TOPICS IN INTERNET TECHNOLOGY

Designing Multiprotocol
Label Switching Networks
Jeremy Lawrence, Cisco Systems

ABSTRACT MPLS NETWORK STRUCTURES

An MPLS network such as that in Fig. 1a consists
Multiprotocol label switching adds to the of edge label switch routers (edge LSRs)1 around
capabilities of IP networks in several ways. a core of label switch routers (LSRs). Customer
Despite new capabilities, MPLS technology has sites are connected to the carrier’s MPLS net-
much in common with ordinary IP networks. In work, or equivalently a large organization’s net-
turn, the design process for MPLS networks has work backbone. Figure 1a shows nine customer
much in common with the design of any IP net- sites and four edge LSRs, but more typically there
work. This article examines MPLS and IP tech- will be hundreds or more customer sites per edge
nology with particular emphasis on what is LSR. The customer premises equipment connect-
common between them. The common design ed to an MPLS network typically runs ordinary IP
steps of MPLS networks and other IP networks forwarding rather than MPLS, and is typically a
are outlined briefly, and those issues specific to router or a LAN switch. Since the customer
MPLS networks are covered in more detail. This equipment typically does not run MPLS, the edge
article emphasizes MPLS point of presence LSRs are part of the carrier’s network and under
design, routing design issues for MPLS, and pro- the carrier’s administration.
visioning of sufficient label space. A carrier’s MPLS network will often be con-
nected to one or more other IP networks as part
of the Internet. An IP connection to another
INTRODUCTION carrier might be an MPLS link, although use of
Multiprotocol label switching (MPLS) is an MPLS on intercarrier links is usually not
extension to the existing Internet Protocol (IP) required. As with any interconnection of carrier
architecture [1]. By adding new capabilities to networks in the Internet, the Border Gateway
the IP architecture, MPLS enables support of Protocol [8] would typically operate over links to
new features and applications, which have been other carriers, to exchange routing information
described previously [2]. The new applications with them. The neighboring IP networks may
include traffic engineering [3, 4], IP virtual pri- use MPLS internally, but not necessarily.
vate networks [5, 6], integration of IP routing The links between customer equipment, edge
and layer 2 or optical switching [7], and other LSRs, and/or LSRs may be of virtually any type.
applications. The previous descriptions of MPLS Traditional non-MPLS switches may be used to
have concentrated, quite rightly, on what is new carry connections to MPLS equipment, typically
in MPLS (i.e., the differences between MPLS using some sort of permanent virtual circuits or
and traditional IP routing and forwarding). optical lightpaths. Figure 1b illustrates several
However, an important consideration is that ways of doing this.
there is much in common between the technolo-
gy of MPLS networks and IP routed networks.
The overlap in technology means that there is PROTOCOLS IN
1
much in common between the process of design IP AND MPLS NETWORKS
Edge LSRs are also of MPLS networks and the design of other IP
known as label edge routed networks. This article starts with a IP PROTOCOLS
routers. Most commer- description of MPLS technology that concen- MPLS networks use a superset of the protocols
cially available LSRs have trates as much on the similarities between used in ordinary IP routers, and most MPLS
at least limited edge LSR MPLS and other IP routed networks as on the networks are fully interoperable with ordinary IP
capability, and vice versa. differences. The design process of MPLS net- equipment. The IP protocol architecture, or IP
The term “edge LSR” works is then outlined, and the new design stack, consists of a large number of protocols, a
conveys this overlap more issues specific to all MPLS networks are subset of which is used on IP routers, as opposed
effectively. described in detail. to hosts or servers. Figure 2a shows the most

134 0163-6804/01/$10.00 © 2001 IEEE IEEE Communications Magazine • July 2001

 Function: CPE MPLS MPLS MPLS
edge edge

ATM: Asynchronous transfer mode

CPE: Customer premises equipment
IP: Internet Protocol
LSR: Label switch router
MPLS: Multiprotocol label switching Other IP
networks:
Device: Edge LSR Label switch routers Edge LSR

(a) MPLS: general architecture

Function: CPE Access Backhaul MPLS Tunneling MPLS Tunneling MPLS MPLS Access CPE
edge edge

Device: Non-MPLS access Non-MPLS Edge LSR Non-MPLS LSR Non-MPLS LSR Edge LSR Non-MPLS access
switch or switch switch switch switch or
concentrator concentrator

(b) MPLS with backhaul and tunneling via ATM, frame relay, optical, or other switches

■ Figure 1. MPLS network structures.

important of these protocols as a basis of com-

parison with the MPLS stack. CR-LDP: Constraint-based routing with LDP MPLS: Multiprotocol label switching
IP: Internet Protocol RSVP: Resource Reservation Protocol
The protocols that run on an IP router con- LDP: Label Distribution Protocol
sist of routing and forwarding protocols, and
others. Routing protocols inform a router about IP routing protocol(s)
the routes to reach the known IP destinations. IP routing protocol(s)
Label-switched path setup:
Each router will, in general, run an interior gate- LDP, RSVP+extensions,
IP explicit quality-of-service and/or CR-LDP
way (routing) protocol, which deals with routing signaling (rare): RSVP
inside a network. There is a choice of standard Label switching
interior gateway protocols: IP packet forwarding
• The Routing Information Protocol has many (Data link and physical layers)
limitations, and is becoming obsolete. (Data link and physical layers)
• Open Shortest Path First is widely used. MPLS network
MPLS network
• Interior System to Interior System is similar a. Ordinary IP router b. Label switch router
to Open Shortest Path First, and is also
common among large providers and carriers.
Typically, all routers under one distinct IP routing protocol(s) IP routing protocol(s)
administration will run the same interior gate- Label-switched path setup:
IP explicit quality-of-service
way protocol. In general, a router used to com- LDP, RSVP+extensions,
and/or CR-LDP signaling (rare): RSVP
municate between different administrations will
MPLS packet
run another routing protocol called the Border Label switching encapsulation IP packet forwarding
Gateway Protocol, which deals with exchange of
(Data link and (Data link and
information between administrative domains. physical layers) physical layers)
Routing protocols automatically build a table
of forwarding information in each router. The MPLS network IP network
forwarding table lists the IP destination prefixes c. Edge label switch router connecting to a traditional IP network
that are known, and which data link on the
router is the “next hop” when moving IP packets ■ Figure 2. Protocol stacks for IP and MPLS. These stacks are simplified for
to each known destination prefix. In routers, the clarity; see the text.
forwarding table is used by IP when moving
packets toward their destination. IP may be
referred to as a forwarding protocol, since it Many other protocols are used by routers.
moves packets rather than determines routes for There are specific data link protocols for various
them. Version 4 of IP (IPv4) is the most widely media, and protocols to perform auxiliary func-
used for IP applications, although use of IPv6 is tions. For example, routing protocols typically
also beginning. use the Transport Control Protocol or Universal

IEEE Communications Magazine • July 2001 135

 Datagram Protocol as part of their communica- IP routing protocol(s) in a network in order to
In both tion process. However, these protocols are not set up LSPs.
used in the main function of routers, namely for-
hop-by-hop warding packets along the correct route to their Label Encapsulation and Packet Forward-
destination. For a complete listing of all proto- ing — When a packet is sent on an LSP, a label
routed MPLS and
cols in the IP stack, their standards status, and is applied to the packet (Fig. 3). There are sev-
explicit routing, references to their RFC documents, see [8]. eral distinct means of carrying labels on packets.
In the generic label encapsulation, which may be
label distribution Terminology: IP Destination Prefixes — A used with any link type, the label is carried in an
works in close destination prefix is a group of IP addresses that extra header applied to each packet between the
may be treated similarly for forwarding purposes. data link header and the layer 3 header. On link
conjunction with For example, the IPv4 destination prefix types that inherently support the concept of a
[Link]/24 represents the range of IP addresses in virtual channel, the label may be carried as virtu-
the IP routing
which the first three octets (24 bits) are 0000 1010, al channel information, typically in a layer 2
protocol(s) in a 0000 0001, and 0000 0001, with any values for the header. On asynchronous transfer mode (ATM)
last octet. In other words, the destination prefix links, the label may be carried as the virtual cir-
network in order [Link]/24 represents any IPv4 address 10.1.1.x, cuit identifier and/or virtual path identifier
to set up LSPs. where x is a value between 0 and 255 inclusive. applied to each cell of the packet. On frame
relay links, the label may be carried as the data
PROTOCOLS IN THE MPLS NETWORK CORE link connection identifier applied to each packet.
IP Routing and Label-Switched Paths — As A similar scheme has also been proposed for
with ordinary routers, LSRs will run one or more synchronous optical network links. MPLS over
IP routing protocols (Fig. 2b). However, instead optical switches similarly uses virtual channels,
of IP packet forwarding, LSRs use a different with each distinct optical wavelength used on a
type of forwarding mechanism called label switch- link constituting a different label [7].
ing. Label switching relies on the setup of switched Furthermore, two distinct types of MPLS can
paths through the network; these are label- be defined:
switched paths (LSPs). An LSP, in the simplest • Packet-based MPLS is where the LSRs have
case, corresponds to an IP destination prefix that full packet-handling capacity, and can exam-
appears in the forwarding tables of several ine layer 3 headers on packets. 3 Packet-
routers. For example, the interior gateway proto- based LSRs typically also have IP router
col instances in a succession of LSRs might create function. Edge LSRs are typically a type of
forwarding table entries for an IPv4 destination packet-based LSR.
prefix [Link]/24. The LSRs then, in effect, com- • Switch-based MPLS is the form of MPLS
pare their forwarding information to find that where LSRs, known as switch-based LSRs,
they all have a route to [Link]/24, and set up an forward packets by means of layer 2 switch-
LSP through themselves in order to carry all traf- ing or optical switching. Switch-based LSRs
fic for the destination prefix [Link]/24. Such have little or no capability to examine layer
setup of an LSP according to IP routing informa- 3 headers.
tion is known as hop-by-hop routed MPLS. Any type of label encapsulation can be sup-
Setup of LSPs is done by a process of label ported by either class of LSR. For example,
distribution. There are several valid combinations packet-based LSRs may have optical links using
of protocols for label distribution: distinct lightpath wavelengths as labels; or there
• In general, all MPLS networks use the Label may exist switch-based LSRs dealing with the
Distribution Protocol, which supports hop- generic MPLS label encapsulation, but without
by-hop routed MPLS. the ability to deal with layer 3 headers. However,
• One alternative for explicitly routed MPLS most switch-based LSRs used to date have been
is the Label Distribution Protocol in combi- based on ATM switches. The use of switch-based
nation with the Resource Reservation Pro- MPLS on ATM switches is also referred to as
tocol plus some extensions.2 ATM MPLS.
• The other alternative for explicitly routed On each link, each distinct label value corre-
MPLS is the Label Distribution Protocol in sponds to a distinct LSP. The effect of carrying
combination with some extensions, called label information with each packet is most easily
2 It is, in principle, possi- constraint-based LSP setup using LDP. understood in switch-based MPLS. Because the
ble to run an MPLS net- Explicit routing is a mode of MPLS operation MPLS protocols associate a virtual channel with
work using the Resource where LSPs are set up to override the normal an LSP at each hop, a virtual channel connection
Reservation Protocol plus hop-by-hop routed path chosen by an IP routing is set up end to end across the network to support
extensions, but without protocol for selected traffic. The normal applica- the LSP. With packet-based MPLS, the generic
the use of the Label Dis- tion is in IP traffic engineering [3, 4]. Although label encapsulation allows labeling to occur over
tribution Protocol. How- explicit routing and traffic engineering override almost any link type, not just those that inherently
ever, it is not clear the routes from the IP routing protocol, they do support virtual channels. Packet-based MPLS
whether any MPLS imple- not ignore the IP routing protocol. The IP rout- equipment implements MPLS forwarding by
mentation supports this. ing protocol information must be consulted in mapping incoming label values to next-hop inter-
order to ensure that the explicitly routed paths faces and outgoing label values; these mappings
3 Packet-based MPLS is do not lead to looping routes. In addition, traffic define a type of switching operation. Again, a
also known as frame- engineering makes use of extensions to the inte- switched path across the network is established
based MPLS. The latter rior gateway protocols to carry information across the network; this is the LSP. The combina-
term is avoided here about traffic loads in the network. In both hop- tion of the protocols in the MPLS protocol stack
because of possible confu- by-hop routed MPLS and explicit routing, label allows LSPs to be established in conjunction with
sion with frame relay. distribution works in close conjunction with the IP routing, and across any type of link.

136 IEEE Communications Magazine • July 2001

 Point-to-Point Protocol links: PPP header Label header Layer 3 header For the purposes
of packet for-
Local area network media: MAC header Label header Layer 3 header warding, the
a. Generic label encapsulation “traditional IP”
and “MPLS” parts

Asynchronous transfer mode links: Payload

of the edge LSR
VPI VCI ATM cell header
protocol stack are
linked by the
Label
process of MPLS
encapsulation.
Frame relay Payload
Frame relay links: DLCI packet header

Label

b. Layer 2 label encapsulation

MAC: Medium access control VCI: Virtual circuit identifier ATM: Asynchronous transfer mode
VPI: Virtual path identifier DLCI: Data link connection identifier PPP: Point-to-Point Protocol

■ Figure 3. Carrying a label on a packet.

PROTOCOLS IN MPLS EDGE DEVICES those of other IP networks. A typical design pro-
cess is as follows:
Edge LSRs usually connect to both MPLS net- 1. Locate and design the points of presence
works and traditional IP networks. Consequently, (PoPs).
edge LSRs typically run all the protocols found 2. Dimension the backbone links in the net-
on a traditional IP router, including IP packet work.
forwarding, in addition to MPLS. This is illustrat- 3. Design the IP routing.
ed in Fig. 2c.4 IP routing protocols make no dis- 4. Dimension the MPLS label space.
tinction between MPLS and traditional IP 5. Configure any MPLS applications or
networks, which means a single routing protocol advanced features, such as traffic engineer-
process in an edge LSR may support both MPLS ing and virtual private networks.
and traditional IP links. For the purposes of pack- The final design step is an ongoing process:
et forwarding, the “traditional IP” and “MPLS” 6. Refine and extend the design once the net-
parts of the edge LSR protocol stack are linked work is operational.
by the process of MPLS encapsulation. With the exception of steps 4 and 5, most of
In simple operation, MPLS encapsulation these steps are not specific to MPLS networks,
involves the application (or “pushing”) of a label and apply to IP routed networks of any type.
onto an IP packet, which is then forwarded These steps are described below, with emphasis
across the network on the corresponding LSP. on those considerations that are new to MPLS.
When the packet reaches the edge of the MPLS Traffic engineering and virtual private networks
network, the label is then removed (or (step 5) are not covered in this article, which 4 In contrast is a recently
“popped”) from the packet, which is then for- concentrates on design issues common to all proposed MPLS applica-
warded as an ordinary IP packet. MPLS networks. The design issues unique to tion where an MPLS net-
MPLS also allows more than one label to be MPLS traffic engineering and virtual private net- work provides generic
carried on each packet: a “stack” of labels of works would require more space to describe packet transport service,
arbitrary size may be carried using the generic than is available here. similar to the service of a
MPLS label header. Multiple labels are used in traditional frame relay or
MPLS traffic engineering [3, 4], and virtual pri- POINT OF PRESENCE DESIGN AND ATM network. In such a
vate networks [5, 6]. CHOICE OF EQUIPMENT service, IP packet for-
warding is not used in the
The design of PoPs for all IP networks, including edge LSRs, which instead
STEPS IN DESIGNING MPLS networks, is constrained by: use a more rudimentary
• The choice of access link type(s) to be sup- forwarding mechanism:
MPLS NETWORKS ported by the network. all packets from a speci-
The previous discussion has shown that MPLS • The choice of core link type(s) to be used in fied customer link are for-
networks and other IP networks have many pro- the network. warded onto a particular
tocols in common. Their common use in MPLS • Requirements for reliability, including use LSP, and vice versa. In
and other IP networks means that MPLS shares of warm or hot-standby redundant trunk such a network, IP routing
many design steps in common with all IP routed cards, redundant processors, etc. These may protocols run only in the
networks. Consequently, most of the design constrain the choice of equipment to be MPLS network core, and
steps for an MPLS network are in common with used in a PoP. not on the customer links.

IEEE Communications Magazine • July 2001 137

 peerings out of the PoP, assisting routing
CPE protocol scalability.
CPE: Customer premises equipment • It will assist in preventing local traffic from
LSR: Label switch router
MPLS: Multiprotocol label switching having to leave the PoP. In other words,
MPLS links to LSRs or traffic for which both the source and edge
other edge LSRs destination edge LSR are at the same PoP
will not be required to leave the PoP.
a. Single edge LSR
• It will often reduce the number of links
required from the PoP.
• In an MPLS network using ATM or frame
relay switching and label merging (see
LSR “Choosing Core LSRs” below), it reduces
the number of virtual circuits required from
the PoP.
MPLS links to other All except the last point also apply to a PoP
LSRs or edge LSRs in a traditional IP network with analogous struc-
ture.
At least two vendors produce devices that
b. Multiple edge LSRs and an LSR integrate the functions of an access switch, sev-
eral edge LSRs, and an LSR. These provide the
same functions as using several discrete devices,
Access switch function LSR function but with the practical advantages of having a sin-
gle device: less space, somewhat lower power
requirements, and so forth.
…
Choosing Core LSRs — LSRs without edge
= links will often be used, either as part of an edge
Physical view
PoP or as standalone devices. Aside from the
lack of edge links, the design considerations are
Functional broadly the same as those listed above for all
Edge LSR instances view MPLS PoPs. Some specific considerations for
c. Integrated device with edge LSRs and an LSR core LSRs are:
• The number, as well as type, of core links to
be supported: It is more important for a
MPLS links core LSR to support many core links than
for an edge LSR.
• Label merging capability: All packet-based
LSRs inherently support merging LSPs to a
common destination, but such capability is
less common for switch-based LSRs. Those
switch-based LSRs that do support label
d. Standalone LSR merging are typically based on ATM or
frame relay switches supporting virtual cir-
■ Figure 4. PoP designs for MPLS. cuit merging. In an ATM or frame relay
network, label merging capability brings a
significant scalability advantage by reducing
• Requirements for services beyond IP or the number of virtual circuits required.
MPLS, and availability of equipment meet-
ing these requirements. For example, sever- Redundancy and Reliability — In a broad
al commercial edge LSRs also have the sense, the reliability issues with IP and MPLS
capability of providing traditional ATM and equipment are the same as for other telecommu-
frame relay edge switch function, and oth- nications equipment. A network typically must
ers offer only IP and MPLS edge function. be robust to device and link failures. However,
• Location of PoPs, which is largely deter- the overall reliability of IP and MPLS networks
mined by where the cities are. is typically less dependent on the reliability of
• The population of end-user sites surround- individual devices than in circuit or virtual circuit
ing each location. networks.
Once these are taken into account, MPLS Some circuit (or virtual circuit) switches in
equipment offers many different options for PoP connection-oriented networks have the ability to
designs. Some typical PoP designs are shown in reroute circuits around failed links or nodes.
Fig. 4. Rerouting of circuits in a connection-oriented
A small PoP might consist of just a single network typically takes a significant duration.
edge LSR with a number of access lines from This is due to the requirement to recalculate
less than 10 to thousands. However, the popula- routes, often on a connection-by-connection
tion of user lines for a PoP often leads to a basis, and reprogram some hardware for each
requirement for a PoP to have several edge changed connection. Consequently, rerouting in
LSRs. In a PoP with several edge LSRs, the use a connection-oriented network is a last resort, to
of an extra LSR as an aggregation device has be used only after redundant links or nodes have
several scalability advantages: failed to prevent an outage. In IP routed net-
• It reduces the number of routing protocol works, on the other hand, large aggregates of

138 IEEE Communications Magazine • July 2001

traffic can be rerouted simultaneously, and less today. Each of the two devices includes two
(or no) hardware programming is typically functionally separate edge LSRs and an Rerouting may
required. Consequently, the reliability of many LSR. In addition, each of the edge devices
IP routed networks has relied primarily on includes a permanent virtual circuit switch- be somewhat
rerouting and network-level redundancy (i.e., ing function that is functionally separate
slower in
having a choice of several routes between each from the LSR function.
source and destination). Although IP routers are • Routing viewpoint: This viewpoint is the switch-based
gaining similar redundancy features to circuit network as it is seen by an IP routing proto-
switches, network designers are still free to rely col. It is derived from the functional view- MPLS networks
more on network-level redundancy in IP net- point as follows: than in
works than in circuit or virtual circuit networks. –Layer 2 switches and permanent virtual cir-
Packet-based MPLS has redundancy options cuit switching functions are invisible to IP packet-based
similar to those of IP routing. Furthermore, routing. If a customer site is connected to a
MPLS or IP routed
there is recent work on protection switching in router by a permanent virtual circuit, the
MPLS, which involves setting up a single LSP as virtual circuit is seen by IP routing as a sin- networks,
an alternate to a failed link that might carry gle-hop direct connection. For example,
many LSPs. This capability, also known as fast note the sites labeled a in Fig. 5c, and because
reroute, is a new type of protection switching, assume that these are connected to edge switch-based
and is unique to MPLS networks [9]. LSR b. Then, in the routing viewpoint, the
Rerouting may be somewhat slower in switch- sites are directly adjacent to router b. MPLS equipment
based MPLS networks than in packet-based –Each edge LSR or LSR constitutes a
typically requires
MPLS or IP routed networks, because switch- router in the routing viewpoint.
based MPLS equipment typically requires more Designing IP routing in an MPLS network is more hardware
hardware programming to reroute LSPs. In addi- generally the same process as designing IP rout-
tion, ATM, frame relay, and optical switching ing for an ordinary IP network [10]. With refer- programming to
often do not support the pushing and popping of ence to the routing viewpoint, a network can be reroute LSPs.
extra labels, as required for rerouting of traffic divided into interior gateway protocol areas,
from a physical link onto an LSP. Together with route summarization can be designed, and so on.
slower connection programming speeds, this
means that switch-based MPLS is not as capable IP Routing Issues Specific to MPLS Net-
as packet-based MPLS of supporting fast works — There are a small number of routing
reroute. Consequently, switch-based MPLS may issues specific to MPLS networks:
sometimes have to rely more on device-level • MPLS affects the choice of interior gateway
redundancy than packet-based MPLS. protocol. Specifically, MPLS traffic engi-
neering uses a link-state protocol such as
LINK DIMENSIONING Open Shortest Path First or Interior System
The dimensioning of bandwidth of links in an to Interior System to distribute information
MPLS network is essentially the same as for any about the traffic loading on links. Distance-
other IP network. A typical process starts with vector protocols such as the Routing Infor-
an estimate of a PoP-to-PoP traffic matrix based mation Protocol do not support MPLS
on user populations at each PoP and an estimate traffic engineering.
of the proportion of traffic destined to the wider • Some IP routing protocol implementations
Internet as opposed to the network in question. allow unnumbered IP links. These are links
Based on this matrix, an initial estimate of the without an IP address: another IP address
required link capacities can be calculated, and on a router, such as a loopback address, is
then used to provision the network. MPLS does used to substitute for the link address for
provide one important benefit with respect to routing protocol purposes. The use of
link dimensioning: MPLS is an effective traffic unnumbered IP links may be desirable in an
engineering tool, which can be used to fit traffic MPLS network, particularly if it is switch-
to available link capacities [3, 4]. based. Unnumbered links reduce the num-
ber of IP destination prefixes known to the
DESIGNING IP ROUTING FOR AN LSRs, and hence may reduce the number of
MPLS NETWORK labels that will be used in the network.
• Route summarization must not be done at a
Every LSR uses an ordinary IP interior gateway switch-based LSR. Multiple interior gate-
routing protocol. As in an ordinary IP network, way protocol areas can be used in a switch-
the interior gateway protocol determines the based MPLS network, as illustrated in Fig.
routes for traffic, which in MPLS is carried in 6. A switch-based LSR may be used as an
LSPs. An important implication is that the inte- area border router, but only if no summa-
rior gateway protocol sees an MPLS network as rization is done at the area border router.
being exactly like an ordinary router network. As It is common for area border routers to be
illustrated in Fig. 5, it is possible to have various configured to summarize routes. For example, an
viewpoints of an MPLS network: area border router may receive several routes
• Physical viewpoint: This viewpoint repre- from within one area to destinations such as
sents the physical devices and links in a net- [Link]/24, [Link]/24, and [Link]/24. In order
work. to reduce routing table sizes, it may be desired to
• Functional viewpoint: Where a product has advertise them into another area by way of a sin-
several functions, these can be shown sepa- gle summary route (e.g., a route to [Link]/16.)
rately. For example, Fig. 5 includes two However, such summarization may not be done if
MPLS edge devices of a type available the area border router is also a switch-based

IEEE Communications Magazine • July 2001 139

 Many of the
issues of
LSR: Label switch router
designing MPLS
networks are Edge LSR

similar to those
of designing Switch-based LSR

ordinary IP Integrated LSRs

edge switches Access
networks. concentrator
However, one
issue in MPLS a. Physical viewpoint

networks that
does not occur in
other IP routed
LSR function
networks is the
provisioning of
label space.

LSR Circuit-switching
function function
Edge LSR
Circuit-switching functions
function
b. Functional viewpoint

Layer 2 switching functions are invisible to IP routing a

c. Deriving the routing viewpoint

a
d. Routing viewpoint

■ Figure 5. Viewpoints of an MPLS network.

140 IEEE Communications Magazine • July 2001

LSR. If route summarization is required at the
area border routers in a switch-based MPLS net-
work, the area border routers must be packet-
based LSRs. This requirement is not a
contradiction, since it is possible for a packet- Area 1 Area 2
based LSR to support switch-based MPLS links.
• The previous rule also applies to the Border
Gateway Protocol. Autonomous system
boundary routers are those routers which
communicate between administrative
domains using the Border Gateway Proto-
col. Such routers often summarize routes.
Consequently, a switch-based LSR may not Area 0
be an autonomous system boundary router,
but a packet-based LSR may be one. Area border routers
The restrictions on summarization exist
because summarization stops some LSPs being
set up end to end. For example, assume that an
area border summarizes reachability for ■ Figure 6. The routing viewpoint of a network with multiple routing areas.
[Link]/24, [Link]/24, and [Link]/24 with a There are restrictions on the use of label switch routers as area border routers.
single route for [Link]/16. Now assume that a
packet with destination IP address [Link]
arrives with a label for [Link]/16. The area bor- • The subnet address prefix of any numbered
der router cannot label switch the packet. It point-to-point link, or any other subnet, is a
must look past the label and examine the IP destination prefix. For this reason, it is
address to find that the packet should go on to sometimes desirable to use unnumbered
[Link]/24. Since switch-based LSRs cannot links, to reduce the number of destination
examine layer 3 headers and IP addresses, they prefixes.
may not do IP route summarization. • Any other IP address of a device in the area
contributes a destination prefix. Such
DIMENSIONING MPLS LABEL SPACE addresses may include host addresses and
Many of the issues in designing MPLS networks loopback addresses on LSRs.
are similar to those in designing ordinary IP net- • There usually will be destination prefixes
works. However, one issue in MPLS networks from outside the area, but advertised in the
that does not occur in other IP routed networks area by area border routers (or autonomous
is the provisioning of label space. system boundary routers). If many address
A certain number of LSPs are carried on prefixes are summarized into a single address
each link in an MPLS network. Many LSRs sup- prefix before advertisement in the area, this
port label merging, which means that LSPs to counts as a single destination prefix.
the same destination can be merged in the LSR, Labels are not necessarily required for all
so only one label is required to that destination destination prefixes in the forwarding table.
on each link. In a network using label merging MPLS implementations may allow filtering of
and hop-by-hop routed MPLS, bounds on label the forwarding table when determining the desti-
usage are quite simple to calculate. For example, nation prefixes for which labels are required.
if labels in a network are required for 1000 dis- This might be used, for example, to prevent
tinct IP destination prefixes, and five distinct labels being set up to the destination prefixes of
labels are required to each destination for carry- numbered links. In general, labels are required
ing different classes of service, no more than for a subset of the destination prefixes in the
5000 labels will be required on each link. forwarding table, of size d.
Every LSR supports a limited number of In networks using label merging, a limit on
labels on its links, and this limit is implementa- the number of active labels required per link is
tion-specific. The generic MPLS label encapsula- straightforward to calculate, as shown above.
tion allows a 20-bit label range, which means The limit is cd, where c is the number of differ-
that a link may support up to 1,048,576 (1M) ent labels required to each of the destination
distinct labels. However, many LSR implementa- prefixes. c = 1 unless distinct labels are required
tions, particularly of switch-based LSRs, support for multiple service classes, in which case it will
far fewer than this. Consequently, an important typically be 2–5. In networks without label merg-
part of MPLS network design is to calculate the ing, the limits are much larger.
label space requirements on each link, and com- If insufficient label space is available on a
pare this to the number of active labels support- link, some required LSPs will be blocked. Such
ed on each link. blockage will generally lead to a lack of com-
plete connectivity in the network and inability to
Label Usage and Hop-by-Hop Routed MPLS deliver some hop-by-hop routed traffic. Conse-
— In hop-by-hop routed MPLS, labels are quently, it is important to ensure that the num-
required to enable LSPs to correspond to the ber of labels available is sufficient.
known IP destination prefixes. A routing table,
as built by an IP routing protocol, lists the desti- Equipment Limits — Once the label space
nation prefixes known in a routing protocol area. requirements have been analyzed, they can be
The forwarding table will list destination prefixes compared to the capabilities of the equipment in
of several types: the network. The nature of the limits on active

IEEE Communications Magazine • July 2001 141

 labels in commercial equipment will vary widely. Most design steps for an MPLS network cor-
MPLS significantly There may be limits per LSR, per link, and/or of respond to those in designing an ordinary IP
other types. For example, some LSRs support a network, but have some extra considerations that
extends the certain number of active labels per line card, have been described in this article. The most
shared among several links. In LSRs supporting important extra considerations are that MPLS
capabilities of IP
label merging, there may be limits on the number introduces new options for designing reliability
networks by of multipoint-to-point trees of LSPs merging onto and redundancy in networks, and MPLS net-
a single link, distinct from the number of labels works have some constraints on IP routing pro-
introducing supported on the link. Optical LSRs may have tocol configuration. On the other hand, MPLS
support of VPNs further constraints on exactly how labels are used. introduces a new design step that is not found in
For example, some optical LSRs may not allow the design of other IP networks: ensuring suffi-
and traffic the label value (i.e., lightpath wavelength) to be cient provision for active labels on links. The
changed when an LSP is switched from link to advantages of MPLS stem from the introduction
engineering, and
link. Other optical LSRs may allow changes only of a general concept of labels and label switch-
with integration by a certain deviation in lightpath wavelength. ing; consequently, MPLS function relies on hav-
However, dealing with such constraints in optical ing sufficient label space available.
of IP routing with LSRs is a matter for further research. In any case,
various “later 2” full details of the constraints on label usage REFERENCES
should be obtained from the manufacturer(s) of [1] E. Rosen, A. Viswanathan, and R. Callon, “Multiprotocol
switching the particular LSRs used in a network. Label Switching Architecture,” RFC 3031, Jan. 2001.
[2] G. Armitage, “MPLS: The Magic Behind the Myths,” IEEE
technologies. ONGOING REFINEMENT Commun. Mag., vol. 38, no. 1, Jan. 2000, pp. 124–31.
[3] D. Awduche, “MPLS and Traffic Engineering in IP Net-
Network design is an ongoing process. Once an works,” IEEE Commun. Mag., vol. 37, no. 12, Dec.
MPLS network is deployed, it has ongoing design 1999, pp. 42-47.
requirements that have much in common with [4] G. Swallow, “MPLS Advantages for Traffic Engineering,” IEEE
ordinary IP networks. One important activity is to Commun. Mag., vol. 37, no. 12, Dec. 1999, pp. 54–57.
[5] E. Rosen and Y. Rekhter, “BGP/MPLS VPNs,” RFC 2547,
measure the actual PoP and link traffic, compare Mar. 1999; E. Rosen et al., “BGP/MPLS VPNs,” Internet
them against the assumptions used in the initial draft, [Link], work in progress.
design, and make modifications as necessary. This [6] K. Muthukrishnan and A. Malis, “Core MPLS IP VPN
process is repeated as traffic flows increase. Architecture,” RFC 2917, Sept. 2000.
[7] A. Banerjee et al., “Generalized Multiprotocol Label
As customers are added and traffic increases, Switching: An Overview of Routing and Management
network changes will be required, including: Enhancements,” IEEE Commun. Mag., vol. 39, no. 1,
• Adding new edge LSRs to PoPs to cope Jan 2001, pp. 144–50.
with increasing numbers of customer links [8] Internet Engineering Task Force, “Internet Official Proto-
col Standards,” RFC 2600. (Note that every 100th RFC
• Adding new links to the network document is an update to this document, and this doc-
• Adjusting traffic engineering parameters ument will automatically be superseded by RFC 2700,
• Checking for allocation of sufficient label 2800, etc. as published.)
space on links [9] V. Sharma, et al., “Framework for MPLS-based Recov-
ery,” Internet draft, draft-ietf-mpls-recovery-frmwrk-
[Link], work in progress.
CONCLUSIONS [10] B. Halabi, Internet Routing Architectures, New Riders
Publishing 1997.
MPLS significantly extends the capabilities of IP
networks by introducing support of virtual pri-
vate networks and traffic engineering, and with BIOGRAPHY
integration of IP routing with various layer 2 JEREMY LAWRENCE (jlawrenc@[Link]) is a consulting engi-
switching technologies. MPLS achieves these neer at Cisco Systems. He received a B.E.(Hons.) degree
from the University of Canterbury, New Zealand, in 1991,
new capabilities by extending the IP protocol and a Ph.D. from the University of Wollongong, Australia,
architecture, rather than replacing it. In general, in 1995. Since 1996 he has worked on MPLS issues at
all LSRs will use IP routing protocols. Much of Cisco Systems. This work started with design of aspects of
the complexity of routed IP network configura- ATM switching support in the MPLS protocols, and design
of MPLS support in the first carrier-class ATM label switch
tion is in the configuration of the routing proto- router. He has since carried out extensive work on the
col(s), and the same is true for the design of design of ATM MPLS networks, both in application to spe-
MPLS networks. cific networks and with regard to the design process itself.

142 IEEE Communications Magazine • July 2001