Network Working Group D. AwdUChe
Request for Comments: 2702 J. Malcolm
Category: Informational J. Agogbua
M. O'Dell
J. McManus
UUNET (MCI Worldcom)
September 1999

Requirements for Traffic Engineering Over 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 (1999). All Rights Reserved.

Abstract

This document presents a set of requirements for Traffic Engineering
over Multiprotocol Label Switching (MPLS). It identifies the
functional capabilities required to implement policies that
facilitate efficient and reliable network operations in an MPLS
domain. These capabilities can be used to optimize the utilization of
network resources and to enhance traffic oriented performance
characteristics.

Table of Contents

1.0 Introduction ............................................. 2
1.1 Terminology .............................................. 3
1.2 Document Organization .................................... 3
2.0 Traffic Engineering ...................................... 4
2.1 Traffic Engineering Performance Objectives ............... 4
2.2 Traffic and Resource Control ............................. 6
2.3 Limitations of Current IGP Control Mechanisms ............ 6
3.0 MPLS and Traffic Engineering ............................. 7
3.1 Induced MPLS Graph ....................................... 9
3.2 The Fundamental Problem of Traffic Engineering Over MPLS . 9
4.0 Augmented Capabilities for Traffic Engineering Over MPLS . 10
5.0 Traffic Trunk Attributes and Characteristics ........... 10
5.1 Bidirectional Traffic Trunks ............................. 11
5.2 Basic Operations on Traffic Trunks ....................... 12
5.3 Accounting and Performance Monitoring .................... 12

5.4 Basic Attributes of Traffic Trunks ....................... 13
5.5 Traffic Parameter Attributes ............................ 14
5.6 Generic Path Selection and Management Attributes ......... 14
5.6.1 Administratively Specified EXPlicit Paths ................ 15
5.6.2 Hierarchy of Preference Rules for Multi-paths ............ 15
5.6.3 Resource Class Affinity Attributes ....................... 16
5.6.4 Adaptivity Attribute ..................................... 17
5.6.5 Load Distribution Across Parallel Traffic Trunks ......... 18
5.7 Priority Attribute ....................................... 18
5.8 Preemption Attribute ..................................... 18
5.9 Resilience Attribute ..................................... 19
5.10 Policing Attribute ...................................... 20
6.0 Resource Attributes ...................................... 21
6.1 Maximum Allocation Multiplier ............................ 21
6.2 Resource Class Attribute ................................ 22
7.0 Constraint-Based Routing ................................ 22
7.1 Basic Features of Constraint-Based Routing ............... 23
7.2 Implementation Considerations ............................ 24
8.0 Conclusion ............................................. 25
9.0 Security Considerations .................................. 26
10.0 References ............................................. 26
11.0 Acknowledgments .......................................... 27
12.0 Authors' Addresses ....................................... 28
13.0 Full Copyright Statement ................................. 29

1.0 Introduction

Multiprotocol Label Switching (MPLS) [1,2] integrates a label
swapping framework with network layer routing. The basic idea
involves assigning short fixed length labels to packets at the
ingress to an MPLS cloud (based on the concept of forwarding
equivalence classes [1,2]). Throughout the interior of the MPLS
domain, the labels attached to packets are used to make forwarding
decisions (usually without recourse to the original packet headers).

A set of powerful constructs to address many critical issues in the
emerging differentiated services Internet can be devised from this
relatively simple paradigm. One of the most significant initial
applications of MPLS will be in Traffic Engineering. The importance
of this application is already well-recognized (see [1,2,3]).

This manuscript is exclusively focused on the Traffic Engineering
applications of MPLS. Specifically, the goal of this document is to
highlight the issues and requirements for Traffic Engineering in a
large Internet backbone. The expectation is that the MPLS
specifications, or implementations derived therefrom, will address

the realization of these objectives. A description of the basic
capabilities and functionality required of an MPLS implementation to
accommodate the requirements is also presented.

It should be noted that even though the focus is on Internet
backbones, the capabilities described in this document are equally
applicable to Traffic Engineering in enterprise networks. In general,
the capabilities can be applied to any label switched network under
a single technical administration in which at least two paths exist
between two nodes.

Some recent manuscripts have focused on the considerations pertaining
to Traffic Engineering and Traffic management under MPLS, most
notably the works of Li and Rekhter [3], and others. In [3], an
architecture is proposed which employs MPLS and RSVP to provide
scalable differentiated services and Traffic Engineering in the
Internet. The present manuscript complements the aforementioned and
similar efforts. It reflects the authors' operational experience in
managing a large Internet backbone.

1.1 Terminology

The reader is assumed to be familiar with the MPLS terminology as
defined in [1].

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 RFC2119 [11].

1.2 Document Organization

The remainder of this document is organized as follows: Section 2
discusses the basic functions of Traffic Engineering in the Internet.
Section 3, provides an overview of the traffic Engineering potentials
of MPLS. Sections 1 to 3 are essentially background material. Section
4 presents an overview of the fundamental requirements for Traffic
Engineering over MPLS. Section 5 describes the desirable attributes
and characteristics of traffic trunks which are pertinent to Traffic
Engineering. Section 6 presents a set of attributes which can be
associated with resources to constrain the routability of traffic
trunks and LSPs through them. Section 7 advocates the introduction of
a "constraint-based routing" framework in MPLS domains. Finally,
Section 8 contains concluding remarks.

2.0 Traffic Engineering

This section describes the basic functions of Traffic Engineering in
an Autonomous System in the contemporary Internet. The limitations of
current IGPs with respect to traffic and resource control are
highlighted. This section serves as motivation for the requirements
on MPLS.

Traffic Engineering (TE) is concerned with performance optimization
of operational networks. In general, it encompasses the application
of technology and scientific principles to the measurement, modeling,
characterization, and control of Internet traffic, and the
application of such knowledge and techniques to achieve specific
performance objectives. The ASPects of Traffic Engineering that are
of interest concerning MPLS are measurement and control.

A major goal of Internet Traffic Engineering is to facilitate
efficient and reliable network operations while simultaneously
optimizing network resource utilization and traffic performance.
Traffic Engineering has become an indispensable function in many
large Autonomous Systems because of the high cost of network assets
and the commercial and competitive nature of the Internet. These
factors emphasize the need for maximal operational efficiency.

2.1 Traffic Engineering Performance Objectives

The key performance objectives associated with traffic engineering
can be classified as being either:

1. traffic oriented or

2. resource oriented.

Traffic oriented performance objectives include the aspects that
enhance the QoS of traffic streams. In a single class, best effort
Internet service model, the key traffic oriented performance
objectives include: minimization of packet loss, minimization of
delay, maximization of throughput, and enforcement of service level
agreements. Under a single class best effort Internet service model,
minimization of packet loss is one of the most important traffic
oriented performance objectives. Statistically bounded traffic
oriented performance objectives (such as peak to peak packet delay
variation, loss ratio, and maximum packet transfer delay) might
become useful in the forthcoming differentiated services Internet.

Resource oriented performance objectives include the aspects
pertaining to the optimization of resource utilization. Efficient
management of network resources is the vehicle for the attainment of

resource oriented performance objectives. In particular, it is
generally desirable to ensure that subsets of network resources do
not become over utilized and congested while other subsets along
alternate feasible paths remain underutilized. Bandwidth is a crucial
resource in contemporary networks. Therefore, a central function of
Traffic Engineering is to efficiently manage bandwidth resources.

Minimizing congestion is a primary traffic and resource oriented
performance objective. The interest here is on congestion problems
that are prolonged rather than on transient congestion resulting from
instantaneous bursts. Congestion typically manifests under two
scenarios:

1. When network resources are insufficient or inadequate to
accommodate offered load.

2. When traffic streams are inefficiently mapped onto available
resources; causing subsets of network resources to become
over-utilized while others remain underutilized.

The first type of congestion problem can be addressed by either: (i)
expansion of capacity, or (ii) application of classical congestion
control techniques, or (iii) both. Classical congestion control
techniques attempt to regulate the demand so that the traffic fits
onto available resources. Classical techniques for congestion control
include: rate limiting, window flow control, router queue management,
schedule-based control, and others; (see [8] and the references
therein).

The second type of congestion problems, namely those resulting from
inefficient resource allocation, can usually be addressed through
Traffic Engineering.

In general, congestion resulting from inefficient resource allocation
can be reduced by adopting load balancing policies. The objective of
such strategies is to minimize maximum congestion or alternatively to
minimize maximum resource utilization, through efficient resource
allocation. When congestion is minimized through efficient resource
allocation, packet loss decreases, transit delay decreases, and
aggregate throughput increases. Thereby, the perception of network
service quality experienced by end users becomes significantly
enhanced.

Clearly, load balancing is an important network performance
optimization policy. Nevertheless, the capabilities provided for
Traffic Engineering should be flexible enough so that network
administrators can implement other policies which take into account
the prevailing cost structure and the utility or revenue model.

2.2 Traffic and Resource Control

Performance optimization of operational networks is fundamentally a
control problem. In the traffic engineering process model, the
Traffic Engineer, or a suitable automaton, acts as the controller in
an adaptive feedback control system. This system includes a set of
interconnected network elements, a network performance monitoring
system, and a set of network configuration management tools. The
Traffic Engineer formulates a control policy, observes the state of
the network through the monitoring system, characterizes the traffic,
and applies control actions to drive the network to a desired state,
in accordance with the control policy. This can be accomplished
reactively by taking action in response to the current state of the
network, or pro-actively by using forecasting techniques to
anticipate future trends and applying action to obviate the predicted
undesirable future states.

Ideally, control actions should involve:

1. Modification of traffic management parameters,

2. Modification of parameters associated with routing, and

3. Modification of attributes and constraints associated with
resources.

The level of manual intervention involved in the traffic engineering
process should be minimized whenever possible. This can be
accomplished by automating aspects of the control actions described
above, in a distributed and scalable fashion.

2.3 Limitations of Current IGP Control Mechanisms

This subsection reviews some of the well known limitations of current
IGPs with regard to Traffic Engineering.

The control capabilities offered by existing Internet interior
gateway protocols are not adequate for Traffic Engineering. This
makes it difficult to actualize effective policies to address network
performance problems. Indeed, IGPs based on shortest path algorithms
contribute significantly to congestion problems in Autonomous Systems
within the Internet. SPF algorithms generally optimize based on a
simple additive metric. These protocols are topology driven, so
bandwidth availability and traffic characteristics are not factors
considered in routing decisions. Consequently, congestion frequently
occurs when:

1. the shortest paths of multiple traffic streams converge on
specific links or router interfaces, or

2. a given traffic stream is routed through a link or router
interface which does not have enough bandwidth to accommodate
it.

These scenarios manifest even when feasible alternate paths with
excess capacity exist. It is this aspect of congestion problems (-- a
symptom of suboptimal resource allocation) that Traffic Engineering
aims to vigorously obviate. Equal cost path load sharing can be used
to address the second cause for congestion listed above with some
degree of success, however it is generally not helpful in alleviating
congestion due to the first cause listed above and particularly not
in large networks with dense topology.

A popular approach to circumvent the inadequacies of current IGPs is
through the use of an overlay model, such as IP over ATM or IP over
frame relay. The overlay model extends the design space by enabling
arbitrary virtual topologies to be provisioned atop the network's
physical topology. The virtual topology is constructed from virtual
circuits which appear as physical links to the IGP routing protocols.
The overlay model provides additional important services to support
traffic and resource control, including: (1) constraint-based routing
at the VC level, (2) support for administratively configurable
explicit VC paths, (3) path compression, (4) call admission control
functions, (5) traffic shaping and traffic policing functions, and
(6) survivability of VCs. These capabilities enable the actualization
of a variety of Traffic Engineering policies. For example, virtual
circuits can easily be rerouted to move traffic from over-utilized
resources onto relatively underutilized ones.

For Traffic Engineering in large dense networks, it is desirable to
equip MPLS with a level of functionality at least commensurate with
current overlay models. Fortunately, this can be done in a fairly
straight forward manner.

3.0 MPLS and Traffic Engineering

This section provides an overview of the applicability of MPLS to
Traffic Engineering. Subsequent sections discuss the set of
capabilities required to meet the Traffic Engineering requirements.

MPLS is strategically significant for Traffic Engineering because it
can potentially provide most of the functionality available from the
overlay model, in an integrated manner, and at a lower cost than the
currently competing alternatives. Equally importantly, MPLS offers

the possibility to automate aspects of the Traffic Engineering
function. This last consideration requires further investigation and
is beyond the scope of this manuscript.

A note on terminology: The concept of MPLS traffic trunks is used
extensively in the remainder of this document. According to Li and
Rekhter [3], a traffic trunk is an aggregation of traffic flows of
the same class which are placed inside a Label Switched Path.
Essentially, a traffic trunk is an abstract representation of traffic
to which specific characteristics can be associated. It is useful to
view traffic trunks as objects that can be routed; that is, the path
through which a traffic trunk traverses can be changed. In this
respect, traffic trunks are similar to virtual circuits in ATM and
Frame Relay networks. It is important, however, to emphasize that
there is a fundamental distinction between a traffic trunk and the
path, and indeed the LSP, through which it traverses. An LSP is a
specification of the label switched path through which the traffic
traverses. In practice, the terms LSP and traffic trunk are often
used synonymously. Additional characteristics of traffic trunks as
used in this manuscript are summarized in section 5.0.

The attractiveness of MPLS for Traffic Engineering can be attributed
to the following factors: (1) explicit label switched paths which are
not constrained by the destination based forwarding paradigm can be
easily created through manual administrative action or through
automated action by the underlying protocols, (2) LSPs can
potentially be efficiently maintained, (3) traffic trunks can be
instantiated and mapped onto LSPs, (4) a set of attributes can be
associated with traffic trunks which modulate their behavioral
characteristics, (5) a set of attributes can be associated with
resources which constrain the placement of LSPs and traffic trunks
across them, (6) MPLS allows for both traffic aggregation and
disaggregation whereas classical destination only based IP forwarding
permits only aggregation, (7) it is relatively easy to integrate a
"constraint-based routing" framework with MPLS, (8) a good
implementation of MPLS can offer significantly lower overhead than
competing alternatives for Traffic Engineering.

Additionally, through explicit label switched paths, MPLS permits a
quasi circuit switching capability to be superimposed on the current
Internet routing model. Many of the existing proposals for Traffic
Engineering over MPLS focus only on the potential to create explicit
LSPs. Although this capability is fundamental for Traffic
Engineering, it is not really sufficient. Additional augmentations
are required to foster the actualization of policies leading to
performance optimization of large operational networks. Some of the
necessary augmentations are described in this manuscript.

3.1 Induced MPLS Graph

This subsection introduces the concept of an "induced MPLS graph"
which is central to Traffic Engineering in MPLS domains. An induced
MPLS graph is analogous to a virtual topology in an overlay model. It
is logically mapped onto the physical network through the selection
of LSPs for traffic trunks.

An induced MPLS graph consists of a set of LSRs which comprise the
nodes of the graph and a set of LSPs which provide logical point to
point connectivity between the LSRs, and hence serve as the links of
the induced graph. it may be possible to construct hierarchical
induced MPLS graphs based on the concept of label stacks (see [1]).

Induced MPLS graphs are important because the basic problem of
bandwidth management in an MPLS domain is the issue of how to
efficiently map an induced MPLS graph onto the physical network
topology. The induced MPLS graph abstraction is formalized below.

Let G = (V, E, c) be a capacitated graph depicting the physical
topology of the network. Here, V is the set of nodes in the network
and E is the set of links; that is, for v and w in V, the object
(v,w) is in E if v and w are directly connected under G. The
parameter "c" is a set of capacity and other constraints associated
with E and V. We will refer to G as the "base" network topology.

Let H = (U, F, d) be the induced MPLS graph, where U is a subset of
V representing the set of LSRs in the network, or more precisely the
set of LSRs that are the endpoints of at least one LSP. Here, F is
the set of LSPs, so that for x and y in U, the object (x, y) is in F
if there is an LSP with x and y as endpoints. The parameter "d" is
the set of demands and restrictions associated with F. Evidently, H
is a directed graph. It can be seen that H depends on the
transitivity characteristics of G.

3.2 The Fundamental Problem of Traffic Engineering Over MPLS

There are basically three fundamental problems that relate to Traffic
Engineering over MPLS.

- The first problem concerns how to map packets onto forwarding
equivalence classes.

- The second problem concerns how to map forwarding equivalence
classes onto traffic trunks.

- The third problem concerns how to map traffic trunks onto the
physical network topology through label switched paths.

This document is not focusing on the first two problems listed.
(even-though they are quite important). Instead, the remainder of
this manuscript will focus on the capabilities that permit the third
mapping function to be performed in a manner resulting in efficient
and reliable network operations. This is really the problem of
mapping an induced MPLS graph (H) onto the "base" network topology
(G).

4.0 Augmented Capabilities for Traffic Engineering Over MPLS

The previous sections reviewed the basic functions of Traffic
Engineering in the contemporary Internet. The applicability of MPLS
to that activity was also discussed. The remaining sections of this
manuscript describe the functional capabilities required to fully
support Traffic Engineering over MPLS in large networks.

The proposed capabilities consist of:

1. A set of attributes associated with traffic trunks which
collectively specify their behavioral characteristics.

2. A set of attributes associated with resources which constrain
the placement of traffic trunks through them. These can also be
viewed as topology attribute constraints.

3. A "constraint-based routing" framework which is used to select
paths for traffic trunks subject to constraints imposed by items
1) and 2) above. The constraint-based routing framework does not
have to be part of MPLS. However, the two need to be tightly
integrated together.

The attributes associated with traffic trunks and resources, as well
as parameters associated with routing, collectively represent the
control variables which can be modified either through administrative
action or through automated agents to drive the network to a desired
state.

In an operational network, it is highly desirable that these
attributes can be dynamically modified online by an operator without
adversely disrupting network operations.

5.0 Traffic Trunk Attributes and Characteristics

This section describes the desirable attributes which can be
associated with traffic trunks to influence their behavioral
characteristics.

First, the basic properties of traffic trunks (as used in this
manuscript) are summarized below:

- A traffic trunk is an *aggregate* of traffic flows belonging
to the same class. In some contexts, it may be desirable to
relax this definition and allow traffic trunks to include
multi-class traffic aggregates.

- In a single class service model, such as the current Internet,
a traffic trunk could encapsulate all of the traffic between an
ingress LSR and an egress LSR, or subsets thereof.

- Traffic trunks are routable objects (similar to ATM VCs).

- A traffic trunk is distinct from the LSP through which it
traverses. In operational contexts, a traffic trunk can be
moved from one path onto another.

- A traffic trunk is unidirectional.

In practice, a traffic trunk can be characterized by its ingress and
egress LSRs, the forwarding equivalence class which is mapped onto
it, and a set of attributes which determine its behavioral
characteristics.

Two basic issues are of particular significance: (1) parameterization
of traffic trunks and (2) path placement and maintenance rules for
traffic trunks.

5.1 Bidirectional Traffic Trunks

Although traffic trunks are conceptually unidirectional, in many
practical contexts, it is useful to simultaneously instantiate two
traffic trunks with the same endpoints, but which carry packets in
opposite directions. The two traffic trunks are logically coupled
together. One trunk, called the forward trunk, carries traffic from
an originating node to a destination node. The other trunk, called
the backward trunk, carries traffic from the destination node to the
originating node. We refer to the amalgamation of two such traffic
trunks as one bidirectional traffic trunk (BTT) if the following two
conditions hold:

- Both traffic trunks are instantiated through an atomic action at
one LSR, called the originator node, or through an atomic action
at a network management station.

- Neither of the composite traffic trunks can exist without the
other. That is, both are instantiated and destroyed together.

The topological properties of BTTs should also be considered. A BTT
can be topologically symmetric or topologically asymmetric. A BTT is
said to be "topologically symmetric" if its constituent traffic
trunks are routed through the same physical path, even though they
operate in opposite directions. If, however, the component traffic
trunks are routed through different physical paths, then the BTT is
said to be "topologically asymmetric."

It should be noted that bidirectional traffic trunks are merely an
administrative convenience. In practice, most traffic engineering
functions can be implemented using only unidirectional traffic
trunks.

5.2 Basic Operations on Traffic Trunks

The basic operations on traffic trunks significant to Traffic
Engineering purposes are summarized below.

- Establish: To create an instance of a traffic trunk.

- Activate: To cause a traffic trunk to start passing traffic.
The establishment and activation of a traffic trunk are
logically separate events. They may, however, be implemented
or invoked as one atomic action.

- Deactivate: To cause a traffic trunk to stop passing traffic.

- Modify Attributes: To cause the attributes of a traffic trunk
to be modified.

- Reroute: To cause a traffic trunk to change its route. This
can be done through administrative action or automatically
by the underlying protocols.

- Destroy: To remove an instance of a traffic trunk from the
network and reclaim all resources allocated to it. Such
resources include label space and possibly available bandwidth.

The above are considered the basic operations on traffic trunks.
Additional operations are also possible such as policing and traffic
shaping.

5.3 Accounting and Performance Monitoring

Accounting and performance monitoring capabilities are very important
to the billing and traffic characterization functions. Performance
statistics obtained from accounting and performance monitoring

systems can be used for traffic characterization, performance
optimization, and capacity planning within the Traffic Engineering
realm..

The capability to obtain statistics at the traffic trunk level is so
important that it should be considered an essential requirement for
Traffic Engineering over MPLS.

5.4 Basic Traffic Engineering Attributes of Traffic Trunks

An attribute of a traffic trunk is a parameter assigned to it which
influences its behavioral characteristics.

Attributes can be explicitly assigned to traffic trunks through
administration action or they can be implicitly assigned by the
underlying protocols when packets are classified and mapped into
equivalence classes at the ingress to an MPLS domain. Regardless of
how the attributes were originally assigned, for Traffic Engineering
purposes, it should be possible to administratively modify such
attributes.

The basic attributes of traffic trunks particularly significant for
Traffic Engineering are itemized below.

- Traffic parameter attributes

- Generic Path selection and maintenance attributes

- Priority attribute

- Preemption attribute

- Resilience attribute

- Policing attribute

The combination of traffic parameters and policing attributes is
analogous to usage parameter control in ATM networks. Most of the
attributes listed above have analogs in well established
technologies. Consequently, it should be relatively straight forward
to map the traffic trunk attributes onto many existing switching and
routing architectures.

Priority and preemption can be regarded as relational attributes
because they express certain binary relations between traffic trunks.
Conceptually, these binary relations determine the manner in which
traffic trunks interact with each other as they compete for network
resources during path establishment and path maintenance.

5.5 Traffic parameter attributes

Traffic parameters can be used to capture the characteristics of the
traffic streams (or more precisely the forwarding equivalence class)
to be transported through the traffic trunk. Such characteristics may
include peak rates, average rates, permissible burst size, etc. From
a traffic engineering perspective, the traffic parameters are
significant because they indicate the resource requirements of the
traffic trunk. This is useful for resource allocation and congestion
avoidance through anticipatory policies.

For the purpose of bandwidth allocation, a single canonical value of
bandwidth requirements can be computed from a traffic trunk's traffic
parameters. Techniques for performing these computations are well
known. One example of this is the theory of effective bandwidth.

5.6 Generic Path Selection and Management Attributes

Generic path selection and management attributes define the rules for
selecting the route taken by a traffic trunk as well as the rules for
maintenance of paths that are already established.

Paths can be computed automatically by the underlying routing
protocols or they can be defined administratively by a network
operator. If there are no resource requirements or restrictions
associated with a traffic trunk, then a topology driven protocol can
be used to select its path. However, if resource requirements or
policy restrictions exist, then a constraint-based routing scheme
should be used for path selection.

In Section 7, a constraint-based routing framework which can
automatically compute paths subject to a set of constraints is
described. Issues pertaining to explicit paths instantiated through
administrative action are discussed in Section 5.6.1 below.

Path management concerns all aspects pertaining to the maintenance of
paths traversed by traffic trunks. In some operational contexts, it
is desirable that an MPLS implementation can dynamically reconfigure
itself, to adapt to some notion of change in "system state."
Adaptivity and resilience are aspects of dynamic path management.

To guide the path selection and management process, a set of
attributes are required. The basic attributes and behavioral
characteristics associated with traffic trunk path selection and
management are described in the remainder of this sub-section.

5.6.1 Administratively Specified Explicit Paths

An administratively specified explicit path for a traffic trunk is
one which is configured through operator action. An administratively
specified path can be completely specified or partially specified. A
path is completely specified if all of the required hops between the
endpoints are indicated. A path is partially specified if only a
subset of intermediate hops are indicated. In this case, the
underlying protocols are required to complete the path. Due to
operator errors, an administratively specified path can be
inconsistent or illogical. The underlying protocols should be able to
detect such inconsistencies and provide appropriate feedback.

A "path preference rule" attribute should be associated with
administratively specified explicit paths. A path preference rule
attribute is a binary variable which indicates whether the
administratively configured explicit path is "mandatory" or "non-
mandatory."

If an administratively specified explicit path is selected with a
"mandatory attribute, then that path (and only that path) must be
used. If a mandatory path is topological infeasible (e.g. the two
endpoints are topologically partitioned), or if the path cannot be
instantiated because the available resources are inadequate, then the
path setup process fails. In other words, if a path is specified as
mandatory, then an alternate path cannot be used regardless of
prevailing circumstance. A mandatory path which is successfully
instantiated is also implicitly pinned. Once the path is instantiated
it cannot be changed except through deletion and instantiation of a
new path.

However, if an administratively specified explicit path is selected
with a "non-mandatory" preference rule attribute value, then the path
should be used if feasible. Otherwise, an alternate path can be
chosen instead by the underlying protocols.

5.6.2 Hierarchy of Preference Rules For Multi-Paths

In some practical contexts, it can be useful to have the ability to
administratively specify a set of candidate explicit paths for a
given traffic trunk and define a hierarchy of preference relations on
the paths. During path establishment, the preference rules are
applied to select a suitable path from the candidate list. Also,
under failure scenarios the preference rules are applied to select an
alternate path from the candidate list.

5.6.3 Resource Class Affinity Attributes

Resource class affinity attributes associated with a traffic trunk
can be used to specify the class of resources (see Section 6) which
are to be explicitly included or excluded from the path of the
traffic trunk. These are policy attributes which can be used to
impose additional constraints on the path traversed by a given
traffic trunk. Resource class affinity attributes for a traffic can
be specified as a sequence of tuples:

<resource-class, affinity><resource-class, affinity>..

The resource-class parameter identifies a resource class for which an
affinity relationship is defined with respect to the traffic trunk.
The affinity parameter indicates the affinity relationship; that is,
whether members of the resource class are to be included or excluded
from the path of the traffic trunk. Specifically, the affinity
parameter may be a binary variable which takes one of the following
values: (1) explicit inclusion, and (2) explicit exclusion.

If the affinity attribute is a binary variable, it may be possible to
use Boolean expressions to specify the resource class affinities
associated with a given traffic trunk.

If no resource class affinity attributes are specified, then a "don't
care" affinity relationship is assumed to hold between the traffic
trunk and all resources. That is, there is no requirement to
explicitly include or exclude any resources from the traffic trunk's
path. This should be the default in practice.

Resource class affinity attributes are very useful and powerful
constructs because they can be used to implement a variety of
policies. For example, they can be used to contain certain traffic
trunks within specific topological regions of the network.

A "constraint-based routing" framework (see section 7.0) can be used
to compute an explicit path for a traffic trunk subject to resource
class affinity constraints in the following manner:

1. For explicit inclusion, prune all resources not belonging
to the specified classes prior to performing path computation.

2. For explicit exclusion, prune all resources belonging to the
specified classes before performing path placement computations.

5.6.4 Adaptivity Attribute

Network characteristics and state change over time. For example, new
resources become available, failed resources become reactivated, and
allocated resources become deallocated. In general, sometimes more
efficient paths become available. Therefore, from a Traffic
Engineering perspective, it is necessary to have administrative
control parameters that can be used to specify how traffic trunks
respond to this dynamism. In some scenarios, it might be desirable to
dynamically change the paths of certain traffic trunks in response to
changes in network state. This process is called re-optimization. In
other scenarios, re-optimization might be very undesirable.

An Adaptivity attribute is a part of the path maintenance parameters
associated with traffic trunks. The adaptivity attribute associated
with a traffic trunk indicates whether the trunk is subject to re-
optimization. That is, an adaptivity attribute is a binary variable
which takes one of the following values: (1) permit re-optimization
and (2) disable re-optimization.

If re-optimization is enabled, then a traffic trunk can be rerouted
through different paths by the underlying protocols in response to
changes in network state (primarily changes in resource
availability). Conversely, if re-optimization is disabled, then the
traffic trunk is "pinned" to its established path and cannot be
rerouted in response to changes in network state.

Stability is a major concern when re-optimization is permitted. To
promote stability, an MPLS implementation should not be too reactive
to the evolutionary dynamics of the network. At the same time, it
must adapt fast enough so that optimal use can be made of network
assets. This implies that the frequency of re-optimization should be
administratively configurable to allow for tuning.

It is to be noted that re-optimization is distinct from resilience. A
different attribute is used to specify the resilience characteristics
of a traffic trunk (see section 5.9). In practice, it would seem
reasonable to expect traffic trunks subject to re-optimization to be
implicitly resilient to failures along their paths. However, a
traffic trunk which is not subject to re-optimization and whose path
is not administratively specified with a "mandatory" attribute can
also be required to be resilient to link and node failures along its
established path

Formally, it can be stated that adaptivity to state evolution through
re-optimization implies resilience to failures, whereas resilience to
failures does not imply general adaptivity through re-optimization to
state changes.

5.6.5 Load Distribution Across Parallel Traffic Trunks

Load distribution across multiple parallel traffic trunks between two
nodes is an important consideration. In many practical contexts, the
aggregate traffic between two nodes may be such that no single link
(hence no single path) can carry the load. However, the aggregate
flow might be less than the maximum permissible flow across a "min-
cut" that partitions the two nodes. In this case, the only feasible
solution is to appropriately divide the aggregate traffic into sub-
streams and route the sub-streams through multiple paths between the
two nodes.

In an MPLS domain, this problem can be addressed by instantiating
multiple traffic trunks between the two nodes, such that each traffic
trunk carries a proportion of the aggregate traffic. Therefore, a
flexible means of load assignment to multiple parallel traffic trunks
carrying traffic between a pair of nodes is required.

Specifically, from an operational perspective, in situations where
parallel traffic trunks are warranted, it would be useful to have
some attribute that can be used to indicate the relative proportion
of traffic to be carried by each traffic trunk. The underlying
protocols will then map the load onto the traffic trunks according to
the specified proportions. It is also, generally desirable to
maintain packet ordering between packets belong to the same micro-
flow (same source address, destination address, and port number).

5.7 Priority attribute

The priority attribute defines the relative importance of traffic
trunks. If a constraint-based routing framework is used with MPLS,
then priorities become very important because they can be used to
determine the order in which path selection is done for traffic
trunks at connection establishment and under fault scenarios.

Priorities are also important in implementations permitting
preemption because they can be used to impose a partial order on the
set of traffic trunks according to which preemptive policies can be
actualized.

5.8 Preemption attribute

The preemption attribute determines whether a traffic trunk can
preempt another traffic trunk from a given path, and whether another
traffic trunk can preempt a specific traffic trunk. Preemption is
useful for both traffic oriented and resource oriented performance

objectives. Preemption can used to assure that high priority traffic
trunks can always be routed through relatively favorable paths within
a differentiated services environment.

Preemption can also be used to implement various prioritized
restoration policies following fault events.

The preemption attribute can be used to specify four preempt modes
for a traffic trunk: (1) preemptor enabled, (2) non-preemptor, (3)
preemptable, and (4) non-preemptable. A preemptor enabled traffic
trunk can preempt lower priority traffic trunks designated as
preemptable. A traffic specified as non-preemptable cannot be
preempted by any other trunks, regardless of relative priorities. A
traffic trunk designated as preemptable can be preempted by higher
priority trunks which are preemptor enabled.

It is trivial to see that some of the preempt modes are mutually
exclusive. Using the numbering scheme depicted above, the feasible
preempt mode combinations for a given traffic trunk are as follows:
(1, 3), (1, 4), (2, 3), and (2, 4). The (2, 4) combination should be
the default.

A traffic trunk, say "A", can preempt another traffic trunk, say "B",
only if *all* of the following five conditions hold: (i) "A" has a
relatively higher priority than "B", (ii) "A" contends for a resource
utilized by "B", (iii) the resource cannot concurrently accommodate
"A" and "B" based on certain decision criteria, (iv) "A" is preemptor
enabled, and (v) "B" is preemptable.

Preemption is not considered a mandatory attribute under the current
best effort Internet service model although it is useful. However, in
a differentiated services scenario, the need for preemption becomes
more compelling. Moreover, in the emerging optical internetworking
architectures, where some protection and restoration functions may be
migrated from the optical layer to data network elements (such as
gigabit and terabit label switching routers) to reduce costs,
preemptive strategies can be used to reduce the restoration time for
high priority traffic trunks under fault conditions.

5.9 Resilience Attribute

The resilience attribute determines the behavior of a traffic trunk
under fault conditions. That is, when a fault occurs along the path
through which the traffic trunk traverses. The following basic
problems need to be addressed under such circumstances: (1) fault
detection, (2) failure notification, (3) recovery and service
restoration. Obviously, an MPLS implementation will have to
incorporate mechanisms to address these issues.

Many recovery policies can be specified for traffic trunks whose
established paths are impacted by faults. The following are examples
of feasible schemes:

1. Do not reroute the traffic trunk. For example, a survivability
scheme may already be in place, provisioned through an
alternate mechanism, which guarantees service continuity
under failure scenarios without the need to reroute traffic
trunks. An example of such an alternate scheme (certainly
many others exist), is a situation whereby multiple parallel
label switched paths are provisioned between two nodes, and
function in a manner such that failure of one LSP causes the
traffic trunk placed on it to be mapped onto the remaining LSPs
according to some well defined policy.

2. Reroute through a feasible path with enough resources. If none
exists, then do not reroute.

3. Reroute through any available path regardless of resource
constraints.

4. Many other schemes are possible including some which might be
combinations of the above.

A "basic" resilience attribute indicates the recovery procedure to be
applied to traffic trunks whose paths are impacted by faults.
Specifically, a "basic" resilience attribute is a binary variable
which determines whether the target traffic trunk is to be rerouted
when segments of its path fail. "Extended" resilience attributes can
be used to specify detailed actions to be taken under fault
scenarios. For example, an extended resilience attribute might
specify a set of alternate paths to use under fault conditions, as
well as the rules that govern the relative preference of each
specified path.

Resilience attributes mandate close interaction between MPLS and
routing.

5.10 Policing attribute

The policing attribute determines the actions that should be taken by
the underlying protocols when a traffic trunk becomes non-compliant.
That is, when a traffic trunk exceeds its contract as specified in
the traffic parameters. Generally, policing attributes can indicate
whether a non-conformant traffic trunk is to be rate limited, tagged,
or simply forwarded without any policing action. If policing is
used, then adaptations of established algorithms such as the ATM
Forum's GCRA [11] can be used to perform this function.

Policing is necessary in many operational scenarios, but is quite
undesirable in some others. In general, it is usually desirable to
police at the ingress to a network (to enforce compliance with
service level agreements) and to minimize policing within the core,
except when capacity constraints dictate otherwise.

Therefore, from a Traffic Engineering perspective, it is necessary to
be able to administratively enable or disable traffic policing for
each traffic trunk.

6.0 Resource Attributes

Resource attributes are part of the topology state parameters, which
are used to constrain the routing of traffic trunks through specific
resources.

6.1 Maximum Allocation Multiplier

The maximum allocation multiplier (MAM) of a resource is an
administratively configurable attribute which determines the
proportion of the resource that is available for allocation to
traffic trunks. This attribute is mostly applicable to link
bandwidth. However, it can also be applied to buffer resources on
LSRs. The concept of MAM is analogous to the concepts of subscription
and booking factors in frame relay and ATM networks.

The values of the MAM can be chosen so that a resource can be under-
allocated or over-allocated. A resource is said to be under-
allocated if the aggregate demands of all traffic trunks (as
expressed in the trunk traffic parameters) that can be allocated to
it are always less than the capacity of the resource. A resource is
said to be over-allocated if the aggregate demands of all traffic
trunks allocated to it can exceed the capacity of the resource.

Under-allocation can be used to bound the utilization of resources.
However,the situation under MPLS is more complex than in circuit
switched schemes because under MPLS, some flows can be routed via
conventional hop by hop protocols (also via explicit paths) without
consideration for resource constraints.

Over-allocation can be used to take advantage of the statistical
characteristics of traffic in order to implement more efficient
resource allocation policies. In particular, over-allocation can be
used in situations where the peak demands of traffic trunks do not
coincide in time.

6.2 Resource Class Attribute

Resource class attributes are administratively assigned parameters
which express some notion of "class" for resources. Resource class
attributes can be viewed as "colors" assigned to resources such that
the set of resources with the same "color" conceptually belong to the
same class. Resource class attributes can be used to implement a
variety of policies. The key resources of interest here are links.
When applied to links, the resource class attribute effectively
becomes an aspect of the "link state" parameters.

The concept of resource class attributes is a powerful abstraction.
From a Traffic Engineering perspective, it can be used to implement
many policies with regard to both traffic and resource oriented
performance optimization. Specifically, resource class attributes can
be used to:

1. Apply uniform policies to a set of resources that do not need
to be in the same topological region.

2. Specify the relative preference of sets of resources for
path placement of traffic trunks.

3. Explicitly restrict the placement of traffic trunks
to specific subsets of resources.

4. Implement generalized inclusion / exclusion policies.

5. Enforce traffic locality containment policies. That is,
policies that seek to contain local traffic within
specific topological regions of the network.

Additionally, resource class attributes can be used for
identification purposes.

In general, a resource can be assigned more than one resource class
attribute. For example, all of the OC-48 links in a given network may
be assigned a distinguished resource class attribute. The subsets of
OC-48 links which exist with a given abstraction domain of the
network may be assigned additional resource class attributes in order
to implement specific containment policies, or to architect the
network in a certain manner.

7.0 Constraint-Based Routing

This section discusses the issues pertaining to constraint-based
routing in MPLS domains. In contemporary terminology, constraint-
based routing is often referred to as "QoS Routing" see [5,6,7,10].

This document uses the term "constraint-based routing" however,
because it better captures the functionality envisioned, which
generally encompasses QoS routing as a subset.

constraint-based routing enables a demand driven, resource
reservation aware, routing paradigm to co-exist with current topology
driven hop by hop Internet interior gateway protocols.

A constraint-based routing framework uses the following as input:

- The attributes associated with traffic trunks.

- The attributes associated with resources.

- Other topology state information.

Based on this information, a constraint-based routing process on each
node automatically computes explicit routes for each traffic trunk
originating from the node. In this case, an explicit route for each
traffic trunk is a specification of a label switched path that
satisfies the demand requirements expressed in the trunk's
attributes, subject to constraints imposed by resource availability,
administrative policy, and other topology state information.

A constraint-based routing framework can greatly reduce the level of
manual configuration and intervention required to actualize Traffic
Engineering policies.

In practice, the Traffic Engineer, an operator, or even an automaton
will specify the endpoints of a traffic trunk and assign a set of
attributes to the trunk which encapsulate the performance
expectations and behavioral characteristics of the trunk. The
constraint-based routing framework is then expected to find a
feasible path to satisfy the expectations. If necessary, the Traffic
Engineer or a traffic engineering support system can then use
administratively configured explicit routes to perform fine grained
optimization.

7.1 Basic Features of Constraint-Based Routing

A constraint-based routing framework should at least have the
capability to automatically obtain a basic feasible solution to the
traffic trunk path placement problem.

In general, the constraint-based routing problem is known to be
intractable for most realistic constraints. However, in practice, a
very simple well known heuristic (see e.g. [9]) can be used to find a
feasible path if one exists:

- First prune resources that do not satisfy the requirements of
the traffic trunk attributes.

- Next, run a shortest path algorithm on the residual graph.

Clearly, if a feasible path exists for a single traffic trunk, then
the above simple procedure will find it. Additional rules can be
specified to break ties and perform further optimizations. In
general, ties should be broken so that congestion is minimized. When
multiple traffic trunks are to be routed, however, it can be shown
that the above algorithm may not always find a mapping, even when a
feasible mapping exists.

7.2 Implementation Considerations

Many commercial implementations of frame relay and ATM switches
already support some notion of constraint-based routing. For such
devices or for the novel MPLS centric contraptions devised therefrom,
it should be relatively easy to extend the current constraint-based
routing implementations to accommodate the peculiar requirements of
MPLS.

For routers that use topology driven hop by hop IGPs, constraint-
based routing can be incorporated in at least one of two ways:

1. By extending the current IGP protocols such as OSPF and IS-IS to
support constraint-based routing. Effort is already underway to
provide such extensions to OSPF (see [5,7]).

2. By adding a constraint-based routing process to each router which
can co-exist with current IGPs. This scenario is depicted
in Figure 1.

------------------------------------------
Management Interface
------------------------------------------

------------ ------------------ --------------
MPLS <-> Constraint-Based Conventional
Routing Process IGP Process
------------ ------------------ --------------

----------------------- --------------
Resource Attribute Link State
Availability Database Database
----------------------- --------------

Figure 1. Constraint-Based Routing Process on Layer 3 LSR

There are many important details associated with implementing
constraint-based routing on Layer 3 devices which we do not discuss
here. These include the following:

- Mechanisms for exchange of topology state information
(resource availability information, link state information,
resource attribute information) between constraint-based
routing processes.

- Mechanisms for maintenance of topology state information.

- Interaction between constraint-based routing processes and
conventional IGP processes.

- Mechanisms to accommodate the adaptivity requirements of
traffic trunks.

- Mechanisms to accommodate the resilience and survivability
requirements of traffic trunks.

In summary, constraint-based routing assists in performance
optimization of operational networks by automatically finding
feasible paths that satisfy a set of constraints for traffic trunks.
It can drastically reduce the amount of administrative explicit path
configuration and manual intervention required to achieve Traffic
Engineering objectives.

8.0 Conclusion

This manuscript presented a set of requirements for Traffic
Engineering over MPLS. Many capabilities were described aimed at
enhancing the applicability of MPLS to Traffic Engineering in the
Internet.

It should be noted that some of the issues described here can be
addressed by incorporating a minimal set of building blocks into
MPLS, and then using a network management superstructure to extend
the functionality in order to realize the requirements. Also, the
constraint-based routing framework does not have to be part of the
core MPLS specifications. However, MPLS does require some interaction
with a constraint-based routing framework in order to meet the
requirements.

9.0 Security Considerations

This document does not introduce new security issues beyond those
inherent in MPLS and may use the same mechanisms proposed for this
technology. It is, however, specifically important that manipulation
of administratively configurable parameters be executed in a secure
manner by authorized entities.

10.0 References

[1] Rosen, E., Viswanathan, A. and R. Callon, "A Proposed
Architecture for MPLS", Work in Progress.

[2] Callon, R., Doolan, P., Feldman, N., Fredette, A., Swallow, G.
and A. Viswanathan, "A Framework for Multiprotocol Label
Switching", Work in Progress.

[3] Li, T. and Y. Rekhter, "Provider Architecture for Differentiated
Services and Traffic Engineering (PASTE)", RFC2430, October
1998.

[4] Rekhter, Y., Davie, B., Katz, D., Rosen, E. and G. Swallow,
"Cisco Systems' Tag Switching Architecture - Overview", RFC
2105, February 1997.

[5] Zhang, Z., Sanchez, C., Salkewicz, B. and E. Crawley "Quality of
Service Extensions to OSPF", Work in Progress.

[6] Crawley, E., Nair, F., Rajagopalan, B. and H. Sandick, "A
Framework for QoS Based Routing in the Internet", RFC2386,
August 1998.

[7] Guerin, R., Kamat, S., Orda, A., Przygienda, T. and D. Williams,
"QoS Routing Mechanisms and OSPF Extensions", RFC2676, August
1999.

[8] C. Yang and A. Reddy, "A Taxonomy for Congestion Control
Algorithms in Packet Switching Networks," IEEE Network Magazine,
Volume 9, Number 5, July/August 1995.

[9] W. Lee, M. Hluchyi, and P. Humblet, "Routing Subject to Quality
of Service Constraints in Integrated Communication Networks,"
IEEE Network, July 1995, pp 46-55.

[10] ATM Forum, "Traffic Management Specification: Version 4.0" April
1996.

11.0 Acknowledgments

The authors would like to thank Yakov Rekhter for his review of an
earlier draft of this document. The authors would also like to thank
Louis Mamakos and Bill Barns for their helpful suggestions, and
Curtis Villamizar for providing some useful feedback.

12.0 Authors' Addresses

Daniel O. Awduche
UUNET (MCI Worldcom)
3060 Williams Drive
Fairfax, VA 22031

Phone: +1 703-208-5277
EMail: awduche@uu.net

Joe Malcolm
UUNET (MCI Worldcom)
3060 Williams Drive
Fairfax, VA 22031

Phone: +1 703-206-5895
EMail: jmalcolm@uu.net

Johnson Agogbua
UUNET (MCI Worldcom)
3060 Williams Drive
Fairfax, VA 22031

Phone: +1 703-206-5794
EMail: ja@uu.net

Mike O'Dell
UUNET (MCI Worldcom)
3060 Williams Drive
Fairfax, VA 22031

Phone: +1 703-206-5890
EMail: mo@uu.net

Jim McManus
UUNET (MCI Worldcom)
3060 Williams Drive
Fairfax, VA 22031

Phone: +1 703-206-5607
EMail: jmcmanus@uu.net

13.0 Full Copyright Statement

Copyright (C) The Internet Society (1999). 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.