Network Working Group Peter Ashwood-Smith (Nortel Networks)
Internet Draft Daniel Awduche (Movaz Networks)
Expiration date: August 2001 Ayan Banerjee (Calient Networks)
Debashis Basak (Accelight Networks)
Lou Berger (Movaz Networks)
Greg Bernstein (Ciena Corporation)
John Drake (Calient Networks)
Yanhe Fan (Axiowave Networks)
Don Fedyk (Nortel Networks)
Gert Grammel (Alcatel)
Kireeti Kompella (Juniper Networks)
Alan Kullberg (NetPlane Systems)
Jonathan P. Lang (Calient Networks)
Fong Liaw (Zaffire, Inc.)
Dimitri Papadimitriou (Alcatel)
Dimitrios Pendarakis (Tellium, Inc.)
Bala Rajagopalan (Tellium, Inc.)
Yakov Rekhter (Juniper Networks)
Debanjan Saha (Tellium, Inc.)
Hal Sandick (Nortel Networks)
Vishal Sharma (Jasmine Networks)
George Swallow (Cisco Systems)
Z. Bo Tang (Tellium, Inc.)
John Yu (Zaffire, Inc.)
Alex Zinin (Cisco Systems)
Eric Mannie (Ebone) - Editor
February 2001
Generalized Multi-Protocol Label Switching (GMPLS) Architecture
draft-many-gmpls-architecture-00.txt
Status of this Memo
This document is an Internet-Draft and is in full conformance with
all provisions of Section 10 of RFC2026 [1].
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
Many Internet-Draft August 2001 1
draft-many-gmpls-architecture-00.txt Feb 2001
The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html.
1. Abstract
Future data and transmission networks will consist of elements such
as routers, switches, DWDM systems, Add-Drop Multiplexors (ADMs),
photonic cross-connects (PXCs) or optical cross-connects (OXCs), etc
that will use Generalized MPLS (GMPLS) to dynamically provision
resources and to provide network survivability using protection and
restoration techniques.
This document describes the architecture of GMPLS. GMPLS extends
MPLS to encompass time-division (e.g. SDH/SONET, PDH, G.709),
wavelength (lambdas) and spatial switching (e.g. incoming port or
fiber to outgoing port or fiber). The main focus of GMPLS is on the
control plane of these various layers since each of them can use
totally different data or forwarding planes. The intention is to
cover both the signaling and the routing part of that control plane.
2. Conventions used in this document
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 [2].
3. Introduction
The architecture presented in this document covers the main building
blocks needed to build a consistent control plane for multiple
switching layers. It does not restrict the way that these layers
work together. Different models can be applied: e.g. overlay,
augmented or integrated. Moreover, each pair of contiguous layer may
work jointly in a different way. It results that a number of
combinations are possible, at the discretion of manufacturers and
operators.
This document generalizes the MPLS architecture [MPLS-ARCH], and in
some cases can differ slightly from that architecture since non
packet-based forwarding planes are now considered. It is not the
intention of this document to describe concepts already described in
the current MPLS architecture. The goal is to describe specific
concepts of GMPLS.
However, some of the concepts described hereafter are not described
in the current MPLS architecture and are applicable to both MPLS and
GMPLS, i.e. link bundling, unnumbered links and LSP hierarchy. Since
they raised from the GMPLS needs and since they are of paramount
importance for an operational GMPLS network, they will be introduced
here.
The following sections will first introduce GMPLS. Then the specific
GMPLS building blocks will be presented and we will explain how they
Many Internet-Draft August 2001 2
draft-many-gmpls-architecture-00.txt Feb 2001
can be combined together. Details about these separate building
blocks can be found in the corresponding documents.
3.1. Acronyms & abbreviations
ABR Area Border Router
AS Autonomous System
ASBR Autonomous System Boundary Router
BGP Border Gateway Protocol
CR-LDP Constraint-based Routing LDP
CSPF Constraint-based Shortest Path First
DWDM Dense Wavelength Division Multiplexing
FA Forwarding Adjacency
GMPLS Generalized Multi-Protocol Label Switching
IGP Interior Gateway Protocol
LDP Label Distribution Protocol
LMP Link Management Protocol
LSA Link State Advertisement
LSR Label Switching Router
LSP Label Switched Path
MIB Management Information Base
MPLS Multi-Protocol Label Switching
RSVP ReSource reserVation Protocol
SDH Synchronous Digital Hierarchy
STM(-N) Synchronous Transport Module (-N)
STS(-N) Synchronous Transport Signal-Level N (SONET)
TE Traffic Engineering
3.2. Multiple Types of Switching and Forwarding Hierarchies
Generalized MPLS differs from traditional MPLS in that it supports
multiple types of switching, i.e. the addition of support for TDM,
lambda, and fiber (port) switching. The support for the additional
types of switching has driven generalized MPLS to extend certain
base functions of traditional MPLS and, in some cases, to add
functionality. These changes and additions impact basic LSP
properties, how labels are requested and communicated, the
unidirectional nature of LSPs, how errors are propagated, and
information provided for synchronizing the ingress and egress LSRs.
The MPLS architecture [MPLS-ARCH] was defined to support the
forwarding of data based on a label. In this architecture, Label
Switching Routers (LSRs) were assumed to have a forwarding plane
that is capable of (a) recognizing either packet or cell boundaries,
and (b) being able to process either packet headers (for LSRs
capable of recognizing packet boundaries) or cell headers (for LSRs
capable of recognizing cell boundaries).
This original architecture is here extended to include LSRs whose
forwarding plane recognizes neither packet, nor cell boundaries, and
therefore, can't forward data based on the information carried in
either packet or cell headers. Specifically, such LSRs include
devices where the forwarding decision is based on time slots,
wavelengths, or physical ports. So, the new set of LSRs, or more
Many Internet-Draft August 2001 3
draft-many-gmpls-architecture-00.txt Feb 2001
precisely interfaces on these LSRs, can be subdivided into the
following classes:
1. Packet-Switch Capable (PSC) interfaces:
Interfaces that recognize packet/cell boundaries and can forward
data based on the content of the packet/cell header. Examples
include interfaces on routers that forward data based on the content
of the "shim" header, interfaces on ATM-LSRs that forward data based
on the ATM VPI/VCI.
2. Time-Division Multiplex Capable (TDM) interfaces:
Interfaces that forward data based on the data's time slot in a
repeating cycle. An example of such an interface is an interface on
a SDH/SONET Cross-Connect (XC), Terminal Multiplexer (TM) or Add-
Drop Multiplexer (ADM). Other examples are an interface implementing
G.709 (the digital wrapper), or a PDH interface.
3. Lambda Switch Capable (LSC) interfaces:
Interfaces that forward data based on the wavelength on which the
data is received. An example of such an interface is an interface
on an Optical Cross-Connect that can operate at the level of an
individual wavelength. Another example is an interface that can
operate at the level of a group of wavelengths, i.e. a waveband.
4. Fiber-Switch Capable (FSC) interfaces:
Interfaces that forward data based on a position of the data in the
real world physical spaces. An example of such an interface is an
interface on a photonic Cross-Connect that can operate at the level
of a single (or multiple) fibers.
A circuit can be established only between, or through, interfaces of
the same type. Depending on the particular technology being used for
each interface, different circuit names can be used, e.g. SDH
circuit, optical trail, light path etc. In the context of GMPLS, all
these circuits are referenced by a common name: Label Switched Path
(LSP).
The concept of nested LSP (LSP within LSP) already available in the
traditional MPLS allows here to build a forwarding hierarchy, i.e. a
hierarchy of LSPs. This hierarchy of LSPs can occur on the same
interface, or between different interfaces.
It can occur on the same interface if this interface is capable of
multiplexing several LSPs from the same technology (layer), e.g. a
lower order SDH/SONET LSP (VC-12) nested in a higher order SDH/SONET
LSP (VC-4). Several levels of signal (LSP) nesting are defined in
the SDH/SONET multiplexing hierarchy.
The nesting can also occur between interfaces. At the top of the
hierarchy are FSC interfaces, followed by LSC interfaces, followed
Many Internet-Draft August 2001 4
draft-many-gmpls-architecture-00.txt Feb 2001
by TDM interfaces, followed by PSC interfaces. This way, an LSP that
starts and ends on a PSC interface can be nested (together with
other LSPs) into an LSP that starts and ends on a TDM interface.
This LSP, in turn, can be nested (together with other LSPs) into an
LSP that starts and ends on an LSC interface, which in turn can be
nested (together with other LSPs) into an LSP that starts and ends
on a FSC interface.
3.3. Extension of the MPLS Control Plane
The establishment of LSPs that span only Packet Switch Capable (PSC)
interfaces is defined for the original MPLS and/or MPLS-TE control
planes. GMPLS extends these control planes to support each of the
four classes of interfaces (i.e. layers) defined in the previous
section.
Note that the GMPLS control plane supports as well an overlay model,
an augmented model or an integrated model. The benefits of using an
augmented or integrated model will have to be clarified and
evaluated in the future. In the mean time, GMPLS is very suitable
for controlling each layer completely independently. This elegant
approach will facilitate the future deployment of other models.
The GMPLS control plane is made of several building blocks that will
be described in more details in the following sections. These
building blocks are indeed well-known IETF signaling and routing
protocols that have been extended and/or modified. They use IPv4
and/or IPv6 addresses. Only one new specialized protocol was
required to support the operations of GMPLS, a signaling protocol
for link management [LMP].
GMPLS is indeed based on the Traffic Engineering (TE) extensions to
MPLS, a.k.a. MPLS-TE. This is because most of the technologies that
can be used below the PSC level require some traffic engineering.
The placement of LSPs at these levels needs in general to take
several constraints into consideration (such as bandwidth,
protection capability, etc) and to bypass the legacy Shortest-Path
First (SPF) algorithm. Note however that this is not mandatory and
that in some cases an SPF routing could be applied.
In order for such a constrained-based SPF routing of LSPs to happen,
the nodes performing LSP establishment need more information about
the links in the network than standard intra-domain routing
protocols provide. These TE attributes are distributed using the
transport mechanisms already available in IGPs and are taken into
consideration by the LSP routing algorithm. Optimization of the LSP
trajectories may also require some external simulations using
heuristics that serve as input for the actual path calculation and
LSP establishment process.
Extensions to traditional routing protocols and algorithms are
needed to uniformly encode and carry TE link information, and
explicit routes (e.g. source routes) are required in the signaling.
In addition, the signaling must now be capable of transporting the
Many Internet-Draft August 2001 5
draft-many-gmpls-architecture-00.txt Feb 2001
required circuit (LSP) parameters such as the bandwidth, the type of
signal, the desired protection, the position in a particular
multiplex, etc. Most of these extensions have already been defined
for PSC (IP) traffic engineering with MPLS. GMPLS mainly adds
additional extensions for TDM, LSC and FSC traffic engineering, by
staying as generic as possible. Only a very few elements are
technology specific.
Thus, GMPLS extends the two signaling protocols defined for MPLS-TE
signaling, i.e. RSVP-TE and CR-LDP. However, GMPLS does not specify
which one of these two signaling protocols must be used. It is the
role of manufacturers and operators to evaluate the two possible
solutions for their own interest.
Since GMPLS is based on RSVP-TE and CR-LDP, it mandates a
downstream-on-demand label allocation and distribution, with an
ingress initiated ordered control. A liberal label retention is
normally used, but a conservative label retention mode could be
used. There is no restriction on the label allocation strategy, it
can be request driven (obvious for circuit switching technologies),
traffic/data driven, or even topology driven. There is no
restriction neither on the route selection, explicit routing is
normally used (strict or loose) but an hop-by-hop routing could be
used as well.
GMPLS extends also two traditional intra-domain routing protocols
already extended for TE, i.e. OSPF-TE and IS-IS-TE. However, if
explicit routing is used, the routing algorithms used by these
protocols don't need to be standardized anymore since they are now
used to compute explicit routes only, and are thus not used anymore
for hop-by-hop routing. Extensions for inter-domain routing (e.g.
BGP) are for further study.
The use of technologies like DWDM (Dense Wavelength Division
Multiplexing) implies that we can now have a very large number of
parallel links between two directly adjacent nodes (hundreds of
wavelengths, or even thousands of wavelengths if multiple fibers are
used). Such a large number of links was not originally considered
for an IP or MPLS control plane. Some slight adaptations of that
control plane are thus required if we want to reuse it in the GMPLS
context.
For instance, the traditional IP routing model assumes the
establishment of a routing adjacency over each link connecting two
adjacent nodes. Having such a large number of adjacencies is not
scalable at all. Each node needs to maintain each of its adjacencies
one by one, and link state routing information must be flooded in
the topology for each link.
To solve this issue the concept of bundling was introduced.
Moreover, the manual configuration and control of these links, even
if they are unnumbered, becomes totally impractical. The Link
Management Protocol (LMP) was specified to solve these problems.
Many Internet-Draft August 2001 6
draft-many-gmpls-architecture-00.txt Feb 2001
LMP runs between data-plane adjacent nodes and is used for both link
provisioning and fault isolation. LMP was defined in the context of
GMPLS, but was specified independently of the GMPLS signaling
specification. It results that LMP can be reused in other contexts,
with non-GMPLS signaling protocols as well.
A unique feature of LMP is that it is able to isolate faults in both
opaque and transparent networks, independent of the encoding scheme
used for the data. LMP will be used to verify connectivity between
nodes; and isolate link, fiber, or channel failures within the
network.
The MPLS signaling and routing protocols require at least one bi-
directional control channel to communicate even if two adjacent node
are connected by unidirectional links. Several control channels can
be used. LMP can be used to establish, maintain and manage these
control channels.
GMPLS does not specify how these control channels must be
implemented, but GMPLS requires IP to transport the signaling and
routing protocols over them. Control channels can be either in-band
or out-of-band, and several solutions can be used to carry IP. Note
also that one type of LMP message is used in-band in the data plane
and may not be transported over IP, but this is a particular case,
needed to verify connectivity in the data plane.
3.4. Key Differences Between MPLS-TE and GMPLS
Some key differences between MPLS-TE and GMPLS are highlighted in
the following. Some of them are key advantages of GMPLS to control
non-PSC layers.
- In MPLS-TE, links traversed by an LSP can include an intermix of
links with heterogeneous label encoding (e.g. links between routers,
links between routers and ATM-LSRs, and links between ATM-LSRs.
GMPLS extends this by including links where the label is encoded as
a time slot, or a wavelength, or a position in the real world
physical space.
- In MPLS-TE, an LSP that carries IP has to start and end on a
router. GMPLS extends this by requiring an LSP to start and end on
similar type of LSRs.
- The type of a payload that can be carried in GMPLS by an LSP is
extended to allow such payloads as SONET/SDH, 1 or 10Gb Ethernet,
etc.
- For non-PSC interfaces, bandwidth allocation for an LSP can be
performed only in discrete units.
- It is expected to have (much) fewer labels on non-PSC links than
on PSC links.
Many Internet-Draft August 2001 7
draft-many-gmpls-architecture-00.txt Feb 2001
- The use of Forwarding Adjacencies (FA), provides a mechanism that
may improve bandwidth utilization, when bandwidth allocation can be
performed only in discrete units, as well as a mechanism to
aggregate forwarding state, thus allowing the number of required
labels to be reduced
- GMPLS allows for a label to be suggested by an upstream node to
reduce the setup latency. This suggestion may be overridden by a
downstream node but, in some cases, at the cost of higher LSP setup
time.
- GMPLS extends on the notion of restricting the range of labels
that may be selected by a downstream node. In GMPLS, an ingress or
other upstream node may restrict the labels that may be used by an
LSP along either a single hop or along the whole LSP path.
- While traditional TE-based (and even LDP-based) LSPs are
unidirectional, GMPLS supports the establishment of bi-directional
LSPs.
- GMPLS supports the termination of an LSP on a specific egress
port, i.e. the port selection at the destination side.
- GMPLS with RSVP-TE supports an RSVP specific mechanism for rapid
failure notification.
4. Routing and addressing model
GMPLS is based on the IP routing and addressing models. This assumes
that IPv4 and/or IPv6 addresses are used to identify interfaces and
that traditional (distributed) IP routing protocols are also reused.
Indeed, the discovery of the topology and the resource state of all
links in a routing domain is achieved via these routing protocols.
Since control and data planes are de-coupled in GMPLS, one cannot do
anymore the assumption that control-plane neighbors (i.e. IGP-learnt
neighbors) are data-plane neighbors, hence mechanisms like LMP are
needed to associate TE links with neighboring nodes.
IP addresses are not used only to identify interfaces of IP hosts
and routers, but more generally to identify any PSC and non-PSC
interfaces. Similarly IP routing protocols are not used only to find
routes for IP datagrams but also to find routes for non-PSC circuits
by using a CSPF algorithm instead of legacy SPF.
However, some additional mechanisms are needed to increase the
scalability of these models and to deal with specific traffic
engineering requirements of non-PSC layers. These mechanisms will be
introduced in the following.
Re-using existing IP routing protocols allows for non-PSC layers to
take advantages of all the valuable developments that toke place
since years for IP routing, in particular in the context of link-
state routing and policy routing.
Many Internet-Draft August 2001 8
draft-many-gmpls-architecture-00.txt Feb 2001
Each particular non-PSC layer can be seen as a set of Autonomous
Systems (ASs) interconnected in an arbitrary way. Similarly to the
traditional IP routing, each AS is managed by a single
administrative authority. For instance, an AS can be an SDH/SONET
network operated by a given carrier. The set of interconnected ASs
being an SDH/SONET Internetwork.
Exchange of routing information between ASs can be done via an
inter-domain routing protocol like BGP-4. There is obviously a huge
value of re-using well-known policy routing facilities provided by
BGP in a non-PSC context. Extensions for BGP traffic engineering in
the context of non-PSC layers are for further study.
Each AS can be subdivided in different routing domains, and each can
run a different intra-domain routing protocol. In turn, each
routing-domain can be divided in areas.
A routing domain is made of GMPLS nodes. These nodes can be either
edge nodes (i.e. hosts, ingress LSRs or egress LSRs), or internal
LSRs. An example of non-PSC host is an SDH/SONET Terminal
Multiplexer (TM). Another example, is an SDH/SONET interface card
within an IP router or ATM switch.
Note that traffic engineering in the intra-domain requires the use
of link-state routing protocols like OSPF or IS-IS.
GMPLS defines extensions to these protocols. These extensions are
needed to disseminate specific non-PSC static and dynamic
characteristics related to nodes and links. The current focus is on
intra-area traffic engineering. However, inter-area traffic
engineering is also under investigation.
4.1 Addressing of PSC and non-PSC layers
The fact that IPv4 and/or IPv6 addresses are used doesn't imply at
all that they should be allocated in the same addressing space than
public IPv4 and/or IPv6 addresses used for the Internet. Each layer
could have a different addressing authority responsible for address
allocation and re-using the full addressing space, completely
independently.
Private IP addresses can be used if they don't require to be
exchanged with any other operator, public IP addresses are otherwise
required. Of course, if an integrated model is used, two layers
could share the same addressing space.
Note that there is a benefit of using public IPv4 and/or IPv6
Internet addresses for non-PSC layers if an integrated model with
the IP layer is foreseen.
If we consider the scalability enhancements proposed in the next
section, the IPv4 (32 bits) and the IPv6 (128 bits) addressing
spaces are both more than sufficient to accommodate any non-PSC
Many Internet-Draft August 2001 9
draft-many-gmpls-architecture-00.txt Feb 2001
layer. We can reasonably expect to have much less non-PSC devices
(e.g. SDH/SONET nodes) than we have today IP hosts and routers.
Other kinds of addressing schemes (e.g. NSAP) are not considered
here since this would imply a modification of the already existing
signaling and routing protocols that uses IPv4 and/or IPv6
addresses. This would be incompatible to our objectives of re-using
existing IP protocols.
4.2 GMPLS scalability enhancements
Non-PSC layers introduce new constraints on the IP addressing and
routing models since several hundreds of parallel physical links
(e.g. wavelengths) can now connect two nodes. Most of the carriers
already have today several tenths of wavelengths per fiber between
two nodes. New generation of DWDM systems will allow several
hundreds of wavelengths.
It becomes rather impractical to associate an IP address to each end
of each physical link, to represent each link as a separate routing
adjacency, and to advertise link states for each of these links. For
that purpose, GMPLS enhances the MPLS routing and addressing models
to increase their scalability.
Two optional mechanisms can be used to increase the scalability of
the addressing and the routing: unnumbered links and link bundling.
These two mechanisms can also be combined. They require extensions
to signaling (RSVP-TE and CR-LDP) and routing (OSPF-TE and IS-IS-TE)
protocols.
4.3 Extensions to IP TE routing protocols
Traditionally, a TE link is advertised as an adjunct to a "regular"
OSPF or IS-IS link, i.e., an adjacency is brought up on the link,
and when the link is up, both the regular IGP properties of the link
(basically, the SPF metric) and the TE properties of the link are
then advertised.
However, GMPLS challenges this notion in three ways:
- first, links that are non-PSC may yet have TE properties; however,
an OSPF adjacency cannot be brought up directly on such links.
- second, an LSP can be advertised as a point-to-point TE link in
the routing protocol, i.e. as a Forwarding Adjacency (FA); thus, an
advertised TE link need no longer be between two OSPF neighbors.
Forwarding Adjacencies (FA) are further described in a separate
section.
- third, a number of links may be advertised as a single TE link
(e.g. for improved scalability), so again, there is no longer a one-
to-one association of a regular adjacency and a TE link.
Many Internet-Draft August 2001 10
draft-many-gmpls-architecture-00.txt Feb 2001
Thus we have a more general notion of a TE link. A TE link is a
logical link that has TE properties, some of which may be configured
on the advertising LSR, others which may be obtained from other LSRs
by means of some protocol, and yet others which may be deduced from
the component(s) of the TE link.
An important TE property of a TE link is related to the bandwidth
accounting for that link. GMPLS will define different accounting
rules for different non-PSC layers. Generic bandwidth attributes are
however defined by the TE routing extensions and by GMPLS, such as
the unreserved bandwidth, the maximum reservable bandwidth, the
maximum LSP bandwidth.
It is expected in a dynamic environment to have frequent changes of
bandwidth accounting information. A flexible policy for triggering
link state updates based on bandwidth thresholds and link dampening
mechanism can be implemented.
TE properties associated with a link should also capture protection
and restoration related characteristics. For instance, shared
protection can be elegantly combined with bundling. Protection and
restoration are mainly generic mechanisms also applicable to MPLS.
It is expected that they will first be developed for MPLS and later
on generalized to GMPLS.
A TE link between a pair of LSRs doesn't imply the existence of an
IGP adjacency between these LSRs. A TE link must also have some
means by which the advertising LSR can know of its liveness (e.g. by
using LMP hellos). When an LSR knows that a TE link is up, and can
determine the TE link's TE properties, the LSR may then advertise
that link to its GMPLS enhanced OSPF or IS-IS neighbors using the TE
objects/TLVs. We call the interfaces over which GMPLS enhanced OSPF
or ISIS adjacencies are established "control channels".
5. Unnumbered links
Unnumbered links (or interfaces) are links(or interfaces) that do
not have IP addresses. Using such links involves two capabilities:
(a) the ability to carry (TE) information about unnumbered links in
IGP TE extensions (ISIS or OSPF), and (b) the ability to specify
unnumbered links in MPLS TE signaling.
The former is covered in ISIS-TE and OSPF-TE. The later requires
extensions to RSVP-TE and CR-LDP since MPLS-TE signaling doesn't
provide support for unnumbered links. GMPLS defines simple
extensions to indicate an unnumbered link in the Explicit Route and
Record Route Objects/TLVs of these protocols, using a new Interface
ID object/TLV.
Since unnumbered links are not identified by an IP address, then for
the purpose of MPLS TE each end need some other identifier, local to
the LSR to which the link belongs. Note that links are directed,
i.e., a link l is from some LSR A to some other LSR B. LSR A chooses
the interface identifier for link l, we call this the "outgoing
Many Internet-Draft August 2001 11
draft-many-gmpls-architecture-00.txt Feb 2001
interface identifier from LSR A's point of view". If there is a
reverse link from LSR B to LSR A, B chooses the outgoing interface
identifier for the reverse link. There is no a priori relationship
between the two interface identifiers. Both ends must also agree on
each of these identifiers.
5.1 Unnumbered Forwarding Adjacencies
If an LSR that originates an LSP advertises this LSP as an
unnumbered FA in IS-IS or OSPF, the LSR must allocate an Interface
ID to that FA. If the LSP is bi-directional, the tail-end LSR
advertises the reverse LSP as an unnumbered FA, the tail-end LSR
must allocate an Interface ID to the reverse FA.
Signaling has been enhanced to carry the Interface IDs. When an LSP
is created which will be advertised as an FA, the head-end LSR
assigns an Interface ID and includes it in the signaling request.
The tail-end LSR responds by assigning and including an Interface ID
in the signaling response.
6. Link bundling
When a pair of LSRs is connected by multiple links, it is possible
to advertise several (or all) of these links as a single link into
OSPF and/or IS-IS. This process is called link bundling, or just
bundling. The resulting logical link is called a bundled link as its
physical links are called component links.
The purpose of link bundling is to improve routing scalability by
reducing the amount of information that has to be handled by OSPF
and/or IS-IS. This reduction is accomplished by performing
information aggregation/abstraction. As with any other information
aggregation/abstraction, this results in losing some of the
information. To limit the amount of losses one need to restrict the
type of the information that can be aggregated/abstracted.
6.1 Restrictions on bundling
The following restrictions are required for GMPLS. All component
links in a bundle must begin and end on the same pair of LSRs, and
share some common characteristics: they must have the same type
(e.g. point-to-point), the same TE metric, the same set of resource
classes, and the same multiplexing capabilities. An FA may be a
component link; in fact, a bundle can consist of a mix of point-to-
point links and FAs.
6.2 Routing considerations for bundling
A bundled link is just another kind of TE link such as those defined
by OSPF-TE or IS-IS-TE. The liveness of the bundled link is
determined by the liveness of each of the component links within the
bundled link. The liveness of a component link can be determined by
any of several means: IS-IS or OSPF hellos over the component link,
Many Internet-Draft August 2001 12
draft-many-gmpls-architecture-00.txt Feb 2001
or RSVP Hello, or LMP hellos, or from layer 1 or layer 2
indications.
Once a bundled link is determined to be alive, it can be advertised
as a TE link and the TE information can be flooded. If IS-IS/OSPF
hellos are run over the component links, IS-IS/OSPF flooding can be
restricted to just one of the component links.
Note that advertising a (bundled) TE link between a pair of LSRs
doesn't imply that there is an IGP adjacency between these LSRs that
is associated with just that link. In fact, in certain cases a TE
link between a pair of LSRs could be advertised even if there is no
IGP adjacency at all between the LSR (e.g. when the TE link is an
FA).
Bandwidth accounting must be clearly defined since an abstraction is
done. Bandwidth information is an important part of a bundle
advertisement. Some attributes can be sums of component
characteristics such as the unreserved bandwidth and the maximum
reservable bandwidth. A GMPLS node with bundled links must apply
admission control on a per-component link basis.
6.3 Signaling considerations
Typically, an LSP's explicit route (contained in an ERO) will choose
the bundled link to be used for the LSP, but not the component
link(s), since information about the bundled link is flooded, but
information about the component links is kept local to the LSR.
The choice of the component link to use is always made by an
upstream node. If the LSP is bidirectional, the upstream node
chooses a component link in each direction.
Three mechanisms for indicating this choice to the downstream node
are possible.
- Mechanism 1: Implicit Indication
This mechanism requires that each component link has a dedicated
signaling channel. The upstream node tells the receiver which
component link to use by sending the message over the chosen
component link's dedicated signaling channel.
- Mechanism 2: Explicit Indication by IP Address
This mechanism requires that each component link has a unique remote
IP address. The upstream node can either send messages addressed to
the remote IP address for the component link or encapsulate messages
in an IP header whose destination address is the remote IP address.
This mechanism does not require each component link to have its own
control channel. In fact, it doesn't even require the whole
(bundled) link to have its own control channel.
- Mechanism 3: Explicit Indication by Component Interface ID
Many Internet-Draft August 2001 13
draft-many-gmpls-architecture-00.txt Feb 2001
With this mechanism, each component link in unnumbered and is
assigned a unique Interface Identifier. These identifiers are
exchanged by the two LSRs at each end of the bundled link. The
choice of a component link is indicated by an upstream node by
including the corresponding identifier in signaling messages.
Discovering Interface Identifiers at bootstrap may be accomplished
by configuration, by means of a protocol such as LMP (preferred
solution), or by means of RSVP/CR-LDP (especially in the case where
a component link is a Forwarding Adjacency). New objects are needed
to indicate Interface Identifiers in signaling, GMPLS defines one
Upstream Interface ID object/TLV and one Downstream Interface ID
object/TLV.
6.4 Unnumbered Bundled Link
A bundled link may itself be numbered or unnumbered independent of
whether the component links are numbered or not. This affects how
the bundled link is advertised in IS-IS/OSPF, and the format of LSP
EROs that traverse the bundled link. Furthermore, unnumbered
Interface Identifiers for all unnumbered outgoing links of a given
LSR (whether component links, Forwarding Adjacencies or bundled
links) MUST be unique in the context of that LSR.
7. UNI and NNI
The interface between an edge GMPLS node and a GMPLS LSR on the
network side may be referred to as a User to Network Interface
(UNI), while the interface between two network side LSRs may be
referred to as a Network to Network Interface (NNI).
GMPLS does not specify separately a UNI and an NNI. Edge nodes are
connected to LSRs on the network side, and these LSRs are in turn
connected between them. Of course, the behavior of an edge node is
not exactly the same as the behavior of an LSR on the network side.
Note also, that an edge node may run a routing protocol, however it
is expected that in most of the cases it will not (see also section
7.2 and the section about signaling with an explicit route).
Conceptually, a difference between UNI and NNI make sense either if
both interface uses completely different protocols, or if they use
the same protocols but with some outstanding differences. In the
first case, separate protocols are often defined successively, with
more or less success.
The GMPLS approach consisted in building a consistent model from day
one, considering both the UNI and NNI interfaces at the same time.
For that purpose a very few specific UNI particularities have been
ignored in a first time. GMPLS is being enhanced to support such
particularities at the UNI by some other standardization bodies,
like the OIF.
7.1 OIF UNI versus GMPLS
Many Internet-Draft August 2001 14
draft-many-gmpls-architecture-00.txt Feb 2001
The current OIF UNI specification [OIF-UNI] defines an interface
between a client SDH/SONET equipment and an SDH/SONET network, each
belonging to a distinct administrative authority. The OIF UNI
defines additional mechanisms on the top of GMPLS for the UNI
For instance, the OIF service discovery procedure is a precursor to
obtaining UNI services. Service discovery allows a client to
determine the static parameters of the interconnection with the
network, including the UNI signaling protocols, the transparency
levels as well as the protection level supported by the network.
Moreover, the only additional mechanism covered by the OIF UNI is
the address allocation process. The corresponding mechanism is
tightly related to the link bundle mechanism as described in LMP
using LinkSummary (and include an address allocation request) and
LinkSummaryAck (and include an address allocation response)
messages.
Since the current OIF UNI interface does not cover photonic
networks, G.709 Digital Wrapper, etc, it is a sub-set of the GMPLS
Architecture.
7.2 Routing at the UNI
This section discusses the selection of an explicit route by an edge
node. The selection of the first LSR by an edge node connected to
multiple LSRs is part of that problem.
An edge node (host or LSR) can participate more or less deeply in
the GMPLS routing. Four different routing models can be supported at
the UNI: configuration based, partial peering, silent listening and
full peering.
- Configuration based: this routing model requires the manual or
automatic configuration of an edge node with a list of neighbor LSRs
sorted by preference order. Automatic configuration can be achieved
using DHCP for instance. No routing information is exchanged at the
UNI, except maybe the ordered list of LSRs. The only routing
information used by the edge node is that list. The edge node sends
by default an LSP request to the preferred LSR. ICMP redirects could
be send by this LSR to redirect some LSP requests to another LSR
connected to the edge node. GMPLS does not preclude that model.
- Partial peering: limited routing information (mainly reachability)
can be exchanged across the UNI using some extensions in the
signaling plane. The reachability information exchanged at the UNI
may be used to initiate edge node specific routing decision over the
network. GMPLS does not have any capability to support this model
today.
- Silent listening: the edge node can silently listen to routing
protocols and take routing decisions based on the information
obtained. An edge node receives the full routing information,
including traffic engineering extensions. One LSR should forward
Many Internet-Draft August 2001 15
draft-many-gmpls-architecture-00.txt Feb 2001
transparently all routing pdus to the edge node. An edge node can
now compute a complete explicit route taking into consideration all
the end-to-end routing information. GMPLS does not preclude this
model.
- Full peering: In addition to silent listening, the edge node
participates within the routing, establish adjacencies with its
neighbors and advertises LSAs. This is useful only if there are
benefits for edge nodes to advertise themselves traffic engineering
information. GMPLS does not preclude this model.
8. Link Management
In the context of GMPLS, a pair of nodes (e.g., a photonic switch)
may be connected by tenths of fibers, and each fiber may be used to
transmit hundreds of wavelengths if DWDM is used. Furthermore,
multiple fibers and/or multiple wavelengths may be combined into one
or more bundled links as explained previously.
Dealing with hundreds or thousands of individual or bundled links
between two nodes requires the help of some signaling tools. In
addition, at least one control channel must be established and
maintained between a node pair, possibly, using some of these links.
Link management is a collection of useful functionality between
adjacent nodes that provide different local services such as control
channel management, link connectivity verification, link property
correlation, and fault isolation. A Link Management Protocol (LMP)
has been defined to fulfill these operations. LMP was initiated in
the context of GMPLS but is indeed a generic toolbox that can be
also used in other contexts.
Control channel management and link connectivity verification are
mandatory mechanisms of LMP. Link property correlation and fault
isolation are optional.
8.1 Control channel
Control plane communications between neighboring nodes need a bi-
directional control channel. The control channel can be used to
exchange MPLS control-plane information such as signaling, routing
and management information.
In GMPLS, the control channel(s) between two adjacent nodes is no
longer required to use the same physical medium as the data-bearing
links between those nodes. For example, a control channel could use
a separate wavelength or fiber, an Ethernet link, or an IP tunnel
through a separate management network. A consequence of allowing the
control channel(s) between two nodes to be physically diverse from
the associated data-bearing links is that the health of a control
channel does not necessarily correlate to the health of the data-
bearing links, and vice-versa. Therefore, new mechanisms must be
developed to manage links, both in terms of link provisioning and
fault isolation.
Many Internet-Draft August 2001 16
draft-many-gmpls-architecture-00.txt Feb 2001
It is essential that a control channel is always available, and in
the event of a control channel failure, an alternate (or backup)
control channel must be made available to reestablish communication
with the neighboring node.
If a primary control channel cannot be established, then an
alternate control channel should be tried. Of course, alternate
control channels should be pre-configured, however, coordinating the
switchover of the control channel to an alternate channel is still
an important issue.
Specifically, if the control channel fails but the node is still
operational (i.e., the data-bearing links are still passing user
data), then both the local and remote nodes should switch to an
alternate control channel. If the bi-directional control channel is
implemented using two separate unidirectional channels, and only one
direction of the control channel has failed, both the local and
remote nodes need to understand that the control channel has failed
so that they can coordinate a switchover. LMP provides a graceful
switchover from one control channel to the other.
8.2 Control channel management
Once a control channel is configured between two neighboring nodes,
a Hello protocol will be used to establish and maintain connectivity
between the nodes and to detect failures. The Hello protocol of LMP
is intended to be a lightweight keep-alive mechanism that will react
to control channel failures rapidly so that IGP Hellos are not lost
and the associated link-state adjacencies are not removed
unnecessarily.
The Hello protocol consists of two phases: a negotiation phase and a
keep-alive phase. The negotiation phase allows negotiation of some
basic Hello protocol parameters, like the Hello frequency. The keep-
alive phase consists of a fast lightweight Hello message exchange.
The failure of a control channel can also be detected by lower
layers (e.g., SONET/SDH) since control channels are electrically
terminated at each node.
8.3 Control channel interfaces
LMP functions to maintain logical control channels between a pair of
nodes via control channel interfaces. Each control channel interface
hides a set of control channels and which of these is actually used
to transport the messages and how this is achieved. This isolate
signaling, routing and management from the actual control channel
management.
LMP does not specify how control channels are implemented, however
it states that messages transported over a control channel must be
IP encoded. Furthermore, since the messages are IP encoded, the link
level encoding is not part of LMP.
Many Internet-Draft August 2001 17
draft-many-gmpls-architecture-00.txt Feb 2001
LMP associates (possibly multiple) link bundles with a control
channel. Multiple control channels may then be configured and
associated with a control channel interface. One control channel is
actually used while the others are backup control channels sorted by
preference order. The control channel interface is announced into
the IGP domain so that messages can be routed to that interface. The
associations between the control channels and the control channel
interface are purely a local matter.
The control channel of a link bundle can be either explicitly
configured or automatically selected, however, GMPLS currently
assume that the control channel is explicitly configured. Once a
link bundle is associated with a control channel, it follows the
failover of that control channel. The association of the control
channel to the control channel interface is configured or
automatically bootstrapped and is a local issue.
Between any two adjacent nodes (from the perspective of link
bundles) there may be multiple active control channel interfaces,
and these control channel interfaces are used for LMP, routing, and
signaling messages. For purposes of flooding routing messages, LMP
messages, and signaling messages, any of the active control channel
interfaces may be used.
8.4 Link property correlation
A link property exchange mechanism allows to dynamically change some
link characteristics. It allows for instance to add data-bearing
links to a link bundle, change a link's protection mechanism, change
port identifiers, or change component identifiers in a bundle. This
mechanism is supported by an exchange of link summary messages.
8.5 Link connectivity verification
Link connectivity verification is an optional procedure that may be
used to verify the physical connectivity of data-bearing links (e.g.
component links of a bundle)as well as to exchange the link
identifiers that will be further used in the RSVP-TE and CR-LDP
signaling.
The use of this procedure is negotiated as part of the configuration
exchange that take place during the negotiation phase of the Hello
protocol. If enabled, the procedure is done initially when a link
bundle is first established, and subsequently, on a periodic basis
for all free component links of a link bundle.
Ping-type Test messages are exchanged over each of the data-bearing
links specified in the bundled link. It should be noted that all LMP
messages except for the Test message are exchanged over the control
channel and that Hello messages continue to be exchanged over the
control channel during the data-bearing link verification process.
The Test message is sent over the data-bearing link that is being
verified. Data-bearing links are tested in the transmit direction as
Many Internet-Draft August 2001 18
draft-many-gmpls-architecture-00.txt Feb 2001
they are uni-directional, and as such, it may be possible for both
nodes to exchange the Test messages simultaneously.
Before exchanging these test messages, the node that initiates the
verification indicates to the adjacent node that it will begin
sending test messages across the data-bearing links of a particular
bundled link. It indicates also the number of data-bearing links
that are to be verified; the interval at which the test messages
will be sent; the encoding scheme, the transport mechanism that are
supported, and data rate for Test messages; and, in the case where
the data-bearing links correspond to fibers, the wavelength over
which the Test messages will be transmitted. The transport mechanism
is negotiated between the two nodes. Furthermore, the local and
remote bundle identifiers are transmitted at this time to perform
the data-bearing link association with the bundle identifiers.
A unique characteristic of photonic switches (all-optical) is that
the data being transmitted over a data-bearing link is not
terminated at the switch, but instead passes through transparently.
This characteristic of PXCs poses a challenge for validating the
connectivity of the data-bearing links.
Therefore, to ensure proper verification of data-bearing link
connectivity in that case, we require that until the links are
allocated, it must be possible to terminate them locally. There is
no requirement that all data-bearing links be terminated
simultaneously, but at a minimum, the data-bearing links must be
able to be terminated one at a time. Furthermore, we assume that the
nodal architecture is designed so that messages can be sent and
received over any data-bearing link. Note that this requirement is
trivial for a digital switch since each data-bearing link is
received electronically before being forwarded to the next switch.
This is an additional requirement for photonic switches.
8.6 Fault localization
Fault localization or isolation is an important requirement from the
operational point of view. When a failure occurs an operator needs
to know where exactly it happened. It can also be used to support
some specific local protection/restoration mechanisms. Logically,
fault localization can occur only after a fault is detected.
Fault detection must be handled at the layer closest to the failure;
for optical networks, this is the physical (optical) layer. One
measure of fault detection at the physical layer is simply detecting
loss of light (LOL). Other techniques for monitoring optical signals
are still being developed and are for further study. However, it
should be clear that the mechanism used to locate the failure is
independent of the mechanism used to detect the failure, but simply
relies on the fact that a failure is detected.
In new technologies such as transparent photonic switching currently
no method is defined to locate a fault, and the mechanism by which
Many Internet-Draft August 2001 19
draft-many-gmpls-architecture-00.txt Feb 2001
the fault information is propagated must be sent ôout of bandö (via
the control plane).
Fault localization is an optional LMP procedure that is used to
rapidly locate link failures. The use of this procedure is also
negotiated as part of the configuration exchange that take place
during the negotiation phase of the Hello protocol. As before, we
assume each link has a bi-directional control channel that is always
available for inter-node communication and that the control channel
spans a single hop between two neighboring nodes.
The mechanism used to rapidly isolate link failures is designed to
work for unidirectional LSPs, and can be easily extended to work for
bi-directional LSPs.
If data-bearing links fail between two photonic switches, the power
monitoring system in all of the downstream nodes will detect LOL and
indicate a failure. To correlate multiple failures between a pair of
nodes, a monitoring window can be used in each node to determine if
a single data-bearing link has failed or if multiple data-bearing
links have failed. As part of the fault localization, a downstream
node that detects data-bearing link failures will send a channel
fail message to its upstream neighbor (bundling together the
notification of all of the failed data-bearing links).
An upstream node that receives the channel fail message will
correlate the failure to see if there is a failure on the
corresponding input and output ports for the LSP()s using this/these
link(s). If there is also a failure on the input port(s) of the
upstream node, the node will return a message to the downstream node
(bundling together the notification of all the data-bearing links),
indicating that it too has detected a failure. If, however, the
fault is clear in the upstream node (e.g., there is no LOL on the
corresponding input channels), then the upstream node will have
localized the failure and will return a specific message to the
downstream node. Once the failure has been localized, the signaling
protocols can be used to initiate span or path
protection/restoration procedures.
9. Generalized Signaling
The GMPLS signaling extends certain base functions of the RSVP-TE
and CR-LDP signaling and, in some cases, add functionality. These
changes and additions impact basic LSP properties, how labels are
requested and communicated, the unidirectional nature of LSPs, how
errors are propagated, and information provided for synchronizing
the ingress and egress.
The GMPLS signaling specification is available in three parts:
1. A signaling functional description [GMPLS-SIG].
2. RSVP-TE extensions [GMPLS-RSVP-TE].
3. CR-LDP extensions [GMPLS-CR-LDP].
Many Internet-Draft August 2001 20
draft-many-gmpls-architecture-00.txt Feb 2001
The following MPLS profile applies to GMPLS:
- Downstream-on-demand label allocation and distribution.
- Ingress initiated ordered control.
- Liberal (typical), or conservative (could) label retention
mode.
- Request, traffic/data, or topology driven label allocation
strategy.
- Explicit routing (typical), or hop-by-hop routing (could).
The GMPLS signaling defines the following new building blocks on the
top of MPLS-TE:
1. A new label request format to encompass non-PSC
characteristics.
2. Labels for non-PSC interfaces, generically known as
Generalized Label.
3. Waveband switching support.
4. Label suggestion by the upstream for optimization purposes
(e.g. latency).
5. Label restriction by the upstream to support some optical
constraints.
6. Bi-directional LSP establishment with contention
resolution.
7. Rapid failure notification to ingress node.
8. Explicit routing with explicit label control for a fine
degree of control.
These building blocks will be described in mode details in the
following. A complete specification can be found in the
corresponding documents.
Note that GMPLS is highly generic and optional. Only building blocks
1 and 2 are mandatory, and only within the specific format that is
needed. Typically building blocks 6 and 8 should be implemented.
Building blocks 3, 4, 5 and 7 are optional.
A typical SDH/SONET switching network would implement building
blocks: 1 (but the SDH/SONET format), 2 (the SDH/SONET label), 6 and
8. It could implement another format of label in case of link
bundling. Building block 7 is optional since the
protection/restoration can be achieved using SDH/SONET overhead
bytes.
A typical wavelength switching network would implement building
blocks: 1 (but the wavelength label), 2 (the generic format), 4, 5,
6, 7 and 8. Building block 3 is only needed in the particular case
of waveband switching.
A typical fiber switching network would implement building blocks: 1
(but the port label), 2 (the generic format), 6, 7 and 8.
A typical MPLS-IP network would not implement any of these building
blocks, since the absence of building block 1 would indicate regular
Many Internet-Draft August 2001 21
draft-many-gmpls-architecture-00.txt Feb 2001
MPLS-IP. Note however that building block 1 can be used to signal
MPLS-IP as well. In that case, the MPLS-IP network can benefit from
the link protection type (not available in CR-LDP, some very basic
form being available in RSVP-TE). Building block 2 is here a regular
MPLS label and no new label format is required.
GMPLS does not specify any profile for RSVP-TE and CR-LDP
implementations that have to support GMPLS - except for what is
directly related to GMPLS procedures. It is to the manufacturer to
decide which are the optional elements and procedures of RSVP-TE and
CR-LDP that need to be implemented. Some optional MPLS-TE elements
can be useful for non-PSC layers, for instance the setup and holding
priorities that are inherited from MPLS-TE.
9.1. Overview: How to Request an LSP
A non-PSC LSP is established by sending a PATH/Label Request message
downstream to the destination. This message contains a Generalized
Label Request with the type of LSP (i.e. the layer concerned), its
payload type and the requested local protection per link. An
Explicit Route (ERO) is also normally added to the message, but this
can be added and/or completed by the first/default LSR.
The requested bandwidth is encoded in the RSVP-TE SENDER_TSPEC and
FLOWSPEC objects, or in the CR-LDP Traffic Parameters TLV. The end-
to-end protection type is for further study. In case of SDH/SONET
concatenation, the requested bandwidth is the total bandwidth and a
field in the Generalized Label Request allows to know the number of
components.
Specific parameters for a given technology are given in the
Generalized Label Request, such as the type of concatenation and/or
transparency for a SDH/SONET LSP.
If the LSP is a bi-directional LSP, an Upstream Label is also
specified in the Path/Label request message. This label will be the
one to use in the downstream to upstream direction.
Additionally, a Suggested Label, a Label Set and a Waveband Label
can also be included in the message. Other operations are defined in
MPLS-TE.
The downstream node will send back a Resv/Label Mapping message
including one Generalized Label object/TLV that can contain several
Generalized Labels. For instance, if a concatenated SDH/SONET signal
is requested, several labels can be returned.
In case of SDH/SONET virtual concatenation, a list of labels is
returned. Each label identifying one element of the virtual
concatenated signal. This limits virtual concatenation to remain
within a single (component) link.
Many Internet-Draft August 2001 22
draft-many-gmpls-architecture-00.txt Feb 2001
In case of any type of SDH/SONET contiguous concatenation, only one
label is returned. That label is the lowest signal of the contiguous
concatenated signal (given an order specified in [GMPLS-SIG].
In case of SDH/SONET bundling, i.e. co-routing of circuits of the
same type but without concatenation, the explicit list of all
signals that take part in the bundling is returned.
9.2. Generalized Label Request
The Generalized Label Request is a new object/TLV to be added in an
RSVP-TE Path message instead of the regular Label Request, or in a
CR-LDP Request message in addition to the already existing TLVs.
Only one label request can be used per message, so a single LSP can
be requested at a time per signaling message.
The Generalized Label Request gives some major characteristics
(parameters) required to support the LSP being requested, such as
the LSP encoding type, the LSP payload type, the desired link
protection.
GMPLS defines a generic Generalized Label Request, and in addition
it can define specialized Generalized Label Requests, if and only if
there are specific characteristics that cannot be signaled by the
generic request, i.e. specific characteristics.
Currently, only one specific Generalized Label Request is defined,
for SDH/SONET. The SDH/SONET Generalized Label Request indicates the
same generic characteristics as the generic request but includes in
addition the requested SDH/SONET concatenation and transparency (if
needed).
Note that it is expected than a specific Generalized Label Request
will be defined in the future for photonic (all optical) switching.
The characteristics described hereafter are generic to all
technologies:
- The LSP encoding type.
- The LSP payload type.
- The link protection type.
The LSP encoding type indicates the type of technology (e.g.
Ethernet, SDH, SONET, fiber, etc) to which this requested LSP
corresponds. It represents the nature of the LSP, and not the nature
of the links that the LSP traverses. A link may support a set of
encoding formats, where support means that a link is able to carry
and switch a signal of one or more of these encoding formats
depending on the resource availability and capacity of the link.
For example, consider an LSP signaled with "photonic" encoding. It
is expected that such an LSP would be supported with no electrical
conversion and no knowledge of the modulation and speed by the
Many Internet-Draft August 2001 23
draft-many-gmpls-architecture-00.txt Feb 2001
transit nodes. Some other formats (electrical) require other
knowledge such as the bandwidth.
The LSP payload type identifies the payload carried by an LSP, i.e.
the client layer of that LSP. This must be interpreted according to
the technology encoding type of the LSP and is used by the nodes at
the endpoints of the LSP to know to which client layer a request is
destined.
The link protection type indicates the desired local link protection
for each link of an LSP. If a particular protection type, i.e., 1+1,
or 1:N, is requested, then a connection request is processed only if
the desired protection type can be honored. Note that GMPLS
advertises the protection capabilities of a link in the routing
protocols. Path computation algorithms may take this information
into account when computing paths for setting up LSPs.
9.3. Generalized Label
The Generalized Label extends the traditional MPLS label by allowing
the representation of not only labels which identify and travel in-
band with associated data packets, but also (virtual) labels which
identify time-slots, wavelengths, or space division multiplexed
positions.
For example, the Generalized Label may identify (a) a single fiber
in a bundle, (b) a single waveband within fiber, (c) a single
wavelength within a waveband (or fiber), or (d) a time-slot within a
wavelength (or fiber). It may also be a generic MPLS label, a Frame
Relay label, or an ATM label (VCI/VPI). The format of a label can be
as simple as an integer value such as a wavelength label or can be
more elaborated such as an SDH/SONET label.
SDH and SONET define each a multiplexing structure. These
multiplexing structures will be used as naming trees to create
unique labels. Such a label will identify the type of a particular
signal (time-slot) and its exact position in a multiplexing
structure (both are related). Since the SONET multiplexing structure
may be seen as a subset of the SDH multiplexing structure, the same
format of label is used for SDH and SONET.
Since the nodes sending and receiving the Generalized Label know
what kinds of link they are using, the Generalized Label does not
identify its type, instead the nodes are expected to know from the
context what type of label to expect.
A Generalized Label only carries a single level of label, i.e., it
is non-hierarchical. When nested LSPs are used, each LSP must be
established separately and has its own label at each local interface
between two nodes at its level.
9.4. Waveband Switching
Many Internet-Draft August 2001 24
draft-many-gmpls-architecture-00.txt Feb 2001
A special case of wavelength switching is waveband switching. A
waveband represents a set of contiguous wavelengths which can be
switched together to a new waveband. For optimization reasons it may
be desirable for an photonic cross-connect to optically switch
multiple wavelengths as a unit. This may reduce the distortion on
the individual wavelengths and may allow tighter separation of the
individual wavelengths. A Waveband label is defined to support this
special case.
Waveband switching naturally introduces another level of label
hierarchy and as such the waveband is treated the same way all other
upper layer labels are treated. As far as the MPLS protocols are
concerned there is little difference between a waveband label and a
wavelength label except that semantically the waveband can be
subdivided into wavelengths whereas the wavelength can only be
subdivided into time or statistically multiplexed labels.
9.5. Label Suggestion by the Upstream
GMPLS allows for a label to be suggested by an upstream node. This
suggestion may be overridden by a downstream node but, in some
cases, at the cost of higher LSP setup time. The suggested label is
valuable when establishing LSPs through certain kinds of optical
equipment where there may be a lengthy (in electrical terms) delay
in configuring the switching fabric. For example micro mirrors may
have to be elevated or moved, and this physical motion and
subsequent damping takes time. If the labels and hence switching
fabric are configured in the reverse direction (the norm) the
MAPPING/Resv message may need to be delayed by 10's of milliseconds
per hop in order to establish a usable forwarding path. It can also
be important for restoration purposes where alternate LSPs may need
to be rapidly established as a result of network failures.
9.6. Label Restriction by the Upstream
An upstream node can optionally restrict (limit) the choice of label
of a downstream node to a set of acceptable labels. This restriction
is done by giving a list of inclusive (acceptable) or exclusive
(unacceptable) labels in a Label Set. If not applied, all labels
from the valid label range may be used. There are four cases where a
label restriction is useful in the "optical" domain.
The first case is where the end equipment is only capable of
transmitting and receiving on a small specific set of
wavelengths/bands.
The second case is where there is a sequence of interfaces which
cannot support wavelength conversion and require the same wavelength
be used end-to-end over a sequence of hops, or even an entire path.
The third case is where it is desirable to limit the amount of
wavelength conversion being performed to reduce the distortion on
the optical signals.
Many Internet-Draft August 2001 25
draft-many-gmpls-architecture-00.txt Feb 2001
The last case is where two ends of a link support different sets of
wavelengths.
The receiver of a Label Set must restrict its choice of labels to
one which is in the Label Set. A Label Set may be present across
multiple hops. In this case each node generates it's own outgoing
Label Set, possibly based on the incoming Label Set and the node's
hardware capabilities. This case is expected to be the norm for
nodes with conversion incapable interfaces.
9.7. Bi-directional LSP
GMPLS allows establishment of bi-directional LSPs. A bi-directional
LSP has the same traffic engineering requirements including fate
sharing, protection and restoration, LSRs, and resource requirements
(e.g., latency and jitter) in each direction. In the remainder of
this section, the term "initiator" is used to refer to a node that
starts the establishment of an LSP and the term "terminator" is used
to refer to the node that is the target of the LSP. For a bi-
directional LSPs, there is only one initiator and one terminator.
Normally to establish a bi-directional LSP when using [RSVP-TE] or
[CR-LDP] two unidirectional paths must be independently established.
This approach has the following disadvantages:
1. The latency to establish the bi-directional LSP is equal to one
round trip signaling time plus one initiator-terminator signaling
transit delay. This not only extends the setup latency for
successful LSP establishment, but it extends the worst-case latency
for discovering an unsuccessful LSP to as much as two times the
initiator-terminator transit delay. These delays are particularly
significant for LSPs that are established for restoration purposes.
2. The control overhead is twice that of a unidirectional LSP. This
is because separate control messages (e.g. Path and Resv) must be
generated for both segments of the bi-directional LSP.
3. Because the resources are established in separate segments, route
selection is complicated. There is also additional potential race
for conditions in assignment of resources, which decreases the
overall probability of successfully establishing the bi-directional
connection.
4. It is more difficult to provide a clean interface for SDH/SONET
equipment that may rely on bi-directional hop-by-hop paths for
protection switching. Note that existing SDH/SONET gear transmits
the control information in-band with the data.
5. Bi-directional optical LSPs (or lightpaths) are seen as a
requirement for many optical networking service providers.
With bi-directional LSPs both the downstream and upstream data
paths, i.e. from initiator to terminator and terminator to
initiator, are established using a single set of signaling messages.
Many Internet-Draft August 2001 26
draft-many-gmpls-architecture-00.txt Feb 2001
This reduces the setup latency to essentially one initiator-
terminator round trip time plus processing time, and limits the
control overhead to the same number of messages as a unidirectional
LSP.
For bi-directional LSPs, two labels must be allocated. Bi-
directional LSP setup is indicated by the presence of an Upstream
Label in the appropriate signaling message.
9.8. Bi-directional LSP Contention Resolution
Contention for labels may occur between two bi-directional LSP setup
requests traveling in opposite directions. This contention occurs
when both sides allocate the same resources (ports) at effectively
the same time. The GMPLS signaling defines a procedure to resolve
that contention, basically the node with the higher node ID will win
the contention. To reduce the probability of contention, some
mechanisms are also suggested.
9.9. Rapid Notification of Failure
GMPLS defines three signaling extensions for RSVP-TE that enable
expedited notification of failures and other events to nodes
responsible for restoring failed LSPs, and modify error handling.
For CR-LDP there is not currently a similar mechanism.
The first extension, identifies where event notifications are to be
sent. The second, provides for general expedited event notification.
Such extensions can be used by fast restoration mechanisms.
The final extension is an RSVP optimization to allow the faster
removal of intermediate states in some cases.
9.10. Explicit Routing and Explicit Label Control
The path taken by an LSP can be controlled more or less precisely by
using an explicit route. Typically, the node at the head-end of an
LSP finds a more or less precise explicit route and builds an
Explicit Route Object (ERO) that contains that route. Possibly, the
edge node don't build any ERO, and just transmit a signaling request
to a default neighbor LSR (as IP hosts today). For instance, an
explicit route could be added to a signaling message by the first
switching node, on behalf of the edge node. Note also that an
explicit route is altered by intermediate LSRs during its
progression towards the destination.
The ERO is originally defined by MPLS-TE as a list of abstract nodes
(i.e. groups of nodes) along the explicit route. Each abstract node
can be an IPv4 address prefix, an IPv6 address prefix, or an AS
number. This capability allows the generator of the explicit route
to have imperfect information about the details of the path. In the
simplest case, an abstract node can be a full IP address that
identify a specific node (called a simple abstract node).
Many Internet-Draft August 2001 27
draft-many-gmpls-architecture-00.txt Feb 2001
MPLS-TE allows strict and loose abstract nodes. The path between a
strict node and its preceding node must include only network nodes
from the strict node and its preceding abstract node. The path
between a loose node and its preceding node may include other
network nodes that are not part of the strict node or its preceding
abstract node.
This ERO was extended to include interface numbers as abstract nodes
to support unnumbered interfaces; and further extended by GMPLS to
include labels as abstract nodes. Having labels in an explicit route
is an important feature that allows to control the placement of an
LSP with a very fine granularity. This is more likely to be used for
non-PSC links.
In particular, the explicit label control in the ERO allows to
terminate an LSP on a particular outgoing port to an egress node.
This can also be used when it is desirable to "splice" two LSPs
together, i.e. where the tail of the first LSP would be "spliced"
into the head of the second LSP.
Another use is when an optimization algorithm is used for an
SDH/SONET network. This algorithm can provide very detailed explicit
routes, including the label (time-slot) to use on a link, in order
to minimize the external fragmentation of the SDH/SONET multiplex on
the corresponding interface.
Another use is when the label indicates a particular component in a
bundle in order to stay diverse with other components of that
bundle, i.e. to control the usage of components in a bundle for
different LSPs.
9.11 LSP modification and LSP re-routing
LSP modification and re-routing are two features already available
in MPLS-TE. GMPLS does not add anything new. Elegant re-routing is
possible with the concept of "make-before-break" whereby an old path
is still used while a new path is set up by avoiding double
reservation of resources. Then, the node performing the re-routing
can swap on the new path and close the old path. This feature is
supported with RSVP-TE (using shared explicit filters) and CR-LDP
(using the action indicator flag).
LSP modification consists in changing some LSP parameters, but
normally without changing the route. It is supported using the same
mechanism as re-routing. However, the semantic of LSP modification
will differ from one technology to the other. For instance, further
studies are required to understand the impact of dynamically
changing some SDH/SONET circuit characteristics such as the
bandwidth, the protection type, the transparency, the concatenation,
etc.
9.12. Route recording
Many Internet-Draft August 2001 28
draft-many-gmpls-architecture-00.txt Feb 2001
In order to improve the reliability and the manageability of the LSP
being established, the concept of the route recording was introduced
in RSVP-TE to function as:
- First, a loop detection mechanism to discover L3 routing loops, or
loops inherent in the explicit route (this mechanism is strictly
exclusive with the use of explicit routing objects).
- Second, a route recording mechanism collects up-to-date detailed
path information on a hop-by-hop basis during the LSP setup process.
This mechanism provides valuable information to the source and
destination nodes. Any intermediate routing change at setup time, in
case of loose explicit routing, will be reported.
- Third, a recorded route can be used as input for an explicit
route. This is useful if a source node receives the recorded route
from a destination node and applies it as an explicit route in order
to "pin down the path".
Within the GMPLS architecture only the second and third functions
are mainly applicable for non-PSC layers.
10. Forwarding Adjacencies (FA)
To improve scalability of MPLS TE (and thus GMPLS) it may be useful
to aggregate multiple LSPs inside a bigger LSP. Intermediate nodes
see the external LSP only, they don't have to maintain forwarding
states for each internal LSP, less signaling messages need to be
exchanged and the external LSP can be somehow protected instead (or
in addition) to the internal LSPs. This can considerably increase
the scalability of the signaling.
The aggregation is accomplished by (a) an LSR creating a TE LSP, (b)
the LSR forming a forwarding adjacency out of that LSP (advertising
this LSP as a link into ISIS/OSPF), (c) allowing other LSRs to use
forwarding adjacencies for their path computation, and (d) nesting
of LSPs originated by other LSRs into that LSP (e.g. by using the
label stack construct in the case of IP).
An LSR may (under its local configuration control) announce an LSP
as a link into ISIS/OSPF. When this link is advertised into the
same instance of ISIS/OSPF as the one that determines the route
taken by the LSP, we call such a link a "forwarding adjacency" (FA).
We refer to the LSP as the "forwarding adjacency LSP", or just FA-
LSP. Note that since the advertised entity is a link in ISIS/OSPF,
both the end point LSRs of the FA-LSP must belong to the same ISIS
level/OSPF area.
In general, creation/termination of a FA and its FA-LSP could be
driven either by mechanisms outside of MPLS (e.g., via configuration
control on the LSR at the head-end of the adjacency), or by
mechanisms within MPLS (e.g., as a result of the LSR at the head-end
of the adjacency receiving LSP setup requests originated by some
other LSRs).
Many Internet-Draft August 2001 29
draft-many-gmpls-architecture-00.txt Feb 2001
ISIS/OSPF floods the information about FAs just as it floods the
information about any other links. As a result of this flooding, an
LSR has in its link state database the information about not just
conventional links, but FAs as well.
An LSR, when performing path computation, uses not just conventional
links, but FAs as well. Once a path is computed, the LSR uses RSVP-
TE/CR-LDP for establishing label binding along the path. FAs needs
simple extensions to signaling and routing protocols.
Forwarding adjacencies may be represented as either unnumbered or
numbered links. A FA can also be a bundle of LSPs between two nodes.
When a FA is created dynamically, its TE attributes are inherited
from the TE LSP which induced its creation. Note that the bandwidth
of the FA-LSP must be at least as big as the LSP that induced it,
but may be bigger if only discrete bandwidths are available for the
FA-LSP. In general, for dynamically provisioned forwarding
adjacencies, a policy-based mechanism may be needed to associate
attributes to forwarding adjacencies.
10.1 Routing and Forwarding Adjacencies
A FA advertisement could contain the information about the path
taken by the FA-LSP associated with that FA. This information may be
used for path calculation by other LSRs. This information is carried
in a new OSPF and IS-IS TLV called the Path TLV.
It is possible that the underlying path information might change
over time, via configuration updates, or dynamic route
modifications, resulting in the change of that TLV.
If forwarding adjacencies are bundled (via link bundling), and if
the resulting bundled link carries a Path TLV, the underlying path
followed by each of the FA-LSPs that form the component links must
be the same.
It is expected that forwarding adjacencies will not be used for
establishing ISIS/OSPF peering relation between the routers at the
ends of the adjacency.
10.2. Signaling aspects
For the purpose of processing the ERO in a Path/Request message of
an LSP that is to be tunneled over a forwarding adjacency, an LSR at
the head-end of the FA-LSP views the LSR at the tail of that FA-LSP
as adjacent (one IP hop away).
10.3 Cascading of Forwarding Adjacencies
With an integrated model several layers are controlled using the
same routing and signaling protocols. A network may then have links
with different multiplexing/demultiplexing capabilities. For
Many Internet-Draft August 2001 30
draft-many-gmpls-architecture-00.txt Feb 2001
example, a node may be able to multiplex/demultiplex individual
packets on a given link, and may be able to multiplex/demultiplex
channels within a SONET payload on other links.
A new OSPF and IS-IS TLV has been defined to advertise the
multiplexing capability of each interface: PSC, TDM, LSC or FSC. The
information carried in this TLV is used to construct LSP regions,
and determine regions' boundaries.
Path computation may take into account region boundaries when
computing a path for an LSP. For example, path computation may
restrict the path taken by an LSP to only the links whose
multiplexing/demultiplexing capability is PSC. When an LSP need to
cross a region boundary, it can trigger the establishment of an FA
at the underlying layer. This can trigger a cascading of FAs between
layers with the following obvious order: TDM, then LSC, and then
finally FSC.
11. Security considerations
GMPLS introduces no new security considerations to the current MPLS-
TE signaling (RSVP-TE, CR-LDP) and routing protocols (OSPF-TE, IS-
IS-TE).
12. Acknowledgements
This draft is the work of numerous authors and consists of a
composition of a number of previous drafts in this area.
Many thanks to Ben Mack-Crane (Tellabs) for all the useful SDH/SONET
discussions that we had together. Thanks also to Pedro Falcao
(Ebone) and Michael Moelants (Ebone) for their SDH/SONET and optical
technical advice and support. Finally, many thanks also to Krishna
Mitra (Calient) and Curtis Villamizar (Avici).
A list of the drafts from which material and ideas were incorporated
follows:
1. draft-ietf-mpls-generalized-signaling-01.txt
Generalized MPLS - Signaling Functional Description
2. draft-ietf-mpls-generalized-rsvp-te-00.txt
Generalized MPLS Signaling - RSVP-TE Extensions
3. draft-ietf-mpls-generalized-cr-ldp-00.txt
Generalized MPLS Signaling - CR-LDP Extensions
4. draft-ietf-mpls-lmp-01.txt
Link Management Protocol (LMP)
5. draft-ietf-mpls-lsp-hierarchy-01.txt
LSP Hierarchy with MPLS TE
6. draft-ietf-mpls-rsvp-unnum-00.txt
Many Internet-Draft August 2001 31
draft-many-gmpls-architecture-00.txt Feb 2001
Signalling Unnumbered Links in RSVP-TE
7. draft-ietf-mpls-crldp-unnum-00.txt
Signalling Unnumbered Links in CR-LDP
8. draft-kompella-mpls-bundle-04.txt
Link Bundling in MPLS Traffic Engineering
9. draft-kompella-ospf-gmpls-extensions-00.txt
OSPF Extensions in Support of Generalized MPLS
10. draft-ietf-isis-gmpls-extensions-01.txt
IS-IS Extensions in Support of Generalized MPLS
13. References
TBD
14. Author's Addresses
Peter Ashwood-Smith Fong Liaw
Nortel Networks Corp. Zaffire Inc.
P.O. Box 3511 Station C, 2630 Orchard Parkway
Ottawa, ON K1Y 4H7 San Jose, CA 95134
Canada USA
Phone: +1 613 763 4534 Email: fliaw@zaffire.com
Email:
petera@nortelnetworks.com
Daniel O. Awduche Eric Mannie (editor)
Movaz Networks Ebone (GTS)
7296 Jones Branch Drive Terhulpsesteenweg 6A
Suite 615 1560 Hoeilaart
McLean, VA 22102 Belgium
USA Phone: +32 2 658 56 52
Phone: +1 703 847-7350 Email: eric.mannie@gts.com
Email: awduche@movaz.com
Ayan Banerjee Dimitri Papadimitriou
Calient Networks Alcatel - IPO NSG
5853 Rue Ferrari Francis Wellesplein, 1
San Jose, CA 95138 B-2018 Antwerpen
USA Belgium
Phone: +1 408 972-3645 Phone: +32 3 240-84-91
Email: abanerjee@calient.net Email:
dimitri.papadimitriou@alcatel.be
Many Internet-Draft August 2001 32
draft-many-gmpls-architecture-00.txt Feb 2001
Debashis Basak Dimitrios Pendarakis
Accelight Networks Tellium, Inc.
70 Abele Road, Bldg.1200 2 Crescent Place
Bridgeville, PA 15017 P.O. Box 901
USA Oceanport, NJ 07757-0901
Phone: +1 412 220-2102 (ext115) USA
email: dbasak@accelight.com Email: DPendarakis@tellium.com
Lou Berger Bala Rajagopalan
Movaz Networks, Inc. Tellium, Inc.
7926 Jones Branch Drive 2 Crescent Place
Suite 615 P.O. Box 901
MCLean VA, 22102 Oceanport, NJ 07757-0901
USA USA
Phone: +1 703 847-1801 Phone: +1 732 923 4237
Email: lberger@movaz.com Email: braja@tellium.com
Greg Bernstein Yakov Rekhter
Ciena Corporation Juniper
10480 Ridgeview Court Email: yakov@juniper.net
Cupertino, CA 94014
USA
Phone: +1 408 366 4713
Email: greg@ciena.com
John Drake Hal Sandick
Calient Networks Nortel Networks
5853 Rue Ferrari Email:
San Jose, CA 95138 hsandick@nortelnetworks.com
USA
Phone: +1 408 972 3720
Email: jdrake@calient.net
Yanhe Fan Debanjan Saha
Axiowave Networks, Inc. Tellium Optical Systems
100 Nickerson Road 2 Crescent Place
Marlborough, MA 01752 Oceanport, NJ 07757-0901
USA USA
Phone: +1 508 460 6969 Ext. 627 Phone: +1 732 923 4264
Email: yfan@axiowave.com Email: dsaha@tellium.com
Don Fedyk Vishal Sharma
Nortel Networks Corp. Jasmine Networks, Inc.
600 Technology Park Drive 3061 Zanker Road, Suite B
Billerica, MA 01821 San Jose, CA 95134
USA USA
Phone: +1-978-288-4506 Phone: +1 408 895 5030
Email: Email:
dwfedyk@nortelnetworks.com vsharma@jasminenetworks.com
Many Internet-Draft August 2001 33
draft-many-gmpls-architecture-00.txt Feb 2001
Gert Grammel George Swallow
Alcatel Cisco Systems, Inc.
Italy 250 Apollo Drive
Email: Chelmsford, MA 01824
gert.grammel@netit.alcatel.it USA
Phone: +1 978 244 8143
Email: swallow@cisco.com
Kireeti Kompella Z. Bo Tang
Juniper Networks, Inc. Tellium, Inc.
1194 N. Mathilda Ave. 2 Crescent Place
Sunnyvale, CA 94089 P.O. Box 901
USA Oceanport, NJ 07757-0901
Email: kireeti@juniper.net USA
Phone: +1 732 923 4231
Email: btang@tellium.com
Alan Kullberg John Yu
NetPlane Systems, Inc. Zaffire Inc.
888 Washington 2630 Orchard Parkway
St.Dedham, MA 02026 San Jose, CA 95134
USA USA
Phone: +1 781 251-5319 Email: jzyu@zaffire.com
Email: akullber@netplane.com
Jonathan P. Lang Alex Zinin
Calient Networks Cisco Systems
25 Castilian 150 W. Tasman Dr.
Goleta, CA 93117 San Jose, CA 95134
Email: jplang@calient.net Email: azinin@cisco.com
Full Copyright Statement
"Copyright (C) The Internet Society (date). All Rights Reserved.
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
Many Internet-Draft August 2001 34
draft-many-gmpls-architecture-00.txt Feb 2001
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."
Many Internet-Draft August 2001 35