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

Versions: 00 01 02 03 04 05                                             
INTERNET-DRAFT                                             Matt Crawford
                                                                Fermilab
<draft-ietf-ipngwg-esd-analysis-00.txt>                   Allison Mankin
                                                                     ISI
                                                           Thomas Narten
                                                                     IBM
                                                    John W. Stewart, III
                                                                     ISI
                                                             Lixia Zhang
                                                                    UCLA
                                                              March 1997
                     IPng Analysis of the GSE Proposal

                  <draft-ietf-ipngwg-esd-analysis-00.txt>


Status of this Memo

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

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

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

   Distribution of this memo is unlimited.

   This Internet Draft expires September, 1997.


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 three portions:
   a globally unique End System Designator (ESD), a Site Topology
   Partition (STP) and a Routing Goop (RG) portion. The STP corresponds
   (roughly) to a site's subnet portion of an IPv4 address, whereas the



draft-ietf-ipngwg-esd-analysis-00.txt                             [Page 1]


INTERNET-DRAFT                                                March 1997


   RG identifies the attachment point to the public Internet. Routers
   use the RG+STP portions of addresses to route packets to the link to
   which the destination is directly attached; the ESD is used to
   deliver the packet across the last hop link. An important idea in GSE
   is that nodes within a Site would not need to know the RG portion of
   their addresses. Border routers residing between a Site and its
   Internet connect point would dynamically replace the RG part of
   source addresses of all outgoing IP datagrams, and the RG part of
   destination addresses on incoming traffic.

   This document provides a detailed analysis of the GSE plan. Much of
   the analysis presented here is an expansion of official meeting
   minutes, though it also includes issues uncovered by the authors in
   the process of fully fleshing out the analysis. In summary, the
   consensus of the attendees of the PAL1 meeting was that having
   routers rewrite the Routing Goop portion of addresses should not be
   adopted, though other parts of the GSE plan should (e.g., having
   globally unique ESDs). After completing the first draft of this
   document, the authors still strongly concur with this outcome.

   As a first draft, this document should not be considered to represent
   the views of the IPng Working Group. Instead, it should be viewed as
   the rough consensus of the PAL1 attendees and the strong consensus of
   the five authors. It is hoped that this first draft of the document
   will be the catalyst for discussions to refine the written analysis,
   and especially the conclusions, so that after some number of
   iterations it will represent the consensus of the working group.
























draft-ietf-ipngwg-esd-analysis-00.txt                             [Page 2]


INTERNET-DRAFT                                                March 1997


   Contents

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

   1.  Introduction.............................................    4

   2.  Addressing and Routing in IPv4...........................    5
      2.1.  The Need for Aggregation............................    6
      2.2.  The Pre-CIDR Internet...............................    7
      2.3.  CIDR and Provider-Based Addressing..................    8
      2.4.  Multihoming and Aggregation.........................   11

   3.  GSE Background...........................................   12
      3.1.  Motivation For GSE..................................   12
      3.2.  GSE Address Format..................................   13
      3.3.  Routing Stuff (RG and STP)..........................   14
      3.4.  End-System Designator...............................   15
      3.5.  Address Rewriting by Border Routers.................   16
      3.6.  Renumbering and Rehoming Mid-Level ISPs.............   17
      3.7.  Support for Multihomed Sites........................   18
      3.8.  Explicit Non-Goals for GSE..........................   19

   4.  Analysis of GSE's Advantages and Disadvantages...........   19
      4.1.  End System Designator...............................   19
         4.1.1.  IP Addresses in the IPv4 Internet..............   19
         4.1.2.  Overloading Addresses: Network Layer Issues....   20
         4.1.3.  Overloading Addresses: Transport Layer Issues..   22
         4.1.4.  Benefits of Globally Unique ESDs...............   23
         4.1.5.  ESD: Network Layer Issues......................   24
         4.1.6.  ESD: Transport Layer Issues....................   25
         4.1.7.  ESD: Application Layer Issues..................   32
         4.1.8.  When ESDs are Not Unique.......................   34
         4.1.9.  DNS PTR Queries................................   36
         4.1.10.  Reverse Mapping of ESDs.......................   38
         4.1.11.  Reverse Mapping of Complete GSE Addresses.....   39
         4.1.12.  The ICMP ``Who Are You'' Message..............   40
      4.2.  Renumbering and Domain Name System (DNS) Issues.....   41
         4.2.1.  How Frequently Can We Renumber?................   41
         4.2.2.  Efficient DNS support for Site Renumbering.....   42
         4.2.3.  Synthesizing AAAA Records......................   43
         4.2.4.  Two-Faced DNS..................................   43
         4.2.5.  Bootstrapping Issues...........................   44
         4.2.6.  DNS PTR RRs Not Needed.........................   45
         4.2.7.  Renumbering and Reverse DNS Lookups............   45
      4.3.  Address Rewriting Routers...........................   46
         4.3.1.  Load Balancing.................................   46
         4.3.2.  End-To-End Argument: Don't Hide RG from Hosts..   47
      4.4.  Multi-homing........................................   47



draft-ietf-ipngwg-esd-analysis-00.txt                             [Page 3]


INTERNET-DRAFT                                                March 1997


   5.  Recommendations..........................................   49

   6.  Security Considerations..................................   50

   7.  Acknowledgments..........................................   50

   8.  References...............................................   51

   9.  Authors' Addresses.......................................   52


1.  Introduction

   In October of 1996, Mike O'Dell published an Internet-Draft (dubbed
   ``8+8'' that proposed significant changes to the IPv6 addressing
   architecture. The 8+8 proposal was the topic of considerable
   discussion at the December, 1996 IETF meeting in San Jose. Because
   the proposal offered both potential benefits (e.g., enhanced routing
   scalability) and risks (e.g., changes to the basic IPv6
   architecture), the IPng Working Group held an interim meeting on
   February 27-28, 1997 to consider adopting the 8+8 proposal. The
   meeting, at which over 45 persons attended, was held at Sun
   Microsystems' PAL1 facility in Palo Alto, CA.

   Shortly before the interim meeting, an updated version of the
   Internet-Draft was produced, in which the name of the proposal was
   changed from ``8+8'' to ``GSE,'' for the three separate components of
   the address: Global, Site and End-System Designator. This last
   version of the GSE proposal was published as an Informational RFC
   [GSE] for historical purposes.

   The stated purpose of the meeting was to evaluate the GSE proposal
   and make a firm decision to either:

     1) Definitely adopt GSE for IPv6,

     2) Adopt GSE contingent upon certain other documents being
        successfully completed by the April, 1997 IETF, or

     3) Definitely don't adopt it.

   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 by the attendees that the GSE proposal as written presented
   too many risks and should not be adopted as the basis for IPv6.
   However, the attendees also concluded that some of the issues
   discussed in the GSE proposal were equally applicable to the current



draft-ietf-ipngwg-esd-analysis-00.txt                             [Page 4]


INTERNET-DRAFT                                                March 1997


   IPv6 provider-based addressing plan and had enough benefit to warrant
   making changes to some existing IPv6 documents. These changes
   include:

     1) Making changes to the IPv6 provider-based addressing document,
        to facilitate increased aggregation.

     2) Creating hard boundaries in IPv6 addresses to clearly
        distinguish between the portions used to identify hosts, for
        routing within a Site, and for routing within the Public
        Internet.

     3) Designating the low-order 8 bytes of IPv6 addresses to be a
        globally unique End System Designator (ESD). This change has
        potential benefits to future transport protocols (e.g., TCPng).

     4) Make a clear distinction between the ``locator'' part of an
        address and the ``identifier'' part of the address. The former
        is used to route a packet to its end point, the latter is used
        to identify an end point, independent of the path used to
        deliver the packet. Although this is a potentially revolutionary
        change to IPv6 addressing model, existing transport protocols
        such as TCP and UDP will not take advantage of the split. Future
        transport protocols (e.g., TCPng), however, may.

     5) Making changes to the way AAAA records are stored within the
        DNS, so that renumbering a Site (e.g., when a Site changes ISPs)
        requires few changes to the DNS database in order to effectively
        change all of a Site's address AAAA RRs.

   The remainder of this document attempts to capture the debate and
   discussion that led to the above changes.


2.  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 the
   benefits of GSE, one must understand what problems in IPv4 it alleges
   to improve or fix.

   The structure and semantics of a network layer protocol's addresses
   are absolutely core to that protocol. Addressing substantially
   impacts the way packets are routed, the ability of a protocol to
   scale and the kinds of functionality higher layer protocols can
   provide. Indeed, addressing is intertwined with both routing and



draft-ietf-ipngwg-esd-analysis-00.txt                             [Page 5]


INTERNET-DRAFT                                                March 1997


   transport layer issues; a change in any one of these can impact
   another. Issues of administration and operation (e.g., address
   allocation and required renumbering), while not part of the pure
   exercise of engineering a network layer protocol, turn out to be
   critical to the viability of that protocol in a global and commercial
   network. The interaction between addressing, routing and especially
   aggregation, is particularly relevant to this document, so some time
   will be spent describing it.

   Addresses in IPv4 serve two purposes:

     1) Unique identification of an interface. That is, the IP address
        by itself identifies which interface a packet should be
        delivered to.

     2) Location information of that interface. Routers extract location
        information from packets in order to route them towards their
        ultimate destination. That is, addresses identify ``where'' the
        intended recipient is located within the Internet topology.

   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
   ``over-loading'' 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 involves those transport protocols with
   routing. This entanglement is inconsistent with a strictly layered
   model in which routing would be a completely independent function of
   the network layer and not directly impact the transport layer.

   In addition to architectural uncleanliness, combining the locator and
   identifier has the practical impact of complicating the support for
   mobility. In a mobile environment, the location of an end-station may
   change even though its identity stays the same; transport connections
   should be able to survive such changes. In IPv4, however, one cannot
   change the locator without also changing the identifier.
   Consequently, conventional wisdom for some time has been that having
   separate values for location and identification could be of
   significant benefit. The GSE proposal attempts to make such a
   separation.


2.1.  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 purpose of adding subnetting was to allow



draft-ietf-ipngwg-esd-analysis-00.txt                             [Page 6]


INTERNET-DRAFT                                                March 1997


   a collection of tens or hundreds of networks located at one site to
   be viewed from afar as being just one IP network (i.e., to aggregate
   all of the individual networks into one bigger network). A practical
   benefit of subnetting was that all of a site's hosts, even if
   scattered among tens or hundreds of LANs, could be reached via a
   single routing table entry in routers located far from the site. In
   contrast, prior to subnetting, a site with ten LANs might advertise
   ten separate routing table entries to the routing subsystem of the
   Internet.

   The benefits of aggregation should be clear. The amount of work
   involved in computing forwarding tables from routing tables 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 individual networks is
   advertised to the global routing subsystem as individual routing
   entries, the complexity of computing forwarding tables can easily be
   an order of magnitude greater than if each site advertised just a
   single route that covered all of the addresses used within the site.


2.2.  The Pre-CIDR Internet

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

















draft-ietf-ipngwg-esd-analysis-00.txt                             [Page 7]


INTERNET-DRAFT                                                March 1997



                +----------------+
                |                |------- 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 addresses. 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 each of the 5 networks to
   Provider B. That is, the routers within Provider B must have explicit
   routing entries for each of Provider A's customers, 5 separate routes
   in Figure 1.

   Experience has shown that this approach scales very poorly. In the
   Default-Free Zone (DFZ) of the Public Internet, where routers must
   maintain routing tables for all reachable destinations, the cost of
   computing forwarding tables quickly becomes unacceptable large. A
   large part of the cost is related to the seemingly redundant
   computations that must be made for each individual network, even
   though the reality is that many reside at the same end site.


2.3.  CIDR and Provider-Based Addressing

   Classless Inter-Domain Routing (CIDR) and its associated provider-
   assigned address allocation policy were introduced (in part) to help
   reduce the size of and cost of computing forwarding tables. In CIDR,
   sites that want to connect to the Internet approach a provider to
   procure both connectivity and a network address; providers have large
   blocks of address space and assign pieces of them out to customers
   such that customers of the same provider have addresses with some
   number of leading bits in common. Note that CIDR started the use of
   the term ``prefix'' to refer to a Classless network. The combination
   of CIDR and provider-based addressing results in the ability for a
   provider to address many hundreds of sites while introducing just



draft-ietf-ipngwg-esd-analysis-00.txt                             [Page 8]


INTERNET-DRAFT                                                March 1997


   *one* network address into the DFZ global routing system, i.e.,
   aggregating all of its customers addresses under one prefix. An
   example of such a topology and addressing scheme is shown in Figure
   2.


                +----------------+
                |                |------- Customer1 (204.1.0.0/19)
                |                |------- Customer2 (204.1.32.0/23)
                |   Provider A   |------- Customer3 (204.1.34.0/24)
                |                |------- Customer4 (204.1.35.0/24)
                |                |------- Customer5 (204.1.36.0/23)
                +----------------+
                        |
                        |
                        |
                        |
                +----------------+
                |   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 network prefix with 16 bits for
   the network part and 16 bits for local use). Provider A has 5
   customers, each of which has been assigned a (longer) prefix
   subordinate to the aggregate. Note that unlike the pre-CIDR days of
   ``classful addressing'' the amount of address space assigned to a
   customer no longer needs to limited to the hard byte-boundary of the
   ``classful days'' In order for Provider B to be able to forward
   traffic to Customers1-5, Provider A need only announce a single
   prefix, 204.1.0.0/16, because that prefix covers all of its
   customers. The benefit for Provider B is that its routers need only a
   single routing table entry to reach all of Provider A's customers.
   Note the difference between the cases described in Figures 1 and 2.
   The important difference in the two Figures is that the latter
   example uses fewer routes to reach the same number of destinations.

   CIDR was a critical step for the Internet: in the early 1990s the
   overhead of computing and constructing forwarding tables in the DFZ
   required to support the Classful Internet was almost more than the
   commercially-available hardware and software of the day could handle.
   The introduction of BGP4's classless routing and provider-based
   address allocation policies resulted in an immediate relief. Having
   said that, however, there are some weaknesses of the system. First,
   the internet addressing model shifted from one of ``address owning''
   to ``address lending.'' In pre-CIDR days sites acquired addresses



draft-ietf-ipngwg-esd-analysis-00.txt                             [Page 9]


INTERNET-DRAFT                                                March 1997


   from a central authority independent of who their network provider
   was, and a site could assume it ``owned'' the address it was given.
   Owning addresses meant that once one had been given a set of network
   addresses, one could always use them and assume that no matter where
   a site connected to the Internet, the prefix for that network could
   be injected into the public routing system, and in particular, into
   the DFZ. With CIDR, however, it is simply no longer possible for each
   individual site to have its own private network prefix be injected
   into the DFZ; there would simply be too many of them. Consequently,
   if a site decides to change providers, then it needs to number itself
   out of space given to it by the new provider and give its old address
   back to the old 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)
         |      |                |------- Customer5 (204.1.36.0/23)
         |      +----------------+
         |                    |
      +----------------+      |
      |   Provider B   |      |
      +----------------+      |
         |                    |
         |      +----------------+
         +------|   Provider C   |------- Customer3 (204.1.34.0/24)
                +----------------+

                                  Figure 3

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




draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 10]


INTERNET-DRAFT                                                March 1997


   The example in Figure 3 explains why sites must renumber if existing
   levels of aggregation are to be maintained. While it is certainly
   clear that one or two ``exceptions'' to the ideal case can be
   tolerated, the reality in today's Internet is that there are
   thousands of providers, many with thousands of individual customers.
   Some renumbering of sites would seem essential to maintain sufficient
   aggregation.

   The empirical cost of renumbering a site in order to maintain
   aggregation has been the subject of much discussion. The practical
   reality, however, is that forcing all sites to renumber is difficult
   given the size and wealth of companies that now depend on the
   Internet for running their business. Thus, although the technical
   community came to consensus that address lending was necessary in
   order for the Internet to continue to operate and grow, the reality
   has been that some of CIDR's benefits have been lost because sites
   refuse to renumber.

   One unfortunate characteristic of CIDR at an architectural level is
   that the pieces of the infrastructure which benefit from the
   aggregation (i.e., the providers which constitute the default-free
   part of the routing infrastructure) are not the pieces that incur the
   cost to achieve the aggregation. The logical corollary of this
   statement is that the pieces of the infrastructure which 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 one could claim that the continued
   operation of the Internet is a benefit, though it is an indirect
   benefit and requires selflessness on the part of the site in order to
   recognize it.


2.4.  Multihoming 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 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 address. When that site's providers would announce the
   site's network into the global routing system, a ``shortest path''
   type of routing would occur so that pieces of the Internet closest to
   the first provider would use the first provider and other pieces
   would use the second provider. This allowed sites to deal with the
   load on their multiple connections with the routing system itself.



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 11]


INTERNET-DRAFT                                                March 1997


   With CIDR, if a multi-homed site is known by a single prefix taken
   from one of its providers, then that prefix is aggregable by the
   provider which assigned the address but *not* aggregable by the other
   providers.

   One way to prevent entropy from taking over under CIDR is to have
   multi-homed sites use address space from all of its providers. Though
   this in itself is not so difficult, it changes the way load-sharing
   is handled, complicating the engineering by requiring much more
   foresight and by introducing complexities like DNS and its caching
   system. So, like those sites which refuse to renumber, many multi-
   homed sites today are known by a single prefix, thus reducing the
   efficiency of the global routing system.

   To be clear, there certainly are ways with CIDR for sites to be
   multi-homed without having a negative impact on the routing
   infrastructure, and there are some sites that do this today. However,
   operational experience to date has shown an unwillingness on the part
   of most sites to do the work necessary to multi-home in a way that is
   CIDR-friendly. Sites have more experience doing load-sharing under
   the pre-CIDR type of multi-homing than post-CIDR, so this is one
   reason for the reluctance. Another reason, however, is that in its
   documentation, CIDR presents several options related to multi-homing,
   but it does not choose one option and fully flesh out related details
   like load-sharing. While the analysis of GSE will end up showing that
   it actually does little to improve the support for multi-homing, it
   should be given great credit for giving the topic significant
   attention as a distinguished service from the beginning, rather than
   as an after-thought.


3.  GSE Background

   This section provides background information about GSE with the
   intent of making this document stand-alone with respect to the GSE
   ``specification.''  Additional details on GSE can be found in [GSE].
   First the motivations behind GSE will be discussed, then the
   important technical details will be described and finally some
   explicit non-goals will be listed.


3.1.  Motivation For GSE

   The primary motivation for GSE is the fact that the chief IPv6 global
   unicast address structure, provider-based, is fundamentally the same
   as IPv4 with CIDR and provider-based aggregation. Many people are not
   satisfied with the scaling factors achieved with CIDR and provider-
   based aggregation and think that better solutions can, and in fact



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 12]


INTERNET-DRAFT                                                March 1997


   must, be found. The GSE draft asserts that IPv4 with CIDR has not
   achieved the aggressive aggregation required for the route
   computation functions of the default-free zone of the Internet to
   scale for IPv4, and that the larger addresses of IPv6 simply
   exacerbate the problem.

   More importantly, a key aspect of provider-based aggregation is its
   requirement that end sites be renumbered in response to topological
   changes (e.g., when an end site switches ISPs). The GSE proposal
   asserts that acceptable aggregation can continue only if renumbering
   is forced, but the future viability of forced renumbering is unclear
   given the increasing dependence on the Internet by litigious
   commercial organizations. This fact 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 forced to
   renumber by forces outside of its control).

   Finally, GSE deals significantly with Sites that have multiple
   Internet connections. In some addressing schemes (e.g., CIDR), this
   ``multi-homing'' can create exceptions to the aggregation and result
   in poor scaling. That is, the public routing infrastructure needs to
   carry multiple distinct routes for the multi-homed site, one for each
   independent path. GSE 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 some benefit
   without impacting global scaling. This includes a specific mechanism
   that providers could use to support multi-homed Sites, presumably at
   a cost that the Site would consider when deciding whether or not to
   become multi-homed.


3.2.  GSE Address Format

   The key departure of GSE from classical IP addressing (both v4 and
   v6) is that rather than over-loading addresses with both locator and
   identifier purposes, it splits the address into two elements: the
   high-order 8 bytes for routing (called ``Routing Stuff'' throughout
   the rest of this document) and the low-order eight 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 4




draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 13]


INTERNET-DRAFT                                                March 1997


3.3.  Routing Stuff (RG and STP)

   The Routing Goop (RG) describes the place in the Internet topology
   where a Site connects, so it is used to route datagrams to the Site.
   RG is structured as follows:

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

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

                                 Figure 5

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

   Having described the decomposition process, we can now examine the
   bits in the RG. After the 3-bit prefix identifying the address as
   GSE, the next 13 bits identify the LS. By limiting the field to 13
   bits, a ceiling is defined on the complexity of the top-most routing
   level. 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 are used for
   routing structure within a Site, similar to subnetting with IPv4,
   though these bits are *not* part of the Routing Goop. The distinction
   between Routing Stuff and Routing Goop is that RG controls routing in
   transit networks, while Routing Stuff includes the RG plus the Site
   Topology Partition (STP). The STP is used for routing structure



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 14]


INTERNET-DRAFT                                                March 1997


   within a Site.

   The GSE proposal formalizes the ideas of Sites and of public versus
   private topology. In the first case, a Site is a set of hosts,
   routers and media which have one 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.
   [Editorial note: An attempt was made to capitalize ``Site'' when
   assuming the GSE model and lower-case ``site'' when referring to the
   less formal idea of a site in IPv4.]

   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) field.


3.4.  End-System Designator

   The End-System Designator (ESD) is an unstructured 8-byte field that
   uniquely identifies that end-system from all others.  The leading
   contender for the role of a 64-bit globally unique ESD is the
   recently defined ``EUI-64'' identifier [EUI64]. These identifiers
   consist of a 24-bit ``company_id'' concatenated with a 40-bit
   ``extension.'' (Company_id is just a new name for the
   Organizationally Unique Identifier (OUI) 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.  The mapping of MAC addresses into EUI-64
   identifiers is as follows: a 48-bit MAC address xx-xx-xx-yy-yy-yy is
   mapped into the 64-bit EUI-64 identifier xx-xx-xx-FF-FE-yy-yy-yy.

   The existence of the reserved range and defined mapping of 48-bit MAC
   addresses makes the EUI-64 a seemingly ideal ESD candidate.  ESDs
   derived from existing IEEE MAC addresses should be compatible with
   future network media that use EUI-64 as station identifiers (e.g.,
   FireWire, Futurebus+, SCI).

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



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 15]


INTERNET-DRAFT                                                March 1997


   possessing a valid company_id could sell identifiers out of its
   allocation, though there may be a requirement that EUI-64's be sold
   only in the form of an electronically-readable part. The IANA, to
   choose one example, could generate identifiers using its company_id
   (00005E hex).

   Another scheme for an IETF-specific ESD space would be to use one or
   both of the two lowest-order bits of the first octet as a flag. In a
   company_id, those two bits are reserved for use as the Global/Local
   and Individual/Group bits, so if either is set, lack of conflict with
   an EUI-64 would seem to be assured.

   The most important feature of the ESD is its global uniqueness. End-
   points of communication would only care about the ESD; as examples,
   TCP peers could be identified by the ESD alone, checksums would
   exclude the RG (the sender doesn't know its RG, so can't include it
   in the checksum), and on receipt of a datagram only the ESD would be
   used in testing whether a packet is intended for local delivery.


3.5.  Address Rewriting by Border Routers

   Another fundamental aspect of GSE is that Site border routers rewrite
   addresses of the packets they forward across the Site/Internet
   boundary. 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
   Internet, the border router inserts the appropriate RG into the
   packet's source address. Likewise, when a packet from the Public
   Internet is forwarded into a Site, the border router replaces the RG
   part of the destination address with the designated Site-Local RG.

   Having border routers rewrite addresses obviates the need for end
   Sites to renumber --- GSE's approach isn't so much to ease
   renumbering as to simply make it completely transparent to end Sites.
   To achieve transparency, the RG(s) by which a Site is known would
   *not* be known to hosts or routers within the core of that Site.
   (Note: RG can be plural in the previous sentence because multi-homed
   Sites are known by multiple RGs.) Instead, the RG(s) for the Site
   would be known only by the exit router(s), either through static
   configuration or through a dynamic protocol with the upstream
   provider. Because end-hosts don't know their RG(s), they don't know
   their entire 16-byte address(es), so they can't specify the full
   address in the source fields of packets they originate. Consequently,
   when 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



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 16]


INTERNET-DRAFT                                                March 1997


   Site is to ensure the changeability of this value without impacting
   the Site itself. It is expected that the RG will need to change
   relatively frequently (e.g., several times a year) in order to
   support scalable aggregation as the topology of the Public Internet
   changes.  A change to a Site's RG would only require a change at the
   Site's egress point (or points, in the case of a multi-homed Site);
   and it's well possible that this change would 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 intra-Site
   communication, however, it is expected that only the Site-Local RG
   would be used (and stored) which would always work for intra-Site
   communication regardless of changes to the Site's external RG.

   In addition to rewriting source addresses upon leaving a Site,
   destination addresses are rewritten upon entering a Site. To
   understand the motivation behind this, consider a Site with three
   connections. 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. Instead, GSE proposes
   replacing the RG in inbound packets with the special ``Site-local
   RG'' value to reduce intra-Site routing tables to the minimum
   necessary.

   To be clear, when a node initiates a flow to a node in another Site,
   the initiating node knows the full 16-byte address for the
   destination through some mechanism like a DNS query. So this
   initiating node places the full 16-byte address in the destination
   address field of the datagram, and that field stays in tact through
   the first Site and through all of the Public Topology; it is only
   when the datagram arrives at the destination Site that the RG portion
   of the destination address is rewritten with the distinguished
   ``Site-Local RG'' value.  When the destination host needs to send
   return traffic, that host will also know the full 16-byte address for
   the destination because it appeared in the source address field of
   the arriving packet.


3.6.  Renumbering and Rehoming Mid-Level ISPs

   One of the most difficult-to-solve components of the renumbering
   problem is that of renumbering mid-level service providers.
   Specifically, if SmallISP1 changes its transit provider from BigISP1



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 17]


INTERNET-DRAFT                                                March 1997


   to BigISP2, then all of SmallISP1's customers would have to renumber
   into address space covered by an aggregate of BigISP2 (if the overall
   size of routing tables is to stay the same). 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 obviously implies a new Resource Record
   type). That provider may define some of the low-order bits of the RG
   and then reference its immediate provider. This chain of reference
   allows mid-level service providers to change transit providers, and
   the customers of that mid-level will simply ``inherit'' the change in
   RG.


3.7.  Support for Multihomed 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 6

   PBR1 is Provider1's border router while PBR2 is Provider2's border
   router.  SBR1 is the Site's border router that connects to Provider1
   while SBR2 is the Site's border router that connects to Provider2.
   Imagine, for example, that the line between Provider1 and the Site
   goes down. Any already existing flows that use a destination address
   including RG1 would stop working. In addition, any DNS queries that
   return addresses including RG1 would not be viable addresses. If PBR1
   and PBR2 knew about each other, however, then in this case PBR1 could



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 18]


INTERNET-DRAFT                                                March 1997


   tunnel packets destined for RG1-prefixed addresses to PBR2, thus
   keeping the communication working.


3.8.  Explicit Non-Goals for GSE

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

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

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

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

     4) It is not known whether the performance cost of having routers
        rewrite portions of the source and destination address in packet
        headers is acceptable.

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


4.  Analysis of GSE's Advantages and Disadvantages


4.1.  End System Designator


4.1.1.  IP Addresses in the IPv4 Internet

   As described earlier, in the IPv4 Public Internet, IP addresses
   contain two pieces of information: a unique identifier and a locator.
   A key aspect of the embedded location information is that it must be
   aggregable, so that a single routing table entry can cover many
   destination addresses. In practice, this means that sites that are
   topologically close to each other must share a common prefix, as
   exemplified in provider-based addressing [RFC 2073] and CIDR
   [RFC1817]. Without sufficient aggregation, routing in the Public
   Internet can not scale [RFC2008].

   Note that embedding location information within an address has the
   side-effect of helping ensure that all addresses are globally unique.



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 19]


INTERNET-DRAFT                                                March 1997


   If interfaces on two different nodes are assigned the same unicast
   address, the routing subsystem will (generally) deliver packets to
   only one of those nodes. The other node will quickly realize that
   something is wrong (since communication using the duplicate address
   fails) and take corrective action (e.g., obtain a proper address).

   Embedding location information within an address also provides some,
   though not much, protection from forged addresses. Although it is
   trivial to forge a source address in today's Internet, the routing
   subsystem will in most cases forward any return traffic sent to that
   address to its proper destination --- not to an arbitrary node
   masquerading as someone else. To masquerade as someone else requires
   subverting the routing subsystem, placing the intruder somewhere on
   the normal routing path between the masqueraded host and its peer,
   etc. Allowing the Routing Stuff and ESD portions of an address to be
   changed independent of each other potentially increases the ease with
   which packets intended for a particular ESD can be misrouted or
   hijacked elsewhere. As discussed in later sections, additional checks
   must be made to reduce the threat of hijacking.


4.1.2.  Overloading Addresses: Network Layer Issues

   Embedding location information within an address has some important
   consequences. At the network layer, a node compares the destination
   address of received packets against the addresses of its attached
   interfaces. Only if the addresses of received packets match are
   packets handed up to higher layer protocols. The entire address
   (including the Routing Stuff part) must match. Otherwise, the packet
   is assumed to be intended for some other node and forwarded on (if
   received by a router) or silently discarded (if received by a host).
   This has subtle but significant implications:

     1) If a receiving host has multiple interfaces, it has multiple IP
        addresses. When a packet addressed to a multi-homed host is
        received on an interface other than the one to which a packet is
        addressed, the host may reject (i.e., silently discard) the
        packet, if it implements the ``Strong ES Model'' defined in
        [RFC1122].

     2) In IPv6 (and recent IPv4 stacks), an interface may have more
        than one unicast IP address assigned to it. Indeed, one way to
        renumber an end site is to phase out an address (i.e.,
        "deprecate" it using RFC 1971 termininology) while
        simultaneously phasing in a new one. Once the deprecated address
        becomes invalid, packets sent to the invalid address will no
        longer be accepted by the node, even though the packet may have
        intuitively reached its intended recipient. Thus, even if a



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 20]


INTERNET-DRAFT                                                March 1997


        packet sent to an invalid address is somehow delivered to the
        intended recipient (e.g., via tunneling), the receiver would
        reject the packet because the address it was sent to no longer
        belongs to any of the node's interfaces. Consequently, any
        communication using the invalid address will fail (e.g., new and
        existing TCP connections). Anyone wishing to communicate with
        the node must learn and switch to the new address.

     3) Because an address also indicates ``where'' the destination
        resides within the Internet, a mobile node that moves from one
        part of the Internet to another must obtain a new address that
        reflects its new location. Moreover, the routing subsystem will
        continue to forward packets sent to the mobile node's previous
        address to the node's previous point of attachment where they
        are likely be discarded. That is, even if a mobile node is
        willing to continue accepting packets addressed to one its
        previous addresses, it is unlikely that they will be received
        (in the absence of something like Mobile IP [RFC2002]).

    4) A multi-homed host has multiple interfaces, each with its own
        address(es). If one of its interfaces fails, packets could, in
        theory, be delivered to one of the host's other interfaces. In
        practice, however, the routing subsystem has no way of knowing
        that the interface to which a packet is addressed has failed and
        what alternate interface addresses the packet could be delivered
        to. Consequently, packets sent to a multi-homed host won't be
        delivered to the intended recipient, even though the node is
        reachable (through an alternate address).

   Note that the above problems fall into two general categories:

     1) Today's routing subsystem is unable to automatically deliver a
        packet to a host's ``alternate'' addresses (if the host is
        multi-homed) or a new address (if the host moves), should there
        be a problem delivering a packet to the destination address
        listed in the packet. It is possible to imagine, however, future
        routing advances addressing this problem (e.g., Mobile IP).

     2) Even if a packet is delivered to its intended destination, the
        packet may still be rejected because the packet's destination
        address does not match any of the addresses assigned to
        destination's interfaces. This problem does not appear to be
        insurmountable and could be rectified (for example) by having a
        host remember its previous addresses.







draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 21]


INTERNET-DRAFT                                                March 1997


4.1.3.  Overloading Addresses: Transport Layer Issues

   Although the problems discussed previously appear to have viable
   solutions, additional complications occur at the transport level.
   Transport protocols such as TCP and UDP use embedded IP addresses to
   identify the end points of a transport connection. Specifically, the
   communicating end points of a transport connection are uniquely
   identified by the sender's source IP address and source port number
   together with the recipient's destination IP address and port number.
   Once a connection has been established, the IP addresses can not
   change. In particular, if a mobile host moves to a new location and
   obtains a new address, packets intended for a TCP connection created
   prior to the move cannot use the new address. TCP will treat any
   packets sent to the new address as belonging to a different TCP
   connection.

   It is possible to imagine changes to TCP that might allow connections
   to change the addresses they are using mid-connection without
   breaking the connection. However, some subtle issues arise:

     1) Packets intended for a pre-existing connection must be
        demultiplexed to that connection as part of any negotiation to
        change the addresses that identify that transport end point.
        However, because the demultiplexing operation uses the transport
        addresses of the pre-existing TCP connection (which is based on
        the previous address), TCP packets sent to a new address won't
        be delivered to the desired transport end point (which still
        uses the previous address). Consequently, packets would need to
        be sent to the previous address. However, by the time a mobile
        node has moved and knows its new address, packets sent to the
        previous address may no longer be delivered (i.e., they may not
        be forwarded to the mobile host's new location).

     2) When a mobile host moves, it could inform its TCP peers that it
        has a new address. However, such a message could not be
        delivered to the remote TCP connection if it was sent using its
        new address for its source address. Just as above, such packets
        would not be demultiplexed to the correct TCP connection. On the
        other hand, there are difficulties if it attempts to send
        packets using its previous address from its new location.
        Because of the danger of spoofing attacks, routers are now
        encouraged to actively look for, and discard traffic from, a
        source address that does not match known addresses for that
        region of the Internet [CERT]. Consequently, such packets cannot
        be expected to be delivered.

   Although the previous discussion used mobile nodes as an example, the
   same problem arises in other contexts. For example, if a site is



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 22]


INTERNET-DRAFT                                                March 1997


   being renumbered in IPv6, it may have two addresses, a previous
   (i.e., deprecated) one being phased out and a new (i.e., preferred)
   one being phased in. At the transport level, the problem of switching
   addresses is similar in many respects to the mobility problem.


4.1.4.  Benefits of Globally Unique ESDs

   An alternate approach is to break an address into two distinct
   portions:

     1) An End System Designator (ESD) that uniquely identifies an end
        point of communication (independent of the interface through
        which that was reached). Such an identifier should be globally
        unique so that a node that receives a packet can definitively
        determine whether the packet is intended for it by comparing
        only the ESD portion of the address.

     2) A ``locator'' or Routing Stuff that is used by the routing
        subsystem to deliver a packet to the appropriate end system
        identified by the ESD.

   Having a clear separation between the Routing Stuff and the ESD
   portion of an address gives protocols some additional flexibility. At
   the network layer, for example, recipients can examine just the ESD
   portion of the destination addresses when determining whether a
   packet is intended for them. This means that if a packet is delivered
   to the correct destination node, the node will accept the packet,
   regardless of how the packet got there, i.e., without regard to the
   Routing Stuff of the address, which interface it arrived on, etc.
   Excluding the Routing Stuff of an address when making address
   comparisons also makes it possible to change the Routing Stuff of an
   address to reflect a mobile node's new location, or an alternate
   interface on a multi-homed host. Such packets would then be delivered
   and accepted by the target host.

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

     1) The network layer could be responsible for doing the mapping.
        The advantage of such a system is that an ESD could be stored



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 23]


INTERNET-DRAFT                                                March 1997


        essentially forever (e.g., in configuration files), but whenever
        it is actually used, network layer software could 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 quantity. Unfortunately, building such a mapping
        mechanism that is scalable is a hard problem.

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

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

   It should be noted that the GSE proposal [GSE] does not embrace the
   general model. Indeed, it proposes the last. The network layer (and
   indeed the transport layer) is always presented both the Routing
   Stuff (RG + STP) and the ESD together in one IPv6 address. It is not
   the network (or transport) layer's job to determine the Routing Stuff
   given only the ESD. When an application has data to send, it queries
   the DNS to obtain the IPv6 AAAA record for a destination. The
   returned AAAA record contains both the Routing Stuff and the ESD of
   the specified destination. While such an approach eliminates the need
   for the lower layers to be able to map ESDs into corresponding
   Routing Stuff, it also means that when presented with an address
   containing an incorrect (i.e., no longer valid) Routing Stuff, the
   network will be unable to deliver the packet to its correct
   destination. It is up to applications themselves to deal with such
   failures. Note that addresses containing invalid Routing Stuff will
   result any time cached addresses are used after the Routing Stuff of
   the address becomes invalid. This may happen if addresses are stored
   in configuration files, or with long-running communication.


4.1.5.  ESD: Network Layer Issues

   Along with the flexibility offered by separating the ESD from the
   Routing Stuff come additional considerations that must be considered
   at the network layer:

     1) If a receiver observes that recent packets are arriving with a
        different Routing Stuff in the source address than before, it



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 24]


INTERNET-DRAFT                                                March 1997


        may want to send return traffic using the new Routing Stuff.
        However, such information should not be accepted without
        appropriate authentication of the new Routing Stuff, otherwise
        it would be trivial to hijack existing transport connections.
        Always using the most recently received Routing Stuff of an
        address to send return traffic without appropriate
        authentication leads to a vulnerability that is equivalent in
        potential danger to ``reversing and using an unauthenticated
        received source route.''

        Note also that in the GSE proposal, since a sender does not know
        their own RG, it is not possible for the sender to compute an
        Authentication Header via IPSEC that covers the RG portion of an
        address. Thus, a recipient of new RG would need to authenticate
        the received information via some alternate (undefined)
        mechanism.

        Finally, receipt of packets from different Routing Stuff than
        before does not necessarily indicate a permanent change. In the
        GSE proposal, for example, when a Site is multi-homed, some of
        its packets may exit via one egress router and other packets via
        a different egress router. Even packets originated from the same
        source may exit through multiple egress routers. Consequently, a
        node may receive traffic from the same sender in which the
        Routing Stuff part changes on every packet.

     2) In general, whenever an address is embedded within a packet
        (including within data), one must consider whether all the bits
        in the address should be used in computations, or whether just
        the ESD portion should be used. Examples where such decisions
        would need to be made include, but are not limited to, Neighbor
        Discovery packets containing Neighbor Solicitations and
        Responses [RFC 1970], IPSEC packets being demultiplexed to their
        appropriate Security Association, IP deciding whether to accept
        an IP datagram (before reaching the transport level), the
        reassembly of fragments, transport layer demultiplexing of
        received packets to end points, etc.


4.1.6.  ESD: Transport Layer Issues

   Previous sections have made clear that the embedding of full IPv6
   addresses (i.e., Routing Stuff) within transport connection endpoint
   identifiers poses problems for mobility and site renumbering. This
   section discusses an alternate approach, in which transport endpoint
   identifiers use ESDs rather than full addresses (with embedded
   Routing Stuff).




draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 25]


INTERNET-DRAFT                                                March 1997


   In the following discussion, it should be kept in mind that the IPng
   Recommendation [RFC 1752] states that a transition to IPv6 cannot
   also require deployment of a ``TCPng'', that is, make changes to TCP
   that decrease its overall robustness. In addition, although we focus
   on TCP, UDP-based protocols also depend on the Routing Stuff in
   similar ways, e.g., starting with the UDP checksum of the peers'
   addresses. Indeed, we believe that TCP is the ``easy'' case to deal
   with, for two reasons. First, TCP is a stateful protocol in which
   both ends of the connection can negotiate with each other. Some UDP-
   based protocols are stateless, and remember nothing from one packet
   to the next. Consequently, changing UDP-based protocols may require
   the introduction of ``session'' features, perhaps as part of a common
   ``library'', for use by applications whose transport protocol is
   relatively stateless.  Second, changes to UDP-based protocols in
   practice mean changing individual applications themselves.


4.1.6.1.  Dumultiplexing Packets to Transport Endpoints

   Connections in GSE are identified by the ESDs rather than full IPv6
   addresses (with embedded Routing Stuff). That is:
   unique TCP connection:          srcaddr dstaddr srcport destport
   unique GSE TCP connection:      srcESD dstESD srcport dstport
   Consequently, when demultiplexing incoming packets, TCP would ignore
   the Routing Stuff portions of addresses when delivering packets to
   their proper end point.

   Although there are potential benefits to this approach (discussed
   below), demultiplexing of ESDs is in fact a requirement with GSE. If
   a site is multihomed, the packets it sends may exit different egress
   border routers during the lifetime of a connection. Because each
   border router will place its own RG into the source addresses of such
   packets, the receiving TCP must ignore (at least) the RG portion of
   addresses when demultiplexing received packets. The alternative would
   be to make TCP less robust with respect to changes in routing, i.e.,
   if the path changed, packets delivered correctly would be discarded
   by the receiving TCP rather than processed.


4.1.6.2.  Pseudo-Header Checksum Calculations

   Having routers rewrite the RG portion of addresses means that TCP
   cannot include the RG in its checksum calculation; the sender does
   not know its own RG. Consequently, upon receipt of a TCP segment, the
   receiver has no way of determining whether the RG portion of an
   address has been corrupted (or modified) in transit (the implications
   of this are discussed below).




draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 26]


INTERNET-DRAFT                                                March 1997


4.1.6.3.  RG Selection When Sending Packets

   When a host has a packet to send, what RG should it use? There are
   three cases. If the host is performing an ``active open'', it queries
   the DNS to obtain the destination address, which contains appropriate
   RG. If the host is responding to an active open from a remote peer,
   the source address of packets from that peer contain usable RG. The
   interesting case is when RG changes mid-connection.


4.1.6.4.  Mid-Connection RG Changes

   During a connection, the RG appearing on subsequent packets is
   susceptible to change through renumbering events, and indeed more
   frequently, to change through 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. Consequently it may be desirable to switch
   to the just-received RG, as the old RG may no longer be valid (e.g.,
   a border router has failed). However, simply using the most-
   recently-received RG makes it trivial to hijack connections.

   The way TCP packets are demultiplexed under GSE, they will be
   delivered to the correct endpoint even though TCP may send to its
   peer at a deprecated RG or one that is less optimal because the
   peer's Border Router has changed.  It would seem highly desireable
   for TCP connections to be able to survive such events. However, the
   completion of renumbering events (so that an earlier RG is now
   invalid) and certain topology changes would require TCP to switch
   sending to a new RG mid-connection.  To explore the whole space, we
   considered ways of allowing this mid-connection RG change.

   If TCP connection identifiers are based on ESDs rather than full
   addresses, traffic from the same ESD would be viewed as coming from
   the same peer, regardless of its source RG. This makes it trivial for
   any Internet host to impersonate another, and have such traffic be
   accepted by TCP.  Because this vulnerability is already present in
   today's Internet (forging full source addresses is trivial), the mere
   delivery of incoming datagrams with the same ESD but a different RG
   does not introduce new vulnerability to TCP.  In today's Internet,
   any node can already originate FINs/RSTs from an arbitrary source
   address and potentially or definitely disrupt the connection.
   Therefore, changing RG for acceptance, or acceptance of traffic
   independent of its source RG, does not significantly worsen existing
   robustness, as far as our analysis has gone.

   We also considered allowing TCP to reply to each segment using the RG



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 27]


INTERNET-DRAFT                                                March 1997


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

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

   One question that arises is what impact corrupted RG would have on
   robustness. Because the RG is not covered by any checksums, it would
   be difficult to detect such corruption. Moreover, once a specific RG
   is in use, it does not change for the duration of a connection. The
   interesting case occurs on the passive side of a TCP connection,
   where a server accepts incoming connections from remote clients. If
   the initial SYN from the client includes corrupted RG, the server TCP
   will create a TCP connection (in the SYN-RECEIVED state) and cache
   the (corrupted) RG with the connection. The second packet of the 3-
   way handshake, the SYN-ACK packet, would be sent to the wrong RG and
   consequently not reach the correct destination. Later, when the
   client retransmits the (unacknowledged) SYN, the server will continue
   to send the SYN-ACK using the bad RG. Eventually the client times
   out, and the attempt to open a TCP connection fails. Figure 7 shows
   the details.















draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 28]


INTERNET-DRAFT                                                March 1997



         TCP A                                                  TCP B

      1. CLOSED                                                 LISTEN

      2. SYN-SENT --> <SRC RG=BITERR><SEQ=100><CTL=SYN>     --> SYN-RECEIVED

      3. <-- <DST RG=BITERR><SEQ=300><ACK=101><CTL=SYN,ACK> <-- SYN-RECEIVED

      4. SYN-SENT --> <SRC RG><SEQ=100><CTL=SYN>             --> SYN-RECEIVED

      5. <-- <DST RG=BITERR><SEQ=300><ACK=101><CTL=SYN,ACK> <-- SYN-RECEIVED

      ... TCP A times out

                                  Figure 7


   We next considered 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 can one
   determine which RG is valid and which is not? That is, for each of
   the RGs a sender attempts to use, how can it determine which RG
   worked and which did not? Solving this problem is more difficult than
   first appears, since one must cover the cases of delayed segments,
   lost segments, simultaneous opens, etc. If a SYN-ACK is retransmitted
   using different RGs, it is not possible to determine which of those
   RGs worked correctly. We concluded that the only way TCP could
   determine that a particular RG was used to deliver segments was if it
   received an ACK for a specific sequence number in which all
   transmissions of that sequence number used the same RG.

   We analyzed multiple cases of RG changing within the time of the
   opening handshake. One example is diagrammed in Figure 8, and it and
   two others are summarized in Table 1. We observed that RG flap and
   large numbers of passive opens may coincide, for instance, when a
   power failure at a server farm affects both internal routers and
   servers.











draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 29]


INTERNET-DRAFT                                                March 1997



       time TCP A                                    time TCP B

       t0  --> <SRC RG=M><SEQ=100><SYN>              t1

       t3  <-- <DST RG=M><SEQ=300><ACK=101><SYN,ACK> t1

       TCP B's SYN,ACK is delayed and crosses with retransmit of TCP A's
       SYN on which RG has changed from M to N

       t2  --> <SRC RG=N><SEQ=100><SYN>              t3

       t4  --> <SRC RG=N><SEQ=101><ACK=301>          t3  ESTABLISHED

       TCP B decides to use DST RG=M for TCP A, because it heard from
       RG=M and was ACK'd on a send to RG=M

                                  Figure 8



                    SYNFROM  SYNACKTO  ACKFROM   SELECT
                    W        W         X         W
                    ------------------------------------
                    W
                    X        W         X         W
                    ------------------------------------
                    W        W
                    X        X         Y         ??

                                  Table 1

   At best, an RG selection algorithm for TCP would be relatively
   straightforward but would require new logic in implementations of
   TCP's opening handshake --- a significant transition 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, and so on.

   In the end, we concluded that although the corrupted SYN case of
   Figure 7 was a potential problem, 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 transition to
   GSE needing a significant TCPng transition.

   Our final conclusion is that transport protocol end points must make
   an early, single choice of the RG to use when sending to a peer and



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 30]


INTERNET-DRAFT                                                March 1997


   stick with that choice for the duration of the connection.
   Specifically:

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

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

     3) For all other cases, use the first RG received with a given ESD
        for all sending. We recommend that a means be found for RGs to
        be checksummed if the GSE address structure is used.


4.1.6.5.  Duplicate ESDs with Differing RGs

   Another interesting case occurs if two different (client) nodes are
   using the same ESD, and they attempt to communicate with a common
   server. In such cases, the RG (or STP) portions of the address may be
   different. However, since only the ESD is used in demultiplexing
   packets to their transport end points, traffic from two different
   hosts may be delivered and processed by one transport endpoint. Given
   the above rules that bind RG to existing connections, only one RG
   will be used, and all traffic from the server will be sent to the
   same client. It would appear that in most cases, only one connection
   would reach ESTABLISHED state, and the others would time out.

   One implication of binding RG information to TCP connection state is
   that we may be opening the door to additional security threats. One
   denial of service attack, for instance, would be for an intruder to
   masquerade as another host and ``wedge'' connections in a SYN-
   RECEIVED state by sending SYN segments containing on invalid RG in
   the source IP address. Subsequent connection attempts to the wedged
   host from the legitimate party (if they used the same TCP port
   numbers) would then not complete, since return traffic would be sent
   to the wrong place.


4.1.6.6.  Summary: ESD and RG Not Strictly Independent

   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 take place if there are assurances
   that only properly authenticated RG is used --- RG authenticated by



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 31]


INTERNET-DRAFT                                                March 1997


   the peer.

   Consequently, we conclude that the rules for transport processing
   when ESDs are present differ from classical IP. Specifically:

     1) The demultiplexing of packets to transport connection end points
        should use ESDs, but should not use the Routing Stuff part of
        addresses.

     2) Once a packet has been delivered to its transport endpoint, that
        packet should not be processed without first examining the
        source RG used. Whether (and how) the information in the
        received packet is used is dependent on the transport protocol
        itself. A protocol could chose to completely ignore the packet,
        it could selectively use parts of the packet (e.g., to attempt
        out-of-band authentication of the RG), or it could process the
        packet in its entirety. It must not, however, use the received
        RG to send subsequent return traffic without first
        authenticating the RG.


4.1.7.  ESD: Application Layer Issues

   In this section we define applications as user processes that must
   exchange data reliably with remote processes. Such distributed
   processes need system support to reliably identify remote processes.
   It is desirable (necessary?) for such end identifications to meet the
   following requirements:

     1) The identifier assigned to each end point should be globally
        unique.

     2) This uniqueness should be easily enforceable because it is
        difficult, or probably impossible, to provide an absolute
        guarantee on the uniqueness of these identifiers. Applications
        relying on this uniqueness must be prepared for duplicate
        detection; at the same time, one must be prepared to detect
        maliciously forged identifiers.

   The last point is becoming increasingly important as the Internet
   continues to grow exponentially, both in the scale of user population
   and in the scope of diverse applications that cover all walks of
   life.

   In the original design of the IPv4 architecture, globally unique IP
   addresses are used as the globally unique identifiers for all
   interfaces reachable on the Internet. Thus, a user process easily
   obtains a globally unique identifier by attaching a local port number



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 32]


INTERNET-DRAFT                                                March 1997


   to the address. One fundamental architectural change suggested by GSE
   is to split the address into completely separate interface
   identifiers and locators (routing stuff). In the rest of this section
   we discuss the pros and cons of each of the two approaches.


4.1.7.1.  The Impacts of Address and Identifier Overloading In IPv4

   Because global IP addresses are unique, using addresses as
   identifiers automatically provides the needed uniqueness property of
   an identifier. Moreover, duplicates can be easily detected. If two
   more interfaces claim the same unicast address, then due to shortest
   path routing, packets destined to that IP address are, generally
   speaking, delivered only to the interface closest to the source.
   Nodes with duplicate addresses can detect the error by observing the
   lack of connectivity to certain locations.

   Furthermore, using addresses as identifiers makes forging difficult.
   If a malicious node inserts a false source address in its outgoing
   packets, although the packets are likely to be delivered to the
   destination host, it is almost impossible for the malicious node to
   receive any reply data. We can further prevent the packets with
   faulty source address from being delivered by making all routers
   perform reverse source address checking, that is, checking each
   incoming packet to see if it comes from the correct interface as
   indicated by the routing table, a practice enforced by all IP
   multicast routing protocols. Such source address checking provides a
   simple, universal and effective enforcement to correct interface
   identifications. Even if some routers are compromised and allow
   packets with false source addresses to pass through, delivering
   packets with forged source addresses to the destination would require
   that all routers along the path be compromised.

   The fundamental disadvantage of overloading addresses with
   identification information is that changes in addresses lead to
   changes in identification, which implies that all hosts will be aware
   of all address changes, an issue GSE is designed to resolve. It is
   worth noting that keeping TCP connections running across renumbering
   is a non-goal of GSE design.


4.1.7.2.  The Impact of Separating Locators and Identifiers

   GSE uses the upper 8 bytes of IPv6 addresses as locators and the
   lower 8 bytes as globally unique identifiers (ESD). The chief
   advantage of this complete separation of Routing Stuff and ESD is a
   stable identifier per interface, independent from all renumbering
   changes in a network with a provider-based addressing architecture.



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 33]


INTERNET-DRAFT                                                March 1997


   One intention of the GSE design is to use the ESD alone in
   establishing and maintaining TCP connection state. The discussions at
   the interim meeting, however, revealed significant resistance to
   doing this. The consensus from the meeting was that it is not
   adequate to use an ESD alone for end point identification due to the
   ease of hijacking a TCP connection. Incoming packets with a wrong ESD
   can easily be detected as coming from an incorrect source, however,
   incoming packets with a correct ESD cannot be easily trusted as being
   from the correct source.

   Various approaches to using RG as part of TCP source verification
   were discussed at the meeting. Using ESDs as end point
   identifications seems to require two steps of processing. In the
   first step, the ESD can be used for PCB lookups. In the second step,
   the entire address (including RG) must be considered because one
   cannot safely take packets with an arbitrary RG. So the purpose of
   the first step is to locate the intended end point of an incoming
   packet. The second step then can make a separate decision as to how
   to act on the received data (accept, reject or perform out-of-band
   authentication on the RG). Due to the conflict between applications'
   desire to use RG information for remote end checking and GSE's desire
   to hide RG from hosts, however, none of the approaches can satisfy
   both desires at the same time.

   Thus the disadvantages of the Routing Stuff and ESD separation comes
   directly from this separation. We believe that neither the global
   uniqueness of ESDs nor their correct use is enforceable, thus easy
   detection of wrong ESDs becomes the key. Unfortunately, short of
   using IPSEC for every IP packet delivery, using ESD alone loses the
   advantage of easy forge detection that comes from the address
   overloading in IPv4 design.


4.1.8.  When ESDs are Not Unique

   In IPv4, communication usually fails quickly when addresses are not
   unique. There are two cases to consider, depending on whether the two
   interfaces assigned 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



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 34]


INTERNET-DRAFT                                                March 1997


   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 will generally deliver packets to only one of the
   nodes because only one of the links has the right ``prefix'' or
   ``subnet part'' 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.

   An important observation is that, with classical IP, when different
   nodes mistakenly assign the same IP address to different interfaces,
   problems become apparent relatively quickly because communication
   with several (if not all) destinations fails. In contrast, failure
   scenarios differ when globally unique ESDs are assumed, but two nodes
   mistakenly select the same one.

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

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




draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 35]


INTERNET-DRAFT                                                March 1997


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

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

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


4.1.9.  DNS PTR Queries

   IPv4 uses the top-level domain ``IN-ADDR.ARPA'' to hold PTR Resource
   Records. PTR RRs allow a client to map IP addresses back into the
   domain name corresponding to that address. IPv4 addresses can be put
   into the DNS because they have hierarchical structure -- the same
   hierarchy used to aggregate routes.

   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 ``cheap'' authentication by using the
        DNS to map a source address of a peer into a DNS name. Then, the
        client queries the DNS a second time, this time asking for the



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 36]


INTERNET-DRAFT                                                March 1997


        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 alternate mechanism, an ICMP ``Who Are You'' message,
        is described in Section 4.1.12.

     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.

   Although DNS PTR records have proven useful in several contexts,
   there is also widespread agreement that, in practice, most of the IP
   addresses in use today are not properly registered in the PTR
   hierarchy. Consequently, more often than not, PTR queries fail to
   return usable information. Thus, the overall utility of PTR records
   is questionable.

   It is also worth noting that the primary reason that so few addresses
   are properly registered in the PTR space is the lack of incentive for
   doing so. With no key piece of the Internet infrastructure depending
   on such mappings being in place (or correct), there is little harm in
   failing to keep it up-to-date.

   Finally, it might appear at first glance that secure DNS [RFC2065]
   provides a means for cryptographically signing a PTR record and
   thereby providing authentication. Things are not so simple, however.
   The signature on a PTR record indicates that the entity owning an
   address has given it a DNS name. It does not mean that the owner of
   the address is authorized to use that specific name. For example,
   anyone owning an address can set up a PTR record indicating that the
   address corresponds to the name ``www.ietf.org''. However, the name
   ``www.ietf.org'' belongs to only one entity, regardless of how many
   PTR records indicate otherwise.



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 37]


INTERNET-DRAFT                                                March 1997


4.1.10.  Reverse Mapping of ESDs

   If an end point is identified via an ESD rather than by its full
   address, do we need the ability map the ESD alone into some other
   meaningful quantity, such as a fully qualified domain name? The
   benefits of being able to perform such a mapping are analogous to
   those described in the preceding section.

   The primary difficulty with constructing such a mapping is that it
   requires that ESDs have sufficient structure to support the
   delegating mechanism of a distributed database such as DNS. 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. Hence, stateless
   autoconfiguration [RFC1971] cannot create addresses with the
   necessary hierarchical property.

   Another possibility would be to define ESDs with sufficient structure
   to permit the construction of a mapping mechanism. However, analysis
   performed during the IPng deliberations concluded that close to 48-
   bits of hierarchy were needed to identify all the possible sites
   30-40 years from now. That would leave only 2 bytes for host
   numbering at a site, a number clearly incompatible with stateless
   autoconfiguration [RFC1971].

   There are several arguments against requiring a global ESD-lookup
   capability. Adding sufficient structure to an 8-byte ESD requires
   more bits than are compatible with stateless autoconfiguration. In
   addition, experience with the IN-ADDR.ARPA domain suggests that the
   required databases will be poorly maintained. Finally, imposing a
   required hierarchical structure on ESDs would also introduce a new
   administrative burden and a new or expanded registry system to manage
   ESD space. While the procedures for assigning ESDs, which need only
   organizational and not topological significance, would be simpler
   than the procedures for managing IPv4 addresses (or DNS names), it is
   hard to imagine such a process being universally well-received or
   without controversy; it seems a laudable goal to avoid the problem
   altogether if possible.

   Finally, there is an argument based on allocation efficiency. One of
   the design criteria for IPng was support for 10^9 networks and 10^12
   hosts ``and preferably much more'' [RFC1726], and estimates as high
   as 10^15 hosts have been published. Since GSE uses 64 bits to
   designate the Site and subnet and the same number to designate the
   end system, the allocation efficiency for the latter assignment
   process must be much greater. In terms of the H ratio [RFC1715], the
   RG+STP portion of the address needs to achieve an H ratio of only
   0.14, which is as low a value as any of the example numbering schemes



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 38]


INTERNET-DRAFT                                                March 1997


   examined in RFC 1715, while the ESD assignment must achieve 0.19 to
   0.23, depending on the total host estimate one accepts. These values
   are in the middle to high range of the schemes examined, indicating
   that they represent a feasible efficiency -- achievable with careful
   management.


4.1.11.  Reverse Mapping of Complete GSE Addresses

   Within a Site, one could imagine maintaining a database keyed on
   unstructured 8-byte ESDs. However, it is a matter of debate whether
   such a database could 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 such a database can be built.

   A mechanism supporting a lookup keyed on a flat-space ESD from an
   arbitrary Site requires having sufficient structure to identify the
   Site that needs to be queried. In practice, an ESD will almost always
   be used in conjunction with Routing Stuff (i.e., a full 16-byte
   address). Since the Routing Stuff is organized hierarchically, it
   becomes feasible to maintain a DNS 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 RG portion of an address is expected to change over
   time, this assumption will not be valid indefinitely. As a
   consequence, a packet trace recorded in the past might not contain
   enough information to identify the off-Site sources of the packets in
   the present. This problem can be addressed by requiring that the
   database of RG delegations be maintained for some period of time
   after the RG is no longer usable for routing packets.

   Finally, it should be noted that the problem where an address's RG
   ``expires'' with the implication that the mapping of ``expired''
   addresses into DNS names may no longer hold is not a problem specific
   to the GSE proposal. With provider-based addressing, the same issue
   arises when a site renumbers into a new provider prefix and releases
   the allocation from a previous block.








draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 39]


INTERNET-DRAFT                                                March 1997


4.1.12.  The ICMP ``Who Are You'' Message

   Although there is widespread agreement of the utility of being able
   to determine the DNS name one is communicating with, there is also
   widespread concern that repeating the experience of the ``in-
   addr.arpa'' domain is undesirable. 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 in
   Section 4.1.9, the ICMP ``Who Are You'' replaces the DNS PTR query
   used to obtain the DNS name of a remote peer. The second DNS query,
   to map the DNS name back into a set of addresses, would be performed
   as before. Because the latter DNS query  provides the strength of the
   authentication, the use of an ICMP ``Who Are You'' message does not
   in any way weaken the strength of the authentication method. Indeed,
   it can only make it more useful in practice, because virtually all
   hosts can be expected to implement the ``Who Are You'' message.

   The ``Who Are You'' message would contain an identifier for matching
   replies to requests, as well as a nonce value to provide resistance
   to spoofing. In order to minimize the number of WRU packets on the
   Internet, the WRU messages should be sent by DNS servers who would
   then cache the answers. This has the pleasant side-effect of reducing
   the impact on existing applications (i.e., they would continue to
   look up addresses using the same API as before). In many cases there
   is a natural TTL that the target node can provide in its reply:
   either the remaining lifetime of a DHCP lease or the remaining valid
   time of a prefix from which the address was derived through stateless
   autoconfiguration.

   The ``Who Are You?'' (WRU) message described in [section ref:
   ``Reverse Mapping of Complete IPv6 Addresses'' under
   Analysis/ESD/WRU] is robust against renumbering, since it follows the
   paths of valid routable prefixes. Essentially, it uses the Internet
   routing system in place of the DNS delegation scheme. It is
   attractive in the context of GSE-style renumbering, since no host or
   DNS server needs to be updated after a renumbering event for WRU-



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 40]


INTERNET-DRAFT                                                March 1997


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

   The requirement that the target node be up and reachable at the time
   of the query makes it very uncertain that one would be able to take
   addresses from a packet log and translate them to correct domain
   names at a later date. 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 is pleased
   to be able to state that his site has been doing that for some years.

   (Speculative note: Proxy servers to answer WRU queries are possible.
   If the boundary between the global and site portions of addresses are
   fixed and/or the boundary between the routing and the end-node
   portions are fixed, then one could define a well-known anycast
   address for proxy WRU service per site and/or per subnet. The low-
   order portion of this address would presumably be created from the
   IANA's IEEE OUI. The WRU client-side interface would have to be
   defined to try this address after or before sending a query to the
   target address itself. Nodes answering to this anycast address could
   reply to WRU queries using a database maintained by private means.
   By carrying a /128 route site-wide or in the site's provider, these
   servers need not even be located within the subnet or site they
   serve. Co-location of the proxy WRU servers with some DNS servers is
   a natural choice in some scenarios.)


4.2.  Renumbering and Domain Name System (DNS) Issues


4.2.1.  How Frequently Can We Renumber?

   One premise of the GSE proposal [GSE] is that an ISP can renumber the
   Routing Goop portion of a Site's addresses transparently to the Site
   (i.e., without coordinating the change with the Site). This would
   make it possible for backbone providers to aggressively renumber the
   Routing Goop part of addresses and achieve a high degree of route
   aggregation. On closer examination, frequent (e.g., daily)



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 41]


INTERNET-DRAFT                                                March 1997


   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 servers, as well as DNS
   servers for ``bar.com'' and ``foo.bar.com''. Each of these servers
   must be reachable from the querying client. Consequently, there must
   be an overlap period during which both the old Routing Stuff and the
   new Routing Stuff can be used simultaneously. During the overlap
   period, DNS ``glue records'' would need to be updated to use the new
   addresses (including Routing Stuff). Only after all relevant DNS
   servers have been updated and older cached RRs containing the old
   addresses have timed out can the old address be deleted.

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


4.2.2.  Efficient DNS support for Site Renumbering

   When a site renumbers to satisfy its ISP, only the site's routing
   prefix needs to change. That is, the prefix reflects where within the
   Internet the site resides. Although some sites may also change the
   numbering of their internal topology when switching providers, this
   is not a requirement. Rather, it may be a convenient time to also
   perform any desired internal renumbering since one practical view is
   that any address renumbering tends to cause disruptions.

   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
   interaction XXX], it may be the case that individual hosts ``own''
   their own A or AAAA records, further complicating the question of who
   is able to update the contents of DNS RRs.

   One change that could reduce the cost of updating the DNS when a site
   is renumbered is to split addresses into two distinct portions: a



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 42]


INTERNET-DRAFT                                                March 1997


   Routing Stuff that reflects where a node attaches to the Internet and
   a ``site internal part'' that is the site-specific part of an
   address. During a renumbering, only the Routing Stuff would change;
   the ``site internal part'' would remain fixed. Furthermore, the two
   parts of the address could be stored in the DNS as separate RRs. That
   way, renumbering a site would only require that the Routing Stuff 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.

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


4.2.3.  Synthesizing AAAA Records

   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. [this section is really
   weak; what more is needed?]


4.2.4.  Two-Faced DNS

   The GSE proposal [GSE] 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 an RG portion equal to ``Site local,''
   whereas a query for the same name from a client located at a Site



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 43]


INTERNET-DRAFT                                                March 1997


   elsewhere in the Internet would return the appropriate RG portion.
   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 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 response 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).

   [what else needs mentioning?]


4.2.5.  Bootstrapping Issues

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

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



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 44]


INTERNET-DRAFT                                                March 1997


4.2.6.  DNS PTR RRs Not Needed

   Both IPv4 and IPv6 include a mechanism for mapping addresses into DNS
   names (i.e., the PTR RR and IN-ADDR.ARPA and IP6.INT domains
   respectively). The need for such a mechanism can be decreased
   significantly through the proposed ICMP ``Who Are You'' message (see
   Section 4.1.12).

   In any case, PTR records can be implemented essentially as today.
   Assuming that the Routing Stuff of addresses remains hierarchical,
   each ISP would be responsible for delegating its part of the Routing
   Stuff to its downstream customers. Eventually, the Routing Stuff
   would reach the ISP to which the end site connects. This ISP would
   have a pointer to the site's DNS servers which be responsible for the
   rest of the mapping.


4.2.7.  Renumbering and Reverse DNS Lookups

   It is certain that many sites will from time to time undergo a
   renumbering event, either through the mechanisms proposed for GSE or
   using the facilities already specified for IPv6. It would be useful
   to an outside node corresponding with such a site to be able to
   distinguish a legitimate renumbering from an attempt to impersonate
   the site. We claim that the DNS IP6.INT zone, without security
   extensions [RFC2065], is of no use in making this determination and
   that even a completely secured IP6.INT zone is of little use compared
   with the ``forward'' DNS zone.

   The first half of the claim is almost self-evident. An impersonator
   can set up an insecure zone at some point in the IP6.INT hierarchy
   and load it with any desired data. This is the reason that current
   applications doing minimal access control follow a reverse lookup
   with a forward lookup.

   With a secured reverse zone, the problem of verifying an apparent
   renumbering of a site can still be quite complex in the general case,
   and will certainly be outside the scope of a transport protocol, if
   survival of long-running sessions is contemplated. Under provider-
   based addressing [RFC2073], renumbering is expected to occur due to a
   change in network topology (e.g., a change in a provider relationship
   at some point in the address aggregation tree). This alters the
   global prefixes in use below the point of the change, and
   correspondingly alters the chain of delegations of the DNS reverse-
   mapping tree. And, although operational experience with secure DNS is
   quite limited, it seems likely that there would also be a change in
   the chain of certifications of the signing key of the leaf zone
   representing the site. It is then problematic to translate



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 45]


INTERNET-DRAFT                                                March 1997


   established trust in the old reverse mapping zone into trust in the
   new zone. Certainly it's simpler to rely on the forward zone only.
   The only function of the reverse zone, then, is to suggest an entry
   point to the forward zone's database. It is this function which we
   propose to achieve by means of a new ICMP message exchange.


4.3.  Address Rewriting Routers

   One of the most novel pieces of GSE is the rewriting of addresses as
   datagrams enter and leave Sites. If only a small number of routers
   know the RG portion of the addresses, then the operational impact of
   renumbering a Site would be small. In fact, assuming that the
   critical security issues are dealt with, one could imagine a dynamic
   protocol that a Site uses with its upstream provider to be told what
   RG to use, so it might even be possible to renumber a Site
   transparently.

   GSE's ability to ensure that the RG portion of a Site's addresses
   reflect the actual location of that Site within the Public Internet
   means that very aggressive aggregation (i.e., better route scaling)
   can be achieved. Both GSE and other route-scaling approaches that use
   provider-based addressing depend on aggressive aggregation, but while
   other schemes rely largely on operational policies, GSE attempts to
   include mechanisms in its core to ensure that aggressive aggregation
   happens in practice.

   GSE has an advantage over other provider-based addressing schemes
   like IPv4's CIDR with respect to the ``fair distribution of work.''
   CIDR addresses the scaling of routing in DFZ portions of the
   Internet, but the cost of carrying out the renumbering to maintain
   the aggregation falls on the shoulders of subscribers who are far
   away from the DFZ; in other words, subscribers must do the work of
   renumbering so that their provider (or possibly even their provider's
   provider) see better aggregation. With GSE, the majority of the cost
   required to make the routing scale would be incurred by the parties
   who reap the benefits.


4.3.1.  Load Balancing

   While not considered a major advantage, with GSE, multi-homed Sites
   can more easily achieve symmetry with respect to which of their links
   is used for a given flow. With GSE, if HostA in multi-homed Site1
   initiates a flow to HostB in Site2, then when the initial packet
   leaves Site1 the source address will be rewritten with an RG that
   identifies the outgoing link used.  As a result, when HostB needs to
   send return traffic, it will use the full 16-byte address from the



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 46]


INTERNET-DRAFT                                                March 1997


   arriving packet and this necessarily means that traffic for this flow
   coming into Site1 will use the same circuit that outgoing traffic for
   that flow took.


4.3.2.  End-To-End Argument: Don't Hide RG from Hosts

   Despite these significant advantages, however, it was felt that
   address rewriting by routers should not be pursued as part of the
   current standardization effort. Although hiding RG knowledge from
   hosts has advantages in simple scenarios, that lack of knowledge also
   makes it difficult to solve important problems.

   For example, a host in a multi-homed site is known by multiple
   addresses, but without knowing its address the host can play no role
   in the source address selection; instead, the host relies on the
   routing infrastructure to magically select the right one, i.e., by
   selecting the egress router the packet uses to leave the site. In
   this particular case, the historically difficult-to-solve problem of
   source address selection is made more difficult by moving it from an
   intra-host decision to a distributed one. Now a site's internal
   routers would have to have sufficient knowledge to decide which
   egress router to forward traffic to, perhaps on a source-by-source
   (or worse) basis. Another end-to-end problem resulting from address
   rewriting has to do with how transport connections should deal with
   the RG portion of the address in incoming packets, particularly when
   the RG changes.  The sections on transport issues deal with the
   subject in much more detail.

   Interesting questions arise about address rewriting when dealing with
   tunnels. It was realized that any node that can terminate a tunnel
   whose other end point is in a different Site must be able to behave
   as a Site border router and do address rewriting. This means that the
   RG may need to be configured in more than just a Site's egress
   router, thus making renumbering more problematic. Another problem
   related to both performance and ``architectural cleanliness'' has to
   do with IPv6's Routing Headers. It may be necessary for addresses
   other than just the simple source and destination to be rewritten.
   And again, this rewriting would need to be done by both egress
   routers and nodes which terminate tunnels that go to other Sites.



4.4.  Multi-homing

   Multi-homing can mean many things. In the context of GSE, multi-
   homing refers to a Site having more than one connection to the
   Internet and being known by multiple RGs. In many ways this is close



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 47]


INTERNET-DRAFT                                                March 1997


   to multi-homing with IPv6 provider-based addressing. It is hard to
   make comparisons to IPv4 because multi-homing has traditionally been
   done in an ad hoc fashion.

   With GSE, the ability of a Site to control the load-sharing over its
   multiple links is not clear, partially because there is little
   operational experience with multi-homed sites known by multiple
   prefixes (with IPv4 the site is generally only known by a single
   prefix). The following analysis is relevant to any scheme where an
   Internet-connected site is known by multiple prefixes. For flows that
   the multi-homed site initiates, load-sharing is impacted by the
   source address used because that is the address that the remote site
   will use for return traffic. If we assume the model of routers
   rewriting source addresses, then the outgoing link selected
   determines the load-sharing because that also determines what RG is
   contained in the source address. If the routers do not rewrite source
   addresses, then the end-host itself will have to make the source
   address selection, and the optimal choice may require knowledge of
   the topology. For flows initiated by someone outside of the multi-
   homed site, the load-sharing is dependent on the destination address
   specified, so the DNS has a large impact on load-sharing. There is
   some amount of operational experience in using DNS to control load on
   servers (e.g., having a Web server resolve to multiple addresses),
   though that is load-sharing of a different resource and at a
   different scope and scale. It is also worth noting that the selection
   of the optimal outgoing link may well depend on the destination,
   which has particularly interesting results on the DNS understanding
   topology (and brings up the question of whether the servers or the
   resolvers are responsible for knowing the topology).

   One advantage that GSE has for multi-homed Sites is symmetry. Because
   the source address is selected based on the outgoing link, and that
   source address is what determines the return path, flows initiated by
   the Site will be symmetric with respect to which of the Site's links
   is used.

   The multi-homing mechanism described in Section 3.2.4 has some
   weaknesses and complexities. First, the mechanism only supports
   healing a failed link and not a router; in other words, referencing
   Figure 6, from Section 3.2.4, if PBR1 were not up at all, then it
   could not tunnel the packets anywhere. One could imagine ways of
   distributing PBR1's knowledge of PBR2 to other routers within
   Provider1 to add more reliability, though this makes the problem
   distributed rather than point-to-point and therefore more difficult.
   Second, in the general case static identification of PBR2 to PBR1,
   and vice-versa, is not adequate. Imagine, for example, that the link
   to PBR1 is much faster than the link to PBR2. In this case, it's
   possible that packets whose destination addresses contain RG1 might



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 48]


INTERNET-DRAFT                                                March 1997


   normally transit PBR2 without going directly to the Site. So there
   seems to be a need for a dynamic protocol between PBR1 and PBR2 to
   notify when PBR2, for example, should forward RG1-prefaced
   destinations directly to the Site as opposed to forwarding it towards
   PBR1.

   Another note about multi-homing is the potential impact of internal
   topology changes in the face of address rewriting. Using the
   previously referenced diagram, if a flow from a host within the Site
   is leaving via SBR1, but then something happens such that SBR2
   becomes the host's closest exit point, then the remote end point of
   the flow will begin seeing different RG. Reasons such as this are why
   the repercussions on the transport layer are so important (e.g.,
   whether or not transport peers pay attention to the RG).


5.  Recommendations

   This section should be viewed as ``proto recommendations'' and not
   final recommendations. It is impossible to have final recommendations
   until there exists an analysis on which there is consensus.

   A straw-man set of recommendations, along with some related open
   questions, is presented below:

     1) Make changes to the IPv6 provider-based addressing document to
        facilitate aggressive aggregation that is also operationally
        realistic.

     2) Create hard boundaries in IPv6 addresses to clearly distinguish
        between the portions used to identify hosts, for routing within
        a site, and for routing within the Public Internet.

     3) Designate the low-order 8 bytes of IPv6 addresses to be a
        globally unique End System Designator (ESD). This change has
        potential benefits to future transport protocols (e.g., TCPng).
        A point of discussion on this topic is whether, in the short-
        term, the ESD will be used alone; if it isn't to be used alone,
        then how important is the global uniqueness?

     4) Make a clear distinction between the ``locator'' part of an
        address and the ``identifier'' part of the address. The former
        is used to route a packet to its end point, the latter is used
        to identify an end point, independent of the path used to
        deliver the packet. Although this is a potentially revolutionary
        change to IPv6 addressing model, existing transport protocols
        such as TCP and UDP will not take advantage of the split. Future
        transport protocols (e.g., TCPng), however, may.



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 49]


INTERNET-DRAFT                                                March 1997


     5) Make changes to the way AAAA records are stored within the DNS,
        so that renumbering a site (e.g., when a site changes ISPs)
        requires few changes to the DNS database in order to effectively
        change all of a site's address AAAA RRs.

     6) Don't hide a node's full address from that node. In a scheme
        where all nodes know their full address, address rewriting
        should not be necessary.

     7) Consider multi-homing and its effect on aggregation and route
        scaling from the beginning. Have a goal of architecting a way to
        do multi-homing that is both scalable and operationally
        practical, and consider related issues such as load-sharing.

     8) Consider the issue of subnetting. For example, how are point-
        to-point links numbered? With IPv4, current practice is to
        number point-to-point links out of ``/30'' subnets. However, do
        network masks longer than 64 bits make sense with the concept of
        the low-order 8 bytes being a globally unique ESD? If not, then
        is it acceptable to either leave point-to-point links un-
        numbered or to use an entire subnet for each point-to-point
        link? Will there need to be an exception for IPv6 host routes
        (i.e., /128s) as a work-around for the bootstrapping issue of
        addressing root DNS servers? If /128s are allowed, but not masks
        between /65 and /127, inclusive, then a possible way to number
        point-to-point links within a backbone is to dedicate a single
        subnet to them and route them as /128s.

     9) Search for ways to minimize the impact that renumbering can have
        on intra-site communication. Renumbering operations that change
        only the RG portion of addresses should not impact existing
        intra-site communication. One possible approach is to encourage
        the use of site-local addresses for all intra-site
        communication.



6.  Security Considerations

   TBD


7.  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 meeting)
   for the planning and execution of the interim meeting. Thanks also
   goes to Mike O'Dell for writing the 8+8 and GSE drafts. By publishing



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 50]


INTERNET-DRAFT                                                March 1997


   these documents and speaking on their behalf, Mike was the catalyst
   for some very valuable discussions that are expected to result in
   improved IPv6 addressing. Special thanks to the attendees of the
   meeting who carried on the high caliber discussions which were the
   source for this document.


8.  References

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

     [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.

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

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

     [DDNS] ``Dynamic Updates in the Domain Name System (DNS UPDATE)'',
             Paul Vixie (Editor), draft-ietf-dnsind-dynDNS-11.txt,
             November, 1996.
     [GSE] ``GSE - An Alternate Addressing Architecture for IPv6', Mike
             O'Dell, draft-ietf-ipngwg-gseaddr-00.txt.

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

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

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

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



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 51]


INTERNET-DRAFT                                                March 1997


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

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

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

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

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

     [RFC2002] ``IP Mobility Support'', 10/22/1996, C. Perkins.

     [RFC2008] ``Implications of Various Address Allocation Policies for
             Internet Routing'', Y. Rekhter, T. Li.

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

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


9.  Authors' Addresses

   Matt Crawford                           John Stewart
   Fermilab MS 368                         USC/ISI
   PO Box 500                              4350 North Fairfax Drive
   Batavia, IL 60510 USA                   Suite 620
   Phone: 708-840-3461                     Arlington, VA  22203 USA
   EMail: crawdad@fnal.gov                 Phone: 703-807-0132
                                           EMail: jstewart@isi.edu

   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: 703-807-0132

   Thomas Narten
   IBM Corporation
   3039 Cornwallis Ave.
   PO Box 12195 - F11/502
   Research Triangle Park, NC 27709-2195



draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 52]


INTERNET-DRAFT                                                March 1997


   Phone: 919-254-7798
   EMail: narten@raleigh.ibm.com

















































draft-ietf-ipngwg-esd-analysis-00.txt                            [Page 53]