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

Versions: 00 01 02 03 04 05                                             
INTERNET-DRAFT                                             Matt Crawford
                                                                Fermilab
<draft-ietf-ipngwg-esd-analysis-01.txt>                   Allison Mankin
                                                                     ISI
                                                           Thomas Narten
                                                                     IBM
                                                    John W. Stewart, III
                                                                     ISI
                                                             Lixia Zhang
                                                                    UCLA
                                                           July 30, 1997

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

                  <draft-ietf-ipngwg-esd-analysis-01.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 January 30, 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



draft-ietf-ipngwg-esd-analysis-01.txt                           [Page 1]


INTERNET-DRAFT                                             July 30, 1997


   Partition (STP) and a Routing Goop (RG) portion. The STP corresponds
   (roughly) to a site's subnet portion of an IPv4 address, whereas the
   RG identifies the attachment point to the public Internet. Routers
   use the RG+STP portions of addresses (called 'Routing Stuff' in this
   document) 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
   do not know the RG portion of their addresses. A border router at the
   site's 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
   working group eventually decided that the full addresses of nodes
   within a site should not be hidden from those nodes, so as a result
   it is not necessary for routers to rewrite the Routing Goop portion
   of addresses.  However, other parts of the GSE plan were adopted
   (e.g., having 64-bit interface identifiers with an option for
   specifying them as globally unique and easing the renumbering of the
   high-order portion of addresses within DNS).

   In addition to analyzing the GSE proposal in particular, the document
   also studies the general issue of separating network layer addresses
   into two separate values satisfying location and identification
   purposes, respectively.























draft-ietf-ipngwg-esd-analysis-01.txt                           [Page 2]


INTERNET-DRAFT                                             July 30, 1997


   Contents

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

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

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

   3.  GSE Background...........................................   14
      3.1.  Motivation For GSE..................................   14
      3.2.  GSE Address Format..................................   15
      3.3.  Routing Stuff (RG and STP)..........................   15
      3.4.  End-System Designator...............................   17
      3.5.  Address Rewriting by Border Routers.................   18
      3.6.  Renumbering and Rehoming Mid-Level ISPs.............   19
      3.7.  Support for Multi-Homed Sites.......................   20
      3.8.  Explicit Non-Goals for GSE..........................   21

   4.  Analysis of GSE's Advantages and Disadvantages...........   21
      4.1.  End System Designator...............................   21
         4.1.1.  Uniqueness Enforcement in the IPv4 Internet....   21
         4.1.2.  Overloading Addresses: Network Layer Issues....   23
         4.1.3.  Overloading Addresses: Transport Layer Issues..   24
         4.1.4.  Potential Benefits of Globally Unique ESDs.....   25
         4.1.5.  ESD: Network Layer Issues......................   26
         4.1.6.  ESD: Transport Layer Issues....................   28
         4.1.7.  On The Uniqueness Of ESDs......................   34
         4.1.8.  DNS PTR Queries................................   35
         4.1.9.  Reverse Mapping of ESDs........................   37
         4.1.10.  Reverse Mapping of Complete GSE Addresses.....   38
         4.1.11.  The ICMP "Who Are You" Message................   39
      4.2.  Renumbering and Domain Name System (DNS) Issues.....   40
         4.2.1.  How Frequently Can We Renumber?................   40
         4.2.2.  Efficient DNS support for Site Renumbering.....   41
         4.2.3.  Two-Faced DNS..................................   42
         4.2.4.  Bootstrapping Issues...........................   43
         4.2.5.  Renumbering and Reverse DNS Lookups............   44
      4.3.  Address Rewriting Routers...........................   44
         4.3.1.  Load Balancing.................................   45
         4.3.2.  End-To-End Argument: Don't Hide RG from Hosts..   45
      4.4.  Multi-Homing........................................   46

   5.  Results..................................................   48




draft-ietf-ipngwg-esd-analysis-01.txt                           [Page 3]


INTERNET-DRAFT                                             July 30, 1997


   6.  Security Considerations..................................   49

   7.  Acknowledgments..........................................   49

   8.  References...............................................   49

   9.  Authors' Addresses.......................................   51


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," to identify 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 purpose of the meeting was to evaluate the GSE proposal and
   decide whether to adopt it in whole or in part or to reject 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
   IPv6 provider-based addressing plan and had enough benefit to warrant
   further consideration apart from the GSE address format. 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 for identifying hosts and



draft-ietf-ipngwg-esd-analysis-01.txt                           [Page 4]


INTERNET-DRAFT                                             July 30, 1997


        for routing.

     3) Having an option to indicate that the low-order 8 bytes of an
        IPv6 address is a globally unique End System Designator (ESD).
        This change has potential benefits to future transport protocols
        (e.g., TCPng).

     4) Making 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.

     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.

   While this document does contain an analysis of the specific
   mechanisms of the GSE proposal, much of document's analysis applies
   to any proposal in which the identifying and locating properties of
   an address (which are combined in IPv4) are split apart into
   separable pieces.


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




draft-ietf-ipngwg-esd-analysis-01.txt                           [Page 5]


INTERNET-DRAFT                                             July 30, 1997


   Addresses in IPv4 serve two purposes:

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

     2) Location information of that interface. Routers extract location
        information from a packet's destination address in order to
        route it towards its ultimate destination. That is, addresses
        identify "where" the intended recipient is located within the
        Internet topology.

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

   What is important to note is that these identification and location
   requirements have been met through the use of the same value, namely
   the IP address. As will be noted repeatedly in this document, the
   "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 couples those transport protocols with routing.
   This entanglement is inconsistent with a strictly layered model in
   which routing would be a completely independent function of the
   network layer and not directly impact the transport layer.

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

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




draft-ietf-ipngwg-esd-analysis-01.txt                           [Page 6]


INTERNET-DRAFT                                             July 30, 1997


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 motivation for adding subnetting was to
   allow a collection 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). The practical benefit
   of subnetting was that all of a site's hosts, even if scattered among
   tens or hundreds of LANs, could be represented via a single routing
   table entry in routers located far from the site. In contrast, prior
   to subnetting, a site with ten LANs would advertise ten separate
   network entries, and all routers would have to maintain ten separate
   entries, even though they contained redundant information..

   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 networks is individually
   advertised to the global routing subsystem, the complexity of
   computing forwarding tables can easily be an order of magnitude
   greater than if each site advertised just a single entry 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 the same 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 (i.e., customers connected to
   only one provider and the same path was used to reach all customers
   of the same provider), the addressing was not, 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-01.txt                           [Page 7]


INTERNET-DRAFT                                             July 30, 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 entries for all reachable destinations, the cost of
   computing forwarding tables quickly becomes unacceptably large. A
   large part of the cost is related to the seemingly redundant
   computations that must be made for each individual network, even
   though the reality is that many reside in the same topological
   location (e.g., the same provider). Looking at Figure 1, the problem
   is that provider B performs 5 separate calculations to construct the
   routing tables needed to reach each of A's customers.




2.3.  CIDR and Provider-Based Addressing

   One of the reasons Classless Inter-Domain Routing (CIDR) and its
   associated provider-assigned address allocation policy were
   introduced was to help reduce the size of and cost of computing
   forwarding tables. CIDR reduces the cost of computing forwarding
   tables by aggressively aggregating addresses. Aggregating addresses
   means structuring them in such a way that the location of the nodes



draft-ietf-ipngwg-esd-analysis-01.txt                           [Page 8]


INTERNET-DRAFT                                             July 30, 1997


   having those addresses can be represented by a single routing entry.
   In CIDR, this means that addresses share a common prefix. The common
   prefix provides location information for all addresses sharing that
   same prefix.

   In CIDR, sites that want to connect to the Internet approach a
   provider to procure both connectivity and a network address;
   individual providers have a large block of address space covered by
   one prefix and assign pieces of their space to customers.
   Consequently, customers of the same provider have addresses that
   share the same prefix. 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 *one* network
   address into the 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)
                +----------------+
                        |
                        |  A announces
                        |  204.1/16 to B
                        |
                +----------------+
                |   Provider B   |
                +----------------+

                                  Figure 2

   In Figure 2, Provider A has been assigned the classless block, or
   "aggregate," 204.1.0.0/16 (i.e., a 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 prefix subordinate to
   the aggregate.  In order for Provider B to be able to reach
   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 slots in the routing table to reach the same number of



draft-ietf-ipngwg-esd-analysis-01.txt                           [Page 9]


INTERNET-DRAFT                                             July 30, 1997


   destinations.

   CIDR was a critical step for the Internet: in the early 1990s the
   size of default-free routing tables required to support the Classful
   Internet was almost more than the commercially-available hardware and
   software of the day could handle.  The introduction of BGP4's
   classless routing and provider-based address allocation policies
   resulted in 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 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. Today, however, it is simply no longer
   possible for each individual site to have its own private prefix
   injected into the DFZ; there would simply be too many of them.
   Consequently, if a site decides to change providers, then it needs to
   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)
        | A announces   |                |---- Customer5 (204.1.36.0/23)
        | 204.1/16 to B +----------------+
        |                     |
      +----------------+      |
      |   Provider B   |      |
      +----------------+      |
        |                     |
        | C announces         |
        | 204.1.34/24         |
        | to B          +----------------+
        +---------------|   Provider C   |---- Customer3 (204.1.34.0/24)
                        +----------------+

                                  Figure 3

   In Figure 3, each of Provider A, B and C are directly connected to



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 10]


INTERNET-DRAFT                                             July 30, 1997


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

   The example in Figure 3 explains why sites must renumber if existing
   levels of aggregation are to be maintained. While it is certainly
   clear that 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.
   It is generally accepted that some renumbering of sites is essential
   for maintaining sufficient aggregation.

   The empirical cost of renumbering a site in order to maintain
   aggregation has been the subject of much discussion. The practical
   reality, however, is that forcing all sites to renumber is difficult
   given the size and wealth of companies that now depend on the
   Internet for running their business. Thus, although the technical
   community came to consensus that address lending was necessary in
   order for the Internet to continue to operate and grow, the reality
   has been that some of CIDR's benefits have been lost because 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 whose major headache is managing
   routing table growth in the DFZ) are not the pieces that incur the
   cost (i.e., the end site). The logical corollary of this statement is
   that the pieces of the infrastructure 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.  Multi-Homing and Aggregation

   As sites become more dependent on the Internet, they have begun to
   install additional connections to the Internet to improve robustness



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 11]


INTERNET-DRAFT                                             July 30, 1997


   and performance. Such sites are called "multi-homed."  Unfortunately,
   when a site connects to the Internet at multiple places, the impact
   on routing can be much like a site that switches providers but
   refuses to renumber.

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

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

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

   It should be noted that the above example simplifies a very complex



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 12]


INTERNET-DRAFT                                             July 30, 1997


   issue. For example, consider the example in Figure 4 again. Provider
   A could choose *not* to propagate a route entry for the longer
   2.4.1.0/19 prefix, advertising only the shorter 204.1/16. In such
   cases, provider C would always select Provider B. Internally,
   Provider A would continue to router traffic from its other customers
   to customer 1 directly. If Provider A had a large enough customer
   base, effective load sharing would achieved.


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


                                Figure 4


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

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

      2) The site could assign each host multiple addresses (i.e., one
        address for each ISP connection).  There are two problems with
        this.  First, it accelerates the consumption of the address
        space.  Second, when the connection to ISP-A goes down,
        addresses numbered out of ISP-A's space become unreachable.
        Remote peers would have to have sufficient intelligence to use
        the second address. For example, when initiating a connection to
        a host, the DNS would return multiple candidate addresses.
        Clients would need to try them all before concluding that a
        destination is unreachable (something not all hosts currently



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 13]


INTERNET-DRAFT                                             July 30, 1997


        do). In addition, a site's hosts would need a significant amount
        of intelligence for choosing the source addresses they use. A
        host shouldn't choose a source address corresponding to a
        addresses that are not reachable from the Public Internet. At
        present, hosts do not have such sophistication.

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



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].
   We begin by reviewing the motivation for GSE. Next we review the
   salient technical details, and we conclude by listing the explicit
   non-goals of the GSE proposal.


3.1.  Motivation For GSE

   The primary motivation for GSE is the fact that the chief IPv6 global
   unicast address structure, provider-based [RFC 2073], is
   fundamentally the same as IPv4 with CIDR and provider-based
   aggregation. Provider-based addressing requires that sites renumber
   when they switch providers, so that sites are always aggregated
   within their provider's prefix. In practice, the cost of renumbering
   (which can only grow as a site grows in size and becomes more
   dependent on the Internet for day-to-day business) is high enough
   that an increasing number of sites refuse to renumber.  This cost is
   particularly relevant in cases where end-users are asked to renumber
   because an upstream provider has changed its transit provider (i.e.,
   the end site is asked to renumber for reasons outside of its control
   and for which it sees no direct benefit).  Consequently, The GSE
   draft 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.

   The GSE proposal does not propose to eliminate the need for



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 14]


INTERNET-DRAFT                                             July 30, 1997


   renumbering. Indeed, it asserts that end sites will have to be
   renumbered more frequently in order to continue scaling the Internet.
   However, GSE proposes to make the cost of such a renumbering so
   small, that sites could be renumbered at essentially any time with
   only minor disruption to the site.

   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 8 bytes for unique
   identification of an end-point. The structure of GSE addresses is:

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

                                 Figure 5



3.3.  Routing Stuff (RG and STP)

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







draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 15]


INTERNET-DRAFT                                             July 30, 1997



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

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

                                 Figure 6

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

   Having described the decomposition process, we can now examine the
   bits in the RG. After the 3-bit prefix identifying the address as
   GSE, the next 13 bits identify the LS. By limiting the field to 13
   bits, a ceiling is defined on the complexity of the top-most routing
   level. In the next 34 bits, a series of subordinate structure(s) are
   identified until finally the leaf subordinate structure is
   identified, at which point the remaining bits identify the individual
   link within that leaf structure. The remaining 14 bits of the Routing
   Stuff comprise the STP and 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 the Public Internet,
   while Routing Stuff includes the RG plus the Site Topology Partition
   (STP). The STP is used for routing structure within a Site.

   The GSE proposal 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.



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 16]


INTERNET-DRAFT                                             July 30, 1997


   A Site can have an arbitrarily complicated topology, but all of that
   complexity is hidden from everyone outside of the Site.  A Site only
   carries packets which originated from, or are destined to, that Site;
   in other words, a Site cannot be a transit network. A Site is private
   topology, while the transit networks form the public topology.

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


3.4.  End-System Designator

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

   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.

   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. Several possible
   mechanisms can be imagined (e.g., the IANA could hand out addresses
   from the company id assigned it has been allocated), but we do not
   explore them in detail here.






draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 17]


INTERNET-DRAFT                                             July 30, 1997


3.5.  Address Rewriting by Border Routers

   GSE Site border routers rewrite addresses of the packets they forward
   across the Site/Public Topology 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 Topology, the border router
   replaces the Site-Local RG portion of packet's source address with an
   appropriate value. Likewise, when a packet from the Public Topology
   is forwarded into a Site, the border router replaces the RG part of
   the destination address with the designated Site-Local RG.

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

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

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

   The point of keeping the RG hidden from nodes within the core of a
   Site is to insure the changeability of 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 example, opening a
   TCP connection, writing the address of the peer to a file and then



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 18]


INTERNET-DRAFT                                             July 30, 1997


   later trying to reestablish a connection to that peer is likely to
   fail.  For intra-Site communication, however, it is expected that
   only the Site-Local RG would be used (and stored) which would
   continue to work for intra-Site communication regardless of changes
   to the Site's external RG. This has the benefit of shielding a site's
   internal traffic from the affects of renumbering changes outside of
   the site.

   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
   connection to three Internet providers. Because each of those
   connections has its own RG, each destination within the Site would be
   known by three different 16-byte addresses. As a result, intra-Site
   routers would have to carry a routing table three times larger than
   expected. 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.

   In summary, 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. The initiating
   node places the full 16-byte address in the destination address field
   of the datagram, and that field stays intact through the first Site
   and through all of the Public Topology.  When the datagram reaches
   the exit border router, the router replaces the RG of the packet's
   source address.  When the datagram arrives at entry router at the
   destination Site, the router replaces the RG portion of the
   destination address with the distinguished "Site-Local RG" value.
   When the destination host needs to send return traffic, that host
   knows the full 16-byte address for the 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
   to BigISP2 (in the CIDR model), 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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 19]


INTERNET-DRAFT                                             July 30, 1997


   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 Multi-Homed Sites

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

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

                                    Figure 7

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







draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 20]


INTERNET-DRAFT                                             July 30, 1997


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

   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

   This section contains the bulk of the GSE analysis and the analysis
   of the general locator/identifier split.


4.1.  End System Designator


4.1.1.  Uniqueness Enforcement 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.
   Embedding location information within an address has the side-effect
   of helping insure that all addresses are globally unique. 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).
   This is important for two reasons. It helps detect misconfigurations
   (use of the wrong address prevents communication from taking place),
   and helps thwart intruders.




draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 21]


INTERNET-DRAFT                                             July 30, 1997


   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
   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.

   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.




draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 22]


INTERNET-DRAFT                                             July 30, 1997


4.1.2.  Overloading Addresses: Network Layer Issues

   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. In IPv4, the entire address 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 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 terminology) 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 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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 23]


INTERNET-DRAFT                                             July 30, 1997


        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 failed interface of a
        multi-homed host won't be delivered, even though the node is
        reachable through alternate interfaces.

   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.


4.1.3.  Overloading Addresses: Transport Layer Issues

   The problems discussed previously create particular complications 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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 24]


INTERNET-DRAFT                                             July 30, 1997


        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, it is infeasible 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
   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.  Potential Benefits of Globally Unique ESDs

   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.
   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 is who is
   responsible for finding the Routing Stuff associated with a given
   ESD? There are a number of possibilities:



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 25]


INTERNET-DRAFT                                             July 30, 1997


     1) The network layer could be responsible for doing the mapping.
        The advantage of such a system is that an ESD could be stored
        essentially forever (e.g., in configuration files), but whenever
        it is actually used, network layer software would automatically
        perform the mapping to determine the appropriate Routing Stuff
        for the destination. Likewise, should an existing mapping become
        invalid, network layer software could dynamically determine the
        updated 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 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 or to validate that the Routing Stuff is correct.  When an
   application has data to send, it queries the DNS to obtain the IPv6
   AAAA record for a destination. The returned AAAA record contains both
   the Routing Stuff and the ESD of the specified destination. While
   such an approach eliminates the need for the lower layers to be able
   to map ESDs into corresponding Routing Stuff, it also means that when
   presented with an address containing an incorrect (i.e., no longer
   valid) Routing Stuff, the network is unable to deliver the packet to
   its correct destination. 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:



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 26]


INTERNET-DRAFT                                             July 30, 1997


     1) Addresses must have a locator embedded within them. It is not
        feasible to route packets solely on an ESD; doing so would make
        it impossible to aggregate routing information in a scalable
        way. The GSE proposal assumes that the locator part of an
        address is filled with an appropriate value by higher layers
        (i.e., the transport or application layer).

     2) If a receiver observes that recent packets are arriving with a
        different Routing Stuff in the source address than before, it
        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
        its 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 while other packets
        exit 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.

     3) 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.





draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 27]


INTERNET-DRAFT                                             July 30, 1997


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 end-point
   identifiers poses problems for mobility and site renumbering. This
   section discusses an alternate approach, in which transport end-point
   identifiers use ESDs rather than full addresses (with embedded
   Routing Stuff).

   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." 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, raising
   deployability questions.


4.1.6.1.  Demultiplexing Packets to Transport Endpoints

   Connections in GSE are identified by the ESDs rather than full IPv6
   addresses (with embedded Routing Stuff). That is:

        unique IPv4 TCP connection:     srcaddr dstaddr srcport destport
        unique GSE TCP connection:      srcESD dstESD srcport dstport

   Consequently, with GSE, 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 on ESDs alone without the RS is, in fact,
   required with GSE. If a site is multi-homed, 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 outgoing packets, the receiving TCP must ignore
   (at least) the RG portion of addresses when demultiplexing received
   packets. The alternative would be to make TCP 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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 28]


INTERNET-DRAFT                                             July 30, 1997


   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).


4.1.6.3.  RG Selection When Sending Packets

   When a host has a packet to send, there are three cases for deciding
   what RG to use in the destination.  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
   contains usable RG. Note that assuming that the RG on an incoming TCP
   connection is "correct" needs qualification. It is "correct" in the
   sense that it corresponds to the site originating the connection.
   Whether the ESD paired with the RG is actually located at that site
   cannot be assumed. The issue of spoofing is discussed in more detail
   later.  The last (and most interesting) case is when RG changes mid-
   connection. Although, the GSE proposal calls for always using the
   first RG learned (and then never switching), we explored the
   possibility of doing so in order to better understand the issues.


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 (in the worst case) that traffic-balancing schemes could
   result in the use of two egress routers, with roughly every other
   packet exiting through a different egress router. 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), but care must be
   taken not to thrash. Moreover, simply using the most-recently-
   received RG makes it trivial for an intruder to hijack connections.

   Because TCP under GSE demultiplexes packets using only ESDs, packets
   will be delivered to the correct end-point regardless of what source
   RG is used. However, return traffic will continue to be sent via the



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 29]


INTERNET-DRAFT                                             July 30, 1997


   "old" RG, even though it may have been deprecated or become less
   optimal because the peer's border router has changed.  It would seem
   highly desirable for TCP connections to be able to survive such
   events. However, the completion of renumbering events (so that an
   earlier RG is now invalid) and certain topology changes would require
   TCP to switch sending to a new RG mid-connection.  To explore the
   whole space, we considered ways of allowing this mid-connection RG
   change to happen.

   If TCP connection identifiers are based on ESDs rather than full
   addresses, traffic from the same ESD would be viewed as coming from
   the same peer, regardless of its source RG. 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 appear to significantly worsen
   existing robustness.

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

   Providing protection from hijacking implies that the RG used to send
   packets must be bound to a connection end-point (e.g., it is part of
   the connection state). Although it may be reasonable to accept
   incoming traffic independent of the source RG, the choice of sending
   RG requires more careful consideration. Indeed, any subsequent change
   in what RG is used for sending traffic must be properly authenticated
   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 seen.

   In summary, changing the RG dynamically in a safe way for a
   connection requires that an originator of traffic be able to
   authenticate a proposed change in the RG before sending to a



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 30]


INTERNET-DRAFT                                             July 30, 1997


   particular ESD via that RG. Such a mechanism would need to be
   invented, as the TCP/IP suite has no obvious candidate that operates
   at or below the transport layer (using the DNS, an application
   protocol that resides above IP, would be problematic due to layering
   circularity considerations).


4.1.6.5.  Passive Opens

   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 8 shows
   the details.


         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 8


   We next consider relaxing the restriction on switching RGs in an
   attempt to avoid the previous failure scenario. The situation is
   complicated by the fact that the RG on received packets may change
   for legitimate reasons (e.g., a multi-homed site load-shares traffic
   across multiple border routers). The key question is how can one



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 31]


INTERNET-DRAFT                                             July 30, 1997


   determine which RG is valid and which is not. That is, for each of
   the RGs a sender attempts to use, how can it determine which RG
   worked and which did not? Solving this problem is more difficult than
   first appears, since one must cover the cases of delayed segments,
   lost segments, simultaneous opens, etc. If a SYN-ACK is retransmitted
   using different RGs, it is not possible to determine which of those
   RGs worked correctly. We conclude that the only way TCP could
   determine that a particular RG 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 (a non-trivial
   addition to TCP).

   We analyze multiple cases of RG changing within the time of the
   opening handshake. One example is diagrammed in Figure 9, and it and
   two others are summarized in Table 1. We observe 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.

       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 9















draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 32]


INTERNET-DRAFT                                             July 30, 1997



                    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 conclude that although the corrupted SYN case of
   Figure 8 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
   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.

   Consequently, there does not appear to be a straightforward way to
   use ESDs in conjunction with mobility or site renumbering (in which
   existing connections survive the renumbering).





draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 33]


INTERNET-DRAFT                                             July 30, 1997


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.

   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. This insures that packets are delivered to their
        intended destination  independent of RG.

     2) Once a packet has been delivered to its transport end-point, a
        separate (i.e., distinct) decision should be made concerning
        whether and how to act upon the received packet. Such a decision
        would be transport-protocol specific. 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.  On The Uniqueness Of ESDs

   The uniqueness requirements for ESDs depends on what purpose they
   serve. In GSE, ESDs identify end systems, requiring  that they be
   globally unique. It does not make sense for two different end systems
   to use the same ESD; every end system must have its own ESD to
   distinguish from other end systems.

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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 34]


INTERNET-DRAFT                                             July 30, 1997


   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.

   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 are delivered to the same
        end-point).


4.1.8.  DNS PTR Queries

   IPv4 uses the 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.



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 35]


INTERNET-DRAFT                                             July 30, 1997


   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
        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 4.1.11.

     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, many IP
   addresses in use today are not properly registered in the IN-
   ADDR.ARPA namespace. Consequently, PTR queries frequently 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 absence 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
   practical harm in failing to keep it up-to-date.



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 36]


INTERNET-DRAFT                                             July 30, 1997


   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.


4.1.9.  Reverse Mapping of ESDs

   It is reasonable to ask if it is necessary or desirable to be able to
   map an 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 having a global ESD-lookup
   capability. Adding sufficient structure to an 8-byte ESD would be
   incompatible with stateless autoconfiguration, which already uses 6
   bytes for its token; two additional bytes for hierarchy are clearly
   insufficient. 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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 37]


INTERNET-DRAFT                                             July 30, 1997


   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.


4.1.10.  Reverse Mapping of Complete GSE Addresses

   Although it seems infeasible to have a global scale, reverse mapping
   of ESDs, within a Site, one could imagine maintaining a database
   keyed on unstructured 8-byte ESDs. However, it is a matter of debate
   whether such a database can be kept up-to-date at reasonable cost,
   without making unreasonable assumptions as to how large sites are
   going to grow, and how frequently ESD registrations will be made or
   updated. Note that the issue isn't just the physical database itself,
   but the operational issues involved in keeping it up-to-date. For the
   rest of this section, however, let us assume 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. The authors are aware of one
   such renumbering in IPv4 where a block of returned addresses was
   reassigned and reused within 24 hours of the renumbering.






draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 38]


INTERNET-DRAFT                                             July 30, 1997


4.1.11.  The ICMP "Who Are You" Message

   Although there is widespread agreement on the utility of being able
   to determine the DNS name one is communicating with, there is also
   widespread concern that repeating the experience of the "IN-
   ADDR.ARPA" domain is undesirable. 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 could contain an identifier for matching
   replies to requests, and perhaps 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 4.1.10 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-based lookups to
   work. It has advantages outside the context of GSE as well, including



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 39]


INTERNET-DRAFT                                             July 30, 1997


   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)
   renumbering turns out to be difficult in practice because of a
   circular dependency between the DNS and routing. Specifically, if a



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 40]


INTERNET-DRAFT                                             July 30, 1997


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

   One change that could reduce the cost of updating the DNS when a site
   is renumbered is to split addresses into two distinct portions: a
   Routing Goop that reflects where a node attaches to the Internet and
   a "site internal part" that is the site-specific part of an address.



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 41]


INTERNET-DRAFT                                             July 30, 1997


   During a renumbering, only the Routing Goop 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 Goop RR of a
   site be updated; the "site-internal part" of individual addresses
   would not change.

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

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

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

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



4.2.3.  Two-Faced DNS

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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 42]


INTERNET-DRAFT                                             July 30, 1997


   faced" DNS servers.

   Some issues that must be considered in this context:

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

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



4.2.4.  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-01.txt                          [Page 43]


INTERNET-DRAFT                                             July 30, 1997


4.2.5.  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
   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.



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 44]


INTERNET-DRAFT                                             July 30, 1997


   GSE's ability to insure 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 insure 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) sees 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 egress link used.  As a result, when HostB needs to
   send return traffic, it will use the full 16-byte address from the
   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. In contrast, if the source address (i.e., Routing
   Stuff) is fixed by the sending host, the same return path is used for
   return traffic coming back to a site, regardless of which egress
   router packets traverse when leaving that site.


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

   Despite these significant advantages, however, the overwhelming
   consensus was 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 some 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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 45]


INTERNET-DRAFT                                             July 30, 1997


   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 closest to the sender. For many sites,
   this is the desired behavior. For others, this is not the desired
   behavior. In those cases, 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 authenticating 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.  Any node that acts as a tunnel for which the other end
   resides 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 therefore being known by multiple RGs. In many ways this
   is close 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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 46]


INTERNET-DRAFT                                             July 30, 1997


   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 DNS 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 3.7 has some
   weaknesses and complexities. First, the mechanism only supports
   healing a failed link and not a router; in other words, referencing
   Figure 7, from Section 3.7, 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
   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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 47]


INTERNET-DRAFT                                             July 30, 1997


   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.  Results

   This section summarizes the results of the GSE deliberations on the
   IPv6 process.

     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) Allow an option for the low-order 8 bytes of IPv6 addresses to
        be designated as 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) 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



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 48]


INTERNET-DRAFT                                             July 30, 1997


        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 has 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

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


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 PAL1
   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 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

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




draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 49]


INTERNET-DRAFT                                             July 30, 1997


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

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

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

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

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

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

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

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

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

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

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

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

     [RFC1752] "The Recommendation for the IP Next Generation Protocol,"



draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 50]


INTERNET-DRAFT                                             July 30, 1997


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





draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 51]


INTERNET-DRAFT                                             July 30, 1997





















































draft-ietf-ipngwg-esd-analysis-01.txt                          [Page 52]