RFC6179

Internet Research Task Force (IRTF) F. Templin, Ed. Request for Comments: 6179 Boeing Research & Technology Category: Experimental March 2011 ISSN: 2070-1721

          The Internet Routing Overlay Network (IRON)

Abstract

Since the Internet must continue to support escalating growth due to increasing demand, it is clear that current routing architectures and operational practices must be updated. This document proposes an Internet Routing Overlay Network (IRON) that supports sustainable growth while requiring no changes to end systems and no changes to the existing routing system. IRON further addresses other important issues including routing scaling, mobility management, multihoming, traffic engineering and NAT traversal. While business considerations are an important determining factor for widespread adoption, they are out of scope for this document. This document is a product of the IRTF Routing Research Group.

Status of This Memo

This document is not an Internet Standards Track specification; it is published for examination, experimental implementation, and evaluation.

This document defines an Experimental Protocol for the Internet community. This document is a product of the Internet Research Task Force (IRTF). The IRTF publishes the results of Internet-related research and development activities. These results might not be suitable for deployment. This RFC represents the individual opinion(s) of one or more members of the Internet Research Task Force (IRTF) Research Group of the Internet Research Task Force (IRTF). Documents approved for publication by the IRSG are not a candidate for any level of Internet Standard; see Section 2 of RFC 5741.

Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at http://www.rfc-editor.org/info/rfc6179.

Copyright Notice

Copyright (c) 2011 IETF Trust and the persons identified as the document authors. All rights reserved.

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document.

  6.5. Mobility, Multihoming, and Traffic Engineering

Appendix A. IRON VPs over Internetworks with Different

Introduction

Growth in the number of entries instantiated in the Internet routing system has led to concerns regarding unsustainable routing scaling [RADIR]. Operational practices such as the increased use of multihoming with Provider-Independent (PI) addressing are resulting in more and more fine-grained prefixes being injected into the routing system from more and more end user networks. Furthermore, depletion of the public IPv4 address space has raised concerns for both increased address space fragmentation (leading to yet further routing table entries) and an impending address space run-out scenario. At the same time, the IPv6 routing system is beginning to see growth [BGPMON] which must be managed in order to avoid the same routing scaling issues the IPv4 Internet now faces. Since the Internet must continue to scale to accommodate increasing demand, it is clear that new routing methodologies and operational practices are needed.

Several related works have investigated routing scaling issues. Virtual Aggregation (VA) [GROW-VA] and Aggregation in Increasing Scopes (AIS) [EVOLUTION] are global routing proposals that introduce routing overlays with Virtual Prefixes (VPs) to reduce the number of entries required in each router's Forwarding Information Base (FIB) and Routing Information Base (RIB). Routing and Addressing in Networks with Global Enterprise Recursion (RANGER) [[[RFC5720]]] examines recursive arrangements of enterprise networks that can apply to a very broad set of use-case scenarios [[[RFC6139]]]. IRON specifically adopts the RANGER Non-Broadcast, Multiple Access (NBMA) tunnel virtual-interface model, and uses Virtual Enterprise Traversal (VET) [INTAREA-VET] and the Subnetwork Adaptation and Encapsulation Layer (SEAL) [INTAREA-SEAL] as its functional building blocks.

This document proposes an Internet Routing Overlay Network (IRON) with goals of supporting sustainable growth while requiring no changes to the existing routing system. IRON borrows concepts from VA and AIS, and further borrows concepts from the Internet Vastly Improved Plumbing (Ivip) [IVIP-ARCH] architecture proposal along with its associated Translating Tunnel Router (TTR) mobility extensions [TTRMOB]. Indeed, the TTR model to a great degree inspired the IRON mobility architecture design discussed in this document. The Network Address Translator (NAT) traversal techniques adapted for IRON were inspired by the Simple Address Mapping for Premises Legacy Equipment (SAMPLE) proposal [SAMPLE].

IRON supports scalable addressing without changing the current BGP [[[RFC4271]]] routing system. IRON observes the Internet Protocol standards [[[RFC0791]]][[[RFC2460]]]. Other network-layer protocols that can be encapsulated within IP packets (e.g., OSI/CLNP (Connectionless Network Protocol) [[[RFC1070]]], etc.) are also within scope.

The IRON is a global routing system comprising virtual overlay networks managed by Virtual Prefix Companies (VPCs) that own and manage Virtual Prefixes (VPs) from which End User Network (EUN) prefixes (EPs) are delegated to customer sites. The IRON is motivated by a growing customer demand for multihoming, mobility management, and traffic engineering while using stable addressing to minimize dependence on network renumbering [[[RFC4192]]][[[RFC5887]]]. The IRON uses the existing IPv4 and IPv6 global Internet routing systems as virtual NBMA links for tunneling inner network protocol packets within outer IPv4 or IPv6 headers (see Section 3). The IRON requires deployment of a small number of new BGP core routers and supporting servers, as well as IRON-aware routers/servers in customer EUNs. No modifications to hosts, and no modifications to most routers, are required.

Note: This document is offered in compliance with Internet Research Task Force (IRTF) document stream procedures [[[RFC5743]]]; it is not an IETF product and is not a standard. The views in this document were considered controversial by the IRTF Routing Research Group (RRG), but the RG reached a consensus that the document should still be published. The document will undergo a period of review within the RRG and through selected expert reviewers prior to publication. The following sections discuss details of the IRON architecture.

Terminology

This document makes use of the following terms:

End User Network (EUN):

  an edge network that connects an organization's devices (e.g.,
  computers, routers, printers, etc.) to the Internet.

End User Network Prefix (EP):

  a more specific inner network-layer prefix derived from a Virtual
  Prefix (VP) (e.g., an IPv4 /28, an IPv6 /56, etc.) and delegated
  to an EUN by a Virtual Prefix Company (VPC).

End User Network Prefix Address (EPA):

  a network-layer address belonging to an EP and assigned to the
  interface of an end system in an EUN.

Forwarding Information Base (FIB):

  a data structure containing network prefixes to next-hop mappings;
  usually maintained in a router's fast-path processing lookup
  tables.

Internet Routing Overlay Network (IRON):

  a composite virtual overlay network that comprises the union of
  all VPC overlay networks configured over a common Internetwork.
  The IRON supports routing through encapsulation of inner packets
  with EPA addresses within outer headers that use locator
  addresses.

IRON Client Router/Host ("Client"):

  a customer's router or host that logically connects the customer's
  EUNs and their associated EPs to the IRON via an NBMA tunnel
  virtual interface.

IRON Serving Router ("Server"):

  a VPC's overlay network router that provides forwarding and
  mapping services for the EPs owned by customer Clients.

IRON Relay Router ("Relay"):

  a VPC's overlay network router that acts as a relay between the
  IRON and the native Internet.

IRON Agent (IA):

  generically refers to any of an IRON Client/Server/Relay.

Internet Service Provider (ISP):

  a service provider that connects customer EUNs to the underlying
  Internetwork.  In other words, an ISP is responsible for providing
  basic Internet connectivity for customer EUNs.

Locator

  an IP address assigned to the interface of a router or end system
  within a public or private network.  Locators taken from public IP
  prefixes are routable on a global basis, while locators taken from
  private IP prefixes are made public via Network Address
  Translation (NAT).

Routing and Addressing in Networks with Global Enterprise Recursion

  (RANGER):
  an architectural examination of virtual overlay networks applied
  to enterprise network scenarios, with implications for a wider
  variety of use cases.

Subnetwork Encapsulation and Adaptation Layer (SEAL):

  an encapsulation sublayer that provides extended packet
  identification and a Control Message Protocol to ensure
  deterministic network-layer feedback.

Virtual Enterprise Traversal (VET):

  a method for discovering border routers and forming dynamic
  tunnel-neighbor relationships over enterprise networks (or sites)
  with varying properties.

Virtual Prefix (VP):

  a prefix block (e.g., an IPv4 /16, an IPv6 /20, an OSI Network
  Service Access Protocol (NSAP) prefix, etc.) that is owned and
  managed by a Virtual Prefix Company (VPC).

Virtual Prefix Company (VPC):

  a company that owns and manages a set of VPs from which it
  delegates EPs to EUNs.

VPC Overlay Network

  a specialized set of routers deployed by a VPC to service customer
  EUNs through a virtual overlay network configured over an
  underlying Internetwork (e.g., the global Internet).

The Internet Routing Overlay Network

The Internet Routing Overlay Network (IRON) is a system of virtual overlay networks configured over a common Internetwork. While the principles presented in this document are discussed within the context of the public global Internet, they can also be applied to any autonomous Internetwork. The rest of this document therefore refers to the terms "Internet" and "Internetwork" interchangeably except in cases where specific distinctions must be made.

The IRON consists of IRON Agents (IAs) that automatically tunnel the packets of end-to-end communication sessions within encapsulating headers used for Internet routing. IAs use the Virtual Enterprise Traversal (VET) [INTAREA-VET] virtual NBMA link model in conjunction with the Subnetwork Encapsulation and Adaptation Layer (SEAL) [INTAREA-SEAL] to encapsulate inner network-layer packets within outer headers, as shown in Figure 1.

                                     +-------------------------+
                                     |    Outer headers with   |
                                     ~     locator addresses   ~
                                     |     (IPv4 or IPv6)      |
                                     +-------------------------+
                                     |       SEAL Header       |
   +-------------------------+       +-------------------------+
   |   Inner Packet Header   |  -->  |   Inner Packet Header   |
   ~    with EP addresses    ~  -->  ~    with EP addresses    ~
   | (IPv4, IPv6, OSI, etc.) |  -->  | (IPv4, IPv6, OSI, etc.) |
   +-------------------------+       +-------------------------+
   |                         |  -->  |                         |
   ~    Inner Packet Body    ~  -->  ~    Inner Packet Body    ~
   |                         |  -->  |                         |
   +-------------------------+       +-------------------------+

      Inner packet before                Outer packet after
       encapsulation                       encapsulation

 Figure 1: Encapsulation of Inner Packets within Outer IP Headers

VET specifies the automatic tunneling mechanisms used for encapsulation, while SEAL specifies the format and usage of the SEAL header as well as a set of control messages. Most notably, IAs use the SEAL Control Message Protocol (SCMP) to deterministically exchange and authenticate control messages such as route redirections, indications of Path Maximum Transmission Unit (PMTU) limitations, destination unreachables, etc. IAs appear as neighbors on an NBMA virtual link, and form bidirectional and/or unidirectional tunnel-neighbor relationships.

The IRON is the union of all virtual overlay networks that are configured over a common underlying Internet and are owned and managed by Virtual Prefix Companies (VPCs). Each such virtual overlay network comprises a set of IAs distributed throughout the Internet to serve highly aggregated Virtual Prefixes (VPs). VPCs delegate sub-prefixes from their VPs, which they lease to customers as End User Network Prefixes (EPs). In turn, the customers assign the EPs to their customer edge IAs, which connect their End User Networks (EUNs) to the IRON.

VPCs may have no affiliation with the ISP networks from which customers obtain their basic Internet connectivity. Therefore, a customer could procure its summary network services either through a common broker or through separate entities. In that case, the VPC can open for business and begin serving its customers immediately

without the need to coordinate its activities with ISPs or other VPCs. Further details on business considerations are out of scope for this document.

The IRON requires no changes to end systems or to most routers in the Internet. Instead, the IRON comprises IAs that are deployed either as new platforms or as modifications to existing platforms. IAs may be deployed incrementally without disturbing the existing Internet routing system and act as waypoints (or "cairns") for navigating the IRON. The functional roles for IAs are described in the following sections.

IRON Client

An IRON client (or, simply, "Client") is a customer's router or host that logically connects the customer's EUNs and their associated EPs to the IRON via tunnels, as shown in Figure 2. Client routers obtain EPs from VPCs and use them to number subnets and interfaces within their EUNs. A Client can be deployed on the same physical platform that also connects the customer's EUNs to its ISPs, but it may also be a separate router or even a standalone server system located within the EUN. (This model applies even if the EUN connects to the ISP via a Network Address Translator (NAT) -- see Section 6.7). Finally, a Client may also be a simple end system that connects a singleton EPA and exhibits the outward appearance of a host.

                       .-.
                    ,-(  _)-.
    +--------+   .-(_    (_  )-.
    | Client |--(_     ISP      )
    +---+----+     `-(______)-'
        |   <= T         \     .-.
       .-.       u        \ ,-(  _)-.
    ,-(  _)-.       n     .-(_    (-  )-.
 .-(_    (_  )-.      n  (_   Internet   )
(_     EUN      )       e   `-(______)-
   `-(______)-'           l          ___
        |                   s =>    (:::)-.
   +----+---+                   .-(::::::::)
   |  Host  |                .-(::::::::::::)-.
   +--------+               (:::: The IRON ::::)
                             `-(::::::::::::)-'
                                `-(::::::)-'

      Figure 2: IRON Client Router Connecting EUN to the IRON

IRON Serving Router

An IRON serving router (or, simply, "Server") is a VPC's overlay network router that provides forwarding and mapping services for the EPs owned by customer Client routers. In typical deployments, a VPC will deploy many Servers around the IRON in a globally distributed fashion (e.g., as depicted in Figure 3) so that Clients can discover those that are nearby.

         +--------+    +--------+
         | Boston |    | Tokyo  |
         | Server |    | Server |
         +--+-----+    ++-------+
 +--------+  \         /
 | Seattle|   \   ___ /
 | Server |    \ (:::)-.       +--------+
 +------+-+  .-(::::::::)------+ Paris  |
         \.-(::::::::::::)-.   | Server |
         (:::: The IRON ::::)  +--------+
          `-(::::::::::::)-'

+--------+ / `-(::::::)-' \ +--------+ | Moscow + | \--- + Sydney | | Server | +----+---+ | Server | +--------+ | Cairo | +--------+

              | Server |
              +--------+

     Figure 3: IRON Serving Router Global Distribution Example

Each Server acts as a tunnel-endpoint router that forms a bidirectional tunnel-neighbor relationship with each of its Client customers. Each Server also associates with a set of Relays that can forward packets from the IRON out to the native Internet and vice versa, as discussed in the next section.

IRON Relay Router

An IRON Relay Router (or, simply, "Relay") is a VPC's overlay network router that acts as a relay between the IRON and the native Internet. Therefore, it also serves as an Autonomous System Border Router (ASBR) that is owned and managed by the VPC.

Each VPC configures one or more Relays that advertise the company's VPs into the IPv4 and IPv6 global Internet BGP routing systems. Each Relay associates with all of the VPC's overlay network Servers, e.g., via tunnels over the IRON, via a direct interconnect such as an Ethernet cable, etc. The Relay role (as well as its relationship with overlay network Servers) is depicted in Figure 4.

                  .-.
               ,-(  _)-.
            .-(_    (_  )-.
           (_   Internet   )
              `-(______)-'   |  +--------+
                    |        |--| Server |
               +----+---+    |  +--------+
               | Relay  |----|  +--------+
               +--------+    |--| Server |
                   _||       |  +--------+
                  (:::)-.  (Ethernet)
              .-(::::::::)

+--------+ .-(::::::::::::)-. +--------+ | Server |=(:::: The IRON ::::)=| Server | +--------+ `-(::::::::::::)-' +--------+

              `-(::::::)-'
                   ||      (Tunnels)
               +--------+
               | Server |
               +--------+

  Figure 4: IRON Relay Router Connecting IRON to Native Internet

IRON Organizational Principles

The IRON consists of the union of all VPC overlay networks configured over a common Internetwork (e.g., the public Internet). Each such overlay network represents a distinct "patch" on the Internet "quilt", where the patches are stitched together by tunnels over the links, routers, bridges, etc. that connect the underlying Internetwork. When a new VPC overlay network is deployed, it becomes yet another patch on the quilt. The IRON is therefore a composite overlay network consisting of multiple individual patches, where each patch coordinates its activities independently of all others (with the exception that the Servers of each patch must be aware of all VPs in the IRON). In order to ensure mutual cooperation between all VPC overlay networks, sufficient address space portions of the inner network-layer protocol (e.g., IPv4, IPv6, etc.) should be set aside and designated as VP space.

Each VPC overlay network in the IRON maintains a set of Relays and Servers that provide services to their Client customers. In order to ensure adequate customer service levels, the VPC should conduct a traffic scaling analysis and distribute sufficient Relays and Servers for the overlay network globally throughout the Internet. Figure 5 depicts the logical arrangement of Relays, Servers, and Clients in an IRON virtual overlay network.

                          .-.
                       ,-(  _)-.
                    .-(_    (_  )-.
                   (__ Internet   _)
                      `-(______)-'

      <------------     Relays      ------------>
                ________________________
               (::::::::::::::::::::::::)-.
           .-(:::::::::::::::::::::::::::::)
        .-(:::::::::::::::::::::::::::::::::)-.
       (:::::::::::   The IRON  :::::::::::::::)
        `-(:::::::::::::::::::::::::::::::::)-'
           `-(::::::::::::::::::::::::::::)-'

      <------------    Servers      ------------>
      .-.                .-.                     .-.
   ,-(  _)-.          ,-(  _)-.               ,-(  _)-.
.-(_    (_  )-.    .-(_    (_  )-.         .-(_    (_  )-.

(__ ISP A _) (__ ISP B _) ... (__ ISP x _)

  `-(______)-'       `-(______)-'            `-(______)-'
       <-----------      NATs        ------------>

       <----------- Clients and EUNs ----------->

          Figure 5: Virtual Overlay Network Organization

Each Relay in the VPC overlay network connects the overlay directly to the underlying IPv4 and IPv6 Internets. It also advertises the VPC overlay network's IPv4 VPs into the IPv4 BGP routing system and advertises the overlay network's IPv6 VPs into the IPv6 BGP routing system. Relays will therefore receive packets with EPA destination addresses sent by end systems in the Internet and direct them toward EPA-addressed end systems connected to the VPC overlay network.

Each VPC overlay network also manages a set of Servers that connect their Clients and associated EUNs to the IRON and to the IPv6 and IPv4 Internets via their associations with Relays. IRON Servers therefore need not be BGP routers themselves; they can be simple commodity hardware platforms. Moreover, the Server and Relay functions can be deployed together on the same physical platform as a unified gateway, or they may be deployed on separate platforms (e.g., for load balancing purposes).

Each Server maintains a working set of Clients for which it caches EP-to-Client mappings in its Forwarding Information Base (FIB). Each Server also, in turn, propagates the list of EPs in its working set to each of the Relays in the VPC overlay network via a dynamic

routing protocol (e.g., an overlay network internal BGP instance that carries only the EP-to-Server mappings and does not interact with the external BGP routing system). Therefore, each Server only needs to track the EPs for its current working set of Clients, while each Relay will maintain a full EP-to-Server mapping table that represents reachability information for all EPs in the VPC overlay network.

Customers establish Clients that obtain their basic Internet connectivity from ISPs and connect to Servers to attach their EUNs to the IRON. Each EUN can connect to the IRON via one or multiple Clients as long as the Clients coordinate with one another, e.g., to mitigate EUN partitions. Unlike Relays and Servers, Clients may use private addresses behind one or several layers of NATs. Each Client initially discovers a list of nearby Servers through an anycast discovery process (described below). It then selects one of these nearby Servers and forms a bidirectional tunnel-neighbor relationship with the server through an initial exchange followed by periodic keepalives.

After the Client selects a Server, it forwards initial outbound packets from its EUNs by tunneling them to the Server, which, in turn, forwards them to the nearest Relay within the IRON that serves the final destination. The Client will subsequently receive redirect messages informing it of a more direct route through a Server that serves the final destination EUN.

The IRON can also be used to support VPs of network-layer address families that cannot be routed natively in the underlying Internetwork (e.g., OSI/CLNP over the public Internet, IPv6 over IPv4-only Internetworks, IPv4 over IPv6-only Internetworks, etc.). Further details for the support of IRON VPs of one address family over Internetworks based on other address families are discussed in Appendix A.

IRON Initialization

IRON initialization entails the startup actions of IAs within the VPC overlay network and customer EUNs. The following sub-sections discuss these startup procedures.

IRON Relay Router Initialization

Before its first operational use, each Relay in a VPC overlay network is provisioned with the list of VPs that it will serve as well as the locators for all Servers that belong to the same overlay network. The Relay is also provisioned with external BGP interconnections -- the same as for any BGP router.

Upon startup, the Relay engages in BGP routing exchanges with its peers in the IPv4 and IPv6 Internets the same as for any BGP router. It then connects to all of the Servers in the overlay network (e.g., via a TCP connection over a bidirectional tunnel, via an Internal BGP (IBGP) route reflector, etc.) for the purpose of discovering EP-to- Server mappings. After the Relay has fully populated its EP-to- Server mapping information database, it is said to be "synchronized" with regard to its VPs.

After this initial synchronization procedure, the Relay then advertises the overlay network's VPs externally. In particular, the Relay advertises the IPv6 VPs into the IPv6 BGP routing system and advertises the IPv4 VPs into the IPv4 BGP routing system. The Relay additionally advertises an IPv4 /24 companion prefix (e.g., 192.0.2.0/24) into the IPv4 routing system and an IPv6 ::/64 companion prefix (e.g., 2001:DB8::/64) into the IPv6 routing system (note that these may also be sub-prefixes taken from a VP). The Relay then configures the host number '1' in the IPv4 companion prefix (e.g., as 192.0.2.1) and the interface identifier '0' in the IPv6 companion prefix (e.g., as 2001:DB8::0), and it assigns the resulting addresses as subnet-router anycast addresses [[[RFC3068]]][[[RFC2526]]] for the VPC overlay network. (See Appendix A for more information on the discovery and use of companion prefixes.) The Relay then engages in ordinary packet-forwarding operations.

IRON Serving Router Initialization

Before its first operational use, each Server in a VPC overlay network is provisioned with the locators for all Relays that aggregate the overlay network's VPs. In order to support route optimization, the Server must also be provisioned with the list of all VPs in the IRON (i.e., not just the VPs of its own overlay network) so that it can discern EPA and non-EPA addresses. (Therefore, the Server could be greatly simplified if the list of VPs could be covered within a small number of very short prefixes, e.g., one or a few IPv6 ::/20's). The Server must also discover the VP companion prefix relationships discussed in Section 5.1, e.g., via a global database such as discussed in Appendix A.

Upon startup, each Server must connect to all of the Relays within its overlay network (e.g., via a TCP connection, via an IBGP route reflector, etc.) for the purpose of reporting its EP-to-Server mappings. The Server then actively listens for Client customers that register their EP prefixes as part of establishing a bidirectional tunnel-neighbor relationship. When a new Client registers its EP prefixes, the Server announces the new EP additions to all Relays; when an existing Client unregisters its EP prefixes, the Server withdraws its announcements.

IRON Client Initialization

Before its first operational use, each Client must obtain one or more EPs from its VPC as well as the companion prefixes associated with the VPC overlay network (see Section 5.1). The Client must also obtain a certificate and a public/private key pair from the VPC that it can later use to prove ownership of its EPs. This implies that each VPC must run its own public key infrastructure to be used only for the purpose of verifying its customers' claimed right to use an EP. Hence, the VPC need not coordinate its public key infrastructure with any other organization.

Upon startup, the Client sends an SCMP Router Solicitation (SRS) message to the VPC overlay network subnet-router anycast address to discover the nearest Relay. The Relay will return an SCMP Router Advertisement (SRA) message that lists the locator addresses of one or more nearby Servers. (This list is analogous to the Intra-Site Automatic Tunnel Addressing Protocol (ISATAP) Potential Router List (PRL) [[[RFC5214]]].)

After the Client receives an SRA message from the nearby Relay listing the locator addresses of nearby Servers, it initiates a short transaction with one of the Servers carried by a reliable transport protocol such as TCP in order to establish a bidirectional tunnel- neighbor relationship. The protocol details of the transaction are specific to the VPC, and hence out of scope for this document.

Note that it is essential that the Client select one and only one Server. This is to allow the VPC overlay network mapping system to have one and only one active EP-to-Server mapping at any point in time, which shares fate with the Server itself. If this Server fails, the Client can select a new one that will automatically update the VPC overlay network mapping system with a new EP-to-Server mapping.

IRON Operation

Following the IRON initialization detailed in Section 5, IAs engage in the steady-state process of receiving and forwarding packets. All IAs forward encapsulated packets over the IRON using the mechanisms of VET [INTAREA-VET] and SEAL [INTAREA-SEAL], while Relays (and in some cases Servers) additionally forward packets to and from the native IPv6 and IPv4 Internets. IAs also use SCMP to coordinate with other IAs, including the process of sending and receiving redirect messages, error messages, etc. (Note however that an IA must not send an SCMP message in response to an SCMP error message.) Each IA operates as specified in the following sub-sections.

IRON Client Operation

After selecting its Server as specified in Section 5.3, the Client should register each of its ISP connections with the Server for multihoming purposes. To do so, it sends periodic beacons (e.g., SRS messages) to its Server via each of its ISPs to establish additional tunnel-neighbor state. This implies that a single tunnel-neighbor identifier (i.e., a "nonce") is used to represent the set of all ISP paths between the Client and the Server. Therefore, the nonce names this "bundle" of ISP paths.

If the Client ceases to receive acknowledgements from its Server via a specific ISP connection, it marks the Server as unreachable from that address and therefore over that ISP connection. (The Client should also inform its Server of this outage via one of its working ISP connections.) If the Client ceases to receive acknowledgements from its Server via multiple ISP connections, it marks the Server as unusable and quickly attempts to register with a new Server. The act of registering with a new Server will automatically purge the stale mapping state associated with the old Server, since dynamic routing will propagate the new client/server relationship to the VPC overlay network Relay Routers.

When an end system in an EUN sends a flow of packets to a correspondent, the packets are forwarded through the EUN via normal routing until they reach the Client, which then tunnels the initial packets to its Server as the next hop. In particular, the Client encapsulates each packet in an outer header with its locator as the source address and the locator of its Server as the destination address. Note that after sending the initial packets of a flow, the Client may receive important SCMP messages, such as indications of PMTU limitations, redirects that point to a better next hop, etc.

The Client uses the mechanisms specified in VET and SEAL to encapsulate each forwarded packet. The Client further uses the SCMP protocol to coordinate with Servers, including accepting redirects and other SCMP messages. When the Client receives an SCMP message, it checks the nonce field of the encapsulated packet-in-error to verify that the message corresponds to the tunnel-neighbor state for its Server and accepts the message if the nonce matches. (Note however that the outer source and destination addresses of the packet-in-error may be different than those in the original packet due to possible Server and/or Relay address rewritings.)

IRON Serving Router Operation

After the Server is initialized, it responds to SRSs from Clients by sending SRAs. When the Server receives a SEAL-encapsulated packet from one of its Client tunnel neighbors, it examines the inner destination address. If the inner destination address is not an EPA, the Server decapsulates the packet and forwards it unencapsulated into the Internet if it is able to do so without loss due to ingress filtering. Otherwise, the Server re-encapsulates the packet (i.e., it removes the outer header and replaces it with a new outer header of the same address family) and sets the outer destination address to the locator address of a Relay within its VPC overlay network. It then forwards the re-encapsulated packet to the Relay, which will, in turn, decapsulate it and forward it into the Internet.

If the inner destination address is an EPA, however, the Server rewrites the outer source address to one of its own locator addresses and rewrites the outer destination address to the subnet-router anycast address taken from the companion prefix associated with the inner destination address (where the companion prefix of the same address family as the outer IP protocol is used). The Server then forwards the revised encapsulated packet into the Internet via a default or more specific route, where it will be directed to the closest Relay within the destination VPC overlay network. After sending the packet, the Server may then receive an SCMP error or redirect message from a Relay/Server within the destination VPC overlay network. In that case, the Server verifies that the nonce in the message matches the Client that sent the original inner packet and discards the message if the nonce does not match. Otherwise, the Server re-encapsulates the SCMP message in a new outer header that uses the source address, destination address, and nonce parameters associated with the Client's tunnel-neighbor state; it then forwards the message to the Client. This arrangement is necessary to allow SCMP messages to flow through any NATs on the path.

When a Server ('A') receives a SEAL-encapsulated packet from a Relay or from the Internet, if the inner destination address matches an EP in its FIB, 'A' re-encapsulates the packet in a new outer header and forwards it to a Client ('B'), which, in turn, decapsulates the packet and forwards it to the correct end system in the EUN. However, if 'B' has left notice with 'A' that it has moved to a new Server ('C'), 'A' will instead forward the packet to 'C' and also send an SCMP redirect message back to the source of the packet. In this way, 'B' can leave behind forwarding information when changing between Servers 'A' and 'C' (e.g., due to mobility events) without exposing packets to loss.

IRON Relay Router Operation

After each Relay has synchronized its VPs (see Section 5.1) it advertises the full set of the company's VPs and companion prefixes into the IPv4 and IPv6 Internet BGP routing systems. These prefixes will be represented as ordinary routing information in the BGP, and any packets originating from the IPv4 or IPv6 Internet destined to an address covered by one of the prefixes will be forwarded to one of the VPC overlay network's Relays.

When a Relay receives a packet from the Internet destined to an EPA covered by one of its VPs, it behaves as an ordinary IP router. In particular, the Relay looks in its FIB to discover a locator of the Server that serves the EP covering the destination address. The Relay then simply encapsulates the packet with its own locator as the outer source address and the locator of the Server as the outer destination address and forwards the packet to the Server.

When a Relay receives a packet from the Internet destined to one of its subnet-router anycast addresses, it discards the packet if it is not SEAL encapsulated. If the packet is an SCMP SRS message, the Relay instead sends an SRA message back to the source listing the locator addresses of nearby Servers then discards the message. The Relay otherwise discards all other SCMP messages.

If the packet is an ordinary SEAL packet (i.e., one that encapsulates an inner packet), the Relay sends an SCMP redirect message of the same address family back to the source with the locator of the Server that serves the EPA destination in the inner packet as the redirected target. The source and destination addresses of the SCMP redirect message use the outer destination and source addresses of the original packet, respectively. After sending the redirect message, the Relay then rewrites the outer destination address of the SEAL- encapsulated packet to the locator of the Server and forwards the revised packet to the Server. Note that in this arrangement, any errors that occur on the path between the Relay and the Server will be delivered to the original source but with a different destination address due to this Relay address rewriting.

IRON Reference Operating Scenarios

The IRON supports communications when one or both hosts are located within EP-addressed EUNs, regardless of whether the EPs are provisioned by the same VPC or by different VPCs. When both hosts are within IRON EUNs, route redirections that eliminate unnecessary Servers and Relays from the path are possible. When only one host is within an IRON EUN, however, route optimization cannot be used. The following sections discuss the two scenarios.

Both Hosts within IRON EUNs

When both hosts are within IRON EUNs, it is sufficient to consider the scenario in a unidirectional fashion, i.e., by tracing packet flows only in the forward direction from source host to destination host. The reverse direction can be considered separately and incurs the same considerations as for the forward direction.

In this scenario, the initial packets of a flow produced by a source host within an EUN connected to the IRON by a Client must flow through both the Server of the source host and a Relay of the destination host, but route optimization can eliminate these elements from the path for subsequent packets in the flow. Figure 6 shows the flow of initial packets from host A to host B within two IRON EUNs (the same scenario applies whether the two EUNs are within the same VPC overlay network or different overlay networks).

             ________________________________________
          .-(                 .-.                    )-.
       .-(                 ,-(  _)-.                    )-.
    .-(          +========+(_    (_  +=====+               )-.
  .(             ||    (_|| Internet ||_) ||                  ).
.(               ||      ||-(______)-||   vv                    ).
 .(        +--------++--+   ||          ||   +------------+          ).
 (     +==>| Server(A)  |   vv          ||   | Server(B)  |====+      )
 (    //   +---------|\-+   +--++----++--+   +------------+    \\     )
 (   //  .-.         | \    |  Relay(B)  |                  .-. \\    )
 (  //,-(  _)-.      |  \   +-v----------+               ,-(  _)-\\   )
 ( .||_    (_  )-.   |   \____|                       .-(_    (_  ||. )
 ( _||  ISP A    .)  |                               (__   ISP B  ||_))
 (  ||-(______)-'    | (redirect)                       `-(______)||  )
 (  ||    |          |                                       |    vv  )

( +-----+-----+ | +-----+-----+ )

 | Client(A) | <--+                                 | Client(B) |
 +-----+-----+              The IRON                +-----+-----+
       |    (   (Overlaid on the Native Internet)     )   |
      .-.     .-(                                .-)     .-.
   ,-(  _)-.      .-(________________________)-.      ,-(  _)-.
.-(_    (_  )-.                                    .-(_    (_  )-.

(_ IRON EUN A ) (_ IRON EUN B )

  `-(______)-'                                       `-(______)-'
       |                                                  |
   +---+----+                                         +---+----+
   | Host A |                                         | Host B |
   +--------+                                         +--------+

          Figure 6: Initial Packet Flow before Redirects

With reference to Figure 6, host A sends packets destined to host B via its network interface connected to EUN A. Routing within EUN A will direct the packets to Client(A) as a default router for the EUN, which then uses VET and SEAL to encapsulate them in outer headers with its locator address as the outer source address and the locator address of Server(A) as the outer destination address. Client(A) then simply forwards the encapsulated packets into its ISP network connection that provided its locator. The ISP will forward the encapsulated packets into the Internet without filtering since the (outer) source address is topologically correct. Once the packets have been forwarded into the Internet, routing will direct them to Server(A).

Server(A) receives the encapsulated packets from Client(A) then rewrites the outer source address to one of its own locator addresses and rewrites the outer destination address to the subnet-router anycast address of the appropriate address family associated with the inner destination address. Server(A) then forwards the revised encapsulated packets into the Internet, where routing will direct them to Relay(B), which services the VPC overlay network associated with host B.

Relay(B) will intercept the encapsulated packets from Server(A) then check its FIB to discover an entry that covers inner destination address B with Server(B) as the next hop. Relay(B) then returns SCMP redirect messages to Server(A) (*), rewrites the outer destination address of the encapsulated packets to the locator address of Server(B), and forwards these revised packets to Server(B).

Server(B) will receive the encapsulated packets from Relay(B) then check its FIB to discover an entry that covers destination address B with Client(B) as the next hop. Server(B) then re-encapsulates the packets in a new outer header that uses the source address, destination address, and nonce parameters associated with the tunnel- neighbor state for Client(B). Server(B) then forwards these re- encapsulated packets into the Internet, where routing will direct them to Client(B). Client(B) will, in turn, decapsulate the packets and forward the inner packets to host B via EUN B.

(*) Note that after the initial flow of packets, Server(A) will have received one or more SCMP redirect messages from Relay(B) listing Server(B) as a better next hop. Server(A) will, in turn, forward the redirects to Client(A), which will establish unidirectional tunnel- neighbor state and thereafter forward its encapsulated packets directly to the locator address of Server(B) without involving either Server(A) or Relay(B), as shown in Figure 7.

             ________________________________________
          .-(                 .-.                    )-.
       .-(                 ,-(  _)-.                    )-.
    .-( +=============> .-(_    (_  )-.======+             )-.
  .(   //              (__ Internet   _)    ||                ).
.(    //                  `-(______)-'      vv                  ).
 .(     //                                   +------------+          ).
 (     //                                    |  Server(B) |====+      )
 (    //                                     +------------+    \\     )
 (   //  .-.                                                .-. \\    )
 (  //,-(  _)-.                                          ,-(  _)-\\   )
 ( .||_    (_  )-.                                    .-(_    (_  ||. )
 ( _||  ISP A    .)                                  (__   ISP B  ||_))
 (  ||-(______)-'                                       `-(______)||  )
 (  ||    |                                                  |    vv  )

( +-----+-----+ The IRON +-----+-----+ )

 | Client(A) |  (Overlaid on the native Internet)   | Client(B) |
 +-----+-----+                                      +-----+-----+
       |    (                                         )   |
      .-.     .-(                                .-)     .-.
   ,-(  _)-.      .-(________________________)-.      ,-(  _)-.
.-(_    (_  )-.                                    .-(_    (_  )-.

(_ IRON EUN A ) (_ IRON EUN B )

  `-(______)-'                                       `-(______)-'
       |                                                  |
   +---+----+                                         +---+----+
   | Host A |                                         | Host B |
   +--------+                                         +--------+

          Figure 7: Sustained Packet Flow after Redirects

Mixed IRON and Non-IRON Hosts

When one host is within an IRON EUN and the other is in a non-IRON EUN (i.e., one that connects to the native Internet instead of the IRON), the IA elements involved depend on the packet-flow directions. The cases are described in the following sub-sections.

From IRON Host A to Non-IRON Host B

Figure 8 depicts the IRON reference operating scenario for packets flowing from host A in an IRON EUN to host B in a non-IRON EUN.

              _________________________________________
           .-(         )-.                             )-.
        .-(      +-------)----+                           )-.
     .-(         |  Relay(A)  |--------------+               )-.
   .(            +------------+               \                ).
 .(     +=======>|  Server(A) |                \                ).

.( // +--------)---+ \ ). ( // ) \ ) ( // The IRON ) \ ) ( // .-. ) \ .-. ) ( //,-( _)-. ) \ ,-( _)-. ) ( .||_ (_ )-. ) The Native Internet .-|_ (_ )-. ) ( _|| ISP A ) ) (_ | ISP B )) ( ||-(______)-' ) |-(______)-' ) ( || | )-. v | )

( +-----+ ----+    )-.                               +-----+-----+ )
  | Client(A) |)-.                                   |  Router B |
  +-----+-----+                                      +-----+-----+
        |  (                                            )  |
       .-.   .-(____________________________________)-.   .-.
    ,-(  _)-.                                          ,-(  _)-.
 .-(_    (_  )-.                                    .-(_    (_  )-.
(_  IRON EUN A  )                                  (_non-IRON EUN B)
   `-(______)-'                                       `-(______)-'
        |                                                  |
    +---+----+                                         +---+----+
    | Host A |                                         | Host B |
    +--------+                                         +--------+

           Figure 8: From IRON Host A to Non-IRON Host B

In this scenario, host A sends packets destined to host B via its network interface connected to IRON EUN A. Routing within EUN A will direct the packets to Client(A) as a default router for the EUN, which then uses VET and SEAL to encapsulate them in outer headers with its locator address as the outer source address and the locator address of Server(A) as the outer destination address. The ISP will pass the packets without filtering since the (outer) source address is topologically correct. Once the packets have been released into the native Internet, routing will direct them to Server(A).

Server(A) receives the encapsulated packets from Client(A) then re- encapsulates and forwards them to Relay(A), which simply decapsulates them and forwards the unencapsulated packets into the Internet. Once the packets are released into the Internet, routing will direct them to the final destination B. (Note that Server(A) and Relay(A) are

depicted in Figure 8 as two halves of a unified gateway. In that case, the "forwarding" between Server(A) and Relay(A) is a zero- instruction imaginary operation within the gateway.)

This scenario always involves a Server and Relay owned by the VPC that provides service to IRON EUN A. Therefore, it imparts a cost that would need to be borne by either the VPC or its customers.

From Non-IRON Host B to IRON Host A

Figure 9 depicts the IRON reference operating scenario for packets flowing from host B in an Non-IRON EUN to host A in an IRON EUN.

              _______________________________________
           .-(         )-.                             )-.
        .-(      +-------)----+                           )-.
     .-(         |  Relay(A)  |<-------------+              )-.
   .(            +------------+               \                ).
 .(     +========|  Server(A) |                \                ).

.( // +--------)---+ \ ). ( // ) \ ) ( // The IRON ) \ ) ( // .-. ) \ .-. ) ( //,-( _)-. ) \ ,-( _)-. ) ( .||_ (_ )-. ) The Native Internet .-|_ (_ )-. ) ( _|| ISP A ) ) (_ | ISP B )) ( ||-(______)-' ) |-(______)-' ) ( vv | )-. | | )

( +-----+ ----+    )-.                               +-----+-----+ )
  | Client(A) |)-.                                   |  Router B |
  +-----+-----+                                      +-----+-----+
        |  (                                            )  |
       .-.   .-(____________________________________)-.   .-.
    ,-(  _)-.                                          ,-(  _)-.
 .-(_    (_  )-.                                    .-(_    (_  )-.
(_  IRON EUN A  )                                  (_non-IRON EUN B)
   `-(______)-'                                       `-(_______)-'
        |                                                  |
    +---+----+                                         +---+----+
    | Host A |                                         | Host B |
    +--------+                                         +--------+

           Figure 9: From Non-IRON Host B to IRON Host A

In this scenario, host B sends packets destined to host A via its network interface connected to non-IRON EUN B. Routing will direct the packets to Relay(A), which then forwards them to Server(A) using encapsulation if necessary.

Server(A) will then check its FIB to discover an entry that covers destination address A with Client(A) as the next hop. Server(A) then (re-)encapsulates the packets in an outer header that uses the source address, destination address, and nonce parameters associated with the tunnel-neighbor state for Client(A). Next, Server(A) forwards these (re-)encapsulated packets into the Internet, where routing will direct them to Client(A). Client(A) will, in turn, decapsulate the packets and forward the inner packets to host A via its network interface connected to IRON EUN A.

This scenario always involves a Server and Relay owned by the VPC that provides service to IRON EUN A. Therefore, it imparts a cost that would need to be borne by either the VPC or its customers.

Mobility, Multihoming, and Traffic Engineering Considerations

While IRON Servers and Relays can be considered as fixed infrastructure, Clients may need to move between different network points of attachment, connect to multiple ISPs, or explicitly manage their traffic flows. The following sections discuss mobility, multihoming, and traffic engineering considerations for IRON client routers.

Mobility Management

When a Client changes its network point of attachment (e.g., due to a mobility event), it configures one or more new locators. If the Client has not moved far away from its previous network point of attachment, it simply informs its Server of any locator additions or deletions. This operation is performance sensitive and should be conducted immediately to avoid packet loss.

If the Client has moved far away from its previous network point of attachment, however, it re-issues the anycast discovery procedure described in Section 6.1 to discover whether its candidate set of Servers has changed. If the Client's current Server is also included in the new list received from the VPC, this provides indication that the Client has not moved far enough to warrant changing to a new Server. Otherwise, the Client may wish to move to a new Server in order to reduce routing stretch. This operation is not performance critical, and therefore can be conducted over a matter of seconds/ minutes instead of milliseconds/microseconds.

To move to a new Server, the Client first engages in the EP registration process with the new Server, as described in Section 5.3. The Client then informs its former Server that it has moved by

providing it with the locator address of the new Server; again, via a VPC-specific reliable transaction. The former Server will then garbage-collect the stale FIB entries when their lifetime expires.

This will allow the former Server to redirect existing correspondents to the new Server so that no packets are lost.

Multihoming

A Client may register multiple locators with its Server. It can assign metrics with its registrations to inform the Server of preferred locators, and it can select outgoing locators according to its local preferences. Therefore, multihoming is naturally supported.

Inbound Traffic Engineering

A Client can dynamically adjust the priorities of its prefix registrations with its Server in order to influence inbound traffic flows. It can also change between Servers when multiple Servers are available, but should strive for stability in its Server selection in order to limit VPC network routing churn.

Outbound Traffic Engineering

A Client can select outgoing locators, e.g., based on current Quality-of-Service (QoS) considerations such as minimizing one-way delay or one-way delay variance.

Renumbering Considerations

As new link-layer technologies and/or service models emerge, customers will be motivated to select their service providers through healthy competition between ISPs. If a customer's EUN addresses are tied to a specific ISP, however, the customer may be forced to undergo a painstaking EUN renumbering process if it wishes to change to a different ISP [[[RFC4192]]][[[RFC5887]]].

When a customer obtains EP prefixes from a VPC, it can change between ISPs seamlessly and without need to renumber. If the VPC itself applies unreasonable costing structures for use of the EPs, however, the customer may be compelled to seek a different VPC and would again be required to confront a renumbering scenario. The IRON approach to renumbering avoidance therefore depends on VPCs conducting ethical business practices and offering reasonable rates.

NAT Traversal Considerations

The Internet today consists of a global public IPv4 routing and addressing system with non-IRON EUNs that use either public or private IPv4 addressing. The latter class of EUNs connect to the public Internet via Network Address Translators (NATs). When a Client is located behind a NAT, it selects Servers using the same procedures as for Clients with public addresses, e.g., it can send SRS messages to Servers in order to get SRA messages in return. The only requirement is that the Client must configure its SEAL encapsulation to use a transport protocol that supports NAT traversal, namely UDP.

Since the Server maintains state about its Client customers, it can discover locator information for each Client by examining the UDP port number and IP address in the outer headers of the Client's encapsulated packets. When there is a NAT in the path, the UDP port number and IP address in each encapsulated packet will correspond to state in the NAT box and might not correspond to the actual values assigned to the Client. The Server can then encapsulate packets destined to hosts in the Client's EUN within outer headers that use this IP address and UDP port number. The NAT box will receive the packets, translate the values in the outer headers, then forward the packets to the Client. In this sense, the Server's "locator" for the Client consists of the concatenation of the IP address and UDP port number.

IRON does not introduce any new issues to complications raised for NAT traversal or for applications embedding address referrals in their payload.

Multicast Considerations

IRON Servers and Relays are topologically positioned to provide Internet Group Management Protocol (IGMP) / Multicast Listener Discovery (MLD) proxying for their Clients [[[RFC4605]]]. Further multicast considerations for IRON (e.g., interactions with multicast routing protocols, traffic scaling, etc.) will be discussed in a separate document.

Nested EUN Considerations

Each Client configures a locator that may be taken from an ordinary non-EPA address assigned by an ISP or from an EPA address taken from an EP assigned to another Client. In that case, the Client is said to be "nested" within the EUN of another Client, and recursive nestings of multiple layers of encapsulations may be necessary.

For example, in the network scenario depicted in Figure 10, Client(A) configures a locator EPA(B) taken from the EP assigned to EUN(B). Client(B) in turn configures a locator EPA(C) taken from the EP assigned to EUN(C). Finally, Client(C) configures a locator ISP(D) taken from a non-EPA address delegated by an ordinary ISP(D). Using this example, the "nested-IRON" case must be examined in which a host A, which configures the address EPA(A) within EUN(A), exchanges packets with host Z located elsewhere in the Internet.

                        .-.
             ISP(D)  ,-(  _)-.
  +-----------+   .-(_    (_  )-.
  | Client(C) |--(_    ISP(D)    )
  +-----+-----+     `-(______)-'
        |   <= T         \     .-.
       .-.       u        \ ,-(  _)-.
    ,-(  _)-.       n     .-(_    (-  )-.
 .-(_    (_  )-.      n  (_   Internet   )
(_    EUN(C)    )       e   `-(______)-'
   `-(______)-'           l          ___
        | EPA(C)           s =>     (:::)-.
  +-----+-----+                 .-(::::::::)
  | Client(B) |              .-(::::::::::::)-.  +-----------+
  +-----+-----+             (:::: The IRON ::::) |  Relay(Z) |
        |                    `-(::::::::::::)-'  +-----------+
       .-.                      `-(::::::)-'        +-----------+
    ,-(  _)-.                                       | Server(Z) |
 .-(_    (_  )-.              +-----------+         +-----------+
(_    EUN(B)    )             | Server(C) |            +-----------+
   `-(______)-'               +-----------+            | Client(Z) |
        | EPA(B)                 +-----------+         +-----------+
  +-----+-----+                  | Server(B) |            +--------+
  | Client(A) |                  +-----------+            | Host Z |
  +-----------+                     +-----------+         +--------+
        |                           | Server(A) |
       .-.                          +-----------+
    ,-(  _)-.  EPA(A)
 .-(_    (_  )-.    +--------+
(_    EUN(A)    )---| Host A |
   `-(______)-'     +--------+

                   Figure 10: Nested EUN Example

The two cases of host A sending packets to host Z, and host Z sending packets to host A, must be considered separately, as described below.

Host A Sends Packets to Host Z

Host A first forwards a packet with source address EPA(A) and destination address Z into EUN(A). Routing within EUN(A) will direct the packet to Client(A), which encapsulates it in an outer header with EPA(B) as the outer source address and Server(A) as the outer destination address then forwards the once-encapsulated packet into EUN(B). Routing within EUN(B) will direct the packet to Client(B), which encapsulates it in an outer header with EPA(C) as the outer source address and Server(B) as the outer destination address then forwards the twice-encapsulated packet into EUN(C). Routing within EUN(C) will direct the packet to Client(C), which encapsulates it in an outer header with ISP(D) as the outer source address and Server(C) as the outer destination address. Client(C) then sends this triple- encapsulated packet into the ISP(D) network, where it will be routed into the Internet to Server(C).

When Server(C) receives the triple-encapsulated packet, it removes the outer layer of encapsulation and forwards the resulting twice- encapsulated packet into the Internet to Server(B). Next, Server(B) removes the outer layer of encapsulation and forwards the resulting once-encapsulated packet into the Internet to Server(A). Next, Server(A) checks the address type of the inner address 'Z'. If Z is a non-EPA address, Server(A) simply decapsulates the packet and forwards it into the Internet. Otherwise, Server(A) rewrites the outer source and destination addresses of the once-encapsulated packet and forwards it to Relay(Z). Relay(Z), in turn, rewrites the outer destination address of the packet to the locator for Server(Z), then forwards the packet and sends a redirect to Server(A) (which forwards the redirect to Client(A)). Server(Z) then re-encapsulates the packet and forwards it to Client(Z), which decapsulates it and forwards the inner packet to host Z. Subsequent packets from Client(A) will then use Server(Z) as the next hop toward host Z, which eliminates Server(A) and Relay(Z) from the path.

Host Z Sends Packets to Host A

Whether or not host Z configures an EPA address, its packets destined to host A will eventually reach Server(A). Server(A) will have a mapping that lists Client(A) as the next hop toward EPA(A). Server(A) will then encapsulate the packet with EPA(B) as the outer destination address and forward the packet into the Internet. Internet routing will convey this once-encapsulated packet to Server(B), which will have a mapping that lists Client(B) as the next hop toward EPA(B). Server(B) will then encapsulate the packet with EPA(C) as the outer destination address and forward the packet into the Internet. Internet routing will then convey this twice- encapsulated packet to Server(C), which will have a mapping that

lists Client(C) as the next hop toward EPA(C). Server(C) will then encapsulate the packet with ISP(D) as the outer destination address and forward the packet into the Internet. Internet routing will then convey this triple-encapsulated packet to Client(C).

When the triple-encapsulated packet arrives at Client(C), it strips the outer layer of encapsulation and forwards the twice-encapsulated packet to EPA(C), which is the locator address of Client(B). When Client(B) receives the twice-encapsulated packet, it strips the outer layer of encapsulation and forwards the once-encapsulated packet to EPA(B), which is the locator address of Client(A). When Client(A) receives the once-encapsulated packet, it strips the outer layer of encapsulation and forwards the unencapsulated packet to EPA(A), which is the host address of host A.

Implications for the Internet

The IRON architecture envisions a hybrid routing/mapping system that benefits from both the shortest-path routing afforded by pure dynamic routing systems and the routing-scaling suppression afforded by pure mapping systems. Therefore, IRON targets the elusive "sweet spot" that pure routing and pure mapping systems alone cannot satisfy.

The IRON system requires a deployment of new routers/servers throughout the Internet and/or provider networks to maintain well- balanced virtual overlay networks. These routers/servers can be deployed incrementally without disruption to existing Internet infrastructure and appropriately managed to provide acceptable service levels to customers.

End-to-end traffic that traverses an IRON virtual overlay network may experience delay variance between the initial packets and subsequent packets of a flow. This is due to the IRON system allowing a longer path stretch for initial packets followed by timely route optimizations to utilize better next hop routers/servers for subsequent packets.

IRON virtual overlay networks also work seamlessly with existing and emerging services within the native Internet. In particular, customers serviced by IRON virtual overlay networks will receive the same service enjoyed by customers serviced by non-IRON service providers. Internet services already deployed within the native Internet also need not make any changes to accommodate IRON virtual overlay network customers.

The IRON system operates between routers within provider networks and end user networks. Within these networks, the underlying paths traversed by the virtual overlay networks may comprise links that

accommodate varying MTUs. While the IRON system imposes an additional per-packet overhead that may cause the size of packets to become slightly larger than the underlying path can accommodate, IRON routers have a method for naturally detecting and tuning out all instances of path MTU underruns. In some cases, these MTU underruns may need to be reported back to the original hosts; however, the system will also allow for MTUs much larger than those typically available in current Internet paths to be discovered and utilized as more links with larger MTUs are deployed.

Finally, and perhaps most importantly, the IRON system provides an in-built mobility management and multihoming capability that allows end user devices and networks to move about freely while both imparting minimal oscillations in the routing system and maintaining generally shortest-path routes. This mobility management is afforded through the very nature of the IRON customer/provider relationship, and therefore requires no adjunct mechanisms. The mobility management and multihoming capabilities are further supported by forward-path reachability detection that provides "hints of forward progress" in the same spirit as for IPv6 Neighbor Discovery (ND).

Additional Considerations

Considerations for the scalability of Internet Routing due to multihoming, traffic engineering, and provider-independent addressing are discussed in [RADIR]. Other scaling considerations specific to IRON are discussed in Appendix B.

Route optimization considerations for mobile networks are found in [[[RFC5522]]].

Related Initiatives

IRON builds upon the concepts of the RANGER architecture [[[RFC5720]]] [[[RFC6139]]], and therefore inherits the same set of related initiatives. The Internet Research Task Force (IRTF) Routing Research Group (RRG) mentions IRON in its recommendation for a routing architecture [[[RFC6115]]].

Virtual Aggregation (VA) [GROW-VA] and Aggregation in Increasing Scopes (AIS) [EVOLUTION] provide the basis for the Virtual Prefix concepts.

Internet Vastly Improved Plumbing (Ivip) [IVIP-ARCH] has contributed valuable insights, including the use of real-time mapping. The use of Servers as mobility anchor points is directly influenced by Ivip's associated TTR mobility extensions [TTRMOB].

[RO-CR] discusses a route optimization approach using a Correspondent Router (CR) model. The IRON Server construct is similar to the CR concept described in this work; however, the manner in which customer EUNs coordinate with Servers is different and based on the redirection model associated with NBMA links.

Numerous publications have proposed NAT traversal techniques. The NAT traversal techniques adapted for IRON were inspired by the Simple Address Mapping for Premises Legacy Equipment (SAMPLE) proposal [SAMPLE].

10. Security Considerations

Security considerations that apply to tunneling in general are discussed in [V6OPS-TUN-SEC]. Additional considerations that apply also to IRON are discussed in RANGER [[[RFC5720]]] [[[RFC6139]]], VET [INTAREA-VET] and SEAL [INTAREA-SEAL].

The IRON system further depends on mutual authentication of IRON Clients to Servers and Servers to Relays. This is accomplished through initial authentication exchanges followed by tunnel-neighbor nonces that can be used to detect off-path attacks. As for all Internet communications, the IRON system also depends on Relays acting with integrity and not injecting false advertisements into the BGP (e.g., to mount traffic siphoning attacks).

Each VPC overlay network requires a means for assuring the integrity of the interior routing system so that all Relays and Servers in the overlay have a consistent view of Client<->Server bindings. Finally, Denial-of-Service (DoS) attacks on IRON Relays and Servers can occur when packets with spoofed source addresses arrive at high data rates. However, this issue is no different than for any border router in the public Internet today.

11. Acknowledgements

The ideas behind this work have benefited greatly from discussions with colleagues; some of which appear on the RRG and other IRTF/IETF mailing lists. Robin Whittle and Steve Russert co-authored the TTR mobility architecture, which strongly influenced IRON. Eric Fleischman pointed out the opportunity to leverage anycast for discovering topologically close Servers. Thomas Henderson recommended a quantitative analysis of scaling properties.

The following individuals provided essential review input: Jari Arkko, Mohamed Boucadair, Stewart Bryant, John Buford, Ralph Droms, Wesley Eddy, Adrian Farrel, Dae Young Kim, and Robin Whittle.

12. References

12.1. Normative References

[[[RFC0791]]] Postel, J., "Internet Protocol", STD 5, RFC 791,

          September 1981.

[[[RFC2460]]] Deering, S. and R. Hinden, "Internet Protocol, Version 6

          (IPv6) Specification", RFC 2460, December 1998.

12.2. Informative References

[BGPMON] net, B., "BGPmon.net - Monitoring Your Prefixes,

          http://bgpmon.net/stat.php", June 2010.

[EVOLUTION]

          Zhang, B., Zhang, L., and L. Wang, "Evolution Towards
          Global Routing Scalability", Work in Progress,
          October 2009.

[GROW-VA] Francis, P., Xu, X., Ballani, H., Jen, D., Raszuk, R., and

          L. Zhang, "FIB Suppression with Virtual Aggregation", Work
          in Progress, February 2011.

[INTAREA-SEAL]

          Templin, F., Ed., "The Subnetwork Encapsulation and
          Adaptation Layer (SEAL)", Work in Progress, February 2011.

[INTAREA-VET]

          Templin, F., Ed., "Virtual Enterprise Traversal (VET)",
          Work in Progress, January 2011.

[IVIP-ARCH]

          Whittle, R., "Ivip (Internet Vastly Improved Plumbing)
          Architecture", Work in Progress, March 2010.

[RADIR] Narten, T., "On the Scalability of Internet Routing", Work

          in Progress, February 2010.

[[[RFC1070]]] Hagens, R., Hall, N., and M. Rose, "Use of the Internet as

          a subnetwork for experimentation with the OSI network
          layer", RFC 1070, February 1989.

[[[RFC2526]]] Johnson, D. and S. Deering, "Reserved IPv6 Subnet Anycast

          Addresses", RFC 2526, March 1999.

[[[RFC3068]]] Huitema, C., "An Anycast Prefix for 6to4 Relay Routers",

          RFC 3068, June 2001.

[[[RFC4192]]] Baker, F., Lear, E., and R. Droms, "Procedures for

          Renumbering an IPv6 Network without a Flag Day", RFC 4192,
          September 2005.

[[[RFC4271]]] Rekhter, Y., Li, T., and S. Hares, "A Border Gateway

          Protocol 4 (BGP-4)", RFC 4271, January 2006.

[[[RFC4548]]] Gray, E., Rutemiller, J., and G. Swallow, "Internet Code

          Point (ICP) Assignments for NSAP Addresses", RFC 4548,
          May 2006.

[[[RFC4605]]] Fenner, B., He, H., Haberman, B., and H. Sandick,

          "Internet Group Management Protocol (IGMP) / Multicast
          Listener Discovery (MLD)-Based Multicast Forwarding
          ("IGMP/MLD Proxying")", RFC 4605, August 2006.

[[[RFC5214]]] Templin, F., Gleeson, T., and D. Thaler, "Intra-Site

          Automatic Tunnel Addressing Protocol (ISATAP)", RFC 5214,
          March 2008.

[[[RFC5522]]] Eddy, W., Ivancic, W., and T. Davis, "Network Mobility

          Route Optimization Requirements for Operational Use in
          Aeronautics and Space Exploration Mobile Networks",
          RFC 5522, October 2009.

[[[RFC5720]]] Templin, F., "Routing and Addressing in Networks with

          Global Enterprise Recursion (RANGER)", RFC 5720,
          February 2010.

[[[RFC5743]]] Falk, A., "Definition of an Internet Research Task Force

          (IRTF) Document Stream", RFC 5743, December 2009.

[[[RFC5887]]] Carpenter, B., Atkinson, R., and H. Flinck, "Renumbering

          Still Needs Work", RFC 5887, May 2010.

[[[RFC6115]]] Li, T., "Recommendation for a Routing Architecture",

          RFC 6115, February 2011.

[[[RFC6139]]] Russert, S., Fleischman, E., and F. Templin, "Routing and

          Addressing in Networks with Global Enterprise Recursion
          (RANGER) Scenarios", RFC 6139, February 2011.

[RO-CR] Bernardos, C., Calderon, M., and I. Soto, "Correspondent

          Router based Route Optimisation for NEMO (CRON)", Work
          in Progress, July 2008.

[SAMPLE] Carpenter, B. and S. Jiang, "Legacy NAT Traversal for

          IPv6: Simple Address Mapping for Premises Legacy Equipment
          (SAMPLE)", Work in Progress, June 2010.

[TTRMOB] Whittle, R. and S. Russert, "TTR Mobility Extensions for

          Core-Edge Separation Solutions to the Internet's Routing
          Scaling Problem,
          http://www.firstpr.com.au/ip/ivip/TTR-Mobility.pdf",
          August 2008.

[V6OPS-TUN-SEC]

          Krishnan, S., Thaler, D., and J. Hoagland, "Security
          Concerns With IP Tunneling", Work in Progress,
          October 2010.

Appendix A. IRON VPs over Internetworks with Different Address Families

The IRON architecture leverages the routing system by providing generally shortest-path routing for packets with EPA addresses from VPs that match the address family of the underlying Internetwork. When the VPs are of an address family that is not routable within the underlying Internetwork, however, (e.g., when OSI/NSAP [[[RFC4548]]] VPs are used within an IPv4 Internetwork) a global mapping database is required to allow Servers to map VPs to companion prefixes taken from address families that are routable within the Internetwork. For example, an IPv6 VP (e.g., 2001:DB8::/32) could be paired with a companion IPv4 prefix (e.g., 192.0.2.0/24) so that encapsulated IPv6 packets can be forwarded over IPv4-only Internetworks.

Every VP in the IRON must therefore be represented in a globally distributed Master VP database (MVPd) that maintains VP-to-companion prefix mappings for all VPs in the IRON. The MVPd is maintained by a globally managed assigned numbers authority in the same manner as the Internet Assigned Numbers Authority (IANA) currently maintains the master list of all top-level IPv4 and IPv6 delegations. The database can be replicated across multiple servers for load balancing, much in the same way that FTP mirror sites are used to manage software distributions.

Upon startup, each Server discovers the full set of VPs for the IRON by reading the MVPd. The Server reads the MVPd from a nearby server and periodically checks the server for deltas since the database was last read. After reading the MVPd, the Server has a full list of VP- to-companion prefix mappings.

The Server can then forward packets toward EPAs covered by a VP by encapsulating them in an outer header of the VP's companion prefix address family and using any address taken from the companion prefix as the outer destination address. The companion prefix therefore serves as an anycast prefix.

Possible encapsulations in this model include IPv6-in-IPv4, IPv4-in- IPv6, OSI/CLNP-in-IPv6, OSI/CLNP-in-IPv4, etc.

Appendix B. Scaling Considerations

Scaling aspects of the IRON architecture have strong implications for its applicability in practical deployments. Scaling must be considered along multiple vectors, including Interdomain core routing scaling, scaling to accommodate large numbers of customer EUNs, traffic scaling, state requirements, etc.

In terms of routing scaling, each VPC will advertise one or more VPs into the global Internet routing system from which EPs are delegated to customer EUNs. Routing scaling will therefore be minimized when each VP covers many EPs. For example, the IPv6 prefix 2001:DB8::/32 contains 2^24 ::/56 EP prefixes for assignment to EUNs; therefore, the IRON could accommodate 2^32 ::/56 EPs with only 2^8 ::/32 VPs advertised in the interdomain routing core. (When even longer EP prefixes are used, e.g., /64s assigned to individual handsets in a cellular provider network, considerable numbers of EUNs can be represented within only a single VP.) Each VP also has an associated anycast companion prefix; hence, there will be one anycast prefix advertised into the global routing system for each VP.

In terms of traffic scaling for Relays, each Relay represents an ASBR of a "shell" enterprise network that simply directs arriving traffic packets with EPA destination addresses towards Servers that service customer EUNs. Moreover, the Relay sheds traffic destined to EPAs through redirection, which removes it from the path for the vast majority of traffic packets. On the other hand, each Relay must handle all traffic packets forwarded between its customer EUNs and the non-IRON Internet. The scaling concerns for this latter class of traffic are no different than for ASBR routers that connect large enterprise networks to the Internet. In terms of traffic scaling for Servers, each Server services a set of the VPC overlay network's customer EUNs. The Server services all traffic packets destined to its EUNs but only services the initial packets of flows initiated from the EUNs and destined to EPAs. Therefore, traffic scaling for EPA-addressed traffic is an asymmetric consideration and is proportional to the number of EUNs each Server serves.

In terms of state requirements for Relays, each Relay maintains a list of all Servers in the VPC overlay network as well as FIB entries for all customer EUNs that each Server serves. This state is therefore dominated by the number of EUNs in the VPC overlay network. Sizing the Relay to accommodate state information for all EUNs is therefore required during VPC overlay network planning. In terms of state requirements for Servers, each Server maintains tunnel-neighbor state for each of the customer EUNs it serves, but it need not keep

state for all EUNs in the VPC overlay network. Finally, neither Relays nor Servers need keep state for final destinations of outbound traffic.

Clients source and sink all traffic packets originating from or destined to the customer EUN. Therefore, traffic scaling considerations for Clients are the same as for any site border router. Clients also retain state for the Servers for final destinations of outbound traffic flows. This can be managed as soft state, since stale entries purged from the cache will be refreshed when new traffic packets are sent.

Author's Address

Fred L. Templin (editor) Boeing Research & Technology P.O. Box 3707 MC 7L-49 Seattle, WA 98124 USA

EMail: [email protected]