Network Working Group J. Boyle
Request for Comments: 3346 PD Nets
Category: Informational V. Gill
AOL Time Warner, Inc.
A. Hannan
RoutingLoop
D. Cooper
Global Crossing
D. AwdUChe
Movaz Networks
B. Christian
Worldcom
W.S. Lai
AT&T
August 2002

Applicability Statement for Traffic Engineering with MPLS

Status of this Memo

This memo provides information for the Internet community. It does
not specify an Internet standard of any kind. Distribution of this
memo is unlimited.

Copyright Notice

Copyright (C) The Internet Society (2002). All Rights Reserved.

Abstract

This document describes the applicability of Multiprotocol Label
Switching (MPLS) to traffic engineering in IP networks. Special
considerations for deployment of MPLS for traffic engineering in
operational contexts are discussed and the limitations of the MPLS
approach to traffic engineering are highlighted.

Table of Contents

1. Introduction....................................................2
2. Technical Overview of ISP Traffic Engineering...................3
3. Applicability of Internet Traffic Engineering...................4
3.1 Avoidance of Congested Resources................................4
3.2 Resource Utilization in Network Topologies with Parallel Links..5
3.3 Implementing Routing Policies using Affinities..................5
3.4 Re-optimization After Restoration...............................6
4. Implementation Considerations...................................6
4.1 Architectural and Operational Considerations....................6
4.2 Network Management ASPects......................................7
4.3 Capacity Engineering Aspects....................................8
4.4 Network Measurement Aspects.....................................8
5. Limitations.....................................................9
6. Conclusion.....................................................11
7. Security Considerations........................................11
8. References.....................................................11
9. Acknowledgments................................................12
10. Authors' Addresses.............................................13
11. Full Copyright Statement.......................................14

1. Introduction

It is generally acknowledged that one of the most significant initial
applications of Multiprotocol Label Switching (MPLS) is traffic
engineering (TE) [1][2] in IP networks. A significant community of
IP service providers have found that traffic engineering of their
networks can have tactical and strategic value [2, 3, 4]. To support
the traffic engineering application, extensions have been specified
for Interior Gateway Protocols (IGP) IS-IS [5] and OSPF [6], and to
signaling protocols RSVP [7] and LDP [8]. The extensions for IS-IS,
OSPF, and RSVP have all been developed and deployed in large scale in
many networks consisting of multi-vendor equipment.

This document discusses the applicability of TE to Internet service
provider networks, focusing on the MPLS-based approach. It augments
the existing protocol applicability statements and, in particular,
relates to the operational applicability of RSVP-TE [9]. Special
considerations for deployment of MPLS in operational contexts are
discussed and the limitations of this approach to traffic engineering
are highlighted.

2. Technical Overview of ISP Traffic Engineering

Traffic engineering (TE) is generally concerned with the performance
optimization of operational networks [2]. In contemporary practice,
TE means mapping IP traffic flows onto the existing physical network
topology in the most effective way to accomplish desired operational
objectives. Techniques currently used to accomplish this include,
but are not limited to:

1. Manipulation of IGP cost (metrics)
2. EXPlicit routing using constrained virtual-circuit
switching techniques such as ATM or Frame Relay SPVCs
3. Explicit routing using constrained path setup techniques
such as MPLS

This document is concerned primarily with MPLS techniques.
Specifically, it deals with the ability to use paths other than the
shortest paths selected by the IGP to achieve a more balanced network
utilization, e.g., by moving traffic away from IGP-selected shortest
paths onto alternate paths to avoid congestion in the network. This
can be achieved by using explicitly signaled LSP-tunnels. The
explicit routes to be used may be computed offline and subsequently
downloaded and configured on the routers using an appropriate
mechanism. Alternatively, the desired characteristics of an LSP
(such as endpoints, bandwidth, affinities) may be configured on a
router, which will then use an appropriate algorithm to compute a
path through the network satisfying the desired characteristics,
subject to various types of constraints. Generally, the
characteristics associated with LSPs may include:

o Ingress and egress nodes
o Bandwidth required
o Priority
o Nodes to include or exclude in the path
o Affinities to include or exclude in the path
o Resilience requirements

Affinities are arbitrary, provider-assigned, attributes applied to
links and carried in the TE extensions for the IGPs. Affinities
impose a class structure on links, which allow different links to be
logically grouped together. They can be used to implement various
types of policies, or route preferences that allow the inclusion or
exclusion of groups of links from the path of LSPs. Affinities are
unique to MPLS and the original requirement for them was documented
in [2].

3. Applicability of Internet Traffic Engineering

As mentioned in [2] and [7], traffic engineering with MPLS is
appropriate to establish and maintain explicitly routed paths in an
IP network for effective traffic placement. LSP-tunnels can be used
to forward subsets of traffic through paths that are independent of
routes computed by conventional IGP Shortest Path First (SPF)
algorithms. This gives network operators significant flexibility in
controlling the paths of traffic flows across their networks and
allows policies to be implemented that can result in the performance
optimization of networks. Examples of scenarios where MPLS-based TE
capabilities are applicable in service provider environments are
given below. The applicability of MPLS is certainly not restricted
to these scenarios.

3.1 Avoidance of Congested Resources

In order to lower the utilization of congested link(s), an operator
may utilize TE methods to route a subset of traffic away from those
links onto less congested topological elements. These types of
techniques are viable when segments of the network are congested
while other parts are underutilized.

Operators who do not make extensive use of LSP-tunnels may adopt a
tactical approach to MPLS TE in which they create LSP-tunnels only
when necessary to address specific congestion problems. For example,
an LSP can be created between two nodes (source and destination) that
are known to contribute traffic to a congested network element, and
explicitly route the LSP through a separate path to divert some
traffic away from the congestion. On the other hand, operators who
make extensive use of LSP-tunnels, either for measurement or
automated traffic control, may decide to explicitly route a subset of
the LSPs that traverse the point of congestion onto alternate paths.
This can be employed to respond quickly when the bandwidth parameter
associated with the LSPs does not accurately represent the actual
traffic carried by the LSPs, and the operator determines that
changing the bandwidth parameter values might not be effective in
addressing the issue or may not have lasting impact.

There are other approaches that measure traffic workloads on LSPs and
utilize these empirical statistics to configure various
characteristics of LSPs. These approaches, for example, can utilize
the derived statistics to configure explicit routes for LSPs (also
known as offline TE [10]). They can also utilize the statistics to
set the values of various LSP attributes such as bandwidths,
priority, and affinities (online TE). All of these approaches can be
used both tactically and strategically to react to periods of
congestion in a network. Congestion may occur as a result of many

factors: equipment or facility failure, longer than expected
provisioning cycles for new circuits, and unexpected surges in
traffic demand.

3.2 Resource Utilization in Network Topologies with Parallel Links

In practice, many service provider networks contain multiple parallel
links between nodes. An example is transoceanic connectivity which
is often provisioned as numerous low-capacity circuits, such as
NxDS-3 (N parallel DS-3 circuits) and NxSTM-1 (N parallel STM-1
circuits). Parallel circuits also occur quite often in bandwidth-
constrained cities. MPLS TE methods can be applied to effectively
distribute the aggregate traffic workload across these parallel
circuits.

MPLS-based approaches commonly used in practice to deal with parallel
links include using LSP bandwidth parameters to control the
proportion of demand traversing each link, explicitly configuring
routes for LSP-tunnels to distribute them across the parallel links,
and using affinities to map different LSPs onto different links.
These types of solutions are also applicable in networks with
parallel and replicated topologies, such as an NxOC-3/12/48 topology.

3.3 Implementing Routing Policies using Affinities

It is sometimes desirable to restrict certain types of traffic to
certain types of links, or to explicitly exclude certain types of
links in the paths for some types of traffic. This might be needed
to accomplish some business policy or network engineering objectives.
MPLS resource affinities provide a powerful mechanism to implement
these types of objectives.

As a concrete example, suppose a global service provider has a flat
(non-hierarchical) IGP. MPLS TE affinities can be used to explicitly
keep continental traffic (traffic originating and terminating within
a continent) from traversing transoceanic resources.

Another example of using MPLS TE affinities to exclude certain
traffic from a subset of circuits might be to keep inter-regional
LSPs off of circuits that are reserved for intra-regional traffic.

Still another example is the situation in a heterogeneous network
consisting of links with different capacities, e.g., OC-12, OC-48,
and OC-192. In such networks, affinities can be used to force some
types of traffic to only traverse links with a given capacity, e.g.
OC-48.

3.4 Re-optimization After Restoration

After the occurrence of a network failure, it may be desirable to
calculate a new set paths for LSPs to optimizes performance over the
residual topology. This re-optimization is complementary to the
fast-reroute operation used to reduce packet losses during routing
transients under network restoration. Traffic protection can also be
accomplished by associating a primary LSP with a set of secondary
LSPs, hot-standby LSPs, or a combination thereof [11].

4. Implementation Considerations

4.1 Architectural and Operational Considerations

When deploying TE solutions in a service provider environment, the
impact of administrative policies and the selection of nodes that
will serve as endpoints for LSP-tunnels should be carefully
considered. As noted in [9], when devising a virtual topology for
LSP-tunnels, special consideration should be given to the tradeoff
between the operational complexity associated with a large number of
LSP-tunnels and the control granularity that large numbers of LSP-
tunnels allow. In other Words, a large number of LSP-tunnels allow
greater control over the distribution of traffic across the network,
but increases network operational complexity. In large networks, it
may be advisable to start with a simple LSP-tunnel virtual topology
and then introduce additional complexity based on observed or
anticipated traffic flow patterns [9].

Administrative policies should guide the amount of bandwidth to be
allocated to an LSP. One may choose to set the bandwidth of a
particular LSP to a statistic of the measured observed utilization
over an interval of time, e.g., peak rate, or a particular percentile
or quartile of the observed utilization. Sufficient over-
subscription (of LSPs) or under-reporting bandwidth on the physical
links should be used to account for flows that exceed their normal
limits on an event-driven basis. Flows should be monitored for
trends that indicate when the bandwidth parameter of an LSP should be
resized. Flows should be monitored constantly to detect unusual
variance from expected levels. If an unpoliced flow greatly exceeds
its assigned bandwidth, action should be taken to determine the root
cause and remedy the problem. Traffic policing is an option that may
be applied to deal with congestion problems, especially when some
flows exceed their bandwidth parameters and interfere with other
compliant flows. However, it is usually more prudent to apply
policing actions at the edge of the network rather than within the
core, unless under exceptional circumstances.

When creating LSPs, a hierarchical network approach may be used to
alleviate scalability problems associated with flat full mesh virtual
topologies. In general, operational experience has shown that very
large flows (between city pairs) are long-lived and have stable
characteristics, while smaller flows (edge to edge) are more dynamic
and have more fluctuating statistical characteristics. A
hierarchical architecture can be devised consisting of core and edge
networks in which the core is traffic engineered and serves as an
aggregation and transit infrastructure for edge traffic.

However, over-aggregation of flows can result in a stream so large
that it precludes the constraint-based routing algorithm from finding
a feasible path through a network. Splitting a flow by using two or
more parallel LSPs and distributing the traffic across the LSPs can
solve this problem, at the expense of introducing more state in the
network.

Failure scenarios should also be addressed when splitting a stream of
traffic over several links. It is of little value to establish a
finely balanced set of flows over a set of links only to find that
upon link failure the balance reacts poorly, or does not revert to
the original situation upon restoration.

4.2 Network Management Aspects

Networks planning to deploy MPLS for traffic engineering must
consider network management aspects, particularly performance and
fault management [12]. With the deployment of MPLS in any
infrastructure, some additional operational tasks are required, such
as constant monitoring to ensure that the performance of the network
is not impacted in the end-to-end delivery of traffic. In addition,
traffic characteristics, such as latency across an LSP, may also need
to be assessed on a regular basis to ensure that service-level
guarantees are achieved.

OBTaining information on LSP behavior is critical in determining the
stability of an MPLS network. When LSPs transition or path changes
occur, packets may be dropped which impacts network performance. It
should be the goal of any network deploying MPLS to minimize the
volatility of LSPs and reduce the root causes that induce this
instability. Unfortunately, there are very few, if any, NMS systems
that are available at this time with the capability to provide the
correct level of management support, particularly root cause
analysis. Consequently, most early adopters of MPLS develop their
own management systems in-house for the MPLS domain. The lack of
availability of commercial network management systems that deal
specifically with MPLS-related aspects is a significant impediment to
the large-scale deployment of MPLS networks.

The performance of an MPLS network is also dependent on the
configured values of bandwidth for each LSP. Since congestion is a
common cause of performance degradation in operational networks, it
is important to proactively avoid these situations. While MPLS was
designed to minimize congestion on links by utilizing bandwidth
reservations, it is still heavily reliant on user configurable data.
If the LSP bandwidth value does not properly represent the traffic
demand of that LSP, over-utilization may occur and cause significant
congestion within the network. Therefore, it is important to
develop, deploy, and maintain a good modeling tool for determining
LSP bandwidth size. Lack of this capability may result in sub-
optimal network performance.

4.3 Capacity Engineering Aspects

Traffic engineering has a goal of ensuring traffic performance
objectives for different services. This requires that the different
network elements be dimensioned properly to handle the expected load.
More specifically, in mapping given user demands onto network
resources, network dimensioning involves the sizing of the network
elements, such as links, processors, and buffers, so that performance
objectives can be met at minimum cost. Major inputs to the
dimensioning process are cost models, characterization of user
demands and specification of performance objectives.

In using MPLS, dimensioning involves the assignment of resources such
as bandwidth to a set of pre-selected LSPs for carrying traffic, and
mapping the logical network of LSPs onto a physical network of links
with capacity constraints. The dimensioning process also determines
the link capacity parameters or thresholds associated with the use of
some bandwidth reservation scheme for service protection. Service
protection controls the QoS for certain service types by restricting
Access to bandwidth, or by giving priority access to one type of
traffic over another. Such methods are essential, e.g., to prevent
starvation of low-priority flows, to guarantee a minimum amount of
resources for flows with expected short duration, to improve the
acceptance probability for flows with high bandwidth requirements, or
to maintain network stability by preventing performance degradation
in case of a local overload.

4.4 Network Measurement Aspects

Network measurement entails robust statistics collection and systems
development. Knowing *what* to do with these measurements is often
where the secret-sauce is. Examples for different applications of
measurements are described in [13]. For instance, to ensure that the
QoS objectives have been met, performance measurements and
performance monitoring are required so that real-time traffic control

actions, or policy-based actions, can be taken. Also, to
characterize the traffic demands, traffic measurements are used to
estimate the offered loads from different service classes and to
provide forecasting of future demands for capacity planning purposes.
Forecasting and planning may result in capacity augmentation or may
lead to the introduction of new technology and architecture.

To avoid QoS degradation at the packet level, measurement-based
admission control can be employed by using online measurements of
actual usage. This is a form of preventive control to ensure that
the QoS requirements of different service classes can be met
simultaneously, while maintaining network efficiency at a high level.
However, it requires proper network dimensioning to keep the
probability for the refusal of connection/flow requests sufficiently
low.

5. Limitations

Significant resources can be expended to gain a proper understanding
of how MPLS works. Furthermore, significant engineering and testing
resources may need to be invested to identify problems with vendor
implementations of MPLS. Initial deployment of MPLS software and the
configurations management aspects to support TE can consume
significant engineering, operations, and system development
resources. Developing automated systems to create router
configurations for network elements can require significant software
development and hardware resources. Getting to a point where
configurations for routers are updated in an automated fashion can be
a time consuming process. Tracking manual tweaks to router
configurations, or problems associated with these can be an endless
task. What this means is that much more is required in the form of
various types of tools to simplify and automate the MPLS TE function.

Certain architectural choices can lead to operational, protocol, and
router implementation scalability problems. This is especially true
as the number of LSP-tunnels or router configuration data in a
network increases, which can be exacerbated by designs incorporating
full meshes, which create O(N^2) number of LSPs, where N is the
number of network-edge nodes. In these cases, minimizing N through
hierarchy, regionalization, or proper selection of tunnel termination
points can affect the network's ability to scale. Loss of scale in
this sense can be via protocol instability, inability to change
network configurations to accommodate growth, inability for router
implementations to be updated, hold or properly process
configurations, or loss of ability to adequately manage the network.

Although widely deployed, MPLS TE is a new technology when compared
to the classic IP routing protocols such as IS-IS, OSPF, and BGP.
MPLS TE also has more configuration and protocol options. As such,
some implementations are not battle-hardened and automated testing of
various configurations is difficult if not infeasible. Multi-vendor
environments are beginning to appear, although additional effort is
usually required to ensure full interoperability.

Common approaches to TE in service provider environments switch the
forwarding paradigm from connectionless to connection oriented.
Thus, operational analysis of the network may be complicated in some
regards (and improved in others). Inconsistencies in forwarding
state result in dropped packets whereas with connectionless methods
the packet will either loop and drop, or be misdirected onto another
branch in the routing tree.

Currently deployed MPLS TE approaches can be adversely affected by
both internal and external router and link failures. This can create
a mismatch between the signaled capacity and the traffic an LSP-
tunnel carries.

Many routers in service provider environments are already under
stress processing the software workload associated with running IGP,
BGP, and IPC. Enabling TE in an MPLS environment involves adding
traffic engineering databases and processes, adding additional
information to be carried by the routing processes, and adding
signaling state and processing to these network elements. Additional
traffic measurements may also need to be supported. In some
environments, this additional load may not be feasible.

MPLS in general and MPLS-TE in particular is not a panacea for lack
of network capacity, or lack of proper capacity planning and
provisioning in the network dimensioning process. MPLS-TE may cause
network traffic to traverse greater distances or to take paths with
more network elements, thereby incurring greater latency. Generally,
this added inefficiency is done to prevent shortcomings in capacity
planning or available resources path to avoid hot spots. The ability
of TE to accommodate more traffic on a given topology can also be
characterized as a short-term gain during periods of persistent
traffic growth. These approaches cannot achieve impossible mappings
of traffic onto topologies. Failure to properly capacity plan and
execute will lead to congestion, no matter what technology aids are
employed.

6. Conclusion

The applicability of traffic engineering in Internet service provider
environments has been discussed in this document. The focus has been
on the use of MPLS-based approaches to achieve traffic engineering in
this context. The applicability of traffic engineering and
associated management and deployment considerations have been
described, and the limitations highlighted.

MPLS combines the ability to monitor point-to-point traffic
statistics between two routers and the capability to control the
forwarding paths of subsets of traffic through a given network
topology. This makes traffic engineering with MPLS applicable and
useful for improving network performance by effectively mapping
traffic flows onto links within service provider networks. Tools
that simplify and automate the MPLS TE functions and activation help
to realize the full potential.

7. Security Considerations

This document does not introduce new security issues. When deployed
in service provider networks, it is mandatory to ensure that only
authorized entities are permitted to initiate establishment of LSP-
tunnels.

8. References

1 Rosen, E., Viswanathan, A. and R. Callon, "Multiprotocol Label
Switching Architecture," RFC3031, January 2001.

2 Awduche, D., Malcolm, J., Agogbua, J., O'Dell, M. and J. McManus,
"Requirements for Traffic Engineering Over MPLS," RFC2702,
September 1999.

3 X. Xiao, A. Hannan, B. Bailey, and L. Ni, "Traffic Engineering
with MPLS in the Internet," IEEE Network, March/April 2000.

4 V. Springer, C. Pierantozzi, and J. Boyle, "Level3 MPLS Protocol
Architecture," Work in Progress.

5 T. Li, and H. Smit, "IS-IS Extensions for Traffic Engineering,"
Work in Progress.

6 D. Katz, D. Yeung, and K. Kompella, "Traffic Engineering
Extensions to OSPF," Work in Progress.

7 Awduche, D., Berger, L., Gan, D.H., Li, T., Srinivasan, V. and G.
Swallow, "RSVP-TE: Extensions to RSVP for LSP Tunnels," RFC3209,
December 2001.

8 Jamoussi, B. (Editor), "Constraint-Based LSP Setup using LDP," RFC
3212, January 2002.

9 Awduche, D., Hannan, A. and X. Xiao, "Applicability Statement for
Extensions to RSVP for LSP-Tunnels," RFC3210, December 2001.

10 Awduche, D., Chiu, A., Elwalid, A., Widjaja, I. and X. Xiao,
"Overview and Principles of Internet Traffic Engineering", RFC
3272, May 2002.

11 W.S. Lai, D. McDysan, J. Boyle, M. Carlzon, R. Coltun, T.
Griffin, E. Kern, and T. Reddington, "Network Hierarchy and
Multilayer Survivability," Work in Progress.

12 D. Awduche, "MPLS and Traffic Engineering in IP Networks," IEEE
Communications Magazine, December 1999.

13 W.S. Lai, B. Christian, R.W. Tibbs, and S. Van den Berghe, "A
Framework for Internet Traffic Engineering Measurement," Work in
Progress.

9. Acknowledgments

The effectiveness of the MPLS protocols for traffic engineering in
service provider networks is in large part due to the experience
gained and foresight given by network engineers and developers
familiar with traffic engineering with ATM in these environments. In
particular, the authors wish to acknowledge the authors of RFC2702
for the clear articulation of the requirements, as well as the
developers and testers of code in deployment today for keeping their
focus.

10. Authors' Addresses

Jim Boyle
Protocol Driven Networks
Tel: +1 919-852-5160
EMail: jboyle@pdnets.com

Vijay Gill
AOL Time Warner, Inc.
12100 Sunrise Valley Drive
Reston, VA 20191
EMail: vijay@umbc.edu

Alan Hannan
RoutingLoop Intergalactic
112 Falkirk Court
Sunnyvale, CA 94087, USA
Tel: +1 408-666-2326
EMail: alan@routingloop.com

Dave Cooper
Global Crossing
960 Hamlin Court
Sunnyvale, CA 94089, USA
Tel: +1 916-415-0437
EMail: dcooper@gblx.net

Daniel O. Awduche
Movaz Networks
7926 Jones Branch Drive, Suite 615
McLean, VA 22102, USA
Tel: +1 703-298-5291
EMail: awduche@movaz.com

Blaine Christian
Worldcom
22001 Loudoun County Parkway, Room D1-2-737
Ashburn, VA 20147, USA
Tel: +1 703-886-4425
EMail: blaine@uu.net

Wai Sum Lai
AT&T
200 Laurel Avenue
Middletown, NJ 07748, USA
Tel: +1 732-420-3712
EMail: wlai@att.com

11. Full Copyright Statement

Copyright (C) The Internet Society (2002). 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
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.

Acknowledgement

Funding for the RFCEditor function is currently provided by the
Internet Society.