[Docs] [txt|pdf] [Tracker] [WG] [Email] [Diff1] [Diff2] [Nits]

Versions: 00 01 02 03 04 05

INTERNET-DRAFT                                             Matt Crawford
                                                                Fermilab
<draft-ietf-ipngwg-esd-analysis-04.txt>                   Allison Mankin
                                                                     ISI
                                                           Thomas Narten
                                                                     IBM
                                                    John W. Stewart, III
                                                                 Juniper
                                                             Lixia Zhang
                                                                    UCLA
                                                       February 12, 1999

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

                  <draft-ietf-ipngwg-esd-analysis-04.txt>


Status of this Memo

   This document is an Internet-Draft and is in full conformance with
   all provisions of Section 10 of RFC2026 except that the right to
   produce derivative works is not granted.

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

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

   The list of current Internet-Drafts can be accessed at
   http://www.ietf.org/ietf/1id-abstracts.txt

   The list of Internet-Draft Shadow Directories can be accessed at
   http://www.ietf.org/shadow.html.


Abstract

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



draft-ietf-ipngwg-esd-analysis-04.txt                           [Page 1]

INTERNET-DRAFT                                         February 12, 1999


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

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

   Contents

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

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

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

   3.  Addressing and Routing in IPv4...........................    5
      3.1.  The Need for Aggregation............................    7
      3.2.  The Pre-CIDR Internet...............................    7
      3.3.  CIDR and Provider-Based Addressing..................    9
      3.4.  Multi-Homed Sites and Aggregation...................   12

   4.  The GSE Proposal.........................................   15
      4.1.  Motivation For GSE..................................   15
      4.2.  GSE Address Format..................................   16
         4.2.1.  Routing Stuff (RG and STP).....................   16
         4.2.2.  End-System Designator..........................   18
      4.3.  Address Rewriting by Border Routers.................   19
      4.4.  Renumbering and Rehoming Mid-Level ISPs.............   20
      4.5.  Support for Multi-Homed Sites.......................   21
      4.6.  Explicit Non-Goals for GSE..........................   22

   5.  Analysis: The Pros and Cons of Overloading Addresses.....   22



draft-ietf-ipngwg-esd-analysis-04.txt                           [Page 2]

INTERNET-DRAFT                                         February 12, 1999


      5.1.  Purpose of an Identifier............................   23
      5.2.  Mapping an Identifier to a Locator..................   25
         5.2.1.  Scalable Mapping of Identifiers to Locators....   27
         5.2.2.  Insufficient Hierarchy Space in ESDs...........   27
      5.3.  Authentication of Identifiers.......................   28
         5.3.1.  Identifier Authentication in IPv4..............   29
         5.3.2.  Identifier Authentication in GSE...............   30
      5.4.  Transport Layer: What Locator Should Be Used?.......   30
         5.4.1.  RG Selection On An Active Open.................   31
         5.4.2.  RG Selection On An Passive Open................   31
         5.4.3.  Mid-Connection RG Changes......................   31
         5.4.4.  The Impact of Corrupted Routing Goop...........   33
      5.5.  On The Uniqueness Of ESDs...........................   34
         5.5.1.  Impact of Duplicate ESDs.......................   34
         5.5.2.  New Denial of Service Attacks..................   35
      5.6.  Summary of Identifier Authentication Issues.........   35

   6.  Conclusion...............................................   37

   7.  Security Considerations..................................   38

   8.  Acknowledgments..........................................   38

   9.  References...............................................   38

   10.  Authors' Addresses......................................   40

   Appendix A: Increased Reliance on Domain Name System (DNS)...   41

   Appendix B: Additional Issues Related to GSE.................   45

   Appendix C: Ideas Incorporated Into IPv6.....................   46

   Appendix D: Reverse Mapping of Complete GSE Addresses........   47


1.  Introduction

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

   Shortly before the interim meeting, an updated version of the



draft-ietf-ipngwg-esd-analysis-04.txt                           [Page 3]

INTERNET-DRAFT                                         February 12, 1999


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

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

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

   Finally, this document's focus on unicast issues should not be
   interpreted to mean that the impact of separating identifier and
   locating functions on non-unicast aspects of routing and addressing
   are well understood or trivial to deal with.  Specifically,
   understanding how multicasting and anycast addressing [ANYCAST,
   RFC1884] fits into such a model requires further work.


2.  Definitions and Terminology

   The following terminology is used throughout this document.

      Routing Goop --- A term defined by the GSE document.  It refers to
                    the first six bytes of a sixteen byte IPv6 GSE
                    address.  The Routing Goop portion of an address
                    identifies where a site connects to the public
                    Internet.  More generally, the term refers to the



draft-ietf-ipngwg-esd-analysis-04.txt                           [Page 4]

INTERNET-DRAFT                                         February 12, 1999


                    portion of an address's routing prefix that
                    identifies where on the public Internet the site
                    housing the address resides.

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

      Routing Stuff --- The part of an address that identifies which
                    link the address resides on.  Within the context of
                    GSE, the Routing Stuff comprises the Routing Goop
                    and Site Topology Partition parts of an address
                    (i.e., the left mots eight bytes).

      identifier --- a value that indicates the sender of a packet, or
                    the intended recipient of a packet.  Within the
                    context of GSE, the ESD portion (i.e., the rightmost
                    eight bytes) of the address is an identifier.

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


3.  Addressing and Routing in IPv4

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

   The structure and semantics of a network layer protocol's addresses
   are absolutely core to that protocol.  Addressing substantially
   impacts the way packets are routed, the ability of a protocol to
   scale and the kinds of functionality higher layer protocols can count
   on.  Indeed, addressing is intertwined with both routing and
   transport layer issues; a change in any one of these can impact
   another.  Issues of administration and operation (e.g., address
   allocation/re-allocation and required renumbering), while not part of
   the pure exercise of engineering a network layer protocol, turn out
   to be critical to the scalability of that protocol in a global and



draft-ietf-ipngwg-esd-analysis-04.txt                           [Page 5]

INTERNET-DRAFT                                         February 12, 1999


   commercial network.  The interaction between addressing, routing and
   especially aggregation is particularly relevant to this document, so
   some time will be spent describing it.

   Addresses in IPv4 serve two purposes:

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

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

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

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

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

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



draft-ietf-ipngwg-esd-analysis-04.txt                           [Page 6]

INTERNET-DRAFT                                         February 12, 1999


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


3.1.  The Need for Aggregation

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

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


3.2.  The Pre-CIDR Internet

   In the early days of the Internet, its topology and addressing were
   orthogonal.  Specifically, when a site wanted to connect to the
   Internet, it approached the central Internet Assigned Numbers
   Authority (IANA) to obtain an address block and then approached a
   provider about procuring connectivity.  This procedure for address
   allocation resulted in a system where the addresses used by customers
   of the same provider bore little relation to the addresses used by



draft-ietf-ipngwg-esd-analysis-04.txt                           [Page 7]

INTERNET-DRAFT                                         February 12, 1999


   other customers of that same provider.  In other words, though the
   actual topology of the Internet was mostly hierarchical, the
   addressing was not.  An example of such a topology and addressing
   scheme is shown in Figure 1.


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

                                 Figure 1

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

   Experience has shown that this approach scales very poorly.  In the
   Default-Free Zone (DFZ) of the Public Internet, where routers must
   maintain routing entries for all reachable destinations, the cost of
   computing forwarding tables quickly becomes unacceptably large.  A
   large part of the cost is related to the seemingly redundant
   computations that must be made for each individual network, even
   though many of them reside in the same topological location (e.g.,
   under the same provider).  Looking at Figure 1, the problem is that
   provider B performs 5 separate calculations to construct the
   forwarding table needed to reach each of A's customers, even though
   it is going to take the same path for all of them; in other words,
   there is an opportunity to do data abstraction.

   Figure 1 shows network numbers using the older "classful" notation.
   Since 1981, the first few bits of an address syntactically identified
   which parts of an address identified the "network" and "local"
   portions of an address.  There were a small number of Class A



draft-ietf-ipngwg-esd-analysis-04.txt                           [Page 8]

INTERNET-DRAFT                                         February 12, 1999


   addresses (intended for very large sites), a medium number of Class B
   addresses (for medium-sized sites) and a very large number of Class C
   addresses (for very small sites).  In practice, the actual size of
   real networks didn't match the original allocation of Class A, B, and
   C addresses.  Class B addresses were bigger than most sites needed
   (and there weren't enough of them), and Class C addresses were too
   small (i.e., typical sites would need to get 10 or more C blocks to
   cover all of the hosts on their networks).  Consequently, classless
   addressing was developed [CIDR], which made the boundaries between
   the network and local parts of an address more flexible.  With
   classless addressing, a separate prefix-length (i.e., network mask)
   specifies how many of the left-most bits of an address identify the
   network part of the address.


3.3.  CIDR and Provider-Based Addressing

   One of the reasons CIDR (Classless Inter-Domain Routing) and its
   associated provider-assigned address allocation policy were
   introduced was to help reduce the cost of computing a routing table
   and the size of the forwarding table computed from the routing table.
   To achieve this goal CIDR aggressively aggregates network addresses.
   Aggregating network addresses means "merging" multiple addresses into
   a single "bigger" one, that is to use a common prefix to provide
   location information for all addresses sharing that same prefix.

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















draft-ietf-ipngwg-esd-analysis-04.txt                           [Page 9]

INTERNET-DRAFT                                         February 12, 1999



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

                                  Figure 2

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

   CIDR was a critical step for the Internet: in the early 1990s the
   size of default-free routing tables required to support the classful
   Internet was almost more than the commercially-available hardware and
   software of the day could handle.  The introduction of BGP4's
   classless routing and provider-based address allocation policies
   resulted in a significant decrease in the growth rate of the routing
   tables.  At the same time, however, CIDR introduced some new
   weaknesses.  First, the Internet addressing model had to shift from
   one of "address owning" to "address lending" [RFC2008].  In pre-CIDR
   days sites acquired addresses from a central authority independent of
   their provider, and a site could assume it "owned" the address block
   it was given.  Owning addresses meant that once one had been given a
   set of network addresses, one could always use them; no matter where
   one's site connected to the Internet, the prefix for that network
   could be injected into the public routing system.  Today, however, it
   is simply not possible for all individual sites to have their own
   prefixes injected into the DFZ; there would be too many of them.
   Consequently, if a site decides to change providers, it needs to



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 10]

INTERNET-DRAFT                                         February 12, 1999


   renumber all of its nodes using address space given to it by the new
   provider.  The "old" addresses it had used are returned back to its
   previous provider.  To understand this, consider if, from Figure 2,
   Customer3 changes its provider from Provider A to Provider C, but
   does not renumber.  The picture would be as follows:


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

                                  Figure 3

   In Figure 3, Providers A, B and C are all directly connected to each
   other.  In order for Provider B to reach Customers 1, 2, 4 and 5,
   Provider A still only announces the 204.1.0.0/16 aggregate.  However,
   in order for Provider B to reach Customer3, Provider C must announce
   the prefix 204.1.34.0/24.  Prefix 204.1.34.0/24 is called a "more-
   specific" of 204.1.0.0/16; another term used is that Customer3 and
   Provider C have "punched a hole" in Provider A's address block.  From
   Provider B's view, the address space underneath 204.1.0.0/16 is no
   longer cleanly aggregated into a single prefix and instead the
   aggregation has been broken because the addressing is inconsistent
   with the topology; in order to maintain reachability to Customer1-5,
   Provider B must carry two prefixes where it used to have to carry
   only one.

   The example in Figure 3 explains why sites must renumber if existing
   levels of aggregation are to be maintained.  While a small number of
   new exceptions could be tolerated, and certain prefixes have been



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 11]

INTERNET-DRAFT                                         February 12, 1999


   grandfathered, the reality in today's Internet is that there are
   thousands of providers, many with thousands of individual customers.
   It is generally accepted that renumbering of sites is essential for
   maintaining sufficient aggregation.

   The empirical cost of renumbering a site in order to maintain
   aggregation has been the subject of much discussion.  The practical
   reality, however, is that forcing all sites to renumber is difficult
   given the size and wealth of companies that now depend on the
   Internet for running their business.  Thus, although the technical
   community came to consensus that, with the current practice of
   provider-based addressing, address lending was necessary in order for
   the Internet to continue to operate and grow, the reality has been
   that some of CIDR's benefits have been lost because not all sites
   renumber.  It is worth noting that a number of providers today do
   route filtering based, in part, on prefix length; as a result, a site
   which does not renumber may have only partial connectivity to the
   Internet.  That is, a site may advertise a long prefix into the
   routing system, but there is no assurance that all parts of the
   Internet will accept the route; some simply ignore it.

   One unfortunate characteristic of CIDR at an architectural level is
   that the pieces of the infrastructure that benefit from the
   aggregation (i.e., the providers which make up the DFZ) are not the
   pieces that incur the renumbering cost (i.e., the end site).  The
   logical corollary of this statement is that the pieces of the
   infrastructure that do incur cost to achieve aggregation (e.g., sites
   which renumber when they change providers) don't directly see the
   benefit. (The word "directly" is used here because the continued
   operation of the Internet is a benefit, though it requires
   selflessness on the part of the site to recognize.) This can lead to
   a "tragedy of the commons" situation, where everyone agrees that some
   sites should renumber, but they themselves want to be one of those
   that do not.


3.4.  Multi-Homed Sites and Aggregation

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

   In the pre-CIDR days, multi-homed sites were typically known by only
   one network prefix, the prefix of their own address block.  When that
   site's providers announced the site's network into the global routing



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 12]

INTERNET-DRAFT                                         February 12, 1999


   system, a "shortest path" type of routing would occur so that pieces
   of the Internet closest to the first provider might use the first
   provider while other pieces of the Internet would use the second
   provider.  This allowed sites to use the routing system itself to
   load balance traffic across their multiple connections.  This type of
   multi-homing assumes that a site's prefix can be propagated
   throughout the DFZ, an assumption that is no longer universally true.

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

   The second approach is for a multi-homed site to receive an
   allocation from one of its providers and just use that single prefix.
   The site would advertise its prefix to all of the providers to which
   it connects.  There are two problems with this approach.  First,
   although the prefix is aggregatable by the provider which made the
   allocation, it is not aggregatable by the other providers.  To the
   other providers, the site's prefix poses the same problem that a
   provider-independent address would.  Second, due to CIDR's rule for
   longest-match routing, it turns out that the site's prefix is not
   always aggregatable in practice even by the provider that made the
   allocation, if you want shortest-path routing load-spreading.
   Consider Figure 4.  Provider C has two paths for reaching Customer1.
   Provider A advertises 204.1/16, an aggregate which includes
   Customer1.  But Provider C will also receive an advertisement for
   prefix 204.1.0/19 from Provider B, and because the prefix match
   through B is longer, C will choose that path.  In order for Provider
   C to be able to choose between the two paths, Provider A would also
   have to advertise the longer prefix for 204.1.0/19 in addition to the
   shorter 204.1/16.  At this point, from the routing perspective, the
   situation is very similar to the general problem posed by the use of
   provider-independent addresses.

   It should be noted that the above example simplifies a very complex
   issue.  For example, consider the example in Figure 4 again.
   Provider A could choose not to propagate a route entry for the longer
   204.1.0/19 prefix, advertising only the shorter 204.1/16.  In such
   cases, provider C would always select Provider B.  Internally,
   Provider A would continue to route traffic from its other customers
   to Customer1 directly.  If Provider A had a large enough customer
   base, effective load sharing might be achieved.




draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 13]

INTERNET-DRAFT                                         February 12, 1999



                                      A advertises
                     +------------+  204.1/16 to C  +------------+
                  ___| Provider A |-----------------| Provider C |
                 /   +------------+                 +------------+
                /                       +----------/
               /                       /
    Customer1 ---                     / B advertises 204.1.0/19 to C
   204.1.0.0/19  |                   /
                 |      +------------+
                  ----- | Provider B |
                        +------------+


                                Figure 4


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

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

      2) The site could assign each host multiple addresses (i.e., one
        address for each ISP connection).  There are two problems with
        this.  First, it accelerates the consumption of the address
        space. While this may be a problem for the (limited) IPv4
        address space, it is not a significant issue in IPv6.  Second,
        when the connection to ISP-A goes down, addresses numbered out
        of ISP-A's space become unreachable.  Remote peers would have to
        have sufficient intelligence to use the second address.  For
        example, when initiating a connection to a host, the DNS would
        return multiple candidate addresses.  Clients would need to try
        them all before concluding that a destination is unreachable
        (something not all network applications currently do).  In
        addition, a site's hosts would need a significant amount of
        intelligence for choosing the source addresses they use.  A host
        shouldn't choose a source address corresponding to a link that



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 14]

INTERNET-DRAFT                                         February 12, 1999


        is down.  At present, hosts do not have such sophistication.

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


4.  The GSE Proposal

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


4.1.  Motivation For GSE

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

   The GSE proposal does not propose to eliminate the need for
   renumbering.  Indeed, it asserts that end sites will have to renumber
   more frequently in order to continue scaling the Internet.  However,
   GSE proposes to make the cost of renumbering small enough that sites
   can be renumbered at essentially any time with little or no
   disruption to its network connectivity, and in particular with no
   impact on communications that are strictly within the site.



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 15]

INTERNET-DRAFT                                         February 12, 1999


   Finally, GSE attempts to address the problem of sites that have
   multiple Internet connections.  In CIDR, the pressure for better
   multi-homing support can create exceptions to route aggregation and
   result in poor scaling.  That is, the public routing infrastructure
   may have to carry multiple distinct routes for some demanding multi-
   homed sites, one for each independent path.  GSE recognizes the
   "special work done by the global Internet infrastructure on behalf of
   multi-homed sites" [GSE], and proposes a way for multi-homed sites to
   gain certain benefit without impacting global scaling.  This includes
   a specific mechanism that providers can use to support multi-homed
   sites, presumably at a cost that the site would consider when
   deciding whether or not to become multi-homed.


4.2.  GSE Address Format

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

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

                                 Figure 5



4.2.1.  Routing Stuff (RG and STP)

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












draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 16]

INTERNET-DRAFT                                         February 12, 1999



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

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

                                 Figure 6

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

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

   The remaining 14 bits of the Routing Stuff (i.e., the low-order 14
   bits of the high-order 8 bytes) comprise the STP and are used for
   routing structure within a site, similar to subnetting with IPv4.
   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.



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 17]

INTERNET-DRAFT                                         February 12, 1999


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

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


4.2.2.  End-System Designator

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

   The leading contender for the role of a 64-bit globally unique ESD is
   the recently defined "EUI-64" identifier [EUI64].  These identifiers
   consist of a 24-bit "company_id" concatenated with a 40-bit
   "extension".  (Company_id is a new name for the "Organizationally
   Unique Identifier" that forms the first half of an 802 MAC address).
   Manufacturers are expected to assign locally unique values to the
   extension field, guaranteeing global uniqueness for the complete 64-
   bit identifier.  A range of the EUI-64 space is reserved to cover
   pre-existing 48-bit MAC addresses, and a defined mapping insures that
   an ESD derived from a MAC address will not duplicate the ESD of a
   device that has a built-in EUI-64.

   In some cases, interfaces may not have an appropriate MAC address or
   EUI-64 identifier.  A globally unique ESD must then be obtained
   through some alternate mechanism.  Several possible mechanisms can be
   imagined (e.g., the IANA could hand out addresses from the company_id
   it has been allocated).  Although we do not explore them in detail
   here, we note that a global coordination structure is required here
   to control the allocation of globally unique identifiers.



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 18]

INTERNET-DRAFT                                         February 12, 1999


4.3.  Address Rewriting by Border Routers

   To obviate the need to renumber devices within sites because of
   changing providers, the GSE design hides the global Routing Goop (RG)
   from hosts in each site by having site border routers rewrite
   addresses of the packets they forward across the boundary between the
   site and public topology.  Within a site, nodes need not know the RG
   associated with their addresses.  They simply use a designated
   "Site-Local RG" value for internal addresses.  When a packet is
   forwarded to the public topology, the border router replaces the
   Site-Local RG portion of the packet's source address with an
   appropriate value.  Likewise, when a packet from the public topology
   is forwarded into a site, the border router replaces the RG part of
   the destination address with the designated Site-Local RG.

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

   GSE's approach to easing renumbering isn't so much to ease
   renumbering as to make it transparent to end users.  The RG by which
   a site is known is hidden from nodes within that site.  Instead, the
   RG for the site would be known only by the exit router, either
   through static configuration or through a dynamic protocol with an
   upstream provider.

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

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

   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



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 19]

INTERNET-DRAFT                                         February 12, 1999


   later trying to reestablish a connection to that peer may well fail.
   For intra-site communication, however, it is expected that only the
   Site-Local RG would be used (and stored) which would continue to work
   for intra-site communication regardless of changes to the site's
   external RG.  This shields a site's intra-site traffic from any
   instabilities resulting from renumbering.

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

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


4.4.  Renumbering and Rehoming Mid-Level ISPs

   One of the most difficult-to-solve components of the renumbering
   problem with CIDR is that of renumbering mid-level service providers.
   Specifically, if SmallISP1 changes its transit provider from BigISP1
   to BigISP2, then in order for the overall size of the routing tables
   to stay the same, all of SmallISP1's customers would have to renumber
   into address space covered by an aggregate of BigISP2.  GSE deals
   with this problem by handling the RG in DNS with indirection.
   Specifically, a site's DNS server specifies the RG portion of its
   addresses by referencing the "name" of its immediate provider, which
   is a resolvable DNS name (this implies a new Resource Record type).
   That provider may define some of the low-order bits of the RG and
   then reference its immediate provider.  This chain of reference
   allows mid-level service providers to change transit providers, and
   the customers of that mid-level will simply "inherit" the change in



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 20]

INTERNET-DRAFT                                         February 12, 1999


   RG.  Note that this mechanism does not depend on the GSE address
   format per se and can also be applied to IPv4 addressing.


4.5.  Support for Multi-Homed Sites

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

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

                                    Figure 7

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








draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 21]

INTERNET-DRAFT                                         February 12, 1999


4.6.  Explicit Non-Goals for GSE

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

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

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

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

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

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


5.  Analysis: The Pros and Cons of Overloading Addresses

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

   The following discussion is organized around three fundamental
   points:

     1) Identifiers indicate who the intended recipient of a packet is.
        At the network layer, an identifier refers to an interface, at
        the transport layer it refers to a process or other endpoint of
        a "connection".

     2) Identifiers must be mapped into a locator that the network layer
        can use to actually deliver a packet to its intended
        destination.

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



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 22]

INTERNET-DRAFT                                         February 12, 1999


5.1.  Purpose of an Identifier

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

   In IPv4, when applications communicate, transport "identifiers"
   consist of addresses and port numbers.  For the purposes of this
   discussion, we use the term "identifier" to mean the identifier of an
   interface.  It is assumed that port numbers will be present when
   higher layer entities communicate; the exact port numbers used are
   not relevant to this discussion.

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

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

   The most obvious example that could benefit from the separation of



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 23]

INTERNET-DRAFT                                         February 12, 1999


   locators and identifiers involves communication with a mobile host.
   Transport protocols such as TCP are unable to keep connections open
   if either of the two endpoint identifiers for an open connection
   changes.  Fundamentally, the endpoint identifiers indicate the two
   endpoint entities that are communicating.  If a node were to receive
   a packet from a node with which it had been communicating previously,
   but the identifier used by the sending node has changed, the
   recipient would be unable to distinguish this case from that of a
   packet received from a completely different node.

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

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

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

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



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 24]

INTERNET-DRAFT                                         February 12, 1999


5.2.  Mapping an Identifier to a Locator

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

     1) The network layer could be responsible for doing the mapping.
        The advantage of such a system is that an ESD could be stored
        essentially forever (e.g., in configuration files), but whenever
        it is actually used, network layer software would automatically
        perform the mapping to determine the appropriate Routing Stuff
        for the destination.  Likewise, should an existing mapping
        become invalid, network layer software could dynamically
        determine the updated value.  Unfortunately, building such a
        mapping mechanism that scales is difficult if not impossible
        with a flat identifier space (e.g., the ESD identifier).

     2) The transport layer could be responsible for doing the mapping.
        It could perform the mapping when a connection is first opened,
        periodically refreshing the binding for long-running
        connections.  Implementing such a scheme would change the
        existing transport layer protocols TCP and UDP significantly.
        However, in the case of TCP, such a scheme would have the
        benefit that applications would probably not need to be
        modified.  For UDP-based applications, this may not hold, since
        most UDP-based protocols are implemented within applications.

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

   The GSE proposal uses the last approach.  The network and transport
   layers are always presented with both the Routing Stuff (RG + STP)
   and the ESD together in one IPv6 address.  It is neither of these
   layers' jobs to determine the Routing Stuff given only the ESD or to
   validate that the Routing Stuff is correct.  When an application has
   data to send, it queries the DNS to obtain the IPv6 AAAA record for a
   destination.  The returned AAAA record contains both the Routing
   Stuff and the ESD of the specified destination.  While such an



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 25]

INTERNET-DRAFT                                         February 12, 1999


   approach eliminates the need for the lower layers to be able to map
   ESDs into corresponding Routing Stuff, it also means that when
   presented with an address containing an incorrect (i.e., no longer
   valid) Routing Stuff, the network is unable to deliver the packet to
   its correct destination.  Note that addresses containing invalid
   Routing Stuff will result any time when cached addresses are used
   after the Routing Stuff of the address becomes invalid.  This may
   happen if addresses are stored in configuration files, a mobile node
   moves to a new location, long-running applications (clients and
   servers) cache the result of DNS queries, a long-running connection
   attempts to continue operating during a site renumbering event, etc.
   Whatever the causes, the failures are fundamentally due to dynamic
   topological changes at the network layer, yet in GSE such failures
   are left to be dealt with at the application level (through DNS),
   because neither the transport nor the network level has the ability
   to re-mapping identifiers to corresponding locators.

   To avoid the above problem a network architecture must provide the
   ability to map an identifier to a locator.  In IPv4, this mapping is
   trivial, since the identifier and locator are combined in a single
   quantity (i.e., the IPv4 address).  GSE does not provide this mapping
   functionality directly.  Instead, GSE assumes that a node's DNS name
   serves as its stable identifier, and uses normal DNS queries to map
   the DNS "identifier" into an IPv6 address.  The IPv6 address contains
   both the ESD identifier together with its Routing Stuff, that is an
   initial binding/mapping between the identifier and locator.  When
   this binding breaks (for example due to dynamic topological changes),
   the ESD identifier cannot be mapped into a new locator by itself.
   Instead one must resort back to application level, hoping another DNS
   query would provide rescue to the broken binding between identifier
   to locator that is needed for network delivery.

   The use of DNS to provide identifier to locator mapping contributes
   to GSE's apparent simplicity.  However, there are two fundamental
   problems with this approach, if the intention is to make it
   transparently easy to change locators over time.  First, the burden
   of performing the mapping from identifier to locator is placed
   directly on the application, because lower layers (i.e., transport
   and network layers) cannot perform the mapping themselves due to
   layering violation concerns (i.e., TCP and UDP can't perform a DNS
   query).  Second, following all RG changes the DNS database must be
   promptly updated and all expired information must be flushed out of
   all DNS caches.  This stringent timing requirement imposed by lower
   level operation would represent a departure from the original DNS
   design, which provides DNS names to address mappings that only change
   slowly over time if at all, and which relies heavily on caching over
   relatively long time periods to scale well.




draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 26]

INTERNET-DRAFT                                         February 12, 1999


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


5.2.1.  Scalable Mapping of Identifiers to Locators

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

   Imposing a hierarchy on identifiers poses the following difficulties:

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

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

      - - Due to the requirement of tying an identifier to the
        delegation structure the identifier of a node cannot be burned
        in during manufacturing.  Instead a mechanism is needed to allow
        a node to learn its identifier.  To be practical, such a
        mechanism would need to be automated and avoid the need for
        manual configuration.


5.2.2.  Insufficient Hierarchy Space in ESDs

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



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 27]

INTERNET-DRAFT                                         February 12, 1999


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

   Finally, imposing a required hierarchical structure on identifiers
   such as an ESD would also introduce a new administrative burden and a
   new or expanded registry system to manage ESD space (i.e., to insure
   that ESDs are globally unique).  While the procedures for assigning
   ESDs, which need only organizational and not topological
   significance, would be simpler than the procedures for managing IPv4
   addresses, it seems a laudable goal to avoid the problem altogether
   if possible.  In addition, it would likely increase the complexity
   for connecting new nodes to the Internet, a goal inconsistent with
   Stateless Address autoconfiguration [RFC1971].

   The topic of mapping full 16-byte GSE addresses to a locator or other
   information is discussed in Appendix D.


5.3.  Authentication of Identifiers

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

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





draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 28]

INTERNET-DRAFT                                         February 12, 1999


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-04.txt                          [Page 29]

INTERNET-DRAFT                                         February 12, 1999


   It should be noted that although an address containing a combined
   identifier and locator can be forged, the routing subsystem
   significantly limits communication using the forged address.  First,
   return traffic will be sent to the correct destination and not the
   originator of the forged address.  This alone prevents certain types
   of spoofing attacks.  For example, if a destination receives an
   unexpected packet corresponding to a TCP connection that it is
   unaware of, it may return at TCP segment resetting the connection.
   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.


5.4.  Transport Layer: What Locator Should Be Used?

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

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




draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 30]

INTERNET-DRAFT                                         February 12, 1999


    - - the sending side of an active open

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

    - - changes to the Routing Stuff during an open connection.


5.4.1.  RG Selection On An Active Open

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


5.4.2.  RG Selection On An Passive Open

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


5.4.3.  Mid-Connection RG Changes

   While packets are flowing as part of an open connection, the RG
   appearing on subsequent packets is susceptible to change through
   renumbering events, or as a result of site-internal routing changes
   that cause the egress point for off-site traffic to change.  It is
   even possible that traffic-balancing schemes could result in the use
   of two egress routers, with roughly every other packet exiting
   through a different egress router.

   Because TCP under GSE demultiplexes packets using only ESDs, newly
   arrived packets will be delivered to the correct end-point regardless



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 31]

INTERNET-DRAFT                                         February 12, 1999


   of whether their source RG have changed.  The GSE proposal calls for
   return traffic to continue to be sent via the "old" RG, even though
   it may have been deprecated or become less optimal because the peer's
   border router has changed.  That is, the RG to use for reaching a
   peer is bound to a connection when the connection is established and
   does not change thereafter.  However, the completion of renumbering
   events (so that an earlier RG is now invalid) and certain topology
   changes would require TCP to switch sending to a new RG mid-
   connection.  To explore the scenario, we consider ways of allowing
   the RG change to be made to existing established connections.

   If TCP connection identifiers are based on ESDs rather than full
   addresses, traffic from the same ESD would be viewed as coming from
   the same peer, regardless of the source RG.  Because this
   vulnerability is already present in today's Internet (forging the
   source address of a packet is trivial), the mere delivery of incoming
   datagrams with the same ESD but a different RG does not introduce new
   vulnerability to TCP.  In today's Internet, any node can already
   originate FINs/RSTs from an arbitrary source address and potentially
   or definitely disrupt the connection.  Therefore, acceptance of
   traffic independent of its source RG does not appear to significantly
   worsen existing robustness.  Note, however, that ingress filtering as
   described in Section 5.3.1, cannot be performed on packets containing
   GSE addresses.  This does make it more difficult to prevent certain
   types of attacks.

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

   Providing protection from hijacking implies that the RG used to send
   packets must be bound to a connection end-point (e.g., it is part of
   the connection state).  Although it may be reasonable to accept
   incoming traffic independent of the source RG, the choice of sending
   RG requires more careful consideration.  Indeed, any subsequent
   change in the RG used for sending traffic must be properly
   authenticated (e.g., using cryptographic means).  In the GSE
   proposal, the is no apparent way to authenticate such a change, since
   the remote peer doesn't even know its own RG.  Consequently, the only
   reasonable approach in GSE is to send to the peer using the first RG
   used for the entire life of a connection.  That is, always use the
   first RG seen, and accept the loss of connectivity whenever the RG
   changes.



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 32]

INTERNET-DRAFT                                         February 12, 1999


5.4.4.  The Impact of Corrupted 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),
   no TCP mechanism can detect such corruption at the receiver.
   Moreover, once a specific RG is in use, it does not change for the
   duration of a connection.  One interesting case occurs on the passive
   side of a TCP connection, where a server accepts incoming connections
   from remote clients.  If the initial SYN from the client includes a
   corrupted RG, the server TCP will create a TCP connection (in the
   SYN-RECEIVED state) and cache the corrupted RG with the connection.
   The second packet of the 3-way handshake, the SYN-ACK packet, would
   be sent to the wrong RG and consequently not reach the correct
   destination.  Later, when the client retransmits the unacknowledged
   SYN, the server will continue to send the SYN-ACK using the bad RG.
   Eventually the client times out, and the attempt to open a TCP
   connection fails.

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

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

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



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 33]

INTERNET-DRAFT                                         February 12, 1999


   having a significant TCPng component, a significant drawback.


5.5.  On The Uniqueness Of ESDs

   Although ESDs are expected to be globally unique, their uniqueness
   property may be violated either due to mistakes in allocation or by
   malicious attacks.  The exact uniqueness requirements for ESDs
   depends on what purpose they serve and how they are used.  If the
   correctness of some applications relies on the global uniqueness of
   ESDs, then active checking and enforcement will be necessary.  On the
   other hand if ESDs are used only to uniquely identify individual
   endpoints within a session, then one may consider global uniqueness
   as unnecessary.



5.5.1.  Impact of Duplicate ESDs

   Consider what happens when two nodes using the same ESD attempt to
   communicate with each other.  In the GSE proposal, a node queries the
   DNS to obtain an IPv6 address.  The returned address includes the
   Routing Stuff of an address (the RG+STP portions).  The sender may
   not notice the destination ESD is the same as its own ESD and may
   well forward the packet to a router that delivers the packet to its
   correct destination (using the information in the Routing Stuff).  On
   receipt of the packet, however, the destination node would extract
   the ESD portion of the destination address and detect the conflict.

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

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

     1) Following the GSE specification, the transport end-point would
        accept the packet, without regard to the Routing Stuff of the
        source address.  This may lead to a number of robustness
        problems (and at best will confuse the application).

     2) The transport end-point could verify that the Routing Stuff of



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 34]

INTERNET-DRAFT                                         February 12, 1999


        the source address matches one of a set of expected values
        before processing the packet further.  If the Routing Stuff
        doesn't match any expected value, the packet could be dropped.
        This would result in a connection from one host operating
        correctly, while a connection from another host (using the same
        ESD) would fail.

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


5.5.2.  New Denial of Service Attacks.

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

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


5.6.  Summary of Identifier Authentication Issues

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

     1) It can't be done on an end-to-end basis in GSE (e.g., via IPSec)



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 35]

INTERNET-DRAFT                                         February 12, 1999


        because the sender doesn't know what value the RG portion of the
        address will have when it reaches the receiver.

     2) It can't be easily done in GSE because there is no mechanism at
        or below the transport layer to map ESDs into a quantity that
        can be used as a key to jump start the authentication process
        (using the DNS would be problematic due to layering circularity
        considerations).

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

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

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

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

     3) Simultaneously, we understand that, with the above rules, there
        are still open issues with regard to invalid RGs, either through
        corruption or through a active hostile attacks.

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

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



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 36]

INTERNET-DRAFT                                         February 12, 1999


6.  Conclusion

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

   Functionally speaking, identifiers and locators each have a logically
   different role to play.  Thus overloading both in one field causes
   problems whenever the location of a node changes but its identity
   does not.  However, our analysis shows that overloading also presents
   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 hierarchy and support stateless address
   autoconfiguration.  Instead, GSE forces applications to supply up-
   to-date locators.  However, relying on the locator provided at the
   time communication is established as GSE does is inadequate when the
   remote locator can change dynamically, precisely the scenario that is
   supposed to benefit from the separation.  That is, the benefits of
   separating the identifier from the locator are largely lost, if the
   changes in the identifier to locator binding are not tracked quickly.

   Secondly, when communicating with a remote site, if the RG changes
   there begins to be uncertainty as to whether a reliable TCP handshake
   is possible (because of the need for passively opened TCP to use the
   RG's it obtains from the packets).  Because the reliability of TCP's
   byte stream is critically dependent on its three-way handshake, this
   is a significant issue.

   Finally, when communicating with a remote site, a receiver must be
   able to insure (with reasonable certainty) that received data does
   indeed come from the expected remote entity.  In IPv4, it is possible
   to receive packets from a forged source, but the potential for
   mischief between communicating peers is significantly limited because
   return traffic will not generally reach the source of the forged
   traffic.  That is, communication involving packets sent in both
   directions will not succeed.  In contrast, architectures like GSE
   that decouple the identifier and locator functions lose the built-in
   protection available in classical IP and thus face great difficulty
   assuring that traffic from a source identified only by an identifier



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 37]

INTERNET-DRAFT                                         February 12, 1999


   actually comes from the correct source.  Short of using cryptographic
   techniques (e.g. IPsec), there is no known mechanism that can use an
   identifier alone to perform this remote entity authentication.  Using
   an identifier alone for authentication of received packets is
   dangerously unsafe.

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


7.  Security Considerations

   The primary security consideration with GSE or, more generally, a
   network layer with addresses split into locator and identifier parts,
   is that of one node impersonating another by copying the
   identification without the location.  Indeed, the main conclusion of
   this paper is that a GSE-like addressing structure introduces new
   security vulnerabilities that are not present in IP, and that those
   problems are serious enough to question the benefits of an
   architecture that separates locaters and identifiers in addresses.


8.  Acknowledgments

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


9.  References

     [ANYCAST] "Host Anycasting Service", C. Partridge, T. Mendez, & W.
             Milliken, RFC 1546.

     [BATES] Scalable support for multi-homed multi-provider
             connectivity, Tony Bates & Yakov Rekhter, RFC 2260,
             January, 1998.



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 38]

INTERNET-DRAFT                                         February 12, 1999


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

     [CIDR] "Classless Inter-Domain Routing (CIDR): an Address
             Assignment and Aggregation Strategy". V. Fuller, T. Li, J.
             Yu, & K. Varadhan, RFC 1519, September 1993.

     [DHCP-DDNS] Interaction between DHCP and DNS, Internet Draft, Yakov
             Rekhter, (Work in Progress.)

     [DDNS] "Dynamic Updates in the Domain Name System (DNS UPDATE)",
             Paul Vixie (Editor), RFC 2136, April, 1997.

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

     [GSE] "GSE - An Alternate Addressing Architecture for IPv6", Mike
             O'Dell, (Work in progress).

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

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

     [IPv6-ADDRESS] "An IPv6 Aggregatable Global Unicast Address
             Format", R. Hinden, M. O'Dell, S. Deering, RFC 2374, July,
             1998.

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

     [RFC1752] "The Recommendation for the IP Next Generation Protocol",
             S. Bradner, A. Mankin, RFC 1752, January, 1995.

     [RFC1788] "ICMP Domain Name Messages", W. Simpson, RFC 1788, April,
             1995.

     [RFC1884] "IP Version 6 Addressing Architecture", R. Hinden & S.
             Deering, Editors, RFC 1884.



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 39]

INTERNET-DRAFT                                         February 12, 1999


     [RFC1958] "Architectural Principles of the Internet", B. Carpenter,
             RFC 1958, June, 1996.

     [RFC1971] "IPv6 Stateless Address Autoconfiguration", S. Thomson,
             T. Narten, RFC 1971, August, 1996.

     [RFC2008] "Implications of Various Address Allocation Policies for
             Internet Routing", Y. Rekhter, T. Li, RFC 2008, October
             1996.

     [RFC2073] An IPv6 Provider-Based Unicast Address Format.  Y.
             Rekhter, P. Lothberg, R. Hinden, S. Deering, J. Postel. RFC
             2073, January, 1997.

     [RFC2267] Network Ingress Filtering: Defeating Denial of Service
             Attacks which employ IP Source Address Spoofing, P.
             Ferguson, D. Senie, RFC 2267.

     [ROUTER-RENUM] "Router Renumbering for IPv6", M. Crawford, draft-
             ietf-ipngwg-router-renum-06.txt.


10.  Authors' Addresses

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

   Allison Mankin                          Lixia Zhang
   USC/ISI                                 UCLA Computer Science Department
   4350 North Fairfax Drive                4531G Boelter Hall
   Suite 620                               Los Angeles, CA 90095-1596 USA
   Arlington, VA  22203 USA                Phone: 310-825-2695
   EMail: mankin@isi.edu                   EMail: lixia@cs.ucla.edu
   Phone: 703-812-3706

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





draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 40]

INTERNET-DRAFT                                         February 12, 1999


Appendix A: Increased Reliance on Domain Name System (DNS)

   As we've discussed in previous sections, the motivation for
   separating identifiers from locators in IP address is to allow the
   locator portion to change more easily.  However because GSE does not
   provide a mapping from an ESD to its locator, whenever the locator
   changes, GSE falls back on DNS to provide such mapping.

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


A.1: Renumbering and DNS: How Frequently Can We Renumber?

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

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




draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 41]

INTERNET-DRAFT                                         February 12, 1999


A.2: Efficient DNS support for Site Renumbering

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

   With GSE, When a site renumbers to satisfy its ISP, only the site's
   routing prefix needs to change.  That is, the prefix reflects where
   within the Internet the site resides.  One DNS modification that
   could reduce the cost of updating the DNS when a site is renumbered
   is to store addresses in two distinct RR's: one for the Routing Goop
   that reflects where a node attaches to the Internet and the other for
   STP-plus-ESD that is the site-specific part of an address.  During a
   renumbering, the Routing Goop would change, but the "site internal
   part" would remain fixed.  That way, renumbering a site would only
   require that the Routing Goop RR of a site be updated; the "site-
   internal part" of individual addresses would not change.

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

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

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

   If AAAA records are comprised of multiple distinct RRs, then one
   question is who should be responsible for synthesizing the AAAA from
   its components: the resolver running on the querying client's machine
   or the queried name server? To minimize the impact on client hosts
   and make it easier to deploy future changes, it is recommended that
   the synthesis of AAAA records from its constituent parts be done on
   name servers rather than in client resolvers.



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 42]

INTERNET-DRAFT                                         February 12, 1999


A.2.1: Two-Faced DNS

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

   Some issues that must be considered in this context:

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

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



A.2.2: Bootstrapping Issues

   If Routing Stuff information is distributed via the DNS, key DNS
   servers must always be reachable.  In particular, the addresses
   (including Routing Stuff) of all root DNS servers are, for all
   practical purposes, well-known and assumed to never change.  It is
   not uncommon for the addresses of root servers to be hard-coded into
   software distributions.  Consequently, the Routing Stuff associated
   with such addresses must always be usable for reaching root servers.



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 43]

INTERNET-DRAFT                                         February 12, 1999


   If it becomes necessary or desirable to change the Routing Stuff of
   an address at which a root DNS server resides, the routing subsystem
   will likely need to continue carrying "exceptions" for those
   addresses.  Because the total number of root DNS servers is
   relatively small, the routing subsystem is expected to be able to
   handle this requirement.

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







































draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 44]

INTERNET-DRAFT                                         February 12, 1999


Appendix B: Additional Issues Related to GSE

   This paper focused primarily on the issues of separating identifiers
   and locators in unicast addresses.  It is worth noting that a number
   of additional issues were identified during the IPng interim meeting
   with respect to the GSE proposal that would need to be considered
   before an architecture such as GSE could be deployed.  Specifically:

      - - it is not known how multicast would work under GSE.  One
        identified issue is that a site with multiple egress routers
        would (by default) inject multicast traffic through each of all
        the egress routers, each would then replace the source Routing
        Goop with a differing value.  This would lead to multiple copies
        of the same packet each carrying a different IPv6 address, thus
        being considered as from different sources.

      - - It would be more difficult to create tunnels.  Any tunnel that
        crosses a site boundary (i.e., the entry and exit points are in
        differing sites) would in effect require that both tunnel
        endpoints be border routers to insure that the addresses in the
        inner headers were rewritten correctly.

      - - In order for the DNS to hide a site's Routing Goop from
        internal nodes yet make it visible to external nodes requires a
        two-faced DNS.  The current DNS model assumes a single global
        database in which all queries are answered the same way,
        irregardless of who issued the query.  It is unclear how to make
        the DNS answer queries in a context-sensitive manner without
        also negatively impacting its caching model.






















draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 45]

INTERNET-DRAFT                                         February 12, 1999


Appendix C: Ideas Incorporated Into IPv6

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

   The unicast address format was changed to improve the aggregability
   of unicast addresses.  Instead of a topologically insignificant
   Registry ID immediately following the Format Prefix [RFC2073], there
   is now a Top-Level Aggregation Identifier [IPv6-ADDRESS].  This field
   identifies a large routable aggregate to which an address belongs
   rather than an administrative unit that assigned the address.  The
   TLA corresponds to the "Large Structure" of GSE.  The IPv6 Next-Level
   Aggregation Identifier (NLA) is roughly the rest of the GSE "Routing
   Goop" and the Site-Level Aggregation Identifier (SLA) is a slightly
   expanded GSE Site Topology Partition.

   The decision to put fixed boundaries between parts of the unicast
   address (TLA, NLA, SLA, Interface Identifier) into IPv6 addresses
   [IPv6-ADDRESS] also came from GSE.  The previous "provider-based"
   addressing architecture for IPv6 [RFC2073] had fluid boundaries
   between Registry ID, Provider ID, Subscriber ID and the Intra-
   Subscriber part, as well as undefined divisions within the Provider-
   ID and Intra-Subscriber part.  (On subnetworks with a MAC-layer
   address, the latter boundary was generally placed to accommodate use
   of that address as an Interface ID.)  The new addressing architecture
   still expects divisions within the NLA portion of the address, placed
   to reflect topological aggregation points.

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

   Another change was the redefinition of the interface identifier to be
   a 64-bit quantity.  In the common case where a node has at least one
   IEEE interface, the interface identifier is constructed from an IEEE
   identifier (i.e., a MAC address) in such a way that there is a very
   high probability that the identifier will be globally unique.  In the
   case where a globally unique identifier can't easily be constructed
   automatically, a bit in the identifier indicates that the address is
   not globally unique.  At present, there are no plans for transport
   protocols such as TCP to exploit interface identifiers, but the door
   has been left open for a future protocol (e.g., TCPng) to take
   advantage of the ESD concept.

   Another change to come out of the GSE discussions relates to reducing



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 46]

INTERNET-DRAFT                                         February 12, 1999


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

   A change in the method of doing DNS address-to-name lookups is also
   in the works.  This may be a change in the form and/or operation of
   the ip6.int domain or some new mechanism which involves participation
   by the routers or the end-nodes themselves.

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


Appendix D: Reverse Mapping of Complete GSE Addresses

   The ability to map an IP address into its corresponding DNS name is
   used in several contexts:

     1) Network packet tracing utilities (e.g., tcpdump) display the
        contents of packets.  Printing out the DNS names appearing in
        those packets (rather than dotted IP addresses) requires access
        to an address-to-name mapping mechanism.

     2) Some applications perform a "poor-man's" authentication by using
        the DNS to map the source address of a peer into a DNS name.
        The client then queries the DNS a second time, this time asking
        for the address(es) corresponding to the peer's DNS name.  Only
        if one of the addresses returned by the DNS matches the peer
        address of the TCP connection is the source of the TCP
        connection accepted as being from the indicated DNS name.

        It is important to note that although two DNS queries are made
        during the above operation, it is the second one --- mapping the
        peer's DNS name back into an IP address --- that provides the
        authentication property.  The first transaction simply obtains
        the peer's DNS name, but no assumption is made that the returned
        DNS name is correct.  Thus, the first DNS query could be
        replaced by an alternate mechanism without weakening the already
        weak authentication check described above.  One possible



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 47]

INTERNET-DRAFT                                         February 12, 1999


        alternate mechanism, an ICMP "Who Are You" message, is described
        below

     3) Applications that log all incoming network connections (e.g.,
        anonymous FTP servers) may prefer logging recognizable DNS names
        to addresses.

     4) Network administrators examining logs or other trace data
        containing addresses may wish to determine the DNS name of some
        addresses.  Note that this may occur sometime after those
        addresses were actually used.

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


D.1: DNS-Like Reverse Mapping of Full GSE Addresses

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

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



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 48]

INTERNET-DRAFT                                         February 12, 1999


   As a consequence, a packet trace recorded in the past might not
   contain enough information to identify the off-Site sources of the
   packets in the present.  This problem can be addressed by requiring
   that the database of RG delegations be maintained, together with
   accurate timing information, for some period of time after the RG is
   no longer usable for routing packets.

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


D.2: The ICMP Who-Are-You Message

   There is widespread agreement on the utility of being able to
   determine the DNS name one is communicating with from the address
   being used.  In addition to the fact that DNS names are more
   meaningful to human users and more stable than addresses, many users
   use this reverse mapping as part of a poor-man's authentication for
   the remote peer; if one can map the obtained DNS name back to the
   same address, one has an increased confidence of the peer being a
   legitimate one.

   In practice, however, the IN-ADDR.ARPA domain is not fully populated
   and poorly maintained.  Consequently, an old proposal to define an
   ICMP Who-Are-You message was resurrected [RFC1788].  A client would
   send such a message to a peer, and that peer would return an ICMP
   message containing its DNS name.  Asking a remote host to supply its
   own name in no way implies that the returned information is accurate.
   However, having a remote peer provide a piece of information that a
   client can use as input to a separate authentication procedure
   provides a starting point for performing strong authentication.  The
   actual strength of the authentication depends on the authentication
   procedure invoked, rather than the untrustable piece of information
   provided by a remote peer.

   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



draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 49]

INTERNET-DRAFT                                         February 12, 1999


   strength of the authentication method.  Indeed, it can only make it
   more useful in practice, because virtually all hosts can be expected
   to implement the Who-Are-You message.

   The Who-Are-You message has advantages outside the context of GSE as
   well, including a more decentralized, and hence more scalable,
   administration and easier upkeep than a DNS reverse-lookup zone.  It
   also has drawbacks: it requires the target node to be up and
   reachable at the time of the query and to know its fully qualified
   domain name.  It is also not possible to resolve addresses once those
   addresses become unroutable.  In contrast, the DNS PTR mirrors, but
   is independent of, the routing hierarchy.  The DNS can maintain
   mappings long after the routing subsystem stops delivering packets to
   certain addresses.

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



























draft-ietf-ipngwg-esd-analysis-04.txt                          [Page 50]


Html markup produced by rfcmarkup 1.111, available from https://tools.ietf.org/tools/rfcmarkup/