrors, but is
   independent of, the routing hierarchy. The DNS can maintain mappings
   long after the routing subsystem stops delivering packets to certain
   addresses.

   The requirement that the target node be up and reachable at the time
   of the query makes it very uncertain that one would be able to take
   addresses from a packet log and translate them to correct domain
   names at a later date. One can argue that this is a design flaw in
   the logging system, as it violates the architectural principle,
   "Avoid any design that requires addresses to be ... stored on non-
   volatile storage."  [RFC1958] A better-designed system would look up
   domain names promptly from logged addresses. Indeed, one of the
   authors has been doing that for some years.



5.3.  Authentication of Identifiers

   The true value of a globally unique identifier lies not on its
   uniqueness but on an ability to use the same identifier repeatedly
   and have it refer to the same end point.  That is, when an identifier
   is used, there is an expectation that repeated and subsequent use of
   the identifier results in continued communication with the same end
   point.  To be useful then, a valid identifier must either be easily
   distinguishable from a fraudulant one, or the system must have a way
   to prevent identifiers from being used in an unauthorized manner.

   The remainder of this section discusses how identifer authentication
   is done in both IPv4 and GSE, and shows how overloading an address
   with both an identifier and a locator provides automatic identifier
   authentication. In contrast, there is essentially no identifier
   authentication in GSE.  It should be noted that the actual strength
   of authentication that would be considered sufficient is a topic in
   its own right, and we do not spent much time on it. Instead, we focus
   on the relative strengths in the two schemes.







draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 29]


INTERNET-DRAFT                                            March 13, 1998


5.3.1.  Identifier Authentication in IPv4

   As described earlier, an IPv4 address simultaneously plays two roles:
   a unique identifier and a locator.  Using an overloaded address as an
   identifier has the side-effect of insuring that (for all practical
   purposes) the identifier is globally unique.  Furthermore, because
   the same number is used both to identify an interface and to deliver
   data to that interface, it is impossible for some interface A to use
   the identification of another interface B in an attempt to receive
   data destined to B without being detected, unless the routing system
   is compromised.

   When both interfaces A and B claim the same unicast address, the
   routing subsystem generally delivers packets to only one of them. The
   other node will quickly realize that something is wrong (since
   communication using the duplicate address fails) and take corrective
   actions, either correcting a misconfiguration or otherwise detecting
   and thwarting the intruder.  To understand how the routing subsystem
   prevents the same address from being used in multiple locations,
   there are two cases to consider, depending on whether the two
   interfaces using duplicate addresses are attached to the same or to
   different links.

   When two interfaces on the same link use the same address, a node
   (host or router) sending traffic to the duplicate address will in
   practice send all packets to one of the nodes. On Ethernets, for
   example, the sender will use ARP (or Neighbor Discovery in IPv6) to
   determine the link-layer address corresponding to the destination
   address. When multiple ARP replies for the target IP address are
   received, the most recently received response replaces whatever is
   already in the cache. Consequently, the destinations a node using a
   duplicate IP address can communicate with depends on what its
   neighboring nodes have in their ARP caches. In most cases, such
   communication failures become apparent relatively quickly, since it
   is unlikely that communication can proceed correctly on both nodes.

   It is also the case that a number of ARP implementations (e.g., BSD-
   derived implementations) log warning messages when an ARP request is
   received from a node using the same address as the machine receiving
   the ARP request.

   When two interfaces on different links use the same address, the
   routing subsystem generally delivers packets to only one of the nodes
   because only one of the links has the right subnet corresponding to
   the IP address. Consequently, the node using the address on the
   "wrong" link will generally never receive any packets sent to it and
   will be unable to communicate with anyone. For obvious reasons, this
   condition is usually detected quickly.



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 30]


INTERNET-DRAFT                                            March 13, 1998


   It should be noted that although an address containing a combined
   identifier and locator can be forged, the routing subsystem
   significantly limits communication using the forged address. First,
   return traffic will be sent to the correct destination and not the
   originator of the forged address. Second, routers performing ingress
   filtering can refuse to forward traffic claiming to originate from a
   source whose claimed address does not match the expected addresses
   (from a topology perspective) for sources located within a particular
   region [RFC 2267].  To effectively masquerade as someone else
   requires subverting the intermediate routing subsystem.



5.3.2.  Identifier Authentication in GSE

   In GSE, it is not possible for the routing subsystem to provide any
   enforcement on the authenticity of identifiers with respect to their
   corresponding Routing Stuff, since the Routing Stuff and ESD portions
   of an address are by definition completely orthogonal quantities.
   This fundamental problem is compounded by the fact that GSE provides
   no way (at the transport or network layer) to map an ESD into its
   corresponding Routing Stuff. Thus, when looking at the source address
   of a received packet, there is no way to ascertain whether the
   Routing Stuff portion of the address corresponds to legitimate
   Routing Stuff with respect to the corresponding ESD. Consequently, it
   becomes trivial in many cases for one node to masquerade as another.


5.3.3.  Transport Layer: What Locator Should Be Used?

   In the following, we focus on what Routing Stuff to use with TCP.
   UDP-based protocols also depend on the Routing Stuff in similar way.
   Indeed, we believe that TCP is the "easier" case to deal with, for
   two reasons. First, TCP is a stateful protocol in which both ends of
   the connection can negotiate with each other. Some UDP-based
   protocols are stateless, and remember nothing from one packet to the
   next. Consequently, changing UDP-based protocols may require the
   introduction of "session" features, perhaps as part of a common
   "library", for use by applications whose transport protocol is
   relatively stateless.  Second, changes to UDP-based protocols in
   practice mean changing individual applications themselves, raising
   deployability questions.

   There are three cases of interest from TCP's perspective:

    - the sending side of an active open

    - the sending side of a passive open (i.e., how to respond to an



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 31]


INTERNET-DRAFT                                            March 13, 1998


      active open)

    - changes to the Routing Stuff during an open connection.


5.3.4.  RG Selection On An Active Open

   If the host is performing a TCP "active open", the application first
   queries the DNS to obtain the destination address, which contains
   appropriate RG. That is, the initiator of communication is assumed to
   provide the correct Routing Stuff when initiating communication to a
   specific destination.


5.3.5.  RG Selection On An Passive Open

   When a server passively accepts connections from arbitrary clients,
   it has no choice but to assume that the Routing Stuff in the source
   address of a received packet that initiated the communication is
   correct, because it has no way to authenticate its validity.  Note
   that the Routing Stuff is "correct" only in the sense that it
   corresponds to the site originating the connection. Whether the
   Routing Stuff paired with the received ESD is actually located at
   that site where the legitimate owner of the ESD currently resides is
   not known.  Because the ESD alone cannot be mapped into a locator (or
   some other quantity that can provide input to an authentication
   procedure), there is no way to determine whether the received Routing
   Stuff corresponds to that legitimately associated with the source
   identifier of the received packet.  The issue of spoofing is
   discussed in more detail later.


5.3.6.  Mid-Connection RG Changes

   While packets are flowing as part of an open connection, the RG
   appearing on subsequent packets is susceptible to change through
   renumbering events, or as a result of site-internal routing changes
   that cause the egress point for off-site traffic to change. It is
   even possible (in the worst case) that traffic-balancing schemes
   could result in the use of two egress routers, with roughly every
   other packet exiting through a different egress router. In GSE, the
   RG does not change once a connection has been opened.

   Because TCP under GSE demultiplexes packets using only ESDs, packets
   will be delivered to the correct end-point regardless of what source
   RG is used. However, in GSE return traffic continues to be sent via
   the "old" RG, even though it may have been deprecated or become less
   optimal because the peer's border router has changed.  It would seem



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 32]


INTERNET-DRAFT                                            March 13, 1998


   highly desirable for TCP connections to be able to survive such
   events.  However, the completion of renumbering events (so that an
   earlier RG is now invalid) and certain topology changes would require
   TCP to switch sending to a new RG mid-connection.  To explore the
   whole space, we considered ways of allowing this mid-connection RG
   change to happen.

   If TCP connection identifiers are based on ESDs rather than full
   addresses, traffic from the same ESD would be viewed as coming from
   the same peer, regardless of its source RG. Because this
   vulnerability is already present in today's Internet (forging full
   source addresses is trivial), the mere delivery of incoming datagrams
   with the same ESD but a different RG does not introduce new
   vulnerability to TCP.  In today's Internet, any node can already
   originate FINs/RSTs from an arbitrary source address and potentially
   or definitely disrupt the connection.  Therefore, changing RG for
   acceptance, or acceptance of traffic independent of its source RG,
   does not appear to significantly worsen existing robustness. (See the
   comment on ingress filtering in Section 5.3.1, however.)

   We also considered allowing TCP to reply to each segment using the RG
   of the most recently-received segment. Although this allows TCP to
   survive some important events (e.g., renumbering), it also makes it
   trivial to hijack connections, unacceptably weakening robustness
   compared with today's Internet. A sender simply needs to guess the
   sequence numbers in use by a given TCP connection [Bellovin 89] and
   send traffic with a bogus RG to hijack a connection to an intruder at
   an arbitrary location.

   Providing protection from hijacking implies that the RG used to send
   packets must be bound to a connection end-point (e.g., it is part of
   the connection state). Although it may be reasonable to accept
   incoming traffic independent of the source RG, the choice of sending
   RG requires more careful consideration. Indeed, any subsequent change
   in what RG is used for sending traffic must be properly authenticated
   (e.g., using cryptographic means). In the GSE proposal, it is not
   clear how to authenticate such a change, since the remote peer
   doesn't even know its own RG.  Consequently, the only reasonable
   approach in GSE is to send to the peer using the first RG used for
   the entire life of a connection. That is, always use the first RG
   seen.


5.3.7.  The Impact of Corrupt Routing Goop

   Another interesting issue that arises is what impact corrupted RG
   would have on robustness. Because the RG is not covered by the TCP
   checksum (the sender doesn't know what source RG will be inserted),



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 33]


INTERNET-DRAFT                                            March 13, 1998


   it would be difficult to detect such corruption at the receiver.
   Moreover, once a specific RG is in use, it does not change for the
   duration of a connection. The interesting case occurs on the passive
   side of a TCP connection, where a server accepts incoming connections
   from remote clients. If the initial SYN from the client includes
   corrupted RG, the server TCP will create a TCP connection (in the
   SYN-RECEIVED state) and cache the corrupted RG with the connection.
   The second packet of the 3-way handshake, the SYN-ACK packet, would
   be sent to the wrong RG and consequently not reach the correct
   destination. Later, when the client retransmits the unacknowledged
   SYN, the server will continue to send the SYN-ACK using the bad RG.
   Eventually the client times out, and the attempt to open a TCP
   connection fails.

   We next consider relaxing the restriction on switching RGs in an
   attempt to avoid the previous failure scenario. The situation is
   complicated by the fact that the RG on received packets may change
   for legitimate reasons (e.g., a multi-homed site load-shares traffic
   across multiple border routers). The key question is how one can
   determine which RG is valid and which is not. That is, for each of
   the RGs a sender attempts to use, how can it determine which RG
   worked and which did not? Solving this problem is more difficult than
   first appears, since one must cover the cases of delayed segments,
   lost segments, simultaneous opens, etc. If a SYN-ACK is retransmitted
   using different RGs, it is not possible to determine which of those
   RGs worked correctly. We conclude that the only way TCP could
   determine that a particular RG is correct is by receiving an ACK for
   a specific sequence number in which all transmissions of that
   sequence number used the same RG (a non-trivial addition to TCP).

   At best, an RG selection algorithm for TCP would be relatively
   straightforward but would require new logic in implementations of
   TCP's opening handshake --- a significant transition/deployment
   issue. We are not certain that a valid algorithm is attainable,
   however. RG changes would have to be handled in all cases handled by
   the opening handshake: delayed segments, lost segments, undetected
   bit errors in RG, simultaneous opens, old segments, etc.

   In the end, we conclude that although the corrupted SYN case
   introduces potential problems, the changes that would need to be made
   to TCP to robustly deal with such corruption would be significant, if
   tractable at all. This would result in a transition to GSE also
   having a significant TCPng component, a significant drawback.








draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 34]


INTERNET-DRAFT                                            March 13, 1998


5.3.8.  On The Uniqueness Of ESDs

   The uniqueness requirements for ESDs depends on what purpose they
   serve and how they are used. In GSE, ESDs identify interfaces,
   requiring that they be globally unique. It does not make sense for
   two different interfaces to use the same ESD; every interface must
   have its own ESD to distinguish it from others.

   If ESDs are only used to identify session endpoints, the situation
   becomes more complex.  At first glance it might appear that two nodes
   using the same ESD cannot communicate. However, this is not
   necessarily the case. In the GSE proposal, for example, a node
   queries the DNS to obtain an IPv6 address. The returned address
   includes the Routing Stuff of an address (the RG+STP portions). Since
   the sending host transmits packets based on the entire destination
   IPv6 address, the sender may well forward the packet to a router that
   delivers the packet to its correct destination (using the information
   in the Routing Stuff). It is only on receipt of a packet that a node
   would extract the ESD portion of a datagram's destination address and
   ask "is this for me?" That is, a sender may not notice the
   destination ESD is the same as the sending ESD because of the Routing
   Stuff part of the address.

   A more problematic case occurs if two nodes using the same ESD
   communicate with a third party. To the third party, packets received
   from either machine might appear to be coming from the same machine
   since they are both using the same ESD. Consequently, at the
   transport level, if both machines choose the same source and
   destination port numbers (one of the ports --- a server's well-known
   port number --- will likely be the same), packets belonging to two
   distinct transport connections will be demultiplexed to a single
   transport end-point.

   When packets from different sources using the same source ESD are
   delivered to the same transport end-point, a number of possibilities
   come to mind:

     1) The transport end-point could accept the packet, without regard
        to the Routing Stuff of the source address. This may lead to a
        number of robustness problems, if data from two different
        sources mistakenly using the same ESD are delivered to the same
        transport or application end-point (which at best will confuse
        the application).

     2) The transport end-point could verify that the Routing Stuff of
        the source address matches one of a set of expected values
        before processing the packet further. If the Routing Stuff
        doesn't match any expected value, the packet could be dropped.



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 35]


INTERNET-DRAFT                                            March 13, 1998


        This would result in a connection from one host operating
        correctly, while a connection from another host (using the same
        ESD) would fail.

     3) When a packet is received with an unexpected Routing Stuff the
        receiver could invoke special-purpose code to deal with this
        case. Possible actions include attempting to verify whether the
        Routing Stuff is indeed correct (the saved values may have
        expired) or attempting to verify whether duplicate ESDs are in
        use (e.g., by inventing a protocol that sends packets using both
        Routing Stuff and verifies that they are delivered to the same
        end-point).


5.3.9.  New Denial of Service Attacks.

   It is clear that there are potential problems if identifiers are not
   globally unique. How common such problems would actually occur in
   practice depends on how many duplicates there are actually are. Thus,
   one might be tempted to make the argument that a scheme for assigning
   identifiers could be made to be "unique enough" in practice. This
   would be a dangerous and naive assumption, because intruders will
   actively impersonate other sites for the sole purpose of invalidating
   the uniqueness assumption. For example, one could deny service to
   host foo.bar.com by querying the DNS for its corresponding ESD, and
   then impersonaiting that ESD.

   As a specific example, one GSE-specific denial-of-service attack
   would be for an intruder to masquerade as another host and "wedge"
   connections in a SYN-RECEIVED state by sending SYN segments
   containing an invalid RG in the source IP address for a specific ESD.
   Subsequent connection attempts to the wedged host from the legitimate
   owner of the ESD (if they used the same TCP port numbers) would then
   not complete, since return traffic would be sent to the wrong place.


5.3.10.  Summary of Identifier Authentication Issues

   In summary, changing the RG dynamically in a safe way for a
   connection requires that an originator of traffic be able to
   authenticate a proposed change in the RG before sending to a
   particular ESD via that RG. This is difficult for several reasons:

     1) It can't be done on an end-to-end basis in GSE (e.g., via IPSec)
        because the sender doesn't know what the RG portion of the
        address will be when it reaches the sender.

     2) It can't easily be done in GSE because there is no mechanism at



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 36]


INTERNET-DRAFT                                            March 13, 1998


        or below the transport layer to map ESDs into a quantity that
        can be used as a key to jump start the authentication process
        (using the DNS would be problematic due to layering circularity
        considerations).

     3) Any scheme that uses the full IPv6 address to do the
        authentication can be used with standard provider-based
        addressing, raising the question of what benefit is retained
        from having separate identifiers and locators.

   Our final conclusion is that with the GSE approach, transport
   protocol end-points must make an early, single choice of the RG to
   use when sending to a peer and stick with that choice for the
   duration of the connection. Specifically:

     1) The demultiplexing of arriving packets to their transport end
        points should use only the ESD, and not the Routing Stuff.

     2) If the application chooses an RG for the remote peer (i.e., an
        active open), use the provided RG for all traffic sent to that
        peer, even if alternative RGs are received on subsequent
        incoming datagrams from the same ESD.

     3) For all other cases, use the first RG received with a given ESD
        for all sending. Simultaneously, we understand that with this
        rule, there are still open issues with regard to invalid RGs,
        either through corruption or through a active hostile attacks.

   With the above recommendation, there does not appear to be a
   straightforward way to use ESDs in conjunction with mobility or site
   renumbering (in which existing connections survive the renumbering).
   This presents a quandry.  The main benefit of separating identifiers
   and locators is the ability to have communication (e.g., a TCP
   connection) continue transparently, even when the Routing Stuff
   associated with a particular ESD changes. However, switching to a new
   Routing Stuff without properly authenticating it makes it trivial to
   hijack connections.

   We cannot emphasize enough that the use of an ESD independent of an
   associated RG can be very dangerous. That is, communicating with a
   peer implies that one is always talking to the same peer for the
   duration of the communication. But as has been described in previous
   sections, such assurance can only come from properly authenticated
   RG.







draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 37]


INTERNET-DRAFT                                            March 13, 1998


5.4.  Miscellaneous


5.4.1.  Renumbering and Domain Name System (DNS) Issues

   Because any mapping scheme is complicated by renumbering, and because
   recent IPv4 experience has shown a requirement for renumbering at
   some frequency, it is worthwhile to explore the general renumbering
   issue.


5.4.2.  How Frequently Can We Renumber?

   One premise of the GSE proposal [GSE] is that an ISP can renumber the
   Routing Goop portion of a site's addresses transparently to the site
   (i.e., without coordinating the change with the site). This would
   make it possible for backbone providers to aggressively renumber the
   Routing Goop part of addresses and achieve a high degree of route
   aggregation. On closer examination, frequent (e.g., daily)
   renumbering turns out to be difficult in practice because of a
   circular dependency between the DNS and routing. Specifically, if a
   site's Routing Stuff changes, nodes communicating with the site need
   to obtain the new Routing Stuff. In the GSE proposal, one queries the
   DNS to obtain this information. However, in order to reach a site's
   DNS servers, the pointers controlling the downward delegation of
   authoritative DNS servers (i.e., DNS "glue records") must use
   addresses with Routing Stuff that are reachable. That is, in order to
   find the address for the web server "www.foo.bar.com", DNS queries
   might need to be sent to a root DNS server, as well as DNS servers
   for "bar.com" and "foo.bar.com". Each of these servers must be
   reachable from the querying client. Consequently, there must be an
   overlap period during which both the old Routing Stuff and the new
   Routing Stuff can be used simultaneously. During the overlap period,
   DNS glue records would need to be updated to use the new addresses
   (including Routing Stuff). Only after all relevant DNS servers have
   been updated and older cached RRs containing the old addresses have
   timed out can the old address be deleted.

   An important observation is that the above issue is not specific to
   GSE: the same requirement exists with today's provider-based
   addressing architecture. When a site is renumbered (e.g., it switches
   ISPs and obtains a new set of addresses from its new provider), the
   DNS must be updated in a similar fashion.








draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 38]


INTERNET-DRAFT                                            March 13, 1998


5.4.3.  Efficient DNS support for Site Renumbering

   When a site renumbers to satisfy its ISP, only the site's routing
   prefix needs to change. That is, the prefix reflects where within the
   Internet the site resides.

   In the current Internet, when a site is renumbered, the addresses of
   all the site's internal nodes change. This requires a potentially
   large update to the RR database for that site. Although Dynamic DNS
   [DDNS] could potentially be used, the cost is likely to be large due
   to the large number of individual records that would need to be
   updated. In addition, when DHCP and DDNS are used together [DHCP-
   DDNS], it may be the case that individual hosts "own" their own A or
   AAAA records, further complicating the question of who is able to
   update the contents of DNS RRs.

   One change that could reduce the cost of updating the DNS when a site
   is renumbered is to split addresses into two distinct portions: a
   Routing Goop that reflects where a node attaches to the Internet and
   a STP-plus-ESD that is the site-specific part of an address. During a
   renumbering, the Routing Goop would change, but the "site internal
   part" would remain fixed. Furthermore, the two parts of the address
   could be stored in the DNS as separate RRs. That way, renumbering a
   site would only require that the Routing Goop RR of a site be
   updated; the "site-internal part" of individual addresses would not
   change.

   To obtain the address of a node from the DNS, a DNS query for the
   name would return two quantities: the "site internal part" and the
   DNS name of the Routing Stuff for the site. An additional DNS query
   would then obtain the specific RR of the site, and the complete
   address would be synthesized by concatenating the two pieces of
   information.

   Implementing these DNS changes increases the practicality of using
   Dynamic DNS to update a site's DNS records as it is renumbered. Only
   the site's Routing Goop RRs would need updating.

   Finally, it may be useful to divide a node's AAAA RR into the three
   logical parts of the GSE proposal, namely RG, STP and ESD. Whether or
   not it is useful to have separate RRs for the STP and ESD portions of
   an address or a single RR combining both is an issue that requires
   further study.

   If AAAA records are comprised of multiple distinct RRs, then one
   question is who should be responsible for synthesizing the AAAA from
   its components: the resolver running on the querying client's machine
   or the queried name server? To minimize the impact on client hosts



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 39]


INTERNET-DRAFT                                            March 13, 1998


   and make it easier to deploy future changes, it is recommended that
   the synthesis of AAAA records from its constituent parts be done on
   name servers rather than in client resolvers.



5.4.4.  Two-Faced DNS

   The GSE proposal attempts to hide the RG part of addresses from nodes
   within a site. If the nodes do not know their own RG, then they can't
   store or use them in ways that cause problems should the site be
   renumbered and its RG change (i.e., the cached RG become invalid). A
   site's DNS servers, however, will need to have more information about
   the RG its site uses. Moreover, the responses it returns will depend
   on who queries the server. A query from a node within the site should
   return an address with a Site Local RG, whereas a query for the same
   name from a client located at a different site should return the
   global scope RG.  This facilitates intra-site communication to be
   more resilient to failures outside of the site.  Such context-
   dependent DNS servers are commonly referred as "two-faced" DNS
   servers.

   Some issues that must be considered in this context:

     1) A DNS server may recursively attempt to resolve a query on
        behalf of a requesting client. Consequently, a DNS query might
        be received from a proxy rather than from the client that
        actually seeks the information. Because the proxy may not be
        located at the same site as the originating client, a DNS server
        cannot reliably determine whether a DNS request is coming from
        the same site or a remote site. One solution would be to
        disallow recursive queries for off-site requesters, though this
        raises additional questions.

     2) Since cached responses are, in general, context sensitive, a
        name server may be unable to correctly answer a query from its
        cache, since the information it has is incomplete. That is, it
        may have loaded the information via a query from a local client,
        and the information has a site-local prefix. If a subsequent
        request comes in from an off-site requester, the DNS server
        cannot return a correct response (i.e., one containing the
        correct RG).









draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 40]


INTERNET-DRAFT                                            March 13, 1998


5.4.5.  Bootstrapping Issues

   If Routing Stuff information is distributed via the DNS, key DNS
   servers must always be reachable. In particular, the addresses
   (including Routing Stuff) of all root DNS servers are, for all
   practical purposes, well-known and assumed to never change. It is not
   uncommon for the addresses of root servers to be hard-coded into
   software distributions. Consequently, the Routing Stuff associated
   with such addresses must always be usable for reaching root servers.
   If it becomes necessary or desirable to change the Routing Stuff of
   an address at which a root DNS server resides, the routing subsystem
   will likely need to continue carrying "exceptions" for those
   addresses. Because the total number of root DNS servers is relatively
   small, the routing subsystem is expected to be able to handle this
   requirement.

   All other DNS server addresses can be changed, since their addresses
   are typically learned from an upper-level DNS server that has
   delegated a part of the name space to them. So long as the delegating
   server is configured with the new address, the addresses of other
   servers can change.



6.  Conclusion

   The GSE proposal provides a concrete example of a network protocol
   design that separates identifiers from locators in addresses.  In
   this paper we compared GSE with IPv4 to better understand the pro's
   and con's of the respective design approaches.

   Functionally speaking, identifiers and locators each have a logically
   different role to play.  Thus overloading both in one field causes
   problems whenever the location of a node changes but its identity
   does not.  However, our analysis shows that overloading also presents
   two critically important benefits.

   First, for network entity A to send data to network entity B, A must
   not only know B's end identifier but also B's locator.  No scalable
   way is known at this time to provide this mapping at the network
   layer, other than overloading the two quantities into an address as
   is done in IPv4. Fundamentally, a scalable mapping algorithm strongly
   suggests that the identifier space be structured hierarchically, yet
   identifiers in GSE are not sufficiently large to both contain
   sufficient heirarchy and support stateless address autoconfiguration.
   Instead, GSE forces applications to supply up-to-date locators.
   However, relying on the locator provided at the time communication is
   established as GSE does is inadequate when the remote locator can



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 41]


INTERNET-DRAFT                                            March 13, 1998


   change dynamically, precisely the scenario that is supposed to
   benefit from the separation. That is, the benefits of separating the
   identifier from the locator are largely lost, if the changes in the
   identifier to locator binding are not tracked quickly.

   Secondly, when communicating with a remote site, a receiver must be
   able to insure (with reasonable certainty) that received data does
   indeed come from the expected remote entity. In IPv4, it is possible
   to receive packets from a forged source, but the potential for
   mischief between communicating peers is significantly limited because
   return traffic will not reach the source of the forged traffic. That
   is, communication involving packets sent in both directions will not
   succeed. In contrast, architectures like GSE that decouple the
   identifier and locator functions have great difficulty assuring that
   traffic from a source identified only by an identifer actually comes
   from the correct source.  Short of using cryptographic techniques in
   which both end points share a private secret (e.g., using IPSec),
   there is no known mechanism that can use an identifier alone to
   perform this remote entity authentication in a scalable way.  That
   is, using an identifier alone for authentication of received packets
   is dangerously unsafe.

   In summary, although overloading the address field with a combined
   identifier and locator leads to difficulties in retaining the
   identity of a node whenever its address changes, analysis in this
   paper suggests that the benefit of the overloading actually out-
   weighs its cost.  Completely separating an identifier from its
   locator renders the identifier untrustworthy, thus useless, in the
   absence of an accompanying authentication system.


7.  Security Considerations

   The primary security consideration with GSE or, more generally, a
   network layer with addresses split into locator and identifier parts,
   is that of one node impersonating another by copying the
   identification without the location.


8.  Acknowledgments

   Thanks go to Steve Deering and Bob Hinden (the Chairs of the IPng
   Working Group) as well as Sun Microsystems (the host for the PAL1
   meeting) for the planning and execution of the interim meeting.
   Thanks also go to Mike O'Dell for writing the 8+8 and GSE drafts; by
   publishing these documents and speaking on their behalf, Mike was the
   catalyst for some valuable discussions, both for IPv6 addressing and
   for addressing architectures in general. Special thanks to the



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 42]


INTERNET-DRAFT                                            March 13, 1998


   attendees of the PAL1 meeting whose high caliber discussions helped
   motivate and shape this document.



9.  References

     [BATES] Scalable support for multi-homed multi-provider
             connectivity, Internet Draft, Tony Bates & Yakov Rekhter,
             draft-bates-multihoming-01.txt.

     [Bellovin 89] "Security Problems in the TCP/IP Protocol Suite",
             Bellovin, Steve, Computer Communications Review, Vol. 19,
             No. 2, pp32-48, April 1989.

     [CERT] CERT(sm) Advisory CA-96.21
             (ftp://info.cert.org/pub/cert_advisories)

     [DANVERS] Minutes of the IPNG working Group, April 1995.
             ftp://ftp.ietf.cnri.reston.va.us/ietf-online-proceedings/
             95apr/area.and.wg.reports/ipng/ipngwg/ ipngwg-minutes-
             95apr.txt.

     [DHCP-DDNS] Interaction between DHCP and DNS, Internet Draft, Yakov
             Rekhter, draft-ietf-dhc-dhcp-dns-04.txt.

     [DDNS] "Dynamic Updates in the Domain Name System (DNS UPDATE)",
             Paul Vixie (Editor), draft-ietf-dnsind-dynDNS-11.txt,
             November, 1996.

     [EUI64] 64-Bit Global Identifier Format Tutorial.
             http://standards.ieee.org/db/oui/tutorials/EUI64.html.
             Note: "EUI-64" is claimed as a trademark by an organization
             which also forbids reference to itself in association with
             that term in a standards document which is not their own,
             unless they have approved that reference. However, since
             this document is not standards-track, it seems safe to name
             that organization: the IEEE.

     [GSE] "GSE - An Alternate Addressing Architecture for IPv6", Mike
             O'Dell, draft-ietf-ipngwg-gseaddr-00.txt.

     [IEEE802] IEEE Std 802-1990, Local and Metropolitan Area Networks:
             IEEE Standard Overview and Architecture.

     [IEEE1212] IEEE Std 1212-1994, Information technology--
             Microprocessor systems: Control and Status Registers (CSR)
             Architecture for microcomputer buses.



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 43]


INTERNET-DRAFT                                            March 13, 1998


     [RFC1122] "Requirements for Internet hosts - communication layers",
             R. Braden, 10/01/1989.

     [RFC1715] The H Ratio for Address Assignment Efficiency.  C.
             Huitema.

     [RFC1726] Technical Criteria for Choosing IP:The Next Generation
             (IPng). F. Kastenholz, C. Partridge.

     [RFC1752] "The Recommendation for the IP Next Generation Protocol,"
             S. Bradner, A. Mankin, 01/18/1995.

     [RFC1788] "ICMP Domain Name Messages", W. Simpson, 04/14/1995

     [RFC1958] Architectural Principles of the Internet.  B. Carpenter.

     [RFC1971] IPv6 Stateless Address Autoconfiguration.  S. Thomson, T.
             Narten.

     [RFC2002] "IP Mobility Support", C. Perkins, RFC 2002, October,
             1996.

     [RFC2008] "Implications of Various Address Allocation Policies for
             Internet Routing", Y. Rekhter, T. Li.

     [RFC2065] Domain Name System Security Extensions. D. Eastlake, C.
             Kaufman.

     [RFC2073] An IPv6 Provider-Based Unicast Address Format.  Y.
             Rekhter, P. Lothberg, R. Hinden, S. Deering, J. Postel

     [RFC2267] Network Ingress Filtering: Defeating Denial of Service
             Attacks which employ IP Source Address Spoofing, P.
             Ferguson, D. Senie, January 1988.


10.  Authors' Addresses

   Matt Crawford                           John Stewart
   Fermilab MS 368                         Juniper Networks, Inc.
   PO Box 500                              385 Ravendale Drive
   Batavia, IL 60510 USA                   Mountain View, CA  94043
   Phone: 630-840-3461                     Phone: +1 650 526 8000
   EMail: crawdad@fnal.gov                 EMail: jstewart@juniper.net

   Allison Mankin                          Lixia Zhang
   USC/ISI                                 UCLA Computer Science Department
   4350 North Fairfax Drive                4531G Boelter Hall



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 44]


INTERNET-DRAFT                                            March 13, 1998


   Suite 620                               Los Angeles, CA 90095-1596 USA
   Arlington, VA  22203 USA                Phone: 310-825-2695
   EMail: mankin@isi.edu                   EMail: lixia@cs.ucla.edu
   Phone: 703-807-0132

   Thomas Narten
   IBM Corporation
   3039 Cornwallis Ave.
   PO Box 12195 - F11/502
   Research Triangle Park, NC 27709-2195
   Phone: 919-254-7798
   EMail: narten@raleigh.ibm.com







































draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 45]


INTERNET-DRAFT                                            March 13, 1998


Appendix B -- Ideas Incorporated Into IPv6

   This section summarizes changes made to IPv6 specifications which
   originated in the GSE proposal or in the discussions arising from it.

   First and most visible was the change to the unicast address format.
   Instead of an topologically insignificant Registry ID immediately
   following the Format Prefix, there is now a Top-Level Aggregation
   Identifier.  This field will identify a large routable aggregate to
   which an address belongs rather than an administrative unit by which
   an address was assigned.  The TLA corresponds to the "Large
   Structure" of GSE.  The IPv6 Next-Level Aggregation Identifier (NLA)
   is roughly the rest of the GSE "Routing Goop" and the Site-Level
   Aggregation Identifier (SLA) is a slightly expanded GSE Site Topology
   Partition.

   The decision to put fixed boundaries between parts of the unicast
   address (TLA, NLA, SLA, Interface Identifier) also came from GSE.
   The previous "provider-based" addressing architecture for IPv6 had
   fluid boundaries between Registry ID, Provider ID, Subscriber ID and
   the Intra-Subscriber part, as well as undefined divisions within the
   Provider-ID and Intra-Subscriber part.  (On subnetworks with a MAC-
   layer address, the latter boundary was generally placed to
   accommodate use of that address as an Interface ID.)  The new
   addressing architecture still expects divisions within the NLA
   portion of the address, placed to reflect topological aggregation
   points.

   Defining a fixed boundary between the routable portion of the address
   and the node-on-link identifier required the specification of an
   Interface Identifier which would be as suitable as possible for all
   subnetwork technologies.  The IEEE "EUI-64" identifier was selected,
   having the advantages of an easy mapping from 48 bit MAC addresses
   and a defined escape flag into locally-administered values.

   The second change to come out of the GSE discussions relates to
   reducing the number of DNS record changes required in the event of
   site renumbering.  This work is not finalized as of this writing, but
   the result may be that individual IPv6 addresses are stored (and
   signed, in the case of Secure DNS) as a partial address and an
   indirect pointer which leads to the high-order part of the address.
   There may be multiple levels of indirection and a changed record at
   any one level would suffice to update the DNS's record of the IPv6
   addresses of every node in a given branch of the addressing
   hierarchy.

   A change in the method of doing DNS address-to-name lookups is also
   in the works.  This may be a change in the form and/or operation of



draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 46]


INTERNET-DRAFT                                            March 13, 1998


   the ip6.int domain or some new mechanism which involves participation
   by the routers or the end-nodes themselves.

   Two other changes arising from GSE will not affect the IPv6 base
   specifications themselves, but do direct additional work.  Those are
   the injection of global prefix information into a site from a
   provider or exchange, and some inter-provider cooperative method of
   providing multihoming to mutual customers with minimal impact on
   routing tables in distant parts of the network.










































draft-ietf-ipngwg-esd-analysis-02.txt                          [Page 47]