Network Working Group                                            E. Lear
Request for Comments: 1627                        Silicon Graphics, Inc.
Category: Informational                                          E. Fair
                                                    Apple Computer, Inc.
                                                              D. Crocker
                                                  Silicon Graphics, Inc.
                                                              T. Kessler
                                                  Sun Microsystems, Inc.
                                                               July 1994

                     Network 10 Considered Harmful
                 (Some Practices Shouldn't be Codified)

Status of this Memo

   This memo provides information for the Internet community.  This memo
   does not specify an Internet standard of any kind.  Distribution of
   this memo is unlimited.


   Re-use of Internet addresses for private IP networks is the topic of
   the recent RFC 1597 [1].  It reserves a set of IP network numbers,
   for (re-)use by any number of organizations, so long as those
   networks are not routed outside any single, private IP network.  RFC
   1597 departs from the basic architectural rule that IP addresses must
   be globally unique, and it does so without having had the benefit of
   the usual, public review and approval by the IETF or IAB.  This
   document restates the arguments for maintaining a unique address
   space.  Concerns for Internet architecture and operations, as well as
   IETF procedure, are explored.


   Growth in use of Internet technology and in attachments to the
   Internet have taken us to the point that we now are in danger of
   running out of unassigned IP network numbers.  Initially, numbers
   were formally assigned only when a network was about to be attached
   to the Internet.  This caused difficulties when initial use of IP
   substantially preceded the decision and permission to attach to the
   Internet.  In particular, re-numbering was painful.  The lesson that
   we learned was that every IP address ought to be globally unique,
   independent of its attachment to the Internet.  This makes it
   possible for any two network entities to communicate, no matter where
   either might be located.  This model is the result of a decades-long
   evolution, through which the community realized how painful it can be
   to convert a network of computers to use an assigned number after

Lear, Fair, Crocker & Kessler                                   [Page 1]

RFC 1627             Network 10 Considered Harmful             July 1994

   using random or default addresses found on computers just out of the
   box.  RFC 1597 abrogates this model without benefit of general IETF
   community discussion and consensus, leaving policy and operational
   questions unasked and unanswered.


   A common -- if not universal -- ideal for the future of IP is for
   every system to be globally accessible, given the proper security
   mechanisms.  Whether such systems comprise toasters, light switches,
   utility power poles, field medical equipment, or the classic examples
   of "computers", our current model of assignment is to ensure that
   they can interoperate.

   In order for such a model to work there must exist a globally unique
   addressing system.  A common complaint throughout the community is
   that the existing security in host software does not allow for every
   (or even many) hosts in a corporate environment to have direct IP
   access.  When this problem is addressed through proper privacy and
   authentication standards, non-unique IP addresses will become a
   bottleneck to easy deployment if the recommendations in RFC 1597 are

   The IP version 4 (IPv4) address space will be exhausted.  The
   question is simply:  when?

   If we assert that all IP addresses must be unique globally, connected
   or not, then we will run out of IP address space soon.

   If we assert that only IP addresses used on the world-wide Internet
   need to be globally unique, then we will run out of IP address space

   It is absolutely key to keep the Internet community's attention
   focused on the efforts toward IP next generation (IPng), so that we
   may transcend the limitations of IPv4.  RFC 1597 produces apparent
   relief from IPv4 address space exhaustion by masking those networks
   that are not connecting to the Internet, today.  However, this
   apparent relief will likely produce two results: complacency on the
   large part of the community that does not take the long term view,
   and a very sudden IP address space exhaustion at some later date.

   Prior to IPng deployment, it is important to preserve all the
   semantics that make both the Internet and Internet technology so very
   valuable for interoperability.  Apple Computer, IBM, and Motorola
   could not collaborate as easily as they have to produce the PowerPC
   without uniquely assigned IP addresses. The same can be said of the
   Silicon Graphics merger with MIPS. There are many, many more examples

Lear, Fair, Crocker & Kessler                                   [Page 2]

RFC 1627             Network 10 Considered Harmful             July 1994

   that can be cited.

   It should be noted that a scheme similar to RFC 1597 can be
   implemented at the time that we actually run out of assignable IPv4
   address space; it simply requires that those organizations which have
   been assigned addresses but are not yet connected to the Internet
   return their addresses to IANA. It is important that the IAB (and
   IANA as its agent) reassert their ownership of the IP address space
   now, to preclude challenges to this type of reassignment.


RFC 1597 Implementations

   Methods are needed to ensure that the remaining addresses are
   allocated and used frugally.  Due to the current problems, Internet
   service providers have made it increasingly difficult for
   organizations to acquire public IP network numbers.  Private networks
   have always had the option of using addresses not assigned to them by
   appropriate authorities.  We do not know how many such networks
   exist, because by their nature they do not interact with the global
   Internet.  By using a random address, a company must take some care
   to ensure it is able to route to the properly registered owner of
   that network.

   RFC 1597 proposes to solve the routing problem by assigning numbers
   that will never be used outside of private environments.  Using such
   standard numbers introduces a potential for clashes in another way.
   If two private networks follow RFC 1597 and then later wish to
   communicate with each other, one will have to renumber.  The same
   problem occurs if a private network wishes to become public.  The
   likely cost of renumbering is linear to the number of hosts on a
   network.  Thus, a large company with 10,000 hosts on a network could
   incur considerable expense if it either merged with another company
   or joined the Internet in such a way as to allow all hosts to
   directly access the outside network.

   The probability of address clashes occurring over time approach 100%
   with RFC 1597.  Picking a random network number reduces the chances
   of having to renumber hosts, but introduces the routing problems
   described above.  Best of all, retrieving assigned numbers from the
   appropriate authority in the first place eliminates both existing and
   potential address conflicts at the cost of using a part of the
   address space.

   Apple Computer once believed that none of its internal systems would
   ever speak IP directly to the outside world, and as such, network
   operations picked IP class A network 90 out of thin air to use.

Lear, Fair, Crocker & Kessler                                   [Page 3]

RFC 1627             Network 10 Considered Harmful             July 1994

   Apple is only now recovering from this error, having renumbered some
   5,000 hosts to provide them with "desktop" Internet access.  Unless
   the Internet community reaffirms its commitment to a globally unique
   address space, we condemn many thousands of organizations to similar
   pain when they too attempt to answer the call of the global Internet.

   Another timely example of problems caused by RFC 1597 is Sun's use of
   Internet multicasting.  Sun selectively relays specific multicast
   conferences.  This has the effect of making many hosts at Sun visible
   to the Internet, even though they are not addressable via IP unicast
   routing.  If they had non-global addresses this would not work at
   all.  It is not possible to predict which machines need global
   addresses in advance.  Silicon Graphics has a similar configuration,
   as is likely for others, as well.

   Some might argue that assigning numbers to use for private networks
   will prevent accidental leaks from occurring through some sort of
   convention a'la Martian packets.  While the proposal attempts to
   create a standard for "private" address use, there is absolutely no
   way to ensure that other addresses are not also used.

   Hence, the "standard" becomes nothing but a misleading heuristic.  In
   fact, it is essential that routers to the global Internet advertise
   networks based only on explicit permission, rather than refusing to
   advertise others based on implicit prohibition, as supported by the
   policy formally created in RFC 1597.

Security Issues

   Administrators will have a hard time spotting unauthorized networks,
   when their network has been breached (either intentionally or
   unintentionally) because the other networks might have the same
   numbers as those normally in the routing tables.  More over, an
   inadvertent connection could possibly have a double whammy effect of
   partitioning two operational networks.

   It is worth emphasizing that IP providers should filter out all but
   authorized networks.  Such a practice would not only prevent
   accidents but also enhance the security of the Internet by reducing
   the potential number of points of attack.

   Internet multicasting adds a new dimension to security.  In some
   cases it may possible to allow multicasting through firewalls that
   completely restrict unicast routing.  Otherwise unconnected networks
   might well need unique addresses, as illustrated in the example

Lear, Fair, Crocker & Kessler                                   [Page 4]

RFC 1627             Network 10 Considered Harmful             July 1994

Problems with Examples

   RFC 1597 gives several examples of IP networks that need not have
   globally unique address spaces.  Each of those cases is plausible,
   but that does not make it legitimate to ENCOURAGE non-uniqueness of
   the addresses.  In fact, it is equally plausible that globally unique
   IP addresses will be required, for every one of the scenarios
   described in RFC 1597:

   - Airport displays are public information and multicasting beyond the
     airport might be useful.

   - An organization's machines which, today, do not need global
     connectivity might need it tomorrow.  Further, merging
     organizations creates havoc when the addresses collide.

   - Current use of firewalls is an artifact of limitations in the
     technology.  Let's fix the problem, not the symptom.

   - Inter-organization private links do not generate benefit from being
     any more correct in guessing which machines want to interact than
     is true for general Internet access.

   This is another point that warrants repetition: the belief that
   administrators can predict which machines will need Internet access
   is quite simply wrong.  We need to reduce or eliminate the penalties
   associated with that error, in order to encourage as much Internet
   connectivity as operational policies and technical security permit.
   RFC 1597 works very much against this goal.

Problems With "Advantages" And More Disadvantages

   RFC 1597 claims that Classless Inter-Domain Routing (CIDR) will
   require enterprises to renumber their networks.  In the general case,
   this will only involve those networks that are routed outside of
   enterprises.  Since RFC 1597 addresses private enterprise networks,
   this argument does not apply.

   The authors mention that DCHP-based tools [2] might help network
   number transition.  However, it is observed that by and large such
   tools are currently only "potential" in nature.

   Additionally, with the onslaught of ISDN, slip, and PPP in host
   implementations, the potential for a workstation to become a router
   inadvertently has never been greater.  Use of a common set of
   addresses for private networks virtually assures administrators of
   having their networks partitioned, if they do not take care to
   carefully control modem connections.

Lear, Fair, Crocker & Kessler                                   [Page 5]

RFC 1627             Network 10 Considered Harmful             July 1994

   Finally, RFC 1597 implies that it may be simple to change a host's IP
   address.  For a variety of reasons this may not be the case, and it
   is not the norm today.  For example, a host may be well known within
   a network.  It may have long standing services such as NFS, which
   would cause problems for clients were its address changed.  A host
   may have software licenses locked by IP address.  Thus, migrating a
   host from private to global addressing may prove difficult.  At the
   very least, one should be careful about addressing well known hosts.


IANA Has Overstepped Their Mandate

   For many years, IANA has followed an assignment policy based on the
   expectation of Internet connectivity for ALL assignees.  As such it
   serves to encourage interconnectivity.  IANA assignment of the
   network numbers listed in RFC 1597 serves to formally authorize
   behavior contrary to this accepted practice.  Further, this change
   was effected without benefit of community review and approval.

   RFC 1597 specifies a new operational requirement explicitly: network
   service providers must filter the IANA assigned network numbers
   listed in RFC 1597 from their routing tables.  This address space
   allocation is permanently removed from being used on the Internet.

   As we read RFC 1601 [3], this action is not within the purview of
   IANA, which should only be assigning numbers within the current
   standards and axioms that underlie the Internet.  IP network numbers
   are assigned uniquely under the assumption that they will be used on
   the Internet at some future date.  Such assignments violate that
   axiom, and constitute an architectural change to the Internet.  RFC
   1602 [4] and RFC 1310 [5] also contain identical wording to this
   effect in the section that describes IANA.

   While RFC 1597 contains a view worthy of public debate, it is not
   ready for formal authorization.  Hence, we strongly encourage IANA to
   withdraw its IP address assignments documented by RFC 1597 forthwith.

   The IAB should review the address assignment policies and procedures
   that compose IANA's mandate, and reaffirm the commitment to a
   globally unique IP address space.


   The Internet technology and service is predicated on a global address
   space.  Members of the Internet community have already experienced
   and understood the problems and pains associated with uncoordinated
   private network number assignments.  In effect the proposal attempts

Lear, Fair, Crocker & Kessler                                   [Page 6]

RFC 1627             Network 10 Considered Harmful             July 1994

   to codify uncoordinated behavior and alter the accepted Internet
   addressing model.  Hence, it needs to be considered much more

   RFC 1597 gives the illusion of remedying a problem, by creating
   formal structure to a long-standing informal practice.  In fact, the
   structure distracts us from the need to solve these very real
   problems and does not even provide substantive aid in the near-term.

   In the past we have all dreaded the idea of having any part of the
   address space re-used.  Numerous luminaries have both written and
   spoke at length, explaining why it is we want direct connections from
   one host to another.  Before straying from the current architectural
   path, we as a community should revisit the reasoning behind the
   preaching of unique addressing.  While RFC 1597 attempts to change
   this model, its costs and limitations for enterprises can be
   enormous, both in the short and long term.


   [1]  Rekhter, Y., Moskowitz, B., Karrenberg, D., and G. de Groot,
        "Address Allocation for Private Internets", T.J. Watson Research
        Center, IBM Corp., Chrysler Corp., RIPE NCC, RFC 1597, March

   [2]  Droms, R., "Dynamic Host Configuration Protocol", RFC 1541,
        Bucknell University, October 1993.

   [3]  Huitema, C., "Charter of the Internet Architecture Board (IAB)",
        RFC 1601, IAB, March 1994.

   [4]  Internet Architecture Board, Internet Engineering Steering
        Group, "The Internet Standards Process -- Revision 2", IAB,
        IESG, RFC 1602, March 1994.

   [5]  Internet Activities Board, "The Internet Standards Process", RFC
        1310, IAB, March 1992.

   [6]  Internet Activities Board, "Summary of Internet Architecture
        Discussion", Notes available from ISI, [
        pub/IAB/IABmins.jan91Arch.txt], IAB, January 1991.


   See the section, "Security Issues".

Lear, Fair, Crocker & Kessler                                   [Page 7]

RFC 1627             Network 10 Considered Harmful             July 1994


   Eliot Lear
   Silicon Graphics, Inc.
   2011 N. Shoreline Blvd.
   Mountain View, CA

   Phone: +1 415 390 2414

   Erik Fair
   Apple Computer, Inc.
   1 Infinite Loop
   Cupertino, CA 95014

   Phone: +1 408 974 1779

   Dave Crocker
   Silicon Graphics, Inc.
   2011 N. Shoreline Blvd.
   Mountain View, CA

   Phone: +1 415 390 1804

   Thomas Kessler
   Sun Microsystems Inc.
   Mail Stop MTV05-44
   2550 Garcia Ave.
   Mountain View, CA 94043

   Phone: +1 415 336 3145

Lear, Fair, Crocker & Kessler                                   [Page 8]