wdiff draft-ietf-grow-rift-00.txt draft-ietf-grow-rift-01.txt

INTERNET-DRAFT D. Meyer (Editor)
draft-ietf-grow-rift-01.txt
Category Informational
Expires: July August 2004 January February 2004

Operational Concerns and Considerations for Routing Protocol
Design -- Risk, Interference, and Fit (RIFT)
<draft-ietf-grow-rift-00.txt>
<draft-ietf-grow-rift-01.txt>

Status of this Document

This document is an Internet-Draft and is in full conformance with
all provisions of Section 10 of RFC2026.

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

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

The list of current Internet-Drafts can be accessed at
http://www.ietf.org/ietf/1id-abstracts.txt

The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html.

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 [RFC 2119].

This document is a product of the RIFT Design Team. Comments should
be addressed to the authors, or the mailing list at
grow@lists.uoregon.edu.

Abstract

The Risk, Interference, and Fit (RIFT) design team was formed to
document the concerns and considerations surrounding the use of
Internet routing protocols for functions not directly related to
routing of IP packets within the Internet and IP networks. This
document is the output of that activity.

Table of Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 5
2. Scope of this Work . . . . . . . . . . . . . . . . . . . . . . 5
3. Problem Statement. . . . . . . . . . . . . . . . . . . . . . . 6
3.1. Risk, Interference, and Application Fit (RIFT) . . . . . . 6
3.1.1. Risk: Software Engineering . . . . . . . . . . . . . . . 7
3.1.2. Interference: Protocol Specification/Dynamic Behavior . 7
3.1.3. Application Fit: Distribution Topology . . . . . . . . . 7
4. Definitions. . . . . . . . . . . . . . . . . . . . . . . . . . 8
4.1. Reachability Information. . . . . . . . . . . . . . . . . . 8
4.2. Layer 3 Routing Information . . . . . . . . . . . . . . . . 8
4.2.1. Standard Routing Information . . . . . . . . . . . . . . 9
4.3. Auxiliary (non-routing) Information . . . . . . . . . . . . 9
4.4. Address Family Identifier (AFI) . . . . . . . . . . . . . . 9
4.5. Subsequent Address Family Identifier (SAFI) . . . . . . . . 9 10
4.6. Network Layer Reachability. . . . . . . . . . . . . . . . . 9 10
4.7. Application . . . . . . . . . . . . . . . . . . . . . . . . 10
4.8. Routing Protocol. . . . . . . . . . . . . . . . . . . . . . 10
4.9. Fate Sharing. . . . . . . . . . . . . . . . . . . . . . . . 10 11
5. Architectural Models . . . . . . . . . . . . . . . . . . . . . 11
5.1. General Purpose Transport Infrastructure (GPT) Model. . . . 11 12
5.2. Special Purpose Transport Infrastructure (SPT) Model. . . . 12
6. Analyzing Risk and Interference. . . . . . . . . . . . . . . . 12 13
6.1. Risk: Code Impact, and Resource Sharing . . . . . . . . . . 13
6.1.1. Code Impact. . . . . . . . . . . . . . . . . . . . . . . 13
6.1.2. Resource Sharing . . . . . . . . . . . . . . . . . . . . 13 14
6.1.2.1. Resource Sharing and Operating System Level Issues . 14
6.2. Interference. . . . . . . . . . . . . . . . . . . . . . . . 14 15
7. GTP and SPT Models: Risk and Interference. . . . . . . . . . . 15
7.1. Risk. . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
7.1.1. Code Impact. . . . . . . . . . . . . . . . . . . . . . . 15 16
7.1.2. Resource Sharing . . . . . . . . . . . . . . . . . . . . 16 17
7.1.3. Multisession BGP . . . . . . . . . . . . . . . . . . . . 17
7.2. Interference. . . . . . . . . . . . . . . . . . . . . . . . 18 19
7.2.1. Multisession BGP . . . . . . . . . . . . . . . . . . . . 19
8. Application Fit. . . . . . . . . . . . . . . . . . . . . . . . 19
8.1. RFC 2547 Style VPNs . . . . . . . . . . . . . . . . . . . . 19 20
8.1.1. RFC 2547 and Label Distribution. . . . . . . . . . . . . 21
8.2. VPWS. . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
8.3. VPLS. 21
8.2.1. Assertion #1 . . . . . . . . . . . . . . . . . . . . . . 22
8.2.2. Counter-Assertion #1 . . . . . 21
9. Operational Implications . . . . . . . . . . . . . 22
8.2.3. Assertion #2 . . . . . . . 22
10. Other Models. . . . . . . . . . . . . . . . 23
8.2.4. Counter-Assertion #2 . . . . . . . . . 22
11. Conclusions . . . . . . . . . 23
8.2.4.1. Assertion #2a . . . . . . . . . . . . . . . . . . . . 23
8.2.4.2. Counter-Assertion #2a . . . . . . . . . . . . . . . . 23
8.2.5. Assertion #3 . . . . . . . . . . . . . . . . . . . . . . 24
8.2.6. Counter-Assertion #3 . . . . . . . . . . . . . . . . . . 25
8.3. VPWS and Recommendations Per-Wire Attributes. . . . . . . . . . . . . . . . 22
12. Intellectual Property 27
8.3.1. Assertion #4 . . . . . . . . . . . . . . . . . . . . 22
13. Design Team . . 27
8.3.2. Counter-Assertion #4:. . . . . . . . . . . . . . . . . . 27
8.3.3. Assertion #5 . . . . . . 22
14. Acknowledgments . . . . . . . . . . . . . . . . 27
8.3.4. Counter-Assertion #5 . . . . . . . 23
15. Security Considerations . . . . . . . . . . . 27
8.3.5. Assertion #6 . . . . . . . . 24
16. IANA Considerations . . . . . . . . . . . . . . 28
8.3.6. Counter-Assertion #6 . . . . . . . 24
17. References. . . . . . . . . . . . 28
8.3.7. Assertion #7:. . . . . . . . . . . . . . . 25
17.1. Normative References . . . . . . . 28
8.3.8. Counter-Assertion #7:. . . . . . . . . . . . . . . 25
17.2. Informative References . . . 29
8.4. VPLS. . . . . . . . . . . . . . . . . . 27
18. Editor's Address. . . . . . . . . . . 29
8.4.1. Assertion #8 . . . . . . . . . . . . . . . . . . . . . . 29
19. Full Copyright Statement.
8.4.2. Counter-Assertion #8 . . . . . . . . . . . . . . . . . . 29

1. Introduction

The stability of the global Internet routing system has been the
subject of much research (see e.g., [RVBIB]) and discussion on
various IETF mailing lists [IETFOL]. Much of the research into the
8.4.3. Assertion #9 . . . . . . . . . . . . . . . . . . . . . . 30
8.4.4. Counter-Assertion #9 . . . . . . . . . . . . . . . . . . 30
9. Operational Implications . . . . . . . . . . . . . . . . . . . 30
9.1. OAM Functionality . . . . . . . . . . . . . . . . . . . . . 30
9.1.1. Assertion #10: . . . . . . . . . . . . . . . . . . . . . 30
9.1.2. Counter-Assertion #10: . . . . . . . . . . . . . . . . . 31
9.2. Full-Mesh Issues. . . . . . . . . . . . . . . . . . . . . . 31
10. Conclusions and Recommendations . . . . . . . . . . . . . . . 31
11. Intellectual Property . . . . . . . . . . . . . . . . . . . . 31
12. Design Team . . . . . . . . . . . . . . . . . . . . . . . . . 31
13. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 32
14. Security Considerations . . . . . . . . . . . . . . . . . . . 33
15. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 33
16. References. . . . . . . . . . . . . . . . . . . . . . . . . . 34
16.1. Normative References . . . . . . . . . . . . . . . . . . . 34
16.2. Informative References . . . . . . . . . . . . . . . . . . 37
17. Editor's Address. . . . . . . . . . . . . . . . . . . . . . . 38
18. Full Copyright Statement. . . . . . . . . . . . . . . . . . . 38

1. Introduction

The stability of the global Internet routing system has been the
subject of much research (see e.g., [RVBIB]) and discussion on
various IETF mailing lists [IETFOL]. Much of the research into the
routing system has centered around the analysis of the dynamics and
stability of the Border Gateway Protocol Version 4 [BGP] (hereafter
referred to as BGP).

However, while the theoretical properties of BGP remains a topic of
great interest, a more recent discussion has focused on effects of
the addition of new types of Network Layer Reachability Information,
or NLRI to BGP. In particular, the advent of two BGP attributes,
Multiprotocol Reachable NLRI (MP_REACH_NLRI), and Multiprotocol
Unreachable NLRI (MP_UNREACH_NLRI) [RFC2858], have made it possible
to encode and transport a wide variety of features and their
associated signaling using the BGP transport infrastructure. Examples
include include IPv6 [RFC2460], flow specification rules [FLOW], IP
VPNs [RFC2547BIS], Virtual Private LAN services [VPLS], Virtual
Private Wire Service [VPWS], and auto-discovery mechanisms for VPNs
in general [BGPVPN],

This document outlines the concerns and issues surrounding using the
BGP infrastructure as a generic feature and signaling transport.
However, the similar concerns apply to the Interior Gateway Protocols
(IGPs) in common use (e.g., ISIS [RFC1142] or OSPF [RFC2328]).

The rest of this document is organized as follows: Section 2 outlines
the scope of this work. Section 3 introduces the problem statement
which is the focus of this document, section 4 provides definitions,
and section 5 outlines the main architectural models that are
discussed. The remaining sections discuss the the implications of
those models.

2. Scope of this Work

It is the intention of the RIFT design team that this document serve
as a guide for both protocol designers and network operators. The
goal is to outline the implications associated with employing
existing routing protocols to enable additional feature sets and
functionality, as contrasted with designing new mechanisms to carry
those feature sets and functionalities.

The issues, concerns and considerations discussed in this document
focus on the implications for BGP [BGP,RFC1771]. It is important to
note that similar issues will arise when considering generalizations
to the information that the IGPs carry.

3. Problem Statement

The advent of the MP_REACH_NLRI and MP_UNREACH_NLRI attributes,
combined with the resulting generalization to the BGP infrastructure,
have created the opportunity to use BGP to transport a wide variety
of data types and their associated signaling. The combination of a
BGP data type and its associated signaling is frequently called an
"application"; example applications include the IPv4 and IPv6
[RFC2460] routing systems, flow specification rules [FLOW], auto-
discovery mechanisms for Layer 3 VPNs [BGPVPN], virtual private LAN
services [VPLS], and virtual private Wire Service [VPWS].

More recently, the discussion in the IETF community has focused on
the use of the BGP as a generalized feature transport infrastructure
[IETFOL]. The debate has recently intensified due to the emergence of
a new class of application that uses the BGP infrastructure to
distribute information that is not directly related to inter-domain
routing. Examples of such applications include the use of the BGP
transport infrastructure to provide auto-discovery for IP VPNs
[RFC2547BIS], the virtual private LAN services mentioned above [VPLS]
and VPNs in general [BGPVPN].

3.1. Risk, Interference, and Application Fit (RIFT)

As mentioned above, much of the debate surrounding these new uses of
the BGP transport infrastructure has focused on the potential
tradeoffs between the stability of the Internet routing system, as
effected by the deployment of new applications, and the desire on the
part of service providers to rapidly deploy these new applications,
and to reduce the operational cost by re-using existing protocols.

These tradeoffs have at times been described in terms of risk,
interference, and application fit. Risk models the software
engineering impact of new applications on a generic implementation,
while interference models the impact of new applications on protocol
definition and behavior. Finally, application fit models the
similarity between an application's data and signaling requirements
and a specific distribution algorithm. Each is described below.

3.1.1. Risk: Software Engineering

Risk attempts to assess the robustness tradeoffs inherent in the
addition of new applications to a given implementation. That is, risk
models the impact of generic software engineering issues on a given
implementation. These issues include the impact of new applications
on existing implementations and on the fate sharing properties of
those implementations.

A second aspect of risk lies in the trade-off of extending an
existing protocol versus designing, implementing, and deploying a new
protocol.

3.1.2. Interference: Protocol Specification/Dynamic Behavior

Interference models the potential for a new application to adversely
effect the operation of an existing implementation at the protocol
level, by inadvertently introducing a detrimental dependency of some
kind. That is, an application is said to "interfere" with an existing
application if, by virtue of the application's protocol extension(s),
one or more fundamental properties of the protocol's operation are
detrimentally altered. For example, could we create a new state which
introduces an unanticipated deadlock situation to occur? Or could we
destabilize the distributed behavior of the protocol? Or might we
simply run out of the attributes or bits available (as happened, for
example, with RADIUS [RFC2138])?

3.1.3. Application Fit: Distribution Topology

Application fit refers to how closely the requirements of the data to
be distributed match the underlying capabilities of a distribution
mechanism. For example, it is clearly inefficient to broadcast data
to all peers that is only required between two peers, just as it is
inefficient to unicast (replicate) data that is required by all peers
when a single broadcast would do.

4. Definitions

4.1. Reachability Information

Reachability information refers to information describing some part
of a network, along with how one can reach it, and perhaps also
containing attributes of the implied path to the network locale.
Typically, this information pertains to IP routing information; an
example of non-IP reachability is VPLS information [VPLS].

4.2. Layer 3 Routing Information

Layer 3 routing information represents either link state information
or network reachability information. Link state information
represents Layer 3 adjacencies and topology. Link state routing
protocols, such as OSPF [RFC2328] and ISIS [RFC1142], flood link
state information throughout an IGP domain, so that each
participating router maintains an identical copy of a database that
is computed to reflect the complete Layer 3 topology.

Layer 3 reachability information expressed as an IP address prefix
represents the set of destinations (systems) whose IP addresses are
contained in the IP address prefix. Distance/path vector routing
protocols, such as BGP, distribute Layer 3 reachability information
among routing domains.

Routers use both types of Layer 3 routing information (link state and
reachability) to produce IP forwarding tables. That, is, for purposes
of this discussion, "routing information" relates to the Layer 3
inter-domain routing data traditionally carried by BGP.

Finally, if one defines routing information as "information used to
forward packets", combined with the above definition of reachability
information, then we can consider information such as described in
[FLOW] (for example) to be routing information (since it is
attempting to add a level of granularity to how an 'aggregate' is
defined). That is, [FLOW] intends to complement to the existing
routing information, and the flow information is dependent on IP4
unicast reachability advertised by the same neighbor.

4.2.1. Standard Routing Information

In the most general terms, then, a routing protocol distributes data
to accomplish the following three functionalities:

(i). To govern the routing decision process (e.g., the
standard BGP decision process)

(ii). To constrain the flow of information (for example, with
BGP communities)

(iii). To tell the recipient how to get packets to the next hop

We will refer to information that falls into this class as "standard
routing information".

4.3. Auxiliary (non-routing) Information

Auxiliary Information is any information that is exchanged by routers
which is neither Layer 3 routing system has centered around the analysis information, nor reachability
information. IS-IS hostname TLVs are an example of the dynamics auxiliary
information [RFC1142].

4.4. Address Family Identifier (AFI)

An Address Family contains addresses that share common structure and
stability of
semantics. An Address Family Identifier (AFI) uniquely identifies
each address family. Several routing protocol messages contain a
field that represents the Border Gateway AFI. The AFI identifies the address type
used by another data item contained in that message. The Routing
Information Protocol Version 4 [BGP] (hereafter
referred to as BGP).

However, while (RIP) [RFC2453], Distance Vector Multicast
Routing Protocol (DVMRP) [RFC1075], and BGP all employ the AFI field.

For example, the theoretical properties of BGP remains MP_REACH_NLRI and MP_UNREACH_NLRI attributes
contain an AFI field. These BGP attributes also contain a topic NLRI field
that enumerates reachable or unreachable subnetworks corresponding to
the associated address family. The AFI field indicates the address
type by which reachable subnetworks are identified. When BGP is used
to distribute Layer 3 routing information, AFIs can indicate the
following address types: IPv4, IPv6, VPNv4 [RFC2547BIS]. When BGP is
used to distribute auxiliary information, AFIs can indicate other
address families.

4.5. Subsequent Address Family Identifier (SAFI)

A Subsequent Address Family Identifier (SAFI) is part of
great interest, the BGP
MP_REACH_NLRI and MP_UNREACH_NLRI attributes. These BGP attributes
also contain a more recent discussion has focused on effects of NLRI field that enumerates reachable or unreachable
subnetworks. The SAFI augments the addition of new types of AFI, carrying additional
information regarding networks enumerated in the NLRI field.

4.6. Network Layer Reachability

Network Layer Reachability Information, or NLRI to BGP. In particular, is the advent data described
by the AFI/SAFI fields [AFI,SAFI]. While these concepts were
originally described for protocols such as DVMRP [RFC1075], the bulk
of two BGP attributes,
Multiprotocol Reachable NLRI (MP_REACH_NLRI), and Multiprotocol
Unreachable NLRI (MP_UNREACH_NLRI) [RFC2858], have made it possible
to encode and transport a wide variety the generalization of features and their
associated signaling using the BGP transport infrastructure. Examples
include include IPv6 [RFC2460], flow specification rules [FLOW], IP
VPNs [RFC2547BIS], Virtual Private LAN services [VPLS], Virtual
Private Wire Service [VPWS], and auto-discovery mechanisms for VPNs NLRI described in general [BGPVPN],

This this document outlines derives
from the concerns and issues surrounding using introduction of the MP_REACH_NLRI and MP_UNREACH_NLRI
attributes to BGP infrastructure as [RFC2858].

4.7. Application

The term application is used in this document to refer to the
combination of a generic feature BGP data type and any signaling transport.
However, data that is carried
by BGP in support of the similar concerns apply to service the Interior Gateway Protocols
(IGPs) in common use (e.g., ISIS [RFC1142] or OSPF [RFC2328]). data type carries. The rest of this document data type
is organized as follows: Section 2 outlines typically described in an AFI/SAFI, while the scope actual data is
frequently contained in both NLRI and BGP community attributes
[RFC1997].

4.8. Routing Protocol

A routing protocol is composed of this work. Section two basic components: a data
distribution algorithm and a decision algorithm. A router typically
obtains Layer 3 introduces routing information via its data distribution
algorithm, and it uses this information to produce an IP forwarding
table (by applying the problem statement
which protocol's decision algorithm to the received
routing data). Note that it is the focus use of this document, section 4 provides definitions,
and section 5 outlines the main architectural models BGP's data distribution
algorithm that are
discussed. The remaining sections discuss the is the implications of
those models.

2. Scope focus of this Work

It is document. However, when judging
application fit, one may also consider whether the intention of decision
algorithms suit the RIFT design team that this document serve
as a guide for both protocol designers and network operators. application.

4.9. Fate Sharing

The
goal is fate sharing principle for end to outline the implications associated with employing
existing routing end network protocols was first
enunciated by Dave Clark [CLARK]. As applied to enable additional feature sets and
functionality, as contrasted with designing new mechanisms to carry
those feature sets and functionalities.

3. Problem Statement

The advent sharing of common resources among a group
of applications. In our case, the MP_REACH_NLRI and MP_UNREACH_NLRI attributes,
combined with particular "fate" of most interest
is the resulting generalization ability of one application, call it application A, to the BGP infrastructure,
have created the opportunity cause an
application with which it is fate sharing, call it application B, to use BGP
experience one or more faults due to transport a wide variety
of data types and their associated signaling. The combination faults in application A. Fate-
sharing can exist at many levels, including between modules on a
system, between routing protocols, between sessions of a
BGP data type and its associated signaling is frequently called an
"application"; example routing
protocols such as BGP, or between applications include the IPv4 and IPv6
[RFC2460] within a routing systems, flow specification rules [FLOW], auto-
discovery mechanisms for Layer 3 VPNs [BGPVPN], virtual private LAN
services [VPLS], and virtual private Wire Service [VPWS].

More recently,
protocol.

5. Architectural Models

In this section, we consider the discussion two architectural models which are
motivated by salient questions considered in this document, namely:

(i). Does the IETF community has focused BGP distribution protocol suit a particular
application (i.e., does an application fit the BGP
distribution protocol)?

(ii). What are the effects on the use global routing system (if
any) of carrying that application using the BGP as a generalized feature transport infrastructure
[IETFOL]. The debate has recently intensified due to distribution
protocol?

These questions must be analyzed in terms of the emergence cost of
a new class protocol and
code development, as well as in terms of application that uses the BGP infrastructure to
distribute information operational expense that is
may be incurred by utilizing (or not directly related to inter-domain
routing. Examples of such applications include utilizing) the use of mechanisms
already present in BGP.

Two models, describing alternate viewpoints, are examined in the
following sections.

5.1. General Purpose Transport Infrastructure (GPT) Model

The GPT model models BGP
transport data distribution infrastructure to provide auto-discovery for IP VPNs
[RFC2547BIS], as a
generic application transport mechanism. As such, it focuses on
application fit, and assumes that the virtual private LAN services mentioned above [VPLS] tradeoffs, both in terms of
risk and VPNs interference can be managed in general [BGPVPN].

3.1. Risk, Interference, and Application Fit (RIFT) an efficient manner. As mentioned above, much of a
result, the debate surrounding GTP models these new uses of
the BGP transport infrastructure has focused on the potential
tradeoffs between the stability issues not in terms of whether the Internet routing system, as
effected by the deployment of new applications,
application and the desire on the signaling data that need to be distributed are part
of service providers to rapidly deploy some particular class (routing, in this case), but rather whether
the requirements for the distribution these new applications,
and attributes are similar
enough to reduce the operational cost by re-using existing protocols.

These tradeoffs have at times been described in terms of risk,
interference, and application fit. Risk models the software
engineering impact distribution mechanisms of new applications on BGP. In those cases when
distribution requirements are sufficiently similar, BGP can be a generic implementation,
while interference models
logical candidate for a transport infrastructure. Note that this is
not because of the impact nature of new applications on protocol
definition and behavior. Finally, application fit models information distributed, but rather due
to the similarity between an application's data and signaling requirements
and in the transport requirements. There are of course
other operational considerations that make BGP a specific distribution algorithm. Each is described below.

3.1.1. Risk: Software Engineering

Risk attempts logical candidate,
including its close to assess ubiquitous deployment in the robustness tradeoffs inherent Internet (as well
as in intra-nets), its policy capabilities, and operator comfort
levels with the technology.

5.2. Special Purpose Transport Infrastructure (SPT) Model

The SPT model, on the
addition of new applications to a given implementation. That is, risk other hand, models the impact of generic software engineering issues on BGP infrastructure as a given
implementation. These issues include the impact of new applications
on existing implementations
special purpose transport designed specifically to transport inter-
domain routing information. As such, it is more sensitive to risk and
interference than to application fit.

There are two basic arguments supporting the SPT model: The first is
based on the fate sharing properties of
those implementations.

A second aspect of perceived risk lies profile involved in the trade-off of extending an
existing protocol versus designing, implementing, and deploying a adding new
protocol.

3.1.2. Interference: Protocol Specification/Dynamic Behavior

Interference models
applications to the potential for a BGP transport infrastructure or new application features to adversely
effect the operation of an
existing implementation at the protocol
level, by inadvertently introducing a detrimental dependency of some
kind. That is, an application BGP applications. The concern here is said that changes to "interfere" with an existing
application if, by virtue of the application's protocol extension(s),
one or more fundamental properties of the protocol's operation are
detrimentally altered. For example, could we create a new state which
introduces an unanticipated deadlock situation BGP
implementations will cause software quality to occur? Or could we degrade, and hence
destabilize the distributed behavior of global routing system. This position is based upon
well understood software engineering principles, and is strengthened
by long-standing experience that there is a direct correlation
between software features and software stability [MULLER1999]. This
concern is augmented by the protocol? Or might we
simply run out fact that in many cases, the existence of
the attributes or bits available (as happened, code for
example, with RADIUS [RFC2138])?

3.1.3. Application Fit: Distribution Topology

Application fit refers to how closely the requirements of these features, even if unused, can also cause
destabilization in the data to routing system, since in many cases software
faults cannot be distributed match isolated.

A second concern is based on interference arguments, notably that the underlying capabilities
increase in complexity of a distribution
mechanism. For example, it is clearly inefficient BGP due to broadcast the number of data
to all peers types that is only required between two peers, just as it is
inefficient to unicast (replicate) data that is required by all peers
when a single broadcast would do.

4. Definitions

4.1. Reachability Information

Reachability information refers to information describing some part
of a network, along with how one
carries can reach it, and perhaps also
containing attributes potentially destabilize the global routing system.

This concern is based on a wide range of concerns, including the implied path to fact
that the network locale.
Typically, this information pertains to IP routing information; an
example interaction of non-IP reachability is VPLS information [VPLS].

4.2. Layer 3 Routing Information

Layer 3 routing information represents either link state information
or network reachability information. Link state information
represents Layer 3 adjacencies BGP dynamics and topology. Link state current deployment practices
are poorly understood, and that the addition of non-routing data
types may adversely effect convergence and other scaling properties
of the global routing
protocols, such as OSPF [RFC2328] system.

6. Analyzing Risk and ISIS [RFC1142], flood link
state information throughout Interference

One way to frame the tradeoffs involved in a model's risk profile is
in terms of the software engineering issues surrounding where an IGP domain, so
implementation might demultiplex among applications. The important
point here is that each
participating router maintains an identical copy implementation's choice of a database that
is computed to reflect demultiplexing point
directly affects the complete Layer 3 topology.

Layer 3 reachability information expressed as an IP address prefix
represents implementation's risk profile due to its effects
on existing code, and on the set system resources it requires to be
shared among those applications.

6.1. Risk: Code Impact, and Resource Sharing

For purposes of destinations (systems) whose IP addresses are
contained in this discussion, then, we consider the IP address prefix. Distance/path vector routing
protocols, such as BGP, distribute Layer 3 reachability information
among routing domains.

Routers use both types risk profile
of Layer 3 routing information (link state the SPT and
reachability) GPT models with respect to produce IP forwarding tables. That, their application
demultiplexing point. The GPT model typically provides a single point
for demultiplexing all applications (i.e., the AFI/SAFI). On the
other hand, the SPT model, provides an application demultiplexing
point above BGP (typically at the TCP port level). That is, in the
GPT model, applications typically share a common transport session,
while the SPT model generally envisions one or more applications per
transport session (see section 7.1.3 for purposes a discussion of the impact
of multisession BGP [MULTISESSION,SOFTNOTIFY] on this discussion, "routing information" relates taxonomy).

Finally, note that these models can have very different risk profiles
with respect to code impact and resource sharing. Some of the
questions relating to risk assessment are considered below.

6.1.1. Code Impact

In this section, we outline the Layer 3
inter-domain routing data traditionally carried by BGP.

Finally, if high-level questions one defines routing information as "information used to
forward packets", combined with might ask in
assessing the above definition of reachability
information, then we can consider information such as described difference in
[FLOW] (for example) risk between GPT model and the SPT model
based on their effect on an existing code base.

o Does the code below the demultiplexing point need to be routing information (since it
changed when a new application is
attempting added?

o Does the code in existing applications have to add be changed when
a level of granularity to how an 'aggregate' new application is
defined). That added (that is, [FLOW] intends to complement to what extent are the existing
routing information, and the flow information is dependent on IP4
unicast reachability advertised by
applications decoupled)?

o Can the same neighbor.

4.3. Auxiliary (non-routing) Information

Auxiliary Information is any information that is exchanged by routers
which is neither Layer 3 routing information, nor reachability
information. IS-IS hostname TLVs are an example of Axillary
information [RFC1142].

4.4. Address Family Identifier (AFI)

An Address Family contains addresses that share common structure code in separate applications be developed, tested,
released, debugged and
semantics. An Address Family Identifier (AFI) uniquely identifies
each address family. Several routing protocol messages contain a
field that represents packaged independently from other
applications?

o Is there significant code below the AFI. The AFI identifies demultiplexing point that
can be shared among all applications?

6.1.2. Resource Sharing

In this section, we outline the address type
used by another data item contained high-level questions one might ask in that message. The Routing
Information Protocol (RIP) [RFC2453], Distance Vector Multicast
Routing Protocol (DVMRP) [RFC1075],
assessing the difference in risk between GPT model and BGP all employ the AFI field.

For example, SPT model
with respect to the BGP MP_REACH_NLRI requirements and MP_UNREACH_NLRI attributes
contain an AFI field. These BGP attributes also contain a NLRI field
that enumerates reachable properties of the system
resource sharing they require. In particular:

o Do applications have to compete for socket buffers, and hence
have the potential to block or unreachable subnetworks corresponding starve each other (at the TCP
port level)?

o Do applications have to compete for possible protocol-level
transport-related buffers and queues, and hence have the associated address family. The AFI field indicates
potential to starve or block each other at the address
type by which reachable subnetworks are identified. When BGP is used protocol
send/receive level?

o Do applications have to distribute Layer 3 routing information, AFIs can indicate compete for a possible per-connection
processing time budget, hence have the
following address types: IPv4, IPv6, VPNv4 [RFC2547BIS]. When BGP is
used potential to distribute auxiliary information, AFIs can indicate starve
each other
address families.

4.5. Subsequent Address Family Identifier (SAFI)

A Subsequent Address Family Identifier (SAFI) is part of at the BGP
MP_REACH_NLRI intra-process scheduling level?

6.1.2.1. Resource Sharing and MP_UNREACH_NLRI attributes. These BGP attributes
also contain a NLRI field that enumerates reachable or unreachable
subnetworks. The SAFI augments Operating System Level Issues

In this section, we outline the AFI, carrying additional
information regarding networks enumerated high-level questions one might ask in
assessing the NLRI field.

4.6. Network Layer Reachability

Network Layer Reachability Information, or NLRI is difference in risk between GPT model and the data described
by SPT model
based on the AFI/SAFI fields [AFI,SAFI]. While these concepts were
originally described for protocols such as DVMRP [RFC1075], affect on resource sharing at the bulk
of operating system
level. In particular:

o Do applications share a common scheduling context? That is,
do applications have to compete for per-process scheduling
budgets?

o What is the generalization degree of fate sharing between applications?

6.2. Interference

Interference models the NLRI described in this document derives
from potential for an application to affect the introduction
behavior of the MP_REACH_NLRI and MP_UNREACH_NLRI
attributes to BGP [RFC2858].

4.7. Application

The term an existing application is used or applications. For example, in this document to refer to
the
combination case of the Internet routing system, one might ask if a BGP data type and any signaling data that is carried certain
application "interferes" with IPv4 Unicast routing by BGP in support affecting some
aspect of the service the data type carries. The data type
is typically described its protocol operation (e.g., convergence time).

Interference in an AFI/SAFI, while the actual data is
frequently contained in both NLRI and BGP community attributes
[RFC1997].

4.8. Routing Protocol

A routing protocol is composed of two basic components: a data
distribution algorithm and a decision algorithm. A router typically
obtains Layer 3 Internet routing information via system has its data distribution
algorithm, roots in the
observation that the routing system itself can be described as highly
self-dissimilar, with extremely different scales and it uses levels of
abstraction. Complex systems with this information property are susceptible to produce an IP forwarding
table (by applying
"coupling", which RFC 3439 [RFC3439] defines as follows:

The Coupling Principle states that as things get larger, they
often exhibit increased interdependence between components.

COROLLARY: The more events that simultaneously occur, the protocol's decision algorithm to larger
the received
routing data). Note likelihood that two or more will interact. This phenomenon
has also been termed "unforeseen feature interaction"
[WILLINGER2002].

That is, interference, if and where it occurs, has its roots in
complexity and is frequently the use of BGP's data distribution
algorithm that is the focus result of this document. However, when judging application fit, one may also consider whether coupling.

7. GTP and SPT Models: Risk and Interference

In this section, we analyze the decision
algorithms suit risk and interference profiles of the application.

4.9. Fate Sharing

The fate sharing principle for end to end network protocols was first
enunciated by Dave Clark [CLARK].
SPT and GPT models.

7.1. Risk

As applied to software systems,
fate sharing refers to mentioned above, risk models the sharing of common resources among a group
of applications. In our case, robustness tradeoffs around
generic software architecture and engineering associated with
protocol implementations, including the particular "fate" of most interest
is impact on existing protocol
implementations, and on the ability of one application, call it application A, to cause an
application with which it is fate sharing, call it application B, to
experience one or more faults due to faults in application A. Fate- sharing can exist at many levels, including between modules on a
system, between routing protocols, between sessions properties of those
implementations. In the following sections we consider these
components of a routing
protocols such as BGP, or between applications within a routing
protocol.

5. Architectural Models risk for both the GPT and SPT models.

7.1.1. Code Impact

In this section, we consider outline the answers to the two architectural models which are
motivated by salient questions considered in this document, namely:

(i). posed above.

o Does the BGP distribution protocol suit code below the demultiplexing point need to be
changed when a particular
application (i.e., does an new application fit the BGP
distribution protocol)?

(ii). What is added?

In theory, such code changes are unlikely to be required in
the effects on SPT model, as the global routing system (if
any) of carrying SPT model envisions that a new
application using will have a new demultiplexing point (port).

The GPT model does not by definition require new code below
the BGP distribution
protocol?

These questions must be analyzed demultiplexing point either. Specifically, it should in terms of
theory be possible to isolate code below the cost of protocol demultiplexing
point with suitable abstraction and
code development, as well constructs such as in terms of the operational expense that
may be incurred by utilizing (or not utilizing)
AFI/SAFI API registries.

o Does the mechanisms
already present code in BGP.

Two models, describing alternate viewpoints, existing applications have to be changed when
a new application is added (that is, to what extent are examined in the
following sections.

5.1. General Purpose Transport Infrastructure (GPT) Model
applications decoupled)?

The GPT SPT model models BGP data distribution infrastructure as a
generic envisions application transport mechanism. independence with respect to
demultiplexing point. As such, it focuses on
application fit, and assumes is unlikely to require such
changes. However, it is important to note that the tradeoffs, both in terms of
risk good software
engineering practices encourage code reuse and interference can be managed in an efficient manner. construction of
general purpose libraries. As a result, if applications share
libraries and/or other code, the GTP models these issues not in terms of whether the
application practical independence
decreases, and signaling data that need to consequently risk increases. The same analysis
can be distributed are part
of some particular class (routing, in this case), but rather whether
the requirements made for the distribution these attributes GPT model, since in this case we are similar
enough to already
demultiplexing on the distribution mechanisms of BGP. In those cases when
distribution requirements are sufficiently similar, BGP can AFI/SAFI fields.

o Can the code in separate applications be a
logical candidate for a transport infrastructure. Note that developed, tested,
released, debugged and packaged independently from other
applications?

While this is
not because of the nature of information distributed, but rather due
to the similarity theoretically possible in the transport requirements. There are of course
other operational considerations that make BGP a logical candidate,
including its close to ubiquitous deployment SPT model (and
possibly more difficult in the Internet (as well
as in intra-nets), its policy capabilities, GPT model) practice and operator comfort
levels with the technology.

5.2. Special Purpose Transport Infrastructure (SPT) Model

The SPT model, on the other hand, models the BGP infrastructure as a
special purpose transport designed specifically to transport inter-
domain routing information. As such, it is more sensitive
experience has shown that achieving this type of independence is
difficult in either model.

7.1.2. Resource Sharing

In this section, we address the questions raised above to assess the
difference in risk between GPT model and
interference than to application fit.

There are two basic arguments supporting the SPT model: The first is model based on the perceived risk profile involved in adding new
effect on resource sharing considerations.

o Do applications have to the BGP transport infrastructure or new features to
existing BGP applications. The concern here is that changes to BGP
implementations will cause software quality to degrade, compete for socket buffers, and hence
destabilize the global routing system. This position is based upon
well understood software engineering principles, and is strengthened
by long-standing experience that there is a direct correlation
between software features and software stability [MULLER1999]. This
concern is augmented by
have the fact that in many cases, potential the existence of to block or starve each other (at the code GPT
level)?

The SPT model does not require applications to compete for these features, even if unused, can
socket level resources. It should also cause
destabilization in the routing system, since in many cases software
faults cannot be isolated.

A second concern is based on interference arguments, notably that the
increase in complexity of BGP due possible to the number of data types that it
carries can also potentially destabilize the global routing system.
This concern is based on a wide range achieve
this type of concerns, including the fact
that application independence in the interaction of BGP dynamics GPT model with
multisession BGP.

o Do applications have to compete for possible protocol-level
transport-related buffers and current deployment practices
are poorly understood, queues, and that hence have the addition of non-routing data
types may adversely effect convergence and
potential to starve or block each other scaling properties
of at the global routing system.

6. Analyzing Risk and Interference

One way to frame protocol
send/receive level?

Again, while the tradeoffs involved in SPT model does not require competition for
transport-level resources, it should be possible to achieve
similar behavior with multisession BGP.

o Do applications have to compete for a model's risk profile is
in terms of the software engineering issues surrounding where an
implementation might demultiplex among applications. The important
point here is that an implementation's choice of demultiplexing point
directly affects possible per-connection
processing time budget, hence have the implementation's risk profile due potential to its effects
on existing code, and on starve
each other at the system resources it requires intra-process scheduling level?

Applications written to be
shared among those applications.

6.1. Risk: Code Impact, and Resource Sharing

For purposes of this discussion, then, we consider the risk profile
of the SPT and GPT models model should not require
this type of resource competition. It should also be possible to
reduce this type of resource competition with respect multisession BGP.

o Do applications have to their application
demultiplexing point. The GPT model typically provides a single point compete for demultiplexing all applications (i.e., resources within the AFI/SAFI). On
network (e.g., bandwidth), when the
other hand, protocol session spans
multiple hops ?

Neither the SPT model, provides an application demultiplexing
point above BGP (typically at model nor the TCP port level). That is, GPT model (again, with
multisession BGP) should require competition for network
resources in this case.

7.1.3. Multisession BGP

Suppose that one makes the
GPT model, applications typically share simplifying assumption that a common transport session,
while GPT
implementation's risk profile is dominated by the SPT model generally envisions probability that an
error in one or more applications per
transport session (see section 7.1.3 for a discussion AFI/SAFI stream will cause some subset of the impact
of multisession BGP [MULTISESSION,SOFTNOTIFY] on other
AFI/SAFI streams to malfunction (e.g., reset). In this taxonomy).

Finally, note that these models can have very different case, risk profiles
with respect to code impact
might be characterized as a function of the model and resource sharing. Some the number of
AFI/SAFI carried. Given this simplification, the
questions relating to risk assessment are considered below.

6.1.1. Code Impact

In this section, profile looks
loosely like

Risk = f(Model, |{AFI,SAFI}|)

where

f:{GPT, SPT} X |{AFI, SAFI}| -> N

Note that we outline the high-level questions one might ask in
assessing assume that

f(SPT,n) = O(f(GPT,n))

where

O(f) = {g:N->R | there exists c > 0 and n such that g(n) < c*f(n)}

That is, that the difference in SPT risk between profile is bounded by the GPT model and risk
profile. Clearly, the SPT model
based on their effect on existence of such an existing code base.

o Does upper bound is an integral
aspect of any argument favoring the code below SPT model.

Note that for the demultiplexing point need to be
changed when SPT model, we can think of the number of AFI/SAFI
that a new application is added?

o Does single session carries as a small constant, call it k. k will
typically be small (close to 1), since by definition the code SPT model
envisions a small number of AFI/SAFI per session (e.g., for AFI/SAFI
IPv4/unicast and IPv6/unicast, k = 2).

When formulated in existing applications have this way, one can see that one objective of
multisession BGP is to be changed when find a new application is added (that value, call it g, such that

f(GPT, g) ~ f(SPT,k), for small values of k (i.e., k close to 1)

where

A(n) ~ B(k) ==> A(n) = B(k) + h(n), h(n) >= 0

That is, A(n) is approximately equal to what extent are the
applications decoupled)?

o Can the code in separate applications be developed, tested,
released, debugged and packaged independently from other
applications?

o Is there significant code below the demultiplexing point that
can be shared among all applications?

6.1.2. Resource Sharing B(k)

In this section, we outline the high-level questions one might ask in
assessing the difference in risk between GPT model and case, g is the SPT model
with respect to size of the requirements multisession AFI/SAFI grouping,
and properties for small values of g, multisession BGP can have a risk profile
that looks very much like the system
resource sharing they require. SPT risk profile. In particular:

o Do applications have to compete particular, for socket buffers, and hence g
= 1, both models would have the potential to block or starve each similar risk profiles. Of course, there
are many other (at components of risk that that are not considered by
this analysis, such as collateral issues resulting from the TCP
port level)?

o Do applications have to compete for possible protocol-level
transport-related buffers and queues, existence
of faulty shared code, operating system process and hence have the
potential to starve or block each other at memory structure,
etc.

7.2. Interference

Interference concerns stem from the protocol
send/receive level?

o Do applications have possibility that application
coupling can lead to compete for a possible per-connection
processing time budget, hence have the potential to starve
each other at destabilization of the intra-process scheduling level?

6.1.2.1. Resource Sharing Internet routing
system in unanticipated and Operating System Level Issues unexpected ways. In this section, section we outline
consider interference properties of the high-level questions GPT and SPT models.

7.2.1. Multisession BGP

Multisession BGP also seeks to reduce the interference profile of the
GPT model by eliminating one might ask potential source of interference,
namely, the potential interference due to presence of multiple
AFI/SAFIs in
assessing a single BGP session. Following the difference analysis presented
in risk between GPT model section 7.1.3, we can see that for small groupings (described as
small values of g in section 7.1.3), the interference profiles of
both models converge.

8. Application Fit

In the following sub-sections, application fit is examined from the
perspective of analyzing the signaling and data distribution needs of
three representative applications, namely:

RFC 2547 Style VPNs
VPWS
VPLS

However, before investigating how the SPT model
based on BGP data distribution mechanism
(and its extensions) fit the affect on resource sharing at requirements of these applications, it
is useful to briefly review the operating system
level. gross characteristics of the BGP data
distribution infrastructure. In particular:

o Do applications share particular, we examine which
distribution topologies can be naturally built using internal BGP (or
iBGP).

iBGP has been described loosely as a broadcast mechanism since an
iBGP speaker sends information to all its peers. This is typically
achieved by means of one or more route reflectors (or RRs); a common scheduling context? That is,
do applications more
direct but less scalable means is for each iBGP speaker to have a BGP
session with each iBGP peer. It may, however, be more accurate to compete for per-process scheduling
budgets?

o What
characterize iBGP as a constrained broadcast mechanism. This is
because the degree use of fate sharing between applications?

6.2. Interference

Interference models the potential for communities in conjunction with import and export
policies allows an application iBGP speaker to affect the
behavior effectively limit its
communication to a subset of an existing application or applications. For example, in the case full set of iBGP peers; the Internet routing system, one might ask if a certain
application "interferes" with IPv4 Unicast routing by affecting some
aspect
efficiency of its protocol operation (e.g., convergence time).

Interference in the Internet routing system has its roots in the
observation that the routing system itself constrained broadcast can be described improved by techniques
such as highly
self-dissimilar, with extremely different scales described in [ORF] and levels of
abstraction. Complex systems with this property [RTCONST].

8.1. RFC 2547 Style VPNs

There are susceptible five classes of information that need to
"coupling", which be distributed for
RFC 3439 [RFC3439] defines as follows:

The Coupling Principle states that as things get larger, they
often exhibit increased interdependence between components.

COROLLARY: The more events that simultaneously occur, the larger
the likelihood that two or more will interact. This phenomenon
has also been termed "unforeseen feature interaction"
[WILLINGER2002].

That is, interference, if and where it occurs, has its roots in
complexity 2547 style VPNs:

(a). Membership (auto-discovery)
(b). Prefixes
(c). Labels
(d). BGP nexthop, and
(e). Path selection attributes

The first of these, membership or auto-discovery, must be sent to all
peers, as a BGP speaker does not know a priori which of its peers are
members of a given VPN. Membership of a given VPN is frequently recognized by
the result use of application coupling.

7. GTP and SPT Models: Risk and Interference

In extended communities called Route Targets. BGP is well-
suited for this section, we analyze the risk and interference profiles mode of distribution.

The next three of these constitute the
SPT and GPT models.

7.1. Risk

As mentioned above, risk models the robustness tradeoffs around
generic software architecture reachability information.
They say what part of a given VPN (b) is reachable, and engineering associated with
protocol implementations, including the impact on existing protocol
implementations, how it is to
be reached (c and on the fate sharing properties d). The final piece of those
implementations. In information is used for
selection if there are multiple paths to a given prefix of a VPN, as
in the following sections we consider case of multi-homing. All of these
components pieces of information need
only be distributed to members of risk for both the GPT and SPT models.

7.1.1. Code Impact

In VPN, i.e., they require a
constrained broadcast mechanism. BGP is reasonably well-suited for
this section, we outline mode of distribution using import and export NLRI filtering.
The addition of the answers mechanism in [RTCONST] makes BGP even better
suited to this.

The encoding of this information as defined in [RFC2547BIS] puts all
of this information in a single NLRI. This seems to imply that a
broadcast mechanism has to be used for the questions posed above.

o Does the code below distribution of RFC 2547
VPN information. However, the demultiplexing point need combination of [RTCONST] and [RFC2918]
allow BGP to be
changed when a new application is added? distribute this information correctly yet efficiently.

In theory, such code changes are unlikely summary, there seems to be required in
the SPT model, as the SPT model envisions little argument that a new the RFC 2547
application will have is a new demultiplexing point (port).

The GPT model does not by definition require new code below routing application. This is because the demultiplexing point either. Specifically, it should information
that gets sent via BGP in
theory be possible RFC 2547 is generally considered to isolate code below be
"routing information". That is, the demultiplexing
point protocol distributes address
prefixes, along with suitable abstraction and constructs such as
AFI/SAFI API registries.

o Does the code in existing applications have their next hops (and of course, some additional
attributes). Finally (and perhaps most importantly), there seems to
be changed when
a new application is added (that is, to what extent are little argument that the
applications decoupled)?

The SPT model envisions information distributed by the RFC 2547
application independence is standard routing information.

8.1.1. RFC 2547 and Label Distribution

One issue that is frequently raised with respect to
demultiplexing point. As such, it is unlikely to require such
changes. However, it whether or not
the RFC 2547 VPN application is important to note that good software
engineering practices encourage code reuse and construction of
general purpose libraries. As a result, if applications share
libraries and/or other code, routing application surrounds the
fact that, in the 2547 application, BGP distributes MPLS labels along
with the practical independence
decreases, and consequently risk increases. routes. The same analysis
can be made for contention then, is that the GPT model, since RFC 2547
application represents more than just a routing application. However,
in this case we are already
demultiplexing on the AFI/SAFI fields.

o Can MPLS label is just a shorthand way of representing
one or more address prefixes. That is, the code assertion is that in separate applications be developed, tested,
released, debugged and packaged independently from other
applications?

While this
case, the label represents "standard routing information".

8.2. VPWS

The question of whether VPWS is theoretically possible in a "good fit" for the SPT model BGP transport
infrastructure is the source of much discussion (and
possibly controversy). In
this section, we will review both positions and their supporting
arguments as a series of assertions and counter-assertions (we will
use this format throughout the rest of this section).

The key debate with respect to VPWS centers around what set of
services are being defined, and how they are to be signaled. One way
to analyze the VPWS application, then is in terms of two of its more difficult
contentious functionalities, namely:

(a). Auto-discovery

Auto-discovery refers to discovery of the set of nodes
that belong in a common L2VPN, and

(b). Signaling

Signaling refers to the GPT model) practice setup and
experience has shown maintenance of the
point-to-point pseudo-wires that achieving this type carry the traffic of independence
the L2VPN.

The next sections examine the various assertions and counter-
assertions around auto-discovery and signaling for VPLS.

8.2.1. Assertion #1

Assertion #1 states VPWS is
difficult in either model.

7.1.2. Resource Sharing

In not a routing application. Those
supporting this section, assertion argue that in the case of VPWS, we are not
distributing address the questions raised above to assess the
difference in risk between GPT model prefixes, and (importantly) unlike the SPT model based on case of
RFC 2547 style VPNs, the BGP decision process is not used (or at
least it is not used in the
effect on resource sharing considerations.

o Do applications have same way). Further, proponents argue that
what we are distributing is state information that corresponds to compete for socket buffers,
point-to-point entities, i.e., pseudo-wires, and hence
have the potential thus argues that
that the VPWS application is completely different.

8.2.2. Counter-Assertion #1

Counter-Assertion #1 states that VPWS is a routing application. More
specifically, this position is outlined in [VPLS] (section 3.4),
namely:

"It is often desired to block or starve each other (at the GPT
level)?

The SPT model does not require applications multi-home a VPLS site, i.e., to compete for
socket level resources. It should also be possible connect
it to achieve
this type of application independence multiple PEs, perhaps even in different ASes. In such a
case, the GPT model with
multisession BGP.

o Do applications have PEs connected to compete for possible protocol-level
transport-related buffers and queues, and hence have the
potential to starve or block each other at same site can either be
configured with the protocol
send/receive level?

Again, while same VE ID or with different VE IDs. In the SPT model does not require competition for
transport-level resources,
latter case, it should be possible is mandatory to achieve
similar behavior with multisession BGP.

o Do applications have run STP on the CE device, and
possibly on the PEs, to compete for construct a possible per-connection
processing time budget, hence have loop-free VPLS topology.

In the potential to starve
each other at case where the intra-process scheduling level?

Applications written PEs connected to the same site are
assigned the SPT model should not require
this type of resource competition. It should also be possible to
reduce this type of resource competition with multisession BGP.

o Do applications have same VE ID, a loop-free topology is constructed by
routing mechanisms, in particular, by BGP path selection. When a
BGP speaker receives two equivalent NLRIs (see below for the
definition), it applies standard path selection criteria such as
Local Preference and AS Path Length to compete for resources within determine which NLRI to
choose; it MUST pick only one.

If the
network (e.g., bandwidth), when chosen NLRI is subsequently withdrawn, the protocol session spans
multiple hops ?

Neither BGP speaker
applies path selection to the SPT model nor remaining equivalent VPLS NLRIs to
pick another; if none remain, the GPT model (again, forwarding information
associated with
multisession BGP) should require competition for network
resources in this case.

7.1.3. Multisession BGP

Suppose that one makes the simplifying assumption that a GPT
implementation's risk profile NLRI is dominated by the probability removed."

8.2.3. Assertion #2

Assertion #2 states that an
error in one AFI/SAFI stream will cause auto-discovery for VPWS requires some subset form
of the other
AFI/SAFI streams constrained broadcast. There doesn't seem to malfunction (e.g., reset). be much controversy
that auto-discovery does require some sort of constrained broadcast
mechanism (which we don't want to be limited to a single AS), and we
may want to be able to optimize it by using a RP (rendezvous point)
like mechanism. BGP route reflectors (RR) provide a convenient and
ubiquitously deployed candidate RP. In this case, risk
might be characterized case (RR as RP), the fit
is good since auto-discovery, like routing, requires an n-party
protocol where each party has no a function priori knowledge of the model and the number existence
or identity of
AFI/SAFI carried. Given this simplification, the risk profile looks
loosely like

Risk = f(Model, |{AFI,SAFI}|)

where

f:{GPT, SPT} X |{AFI, SAFI}| -> N

Note that we assume other n-1 parties.

8.2.4. Counter-Assertion #2

There is no real counter-position to Assertion #2, as it simply
states that

f(SPT,n) = O(f(GPT,n))
where

O(f) = {g:N->R | VPWS auto-discovery requires some form of constrained
broadcast (about which there exists c > 0 and n such that g(n) < c*f(n)}

That is, is some controversy; see Assertion #2a
below).

8.2.4.1. Assertion #2a

Assertion #2a states that auto-discovery is not needed for VPWS.
Further, the SPT risk profile Assertion #2a states that there is not a validated need
for VPWS auto-discovery, since auto-discovery is useful only when
creating full mesh layer 2 topologies, which are undesirable due to
their (well-understood) poor scaling properties; hence auto-discovery
for VPWS is not useful.

8.2.4.2. Counter-Assertion #2a

In summary, with the existence exception of such an upper bound Assertion #2a, the major
controversy surrounding VPWS is an integral
aspect in signaling piece of any argument favoring the SPT model.

Note
application. The "VPWS is not a routing application" camp argues that for
the SPT model, we can think of VPWS signaling requirements do not fit the number of AFI/SAFI BGP distribution
infrastructure, while the "VPWS is a routing application" camp
believes that BGP is a single session carries as good fit. The next sections examine these
assertions.

8.2.5. Assertion #3

Assertion #3 states that VPWS applications are not a small constant, call it k. k will
typically good fit for
BGP. This argument is based on the assertion that BGP is poorly
suited to the VPWS signaling requirements because pseudo-wires are
inherently point-to-point (see, for example [L2VPNSIG]). Further, the
assertion is that VPWS signaling is qualitatively different than in
routing or auto-discovery, in which each piece of information must be small (close
distributed to 1), since by definition the SPT model
envisions n participants. The conclusion here is that BGP's
distribution mechanisms are a small number of AFI/SAFI per session (e.g., poor match for AFI/SAFI
IPv4/unicast and IPv6/unicast, k = 2).

When formulated in VPWS signaling. Another
way to think about this way, one can see is that one objective of
multisession BGP generally works from a single
database, and then applies some filtering on a per-connection basis;
this only makes sense if most of the information is going to go to find a value, call it g, such that

f(GPT, g) ~ f(SPT,k), for small values
lot of k (i.e., k close to 1)

where

A(n) ~ B(k) ==> A(n) = B(k) + h(n), h(n) >= 0

That is, A(n) places.

For example, suppose that a RR is approaches B(k) used for VPWS signaling, and there
is the need to set up n pseudo-wires. In this case, g is instead of
sending n setup messages, one sends one large "meta-setup" message
with all the info that would have been in the n setup messages. That
is, let

n = number of pseudo-wires
l = the size of the multisession AFI/SAFI grouping,
and for small values per-wire label information
k = the size of g, multisession BGP can the per-wire information

In this case, the meta-setup message will be of size O((l + k) * n).

After receiving the setup message, the RR then must send the n
messages that could have a risk profile been sent by the endpoint (note that looks very much like this is
almost true; the SPT risk profile. In particular, for g
= 1, both models endpoint would have similar risk profiles. Of course, there
are many other components to send n messages of risk that size (l +
k), but the RR will have to send n copies of the larger setup
message).

8.2.6. Counter-Assertion #3

Counter-Assertion #3 states that are not considered by the VPWS application is a good fit
for BGP (see, for example [L2VPNT]). In particular, this analysis, such camp
suggests that a RR really only needs to distribute the label-range
[LABELRANGE], so the setup message isn't really n times as collateral issues resulting from large, but
rather is analyzed as follows:

Let n = number of pseudo-wires
m = the existence size of faulty shared code, operating system process the label-range data
k = the size of the per-wire information

Then the messages will be of size O(m + (k * n)), and memory structure,
etc.

7.2. Interference

Interference concerns stem from most
importantly for the possibility label-range argument:

O(m + (k * n)) < O((l + k) * n)

That is, the label-range concept reduces the size of the
messages that application
coupling can lead need to be sent to and by the destabilization of RR.

However, some will argue that the label-range concept is efficient if
and only if:

(a). A large enough label range is preallocated to accommodate
all the systems you might ever want to add to the Internet routing
system in unanticipated
VPLS/VPWS (assuming that service interruption is not
acceptable), and unexpected ways.

(b). There is no per-wire information other than labels that
needs to distributed

In this section we
consider interference properties of these cases, the GPT and SPT models.

7.2.1. Multisession BGP

Multisession BGP also seeks to label range approach can reduce the interference profile of the
GPT model by eliminating one potential source size of interference,
namely, the potential interference due to presence of multiple
AFI/SAFIs in a single BGP session. Following
setup messages as analyzed above. However, the analysis presented
in section 7.1.3, we can see counter argument is
that for small groupings (described any such reduction will become a second-order effect as
small values of g in section 7.1.3), the interference profiles soon as
some other piece of
both models converge.

8. Application Fit per-wire status or configuration (e.g., MTU)
information must be distributed. In addition, the following sub-sections, application fit is examined from the
perspective of analyzing the data distribution needs of three
representative classes idea of application, namely:

RFC 2547 Style VPNs
VPWS
VPLS

8.1. RFC 2547 Style VPNs

First, it is useful pre-
allocating a large enough label range to review the distribution mechanisms available
in BGP, in particular, accommodate future
expansion, while saving bits in i-BGP. i-BGP the setup messages, has been described loosely as other costs
which may be large. In particular:

(a). Until the future expansion takes place (if it ever does),
one may be wasting quite a broadcast mechanism since an i-BGP speaker sends information lot of labels (noting that
that each label you distribute requires you to all
its peers. This is typically achieved by means allocate a
piece of high-speed memory in your forwarding engine;
putting some of it aside for possible later use seems
very costly. Each one you put aside is, e.g., one or more route
reflectors; a more direct but less scalable means is
RFC2547 route you can support).

However, if you don't preallocate enough contiguous label
space for each i-BGP
speaker to future expansion, then if the expansion occurs
you must start adding additional labels or label ranges,
and your setup messages start getting longer anyway (in
theory, you could just carve a new set of label ranges,
instead of adding new ones; counter-position: if you did
that, you'd have to bring down your whole VPWS (and
possibly VPLS) every time you add a BGP session with each i-BGP peer.

However, new endpoint).

(b). Fragmentation of the label space, which can result from
this preallocation, has real impact on label switching
implementations (as the MPLS architecture explicitly
leaves it is more accurate to characterize i-BGP as a constrained
broadcast mechanism. This is because the use of communities in
conjunction with import and export policies allows an i-BGP speaker implementation to effectively limit develop its communication to own label
assignment strategies). So if, for example, a subset of the full set of
i-BGP peers; the efficiency of constrained broadcast hardware
designer thinks s/he can be improved improve performance by techniques such as described using,
say, prime numbered labels first, s/he should have the
ability to design her/his system in [ORF] this way. If an
application is going to come along and [RTCONST].

There are five classes of information demand that need labels
be assigned in contiguous groups, implementations which
are perfectly conformant to the architecture may not be distributed for
RFC 2547 style VPNs:

(a). Membership (auto-discovery)
(b). Prefixes
able to support that application.

(c). Labels
(d). BGP nexthop, and
(e). Path selection attributes

The first For diagnosis of these, membership or auto-discovery, must be sent to all
peers, as a BGP speaker does network problems, the label-range
approach may have the additional issue that the operator
may not know a priori (a priori) which of its label(s) were assigned to
which endpoint(s).

(d). Finally, one may argue that label-range allocation is
sub-optimal for non-full mesh topologies, since all peers are
members
of the VPN must hear about the a given VPN. Membership of label-range withdraw, and
(in a given VPN is recognized by non-full mesh topology), not all peers need to know
about it.

In any event, one may argue that the use scaling benefits of certain extended communities called Route Targets. BGP using a RR
in routing is
clearly eminently well-suited for this mode of distribution.

The next three that the RR pre-digests all the received info; it runs
the (BGP) decision process, and only forwards the results of these constitute the reachability information.
They say what part
decision process, rather than forwarding all the raw data. In the
case of a given VPN (b) VPWS (and possibly VPLS), the argument is reachable, that this advantage
is absent (i.e., we don't run BGP path selection), and how as a result,
the RR doesn't help with scaling in the same way it does with
routing. Of course, the counter position is to
be reached (c and d). The final piece that some form of information BGP
path selection is used used; see discussion above). Finally, one may argue
that using the RR will introduce some latency into the label withdraw
procedure.

8.3. VPWS and Per-Wire Attributes

While several per-wire attributes have been defined (see [L2TPV3],
for
selection if there example), the need for per-wire attributes for VPWS remains
controversial. The following sections examine those controversies.

8.3.1. Assertion #4

Assertion #4 is that VPWS requires various per-wire parameters. These
may include (but are multiple paths not limited to) MTU, whether to a given prefix use sequencing
capabilities, bandwidth capabilities, and QoS. In addition, during
the lifetime of a VPN, pseudo-wire, there are per-wire status indications
that may need to be passed to the other endpoint.

8.3.2. Counter-Assertion #4:

Counter-Assertion #4 states that it has not been demonstrated that
VPWS needs per-wire attributes as few (per-wire attributes) have as
yet been defined (see, e.g., [MARTINI]).

8.3.3. Assertion #5

Assertion #5 states that passing per-wire attributes through an RR
will likely be inefficient. The argument here is that in the case of multi-homing. All of event
that per-wire attributes are required, passing these pieces of information need
only (per-wire)
attributes through a RR will be distributed to members of sub-optimal as the VPN, i.e., they require a
constrained broadcast mechanism. BGP is reasonably well-suited for
this mode of distribution using import and export NLRI filtering.
The addition of RR will forward
the mechanism in [RTCONST] makes BGP even better
suited status to this.

The encoding of this information as defined in [RFC2547BIS] puts all
of this information in a single NLRI. This seems the VPWS members, not just to imply the one endpoint that a
broadcast mechanism
is interested in it. For attributes like sequence numbers, it may
even more difficult as one has to be used for make sure the distribution of RFC 2547
VPN information. However, sequence numbers
resynchronize properly when the combination of [RTCONST] and [RFC2918]
allow BGP pseudo-wire flaps. This seems
somewhat difficult to distribute this information correctly yet efficiently.

Finally, achieve through a BGP RR.

8.3.4. Counter-Assertion #5

The counter assertion here is that, since few (or no) per-wire
attributes have been defined (counter-assertion #4), the fact that it
is useful inefficient to observe that standard use a RR for distribution is irrelevant.

8.3.5. Assertion #6

Assertion #6 states that, while still an open issue, pseudo-wire
congestion control may require regular point-to-point control message
exchanges, something which BGP path selection
mechanisms (local pref, MED, AS path length, etc.) can be applied would seem ill-equipped to handle.

8.3.6. Counter-Assertion #6

In this case, the information in (e).

The conclusion counter assertion is that BGP since few (or no) per-
wire attributes have been defined (see counter-assertion #4), and
further, since congestion control for pseudo-wires is quite well-suited still an open
issue, arguing fit is premature.

8.3.7. Assertion #7:

Assertion #7 states that the primary motivation for VPWS is to
deliver existing service models (i.e., Frame Relay and ATM) over a
packet infrastructure (this is as opposed to some new service). In
this application,
and, case, common deployments involve partial mesh topologies (more
specifically multiple hub and spoke connections, with some hub to hub
connectivity that makes sense for the addition enterprise traffic profile). In
addition, some of mechanisms the connections in such as [RTCONST] and [RFC2918], deployments require per-
wire characteristics (e.g., guaranteed throughput for voice, etc).

In other words, the fit argument here is even closer.

8.2. that a VPWS

A VPN based on service designed to
support so-called legacy services (Frame Relay and ATM) will require
point-to-point signaling due to existing topologies and the need for
per-wire attributes. Further, for new VPWS services that require
full-mesh auto-provisioning, the "Colored Pools PW Provisioning
Model" [L2VPN] suggests a Virtual Private Wire Service [VPWS] method to support such provisioning while
retaining the point-to-point signaling required to support per-wire
attributes.

8.3.8. Counter-Assertion #7:

8.4. VPLS

A VPLS service connects a
number number of sites by an emulated LAN segment.
In the next sections, we examine whether VPLS maybe be considered to
be a routing application, and hence whether BGP is a good fit for its
distribution requirements.

8.4.1. Assertion #8

Assertion #8 states that VPLS is a routing application, since the
notion of sites by virtual wires (or pseudo-wires). The information
needed "VPLS site identification" is analogous to create such a VPN comprises:

(a). Membership (auto-discovery)
(b). VPN site identification
(c). Labels
(d). BGP nexthop
(e). Path selection attributes, and
(f). Per-wire information

The
identifier for VPWS (which this camp also views as a routing
application). As a result, the analysis of the first distribution needs of
these five items is exactly as for RFC 2547 VPNs,
with the slight change that and the definition of a 'part of a VPN' conclusion
is no
longer an IP prefix, but that BGP is a VPN site identifier, which can be
viewed as the VPWS prefix. The distribution requirements reasonably well-suited for this application, and the fit with BGP distribution mechanisms is identical to RFC 2547.

The one major change is
the potential for 'per-wire' attributes, such
as bandwidth for a given site-to-site connection. This information
should be distributed on a point-to-point basis. BGP mechanisms are
not efficient for point-to-point distribution. However, it is an
open question whether such 'per-wire' attributes really need to be
exchanged, as evidenced by addition of [RTCONST] and [REFRESH], the fact that LDP signaling for pseudo-
wires [MARTINI] has not defined any such attributes. If per-wire
information fit is indeed not necessary, BGP distribution mechanisms are
as well-suited for VPWS VPNs as for RFC 2547 VPNs.

Note even better.
Finally, note that existing BGP path selection mechanisms can be used
as is for VPWS, VPLS, and can prove useful for multi-homed sites.

8.3. VPLS

8.4.2. Counter-Assertion #8

Counter-Assertion #8 states that VPLS connects is not a number of sites by an emulated LAN segment. The
information needed routing application.
In particular, the contention here is that while the VPLS NLRI are
used to create identify that a VPLS consists of:

(a). Membership (auto-discovery)
(b). VPLS site identification
(c). Labels
(d). BGP nexthop, and
(e). Path selection attributes

The notion of 'VPLS site identification' is analogous particular PE belongs to a VPN site
identifier for VPWS. The analysis of particular VPLS
instance (as described in Assertion #8),the path which data traffic
follows will depend on the distribution needs of these
five items is exactly as for RFC 2547 VPNs, route to that PE, and that route is
determined by the conclusion ordinary IP routing. As a result, it is not
relevant which neighbor a VPLS NLRI was received from, and hence is
not routing.

8.4.3. Assertion #9

Assertion #9 is that BGP constrained or true broadcast is reasonably well-suited not valuable
for this application, and with VPLS, since the
addition of [RTCONST] and [REFRESH], same label can not be used by all peers. In
particular, the fit is even better.

Note that existing BGP path selection mechanisms same label can not be used as by all peers since MAC
address learning is
for VPLS, and can prove useful for multi-homed sites. performed in the data plane.

8.4.4. Counter-Assertion #9

9. Operational Implications

In this section we examine the operational implications of the
various choices in the design spaces described in this document.

9.1. OAM Functionality

A service provider (SP) may want to know exactly where a particular
pseudo-wire leaves its domain, and in addition may want to keep
various counts and bits of status at that point. Further, the SP may
want to be able to do data path testing to that point. That is, a SP
may want point-to-point pseudo-wire state to be maintained at its
border routers.

9.1.1. Assertion #10:

Assertion #10 states that it may be difficult for service providers
to maintain point-to-point pseudo-wire state at their border routers
with the proposed BGP signaling mechanisms. This is because those
mechanisms provide no way to ensure that a pseudo-wire data path will
leave the network at a node which has state information for that
pseudo-wire.

9.1.2. Counter-Assertion #10:

9.2. Full-Mesh Issues

10. Other Models

11. Conclusions and Recommendations

12.

11. Intellectual Property

The IETF takes no position regarding the validity or scope of any
intellectual property or other rights that might be claimed to
pertain to the implementation or use of the technology described in
this document or the extent to which any license under such rights
might or might not be available; neither does it represent that it
has made any effort to identify any such rights. Information on the
IETF's procedures with respect to rights in standards-track and
standards-related documentation can be found in BCP-11 [RFC2028].
Copies of claims of rights made available for publication and any
assurances of licenses to be made available, or the result of an
attempt made to obtain a general license or permission for the use of
such proprietary rights by implementors or users of this
specification can be obtained from the IETF Secretariat.

The IETF invites any interested party to bring to its attention any
copyrights, patents or patent applications, or other proprietary
rights which may cover technology that may be required to practice
this standard. Please address the information to the IETF Executive
Director.

13.

12. Design Team

The design team that produced this document consisted of Daniel
Awduche (awduche@awduche.com), Ron Bonica (Ronald.P.Bonica@mci.com),
Hank Kilmer (hank@rem.com), Kireeti Kompella (kireeti@juniper.net),
Chris Lewis (chrlewis@cisco.com), Danny McPherson (danny@tcb.net),
David Meyer (dmm@1-4-5.net) and Peter Whiting
(pwhiting@vericenter.com).

14.

13. Acknowledgments

David Ball, Peter Gutierrez, Susan Harris, Pedro Marques, Eric Rosen,
Pekka Savola, and Mark Townsley have all made many insightful
comments on earlier versions of this document.

15.

14. Security Considerations

This document specifies neither a protocol nor an operational
practice, and as such, it creates no new security considerations.

16.

15. IANA Considerations

This document creates a no new requirements on IANA namespaces
[RFC2434].

17.

16. References

17.1.

16.1. Normative References

[AFI] http://www.iana.org/assignments/address-family-numbers

[BGP] Rekhter, Y, T.Li, and S. Hares, "A Border Gateway
Protocol 4 (BGP-4)", draft-ietf-idr-bgp4-23.txt.
Work in progress.

[BGPVPN] Ould-Brahim, H., E. Rosen, and Y. Rekhter, "Using
BGP as an Auto-Discovery Mechanism for
Provider-provisioned VPNs",
draft-ietf-l3vpn-bgpvpn-auto-00.txt. Work in
progress.

[CLARK] Clark, D., "Design Philosophy of the DARPA Internet
Protocols", Computer Communication Review, volume
25, number 1, January 1995. ISSN # 0146-4833.

[EXTCOMM] Sangali, S., D. Tappan, and Y. Rekhter, "BGP
Extended Communities Attribute",
draft-ietf-idr-bgp-ext-communities-06.txt. Work
in progress.

[FLOW] Marques, P, et. al., "Dissemination of flow
specification rules",
draft-marques-idr-flow-spec-00.txt. Work in
progress.

[LABELRANGE] What is the cite here?

[L2VPN] Andersson, L. and E. Rosen, "L2VPN Framework",
draft-ietf-l2vpn-l2-framework-03.txt. Work in
Progress.

[L2VPNSIG] Rosen, E. and V. Rodoaca, "Provisioning Models
and Endpoint Identifiers in L2VPN Signaling",
draft-ietf-l2vpn-signaling-00.txt. Work in
Progress.

[L2VPNT] Kompella, K. (Editor), "Layer 2 VPNs Over
Tunnels", draft-kompella-l2vpn-l2vpn-00.txt.
Work in Progress.

[L2TPv3] Lau, J., M. Townsley and I. Goyret (Editors),
"Layer Two Tunneling Protocol (Version
3)", draft-ietf-l2tpext-l2tp-base-11.txt. Work in
progress.
Progress.

[MARTINI] Martini, L., E.Rosen, and T. Smith, "Pseudowire
Setup and Maintenance using LDP",
draft-ietf-pwe3-control-protocol-05.txt. Work in
progress.

[MULLER1999] Muller, R. et. al., "Control System Reliability
Requires Careful Software Installation
Procedures", International Conference on
Accelerator and Largeand Large Experimental
Physics Systems, 1999, Trieste, Italy.

[MULTISESSION] Scudder, J. and C. Appanna, "Multisession BGP,
draft-scudder-bgp-multisession-00.txt. Work in
progress.

[ORF] Chen, E., and Rekhter, Y., "Cooperative Route
Filtering Capability for BGP-4",
draft-ietf-idr-route-filter-09.txt. Work in
progress.

[RTCONST] Bonica, R. et al, "Constrained VPN route
distribution",
draft-marques-ppvpn-rt-constrain-01.txt. Work in
progress.

[SOFTNOTIFY} Nalawade, G., K. Patel, J. Scudder, and D. Ward,
"BGPv4 Soft-Notification Message",
draft-nalawade-bgp-soft-notify-00.txt., Work in
progress.

[RFC1075] Waitzman, D., C. Partridge, and S. Deering,
"Distance Vector Multicast Routing Protocol", RFC
1075, November, 1988.

[RFC1142] Oran, D. Editor, "OSI IS-IS Intra-domain Routing
Protocol", RFC 1142, February, 1990.

[RFC1771] Rekhter, Y., and T. Li, "A Border Gateway
Protocol 4 (BGP-4)", RFC 1771, March 1995.

[RFC1958] Carpenter, B., "Architectural principles of the
Internet", Editor. RFC 1958, June 1996.

[RFC1997] Chandra, R., P. Traina, and T. Li, "BGP
Communities Attribute", RFC 1997, August, 1996.

[RFC2138] Rigney, C., et. al., "Remote Authentication Dial
In User Service (RADIUS)", RFC 2138, April, 1997.

[RFC2328] Moy, J., "OSPF Version 2", RFC 2328, April, 1998.

[RFC2453] Malkin, G., "RIP Version 2", RFC 2453, November,
1998.

[RFC2460] Deering, S. and R. Hinden, "Internet Protocol,
Version 6 (IPv6) Specification", RFC 2460,
December, 1998.

[RFC2547BIS] Rosen, E., et. al., "BGP/MPLS IP VPNs",
draft-ietf-l3vpn-rfc2547bis-00.txt. Work in
progress.

[RFC2858] Bates, T., et. al., "Multiprotocol Extensions
for BGP-4", RFC 2858, June 2000.

[RFC2918] Chen, E., "Route Refresh Capability for BGP-4",
RFC 2918, September 2000.

[RFC3036] Anderson, L., et. al., "LDP Specification", RFC
3036, January 2001.

[RFC3439] Bush, R. and D. Meyer, "Some Internet
Architectural Guidelines and Philosophy", RFC
3439, December, 2002.

[SAFI] http://www.iana.org/assignments/safi-namespace

[VLPS]

[VPLS] Kompella, K., et. al. "Virtual Private LAN
Service", draft-ietf-l2vpn-vpls-bgp-02.txt. draft-ietf-l2vpn-vpls-bgp-01.txt.
Work in progress.

[VPWS] Kompella, K. et.al. "Layer 2 VPNs Over Tunnels",
draft-kompella-ppvpn-l2vpn-04.txt. Work in
progress.

17.2.

16.2. Informative References

[IETFOL] https://www1.ietf.org/mailman/listinfo/routing-discussion

[RFC2119] Bradner, S., "Key words for use in RFCs to
Indicate Requirement Levels", RFC 2119, March,
1997.

[RFC2026] Bradner, S., "The Internet Standards Process --
Revision 3", RFC 2026/BCP 9, October, 1996.

[RFC2028] Hovey, R. and S. Bradner, "The Organizations
Involved in the IETF Standards Process", RFC
2028/BCP 11, October, 1996.

[RFC2434] Narten, T., and H. Alvestrand, "Guidelines for
Writing an IANA Considerations Section in RFCs",
RFC 2434/BCP 26, October 1998.

[RVBIB] http://www.routeviews.org/papers

[WILLINGER2002] Willinger, W., and J. Doyle, "Robustness and the
Internet: Design and evolution", 2002.

18.

17. Editor's Address

David Meyer
Email: dmm@1-4-5.net

19.

18. Full Copyright Statement

This document and translations of it may be copied and furnished to
others, and derivative works that comment on or otherwise explain it
or assist in its implementation may be prepared, copied, published
and distributed, in whole or in part, without restriction of any
kind, provided that the above copyright notice and this paragraph are
included on all such copies and derivative works. However, this
document itself may not be modified in any way, such as by removing
the copyright notice or references to the Internet Society or other
Internet organizations, except as needed for the purpose of
developing Internet standards in which case the procedures for
copyrights defined in the Internet Standards process must be
followed, or as required to translate it into languages other than
English.

The limited permissions granted above are perpetual and will not be
revoked by the Internet Society or its successors or assigns.

This document and the information contained herein is provided on an
"AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING
TASK FORCE DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING
BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION
HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF
MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.

Meyer, et. al.