Cisco Systems

Santa Barbara 93117 California USA +1-408-526-4257 +1-413-473-2403 fred@cisco.com

Transport behave This note describes a framework for IPv4/IPv6 translation. This is in the context of replacing NAT-PT, which was deprecated by RFC 4966, and to enable networks to have IPv4 and IPv6 coexist in a somewhat rational manner while transitioning to an IPv6 network. The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

This note describes a framework for IPv4/IPv6 translation. This is in the context of replacing NAT-PT, which was deprecated by , and to enable networks to have IPv4 and IPv6 coexist in a somewhat rational manner while transitioning to an IPv6-only network. Deprecation of NAT-PT wasn't intended to say that NAT-PT was "bad", nor did the IETF think that deprecating the technology would stop people from using it. As with the 1993 deprecation of the RIP routing protocol at the time the Internet was converting to CIDR, the point was to inform the community that NAT-PT had operational issues and was not considered a viable medium or long term strategy for either coexistence or transition. The point was to encourage network operators to actually move in the direction of transition. describes the IETF's view of the most sensible transition model. The IETF recommends, in short, that network operators (transit providers, service providers, enterprise networks, small and medium business, SOHO and residential customers, and any other kind of network that may currently be using IPv4) obtain an IPv6 prefix, turn on IPv6 routing within their networks and between themselves and any peer, upstream, or downstream neighbors, enable it on their computers, and use it in normal processing. This should be done while leaving IPv4 stable, until a point is reached that any communication that can be carried out could use either protocol equally well. At that point, the economic justification for running both becomes debatable, and network operators can justifiably turn IPv4 off. This process is comparable to that of , which describes how to renumber a network using the same address family without a flag day. While running stably with the older system, deploy the new. Use the coexistence period to work out such kinks as arise. When the new is also running stably, shift production to it. When network and economic conditions warrant, remove the old, which is now no longer necessary. The question arises: what if that is infeasible due to the time available to deploy or other considerations? What if the process of moving a network and its components or customers is starting too late for contract cycles to effect IPv6 turn-up on important parts at a point where it becomes uneconomical to deploy global IPv4 addresses in new services? How does one continue to deploy new services without balkanizing the network? This set of documents describes translation as one of the tools networks might use to facilitate coexistence and ultimate transition.

Besides dual stack deployment, there are two fundamental approaches one could take to interworking between IPv4 and IPv6: tunneling and translation. One could - and in the 6NET we did - build an overlay network using the new protocol inside tunnels. Various proposals take that model, including 6to4, Teredo, ISATAP,and DS-Lite. The advantage of doing so is that the new is enabled to work without disturbing the old protocol, providing connectivity between users of the new protocol. There are two disadvantages to tunneling: operators of those networks are unable to offer services to users of the new architecture, and those users are unable to use the services of the underlying infrastructure - it is just bandwidth, and it doesn't enable new protocol users to communicate with old protocol users. As noted, in this work, we look at Internet Protocol translation as a transition strategy. forcefully makes the point that many of the reasons people use Network Address Translators are met as well by routing or protocol mechanisms that preserve the end to end addressability of the Internet. What it did not consider is the case in which there is an ongoing requirement to communicate with IPv4 systems, but configuring IPv4 routing is not in the network operator's view the most desirable strategy, or is infeasible due to a shortage of global address space. Translation enables the client of a network, whether a transit network, an access network, or an edge network, to access the services of the network and communicate with other network users regardless of their protocol usage - within limits. Like NAT-PT, IPv4/IPv6 translation under this rubric is not a long term support strategy, but it is a medium term coexistence strategy that can be used to facilitate a long term program of transition.

The following terminology is used in this document and other documents related to it. The IPv4 prefix, if any, subdivided into Mapped IPv4 Prefixes in the IPv6-only domain. This is advertised in routing in the IPv4 domain to attract traffic intended for mapped IPv4 addresses in the IPv6-only domain. A Dual Stack implementation, in this context, comprises an enabled end system stack plus routing in the network. It implies that two application instances are capable of communicating using either IPv4 or IPv6 - they have stacks, they have addresses, and they have any necessary network support including routing. An IPv4-only implementation, in this context, comprises an enabled end system stack plus routing in the network. It implies that two application instances are capable of communicating using either IPv4 but not IPv6 - they have an IPv4 stack, addresses, and network support including IPv4 routing and potentially IPv4/IPv4 translation, but some element is missing that prevents communication using IPv6. An IPv6-only implementation, in this context, comprises an enabled end system stack plus routing in the network. It implies that two application instances are capable of communicating using either IPv6 but not IPv4 - they have an IPv6 stack, addresses, and network support including routing in IPv6, but some element is missing that prevents communication using IPv4. The IPv6 prefix assigned by the network operator for direct mapping of IPv6 addresses to IPv4. See Local Internet Registry. A Local Internet Registry (LIR) is an organization which has received an IP address allocation from a Regional Internet Registry (RIR), and which may assign parts of this allocation to its own internal network or those of its customers. An LIR is thus typically an Internet service provider or an enterprise network. An IPv6 address within a Mapped IPv4 Prefix. An IPv6 prefix constructed from an LIR prefix and an IPv4 prefix. Zero or more IPv4 addresses used in stateful translation. "State" refers to dynamic information that is stored in a network element. For example, if two systems are connected by a TCP connection, each stores information about the connection, which is called "connection state". In this context, the term refers to correlations between IP addresses on either side of a translator, or {IP Address, Transport type, transport port number} tuples on either side of the translator. Of stateful algorithms, there are at least two major flavors depending on the kind of state they maintain: the existence of this state is unknown outside the network element that contains it. the existence of this state is known by other network elements. A translation algorithm may be said to "require state in a network element" or be "stateful" if the transmission or reception of a packet creates or modifies a data structure in the relevant network element. A translation algorithm that is not "stateful" is "stateless". It may require configuration of a translation table, or may derive its needed information algorithmically from the messages it is translating.

In any translation model, there is a question of objectives. Ideally, one would like to make any system and any application running on it able to "talk with" - exchange datagrams supporting applications - with any other system running the same application regardless of whether they have an IPv4 stack and connectivity or IPv6 stack and connectivity. That was the model NAT-PT, and the things it necessitated led to scaling and operational difficulties. So the question comes back to what different kinds of connectivity can be easily supported and what are harder, and what technologies are needed to at least pick the low-hanging fruit. We observe that applications today fall into three main categories: Per whatis.com, "'Client/server' describes the relationship between two computer programs in which one program, the client, makes a service request from another program, the server, which fulfills the request." In networking, the behavior of the applications is that connections are initiated from client software and systems to server software and systems. Examples include mail handling between an end user and his mail system (POP3, IMAP, and MUA->MTA SMTP), FTP, the web, and DNS name translation. Peer to peer applications are those that transfer information directly, rather than through the use of an intermediate repository such as a bulletin board or database. In networking, any system (peer) might initiate a session with any other system (peer) at any time. These in turn fall broadly into two categories: Examples of "infrastructure applications" include SMTP between MTAs, Network News, and SIP. Any MTA might open an SMTP session with any other at any time; any SIP Proxy might similarly connect with any other SIP Proxy. An important characteristic of these applications is that they use ephemeral sessions - they open sessions when they are needed and close them when they are done. Examples of these include Limewire, BitTorrent, and UTorrent. These are applications that open some sessions between systems and leave them open for long periods of time, and where ephemeral sessions are important, are able to learn about the reliability of peers from history or by reputation. They use the long term sessions to map content availability. Short term sessions are used to exchange content. They tend to prefer to ask for content from servers that they find reliable and available. NAT-PT is an example of a facility with known state - at least two software components (the data plane translator and the DNS Application Layer Gateway, which may be implemented in the same or different systems) share and must coordinate translation state. A typical IPv4/IPv4 NAT implements an algorithm with hidden state. Obviously, stateless translation requires less computational overhead than stateful translation, and less memory to maintain the state, because the translation tables and their associated methods and processes exist in a stateful algorithm and don't exist in a stateless one. If the key questions are the ability to open connections between systems, then one must ask who opens connections. We need a technology that will enable systems that act as clients to be able to open sessions with other systems that act as servers, whether in the IPv6->IPv4 direction or the IPv4->IPv6 direction. Ideally, this is stateless; especially in a carrier infrastructure, the preponderance of accesses will be to servers, and this optimizes access to them. However, a stateful algorithm is acceptable if the complexity is minimized and a stateless algorithm cannot be constructed. We also need a technology that will allow peers to connect with each other, whether in the IPv6->IPv4 direction or the IPv4->IPv6 direction. Again, it would be ideal if this was stateless, but a stateful algorithm is acceptable if the complexity is minimized and a stateless algorithm cannot be constructed. In the case of infrastructure applications, which know nothing of choosing among peers by reputation, the IPv4->IPv6 direction is a stronger requirement. Peer to peer file exchange applications, however, may be more forgiving - it may well be adequate to make a subset of IPv4->IPv6 connections work instead of all. (EDITOR'S NOTE: I would be very interested in comments on this assertion) We do not need an algorithm that enables clients to connect to clients, because they don't connect. The complexity arguments bring us in the direction of hidden state: if state must be shared between the application and the translator or between translation components, complexity and deployment issues are greatly magnified. We would very much prefer that any software changes be confined to the translator.

While IPv6 was "by design" incompatible with IPv4, the designers intended that it would coexist with IPv4 during a period of transition. The primary mode of coexistence was dual-stack operation - routers would be dual-stacked so that the network could carry both address families, and IPv6-capable hosts could be dual-stack to maintain access to IPv4-only partners. The goal was that the preponderance of hosts and routers in the Internet would be IPv6-capable long before IPv4 address space allocation was completed. At this time, it appears the exhaustion of IPv4 address space will occur before significant IPv6 adoption. Curran's "A Transition Plan for IPv6" proposes a three-phase progression: characterized by pilot use of IPv6, primarily through transition mechanisms defined in RFC 4213, and planning activities. characterized by general availability of IPv6 in provider networks which SHOULD be native IPv6; organizations SHOULD provide IPv6 connectivity for their Internet-facing servers, but SHOULD still provide IPv4-based services via a separate service name. characterized by a preponderance of IPv6-based services and diminishing support for IPv4-based services. In each of these phases, the coexistence problem and solution space has a different focus: Coexistence tools are needed to facilitate early adopters by removing impediments to IPv6 deployment, and to assure that nothing is lost by adopting IPv6, in particular that the IPv6 adopter has unfettered access to the global IPv4 Internet regardless of whether they have a global IPv4 address (or any IPv4 address or stack at all.) While it might appear reasonable for the cost and operational burden to be borne by the early adopter, the shared goal of promoting IPv6 adoption would argue against that model. Additionally, current IPv4 users should not be forced to retire or upgrade their equipment and the burden remains on service providers to carry and route native IPv4. While IPv6 adoption can be expected to accelerate, there will still be a significant portion of the Internet operating in IPv4-only or preferring IPv4. During this phase the norm shifts from IPv4 to IPv6, and coexistence tools evolve to ensure interoperability between domains that may be restricted to IPv4 or IPv6. In this phase, IPv6 is ubiquitous and the burden of maintaining interoperability shifts to those who choose to maintain IPv4-only systems. While these systems should be allowed to live out their economic life cycles, the IPv4-only legacy users at the edges should bear the cost of coexistence tools, and at some point service provider networks should not be expected to carry and route native IPv4 traffic. The choice between the terms "transition" versus "coexistence" has engendered long philosophical debate. "Transition" carries the sense that we are going somewhere, while "coexistence" seems more like we are sitting somewhere. Historically with IETF, "transition" has been the term of choice , and the tools for interoperability have been called "transition mechanisms". There is some perception or conventional wisdom that adoption of IPv6 is being impeded by the deficiency of tools to facilitate interoperability of nodes or networks that are constrained (in some way, fully or partially) from full operation in one of the address families. In addition, it is apparent that transition will involve a period of coexistence; the only real question is how long that will last. Thus, coexistence is an integral part of the transition plan, not in conflict with it, but there will be a balancing act. It starts out being a way for early adopters to easily exploit the bigger IPv4 Internet, and ends up being a way for late/never adopters to hang on with IPv4 (at their own expense, with minimal impact or visibility to the Internet). One way to look at solutions is that cost incentives (both monetary cost and the operational overhead for the end user) should encourage IPv6 and discourage IPv4. That way natural market forces will keep the transition moving - especially as the legacy IPv4-only stuff ages out of use. There will come a time to set a date after which no one is obligated to carry native IPv4 but it would be premature to attempt to do so yet. The end goal should not be to eliminate IPv4 by fiat, but rather render it redundant through ubiquitous IPv6 deployment. IPv4 may never go away completely, but rational plans should move the costs of maintaining IPv4 to those who insist on using it after wide adoption of IPv6.

There are several potential uses of translation. They are all easily described in terms of "interoperation between a set of systems that only communicate using IPv4 and a set of systems that only communicate using IPv6", but the differences at a detail level make them interesting. At minimum, these include: Connection of IPv4-only islands to an IPv6-only network Connection of IPv6-only islands to an IPv4-only network Connecting IPv4-only devices with IPv6-only devices regardless of network type Connections between IPv4-only networks and IPv6-only networks, especially as a service within a large network such as an enterprise or ISP network or between peer networks.

While the basic issue is the same, there are at least two interesting special cases of this: connecting a small pool of legacy equipment with a view to eventual obsolescence, and connecting a legacy network with a view to eventual transition.

In the first case, , one might have a pool of equipment (printers, perhaps) that is IPv4-capable, but either the network it serves or some equipment in that network is IPv6-only. One pools the IPv4-only devices behind a translator, which enables IPv6-only systems to connect to the IPv4-only equipment. If the network is dual stack and only some of the equipment is IPv6-only, the translator should be a function of a router, and the router should provide normal IPv4 routing services as well as IPv6->IPv4 translation.

creates transition options to a customer network connected to an IPv6-only ISP, or some equivalent relationship. The customer might internally be using traditional IPv4 with NAT services, and the ISP might change its connection to an IPv6-only network and encourage it to transition. If the ISP assigns a /60 prefix to a SOHO, for example, the CPE router in the SOHO could distribute several dual stack subnets internally, one for wireless and one for each of several fixed LANs (the entertainment system, his office, her office, etc). One of the /64 prefixes would be dedicated to representing the SOHO's IPv4 addresses in the ISP or the IPv4 network beyond it, and the other prefixes for the various internal subnets. Internally, the subnets might carry prefix pairs 192.168.n.0/24 and 2001:db8:0.n::/64 for n in 1..15 (1..0xF), and externally might appear as 2001:db8:0:n::/64 for the IPv6 subnets and 2001:db8::192.168.n.0/120 for the IPv4 devices. Note that to connect to an IPv4-only network beyond, RFC 1918 addresses would have to be statefully mapped using traditional IPv4 mechanisms somewhere; if this is done by the ISP, collusion on address mapping is required, and the case in is probably a better choice. In this environment, the key issue is that one wants a prefix that enables the entire address space to be embedded in a single /64 prefix, with the assumption that any routing structure behind the translator is managed by IPv4 routing.

To be completed

To be completed

In this case (see ) we presume that a service provider or equivalent is offering a service in a network in which IPv4 routing is not supported, but customers are allocated relatively large pools of general IPv6 addresses, suitable for clients of IPv4 or IPv6 hosts, and relatively small pools of addresses mapped to global IPv4 addresses that are intended to be accessible to IPv4 peers and clients through translation. Presumably, there are a number of such customers, and the administration wishes to use normal routing to manage the issues. As a carrier offering, there is also a need for stateless translation.

Since specifies that IPv6 prefixes are 64 bits or shorter apart from host routes, one wishes to allocate each customer a /64 mapped to a few IPv4 addresses and a shorter prefix for his general use. The customer's CPE advertises the two prefixes into the IPv6 routing domain to attract relevant traffic. The translator advertises the mapped equivalent of an IPv4 default route into the IPv6 domain to attract all other traffic to it, for translation into the IPv4 routing domain. It also advertises an appropriate IPv4 prefix aggregating the mapped prefixes into the IPv4 domain to attract traffic intended for these customers. In this case, the LIR prefix MUST be within /32../63; a /64 puts the entire IPv4 address space into the host part, which is equivalent to the case in , and a prefix shorter than /32 wastes space with no redeeming argument. In general, the LIR prefix should be 64 bits less the length of IPv4 prefixes it allocates to its IPv4-mapped customers. For example, if it is allocating a mapped IPv4 /24 to each customer, the LIR prefix used for mapping between IPv4 and IPv6 addresses should be a /40, and the least significant bits in the IPv4 address form the host part of the address.

Having laid out the preferred transition model and the options for implementing it (), defined terms(), considered the requirements (), considered the transition model (), and considered the kinds of networks the facility would support (), we now turn to a framework for IPv4/IPv6 translation. This framework has three main parts: The recommended address format The functional components of a translation solution, which include A DNS Application Layer Gateway, An optional stateless translator, and An optional stateful translator. The operational characteristics of the solution.

|<--- host part --->| ]]>

As shown in , the mapped address format has three components: An LIR-specified prefix, either 32..63 bits long or 96 bits long, An embedded IPv4 address. Except in the case of a 96 bit prefix, this address intentionally straddles the boundary between 's 64 bit "subnet" locator and its 64 bit host identifier. The intention is that the /64 be used in routing and the bits in the host part be used for host identification as described in the address architecture. Entirely zero; note that if n=96, this is null. The length of the LIR-specified prefix is itself specified by the LIR to achieve its objectives. There are some obvious values that might be popular, including /40, /44, and /96, but there is no requirement than any of them be used; this is left to the operator's discretion.

As noted in ,translation involves several components. An IPv4 client or peer must be able to determine the address of its server by obtaining an A record from DNS even if the server is IPv6-only - only has an IPv6 stack, or is in an IPv6-only network. Similarly, an IPv6 client or peer must be able to determine the address of its server by obtaining an AAAA record from DNS even if the server is IPv4-only - only has an IPv4 stack, or is in an IPv4-only network. Given the address, the client/peer must be able to initiate a connection to the server/peer, and the server/peer must be able to reply. It would be very nice if this scaled to the size of regional networks with straightforward operational practice. To that end, we describe four subsystems: A Domain Name System Translator A stateless IPv4/IPv6 translator A stateful IPv4/IPv6 translator Application Layer Gateways for some applications

describes the mechanisms by which a DNS Translator is intended to operate. It is designed to operate on the basis of known but fixed state: the resource records, and therefore the names and addresses, that it translates are known to the network outside of the translator, but the process of serving them to applications does not interact with the translator in any way. There are at least three possible implementations of a DNS Translator: One could literally program DNS with corresponding A and AAAA records. This is most appropriate for stub services such as access to a legacy printer pool. In more general operation, the expected behavior is for the application to request both A and AAAA records, and for an A record to be (retrieved and) translated by the DNS translator if and only if no reachable AAAA record exists. This has ephemeral issues with cached translations, which can be dealt with by caching only the source record and forcing it to be translated whenever accessed. In Dynamic DNS usage, a system could potentially report the translation of a name using a Mapped IPv4 Address, or using both a Mapped IPv4 Address and some other address. The DNS translator has several options; it could store a AAAA record for the Mapped IPv4 Address and depend on translation of that for A records inline, it could store both an A and a AAAA record, or (when there is another IPv6 address as well which is stored as the AAAA record) it could store only the A record.

describes and defines the behavior of a stateless translator. This is an optional facility; one could implement or deploy only the stateful mode described in . Stateless translation enables IPv4-only clients and peers to initiate connections to IPv6-only servers or peers equipped with Mapped IPv4 Addresses, as described in . It also enables scalable coordination of IPv4-only stubs of larger enterprise or ISP IPv6-only networks as described in . In addition, since address selection would select a Mapped IPv4 Address when it is available, stateless translation enables IPv6 clients and peers with Mapped IPv4 Addresses to open connections with IPv4 servers and peers in a scalable fashion, supporting aysnchronous routes.

also describes and defines the behavior of the data plane component of a stateful translator. describes the management of the state tables necessitated by stateful translation. Like stateful translation, this is an optional facility; one could implement or deploy only the stateful mode described in . Stateful translation is defined to enable IPv6 clients and peers without Mapped IPv4 Addresses to connect to IPv4-only servers and peers. Stateful translation could be defined to enable IPv4 clients and peers to connect to IPv6-only servers and peers without Mapped IPv4 Addresses. This is far more complex, however, and is out of scope in the present work.

In addition, some applications require special support. An example is FTP. FTP's active mode doesn't work well across NATs without extra support such as SOCKS. Across NATs, it generally uses passive mode. However, the designers of FTP inexplicably wrote different and incompatible passive mode implementations for IPv4 and IPv6 networks. Hence, either they need to fix FTP, or a translator must be written for the application. Other applications may be similarly broken.

Just say "multicast"; this framework could support multicast, but at this point does not. This is a place for future work. As noted, IPv4 client/peer access to IPv6 servers and peers lacking Mapped IPv4 Addresses is not solved. Interoperation between IPv4-only clients and IPv6-only clients is not supported, and is not believed to be needed.

This memo requires no parameter assignment by the IANA. Note to RFC Editor: This section will have served its purpose if it correctly tells IANA that no new assignments or registries are required, or if those assignments or registries are created during the RFC publication process. From the author's perspective, it may therefore be removed upon publication as an RFC at the RFC Editor's discretion.

One "security" issue has been raised, with an address format that was considered and rejected for that reason. At this point, the editor knows of no other security issues raised by the address format that are not already applicable to the addressing architecture in general.

This is under development by a large group of people. Those who have posted to the list during the discussion include Andrew Sullivan, Andrew Yourtchenko, Brian Carpenter, Dan Wing, Ed Jankiewicz, Fred Baker, Hiroshi Miyata, Iljitsch van Beijnum, John Schnizlein, Kevin Yin, Magnus Westerlund, Marcelo Bagnulo Braun, Margaret Wasserman, Masahito Endo, Phil Roberts, Philip Matthews, Remi Denis-Courmont, Remi Despres, and Xing Li. The appendix is largely derived from Hiroshi Miyata's analysis, which is in turn based on documents by many of those just named. Ed Jankiewicz described the transition plan. The definition of a "Local Internet Registry" came from the Wikipedia, and was slightly expanded to cover the present case. (EDITOR'S QUESTION: Would it be better to describe this as an "operator-defined prefix"?)

University of Madrid CERNET Center/Tsinghua University

Room 225, Main Building, Tsinghua University Beijing 100084 China +86 62785983 xing@cernet.edu.cn

CERNET Center/Tsinghua University

Room 225, Main Building, Tsinghua University Beijing 100084 China +86 62785983 congxiao@cernet.edu.cn

Cisco Systems

This appendix summarizes and analyzes the several proposals that have been made for a mapped IPv4 address. These prefixes fall into two broad categories: those that embed the IPv4 address into a well-known prefix, and those that embed it into a prefix defined by the network operator. and define different well-known prefixes, and and define different forms of operator-defined prefixes.

and define two slightly different formats of address that map between IPv4 and IPv6. In both cases, there is a defined 96 bit prefix, and the IPv4 address is inserted into bits 96..127 of the IPv6 address. 's address formats are as follows: An address of the form 0::ffff:a.b.c.d which refers to a node that is not IPv6-capable. In addition to its use in the API this protocol uses IPv4-mapped addresses in IPv6 packets to refer to an IPv4 node. An address of the form 0::0:a.b.c.d which refers to an IPv6/IPv4 node that supports automatic tunneling. Such addresses are not used in this protocol. An address of the form 0::ffff:0:a.b.c.d which refers to an IPv6-enabled node. Note that the prefix 0::ffff:0:0:0/96 is chosen to checksum to zero to avoid any changes to the transport protocol's pseudo header checksum.

The Well-Known Prefix allows automatic IPv6 address mapping to IPv4. One Well-Known Prefix can represent entire IPv4 network address. It is straightforward to ensure that an application prefers native addressing to mapped addressing in selecting an address for its peer or server, as the tables can come configured that way from the manufacturer. It will choose its source address by rules, which prefer the most similar prefix first. Hence, a system with a mapped address communicating through a translator will prefer its own mapped address as a source. If the application wants to know whether the address has been synthesized, this is straightforward.

In interdomain routing, there can be problems similar to those considered in . For example, consider two routing administrations that interconnect using IPv6 and each offer independent IPv4 domains. If an IPv4 client of one administration accesses an IPv6 server in the other network, the replies will be routed to the other network's domain. Using a standard prefix for all IPv4 space means that all IPv4 access is through that system or through the topologically nearest instance of them. If the IPv4 address space is fragmented, and especially if it is duplicated as is done with space, it is impossible to distinguish the access points in the IPv6 network. Even in intradomain routing, control issues can arise in routing if there is more than one translator.

To use DNS re-writing function, the IPv6 node should be configured to send DNS query to appropriate DNS server somehow. But it is same as ordinary DNS configuration. Therefore, no special configuration is required for both IPv6 and IPv4 hosts. No special configuration is required of routers. Each gateway needs to know the Well-Known Prefix, whether that means configuration of the prefix or simply configuration of the translation function. The Gateway must also be configured to advertise the Well-Known Prefix in the IPv6 network and the relevant prefix(es) in the IPv4 network. This must be performed once for each gateway. If the addresses are mapped in statically, each mapping must be configured in the appropriate gateway. This configuration must be performed [number of mapped prefixes] * [number of sharing gateway] times. The DNS re-writing function must be configured with the Well-Known Prefix to synthesize AAAA records from A records for IPv6 clients, but it may be configured by default. This configuration must be performed [number of Local Prefix] times.

Small scale translation (Home Network). Less redundant translation service (no load balancing). Stub IPv6 network. Sample configurations include: To provide the access from IPv6 client in stub IPv6 network to global IPv4 server, place the gateway at the edge of IPv6 stub site.

---[Gateway]----->----+------------[IPv4 Server] | +------------[IPv4 Server]]] ]]>

To provide the access from IPv6 client in stub IPv6 network to private/global IPv4 server (IPv6 stub network attached to a private IPv4 network), place the gateway at the edge of IPv6 stub network.

---[Gateway]----->----+------------[IPv4 Server] | [NAT] | +------------[IPv4 Server] ]]>

Two forms of network operator specified addresses have been proposed, one of them in several minor variations. In short, both have the network operator specify a prefix into which an IPv4 address is embedded, either in bits 96..127 or following a shorter prefix. Since one is a special case of the other (the LIR prefix is 96 bits as opposed to being variable), it would be tempting to comment on the two together. The operational similarities will be great, and the differences will revolve around the economics of the prefix that the IPv4 address is embedded into But to make them clear, we will review them separately.

The IVI Address , shown in , has a variable length prefix specified by the operator followed by the IPv4 address, and the remainder filled with zero. Observing 's requirement that an operator-specified prefix should have 64 bits of subnet locator and 64 bits of host interface identifier, IVI suggests that the operator divide the mapped IPv4 prefix into a subnet part and a host part, and assign a prefix from its allocation that with the subnet part fills 64 bits.

|<--- host part --->| ]]>

The impact of the variability of the LIR Prefix has to do with service offerings. If a network operator wishes to offer customers a general IPv6 prefix such as a /48 plus a smaller IPv4-mapped set of addresses for IPv4-accessible servers, such as an IPv4 /24, he might literally design a service in which each customer gets a general /48 prefix and an IPv4-mapped /64 prefix. To accomplish this, the operator would allocate a /40 for the LIR prefix and embed the IPv4 address space into it. His Advertised IPv4 Prefix would aggregate the Mapped IPv4 Prefixes that he in turn assigns to his customers. If, however, he wanted to assign smaller units, such as /28s to each customer, he would allocate a shorter prefix such as a /36 as the LIR Prefix. There are clear trade-offs here; the point is to enable the network operator to optimize them for the service he wants to offer.

A Mapped IPv4 Address format allows a stateless IPv6 address mapping between an IPv4 address and its mapped IPv6 counterpart. One such prefix can represent the entire IPv4 address space, and if desired multiple prefixes can represent multiple instances of it or accesses to it. selection rules select the source address most similar to the destination address in question, which is to say matching the longest prefix. In general, one would expect a system with an address of this type to prefer IPv6 source addresses derived from IPv4 addresses when they are available. If the tables in a host are configured with the administration's translation prefix, a policy can be made to prefer native IPv4 to translation, or to prefer any other IPv6 address to a translated address. The administration has the option of using the same prefix on multiple gateways, or of using different prefixes. Differing administrations will almost assuredly use different prefixes. This enables the administration to distinguish between distinct address spaces such as separate instances of the address space. It also enables multiple gateways to be used to interconnect between public IPv4 and IPv6 networks without having to manage the state maintained by such translation gateways. Due to the ability to support multiple gateways between the same two domains statelessly and the ability to identify multiple instances of the same IPv4 address space when appropriate, a network operator specified prefix is scalable through normal routing structures. Since the prefix choice is under the network's control, routing is managed relatively easily. These prefixes were chosen to make it unnecessary to adjust TCP/UDP checksums.

By default, hosts are unlikely to come configured with the administration's translation prefix in their tables, and so are unlikely to be able to distinguish such addresses from other IPv6 addresses. Multiple small (and perhaps overlapping) address spaces are readily supported in what might be called a network model. However, these consume much larger blocks of IPv6 address space than the appliance model of a Local Prefix () does. As noted above, IVI is less efficient than the NAT64 model in enumerating small IPv4 islands, and having a prefix per network operator is less efficient on a global basis than having a single well-known prefix. If one uses both stateless and stateful translation in the same network, assigning a normal IPv6 prefix to all systems and additionally mapped addresses to servers, then one needs two routes, one for each prefix. Reducing this burden requires either the total use of stateful translation, disabling IPv4 clients access to IPv6 servers, or total use of stateless translation, meaning that one effectively assigns an IPv4 address to every host. One would generally expect an IVI address to be used in an ISP service, as it requires a 40 bit prefix assigned by the operator in most cases. It could be used with a ULA in an edge network at the cost of losing global routability. Using an operator-specified prefix requires the translator to adjust TCP and UDP checksums.

In general, one would expect a mapped IPv4 address to be assigned in the same way that IPv4 addresses are assigned; this would call for the use of DHCPv6 or manual configuration. If one or more hosts on a LAN are assigned mapped IPv4 addresses, one or more routers on the LAN needs configuration of the corresponding Mapped IPv4 Prefix, and to have that advertised as a route in the IPv6 domain. The gateway needs to advertise three prefixes: The Advertised IPv4 Prefix is advertised into the IPv4 domain to attract traffic that needs translation to IPv6. The Overlay IPv4 Prefix, if stateful translation is in use, is advertised into the IPv4 domain to attract traffic using that translation facility. The LIR Prefix is advertised into the IPv6 domain to attract traffic that needs translation to IPv4. The DNS re-writing function must be configured with the LIR Prefix to synthesize the AAAA records for IPv6 nodes when appropriate. The DNS server needs to be configured with the information to develop A records when appropriate. This may be accomplished using Dynamic DNS or manual configuration. This may mean configuration of IPv4 A records that get translated to AAAA records, or configuration of IPv6 AAAA records that are recognized by the DNS server.

The IVI address is appropriate to large scale, ISP grade, translation, while the NAT64 address is more flexible. Highly redundant translation service. Places where IPv4 clients need to access IPv6 servers. Places where IPv6 clients and peers need to access IPv4 servers and peers.

NAT64 and SNATPT each specify an address that, like the well-known addresses of and the "Dummy Address" of and , has 96 bits of operator-specified prefix and the IPv4 address in bits 96..127. This is shown in . In some proposals, the "IDENT" field is always zero, and in others it enumerates different instances of the IPv4 address space.

|<-identifier-->| ]]>

A similar address format, with an "IDENT" based on the IANA OUI, is used by ISATAP; if a globally-unique "IDENT" field is selected, it needs to differ from that value.

A Mapped IPv4 Address format allows a stateless IPv6 address mapping between an IPv4 address and its mapped IPv6 counterpart. One such prefix can represent the entire IPv4 address space, and if desired multiple prefixes can represent multiple instances of it or accesses to it. selection rules select the source address most similar to the destination address in question, which is to say matching the longest prefix. In general, one would expect a system with an address of this type to prefer IPv6 source addresses derived from IPv4 addresses when they are available. If the tables in a host are configured with the administration's translation prefix, a policy can be made to prefer native IPv4 to translation, or to prefer any other IPv6 address to a translated address. Multiple small (and perhaps overlapping) address spaces are readily supported in what might be called an appliance model; for example, if SOHOs are using IPv4 internally, the IPv6 ISP can give a /64 to each and manage them easily. The administration has the option of using the same prefix on multiple gateways, or of using different prefixes. This approach enables multiple gateways to be used to interconnect between IPv4 and IPv6 networks without having to manage the state maintained by such translation gateways. Due to the ability to support multiple gateways between the same two domains statelessly and the ability to identify multiple instances of the same IPv4 address space when appropriate, a network operator specified prefix is scalable through normal routing structures. Since the prefix choice is under the network's control, routing is managed relatively easily.

By default, hosts are unlikely to come configured with the administration's translation prefix in their tables, and so are unlikely to be able to distinguish such addresses from other IPv6 addresses. If one uses both stateless and stateful translation in the same network, assigning a normal IPv6 prefix to all systems and additionally mapped addresses to servers, then one needs two routes, one for each prefix. Reducing this burden requires either the total use of stateful translation, disabling IPv4 clients access to IPv6 servers, or total use of stateless translation, meaning that one effectively assigns an IPv4 address to every host. One would generally expect an IVI address to be used in an ISP service, as it requires a 40 bit prefix assigned by the operator in most cases. It could be used with a ULA in an edge network at the cost of losing global routability. The NAT64 address, on the other hand, has no such issue. It is difficult to identify a mapped IPv4 address without knowledge that the mapping algorithm is used with a specific prefix. Since IPv4 addresses are allocated by DHCP servers or manually, it is inappropriate to mix Local Prefix IPv4-mapped addresses with Address Autoconfiguration in the same prefix. This may not be obvious to a provider that thinks of itself as simply assigning a /64 IPv6 prefix to the SOHO (regarding which see ). Routing is readily handled in the IPv4 network. However, if routing of IPv4-mapped prefixes is desired in the IPv6 network, we are forced to use prefixes in the neighborhood of /96../128. Apart from routing host addresses, frowns on this, preferring routing prefixes to be 64 bits or shorter and leaving a 64 bit host ID. Using an operator-specified prefix requires the translator to adjust TCP and UDP checksums.

In general, one would expect a mapped IPv4 address to be assigned in the same way that IPv4 addresses are assigned; this would call for the use of DHCPv6 or manual configuration. If one or more hosts on a LAN are assigned mapped IPv4 addresses, one or more routers on the LAN needs configuration of the corresponding Mapped IPv4 Prefix, and to have that advertised as a route in the IPv6 domain. The gateway needs to advertise two prefixes: The Advertised IPv4 Prefix is advertised into the IPv4 domain to attract traffic that needs translation to IPv6. The LIR Prefix is advertised into the IPv6 domain to attract traffic that needs translation to IPv4. The DNS re-writing function must be configured with the LIR Prefix to synthesize the AAAA records for IPv6 nodes when appropriate. The DNS server needs to be configured with the information to develop A records when appropriate. This may be accomplished using Dynamic DNS or manual configuration. This may mean configuration of IPv4 A records that get translated to AAAA records, or configuration of IPv6 AAAA records that are recognized by the DNS server.

Local Prefixes are appropriate to small networks that have little internal IPv6 structure, such as server pools or SOHO clients. The structure in the IPv4 network is as with any IPv4 network, making this appropriate to medium sized IPv4 domains. Highly redundant translation service. Places where IPv4 clients need to access IPv6 servers. Places where IPv6 clients and peers need to access IPv4 servers and peers. Sample configurations include: Place the gateway at the edge of IPv6 stub site.

---[Gateway]----->----+------------[IPv4 Server] | +------------[IPv4 Server] ]]>

Place the gateway in front of IPv4 server.

------[Gateway]---->---[IPv4 Server] | [IPv6 Client]---------+ ]]>

to provide the access from IPv6 client to private IPv4 server, place the gateway in front of IPv4 private network.

------[Gateway]---->---[IPv4 Server] | [IPv6 Client]---------+ ]]>

To provide the access from IPv4 client to IPv6 server by static 1:1 mapping, place the gateway at the edge of IPv4 stub site.