Network-Wide BGP Route Prediction for Traffic Engineering
Nick Feamstera and Jennifer Rexfordb
a Laboratory for Computer Science, Massachusetts Institute of Technology, Cambridge, MA, USA
b Internet and Networking Systems, AT&T Labs–Research, Florham Park, NJ, USA
ABSTRACT
The Internet consists of about 13,000 Autonomous Systems (AS’s) that exchange routing information using the Border
Gateway Protocol (BGP). The operators of each AS must have control over the flow of traffic through their network
and between neighboring AS’s. However, BGP is a complicated, policy-based protocol that does not include any direct
support for traffic engineering. In previous work, we have demonstrated that network operators can adapt the flow of
traffic in an efficient and predictable fashion through careful adjustments to the BGP policies running on their edge
routers.

Nevertheless, many details of the BGP protocol and decision process make predicting the effects of these
policy changes difficult. In this paper, we describe a tool that predicts traffic flow at network exit points based on the
network topology, the import policy associated with each BGP session, and the routing advertisements received from
neighboring AS’s. We present a linear-time algorithm that computes a network-wide view of the best BGP routes for each
destination prefix given a static snapshot of the network state, without simulating the complex details of BGP message
passing. We describe how to construct this snapshot using the BGP routing tables and router configuration files available
from operational routers. We verify the accuracy of our algorithm by applying our tool to routing and configuration data
from AT&T’s commercial IP network. Our route prediction techniques help support the operation of large IP backbone
networks, where interdomain routing is an important aspect of traffic engineering.
Keywords: Traffic engineering, IP routing, BGP, router configuration
1. INTRODUCTION
Traffic engineering involves adapting the operation of a network according to the prevailing traffic conditions in order
to improve performance and use resources efficiently. In practice, traffic engineering involves adjusting the resource
allocation policies for path selection, buffer management, and link scheduling at the individual routers in the network.
For example, if some traffic is experiencing high delay or packet loss due to a congested link, operators can adjust the
configuration of the routing protocol to divert part of the traffic to other paths. Alternatively, an operator may be able to
improve performance by reconfiguring the buffer management policy at the router; one approach might be to selectively
mark or discard packets (e.g., by tuning the Random Early Detection (RED) parameters) to encourage some of the TCP
senders to reduce their transmission rates before the buffer becomes full.


If the link carries multiple classes of traffic,
the operator can also reconfigure the link-scheduling parameters to devote more bandwidth to some portion of the traffic.
Selecting the appropriate values for these parameters requires an accurate, up-to-date view of the offered traffic, net-
work topology, and router configuration, which a well-designed network monitoring infrastructure can provide. Effective
traffic engineering also depends on the ability to predict the outcome of possible changes to the router configuration. Eval-
uating “what-if” scenarios requires network management tools that simulate the network protocols and mechanisms

or explicitly model their effects on the traffic.

In some cases, such as capturing the influence of RED parameters on
TCP traffic over an entire network, simulation may be the only feasible alternative. In contrast, predicting the effects of
routing changes does not require a complex simulation of the messages exchanged in the routing protocol. Nevertheless,
deriving a closed-form analytic expression for the optimal parameter settings may prove difficult. Instead, we provide a
way to explore many different parameter values to allow operators to select a good configuration that makes efficient use
of network resources.
Most of the recent research and standards work on traffic engineering has focused on the Interior Gateway Protocols
(IGPs), such as OSPF (Open Shortest Path First) or IS-IS (Intermediate System-Intermediate System), which control the
selection of paths within a single Autonomous System (AS).  
Because network operators manage all of the routers that
participate in the IGP for a given network, they have complete control over intradomain routing. For example, network

operators can configure the link weights that control the selection of shortest paths in OSPF or IS-IS routing. However,
most of the traffic carried by a large IP backbone network traverses multiple AS’s, which makes interdomain routing an
important aspect of traffic engineering. Additionally, the links between AS’s are common points of congestion, largely
because the control of these links is shared between two or more (sometimes competing) parties. Careful control over
interdomain routing is important for improving end-to-end performance and making efficient use of network resources.
In this paper, we focus on traffic engineering in the context of the existing interdomain routing protocol—the Border
Gateway Protocol (BGP). 
 Thus, our traffic engineering solutions do not require any modifications to the existing IP
infrastructure. However, BGP is a complex, policy-based protocol with a large number of configuration options. Because
changes to BGP routing policies can affect routing stability and the flow of traffic in the Internet as a whole, network
operators should understand the potential impact of changes in routing policy before reconfiguring the operational routers.
In this paper, we describe how to predict the influence of configuration changes, based on a snapshot of the state of the
network. This allows a network operator to evaluate possible changes to BGP policies and compare their impact on the
flow of traffic. Specifically, we present three main contributions:
 Network-wide model: We propose a model of the network state required to predict the influence of changes in
BGP policies on path selection. The model incorporates the BGP routes advertised by neighboring domains and the
BGP import policies configured by network operators. The model specifies the inputs to existing tools that capture
the influence of the IGP configuration.

 Route prediction algorithm: We present a linear-time, centralized algorithm that computes the best BGP routes
chosen by the various routers in the AS based on the routing policies and BGP advertisements. We show that such
an algorithm can predict routes without simulating the passing of BGP messages between routers. Additionally, we
prove that our algorithm accurately represents the BGP decision process implemented on IP routers.
 Prototype implementation: We describe how we populated our network model using the data available from the
routers in AT&T’s commercial IP network. We describe a prototype implementation of our tool that accurately
predicts the effects of BGP import policy changes on path selection.
We present these topics in three separate sections after a brief background section that describes the BGP protocol and
decision process. The paper concludes with a summary of our approach and a discussion of future research directions.
2. BORDER GATEWAY PROTOCOL
In this section, we first present an overview of the Border Gateway Protocol (BGP) and the attributes associated with
BGP advertisements. Next, we describe the BGP decision process, which governs the selection of the best route for each
destination prefix at each router. Finally, we briefly explain how a router constructs a forwarding table based on its best
BGP route and the IGP parameters.
2.1. BGP Protocol
Internet routing and forwarding operate at the level of prefixes, which represent blocks of contiguous IP addresses. A prefix
is represented by a  -bit address and a mask length. For example,   ff flfiffi  specifies ffi!ffi" addresses ranging from

 #  
to 
 ffffffi!!
. Neighboring AS’s exchange routing information by configuring a BGP session between a pair
of edge routers. The two routers establish a session and exchange update messages as they acquire new information about
how to reach individual destination prefixes. For a given prefix, a router sends an advertisement to inform its neighbor of
a new route to the destination prefix or a withdrawal to indicate that the route to that prefix is no longer available. Each
advertisement includes an AS path that identifies the list of AS’s en route to the origin AS that announced the destination
prefix; for this reason, BGP is called a path-vector protocol. Before accepting an advertisement, the receiving router
discards any routes that contain its own AS number in the AS path to prevent the formation of routing loops.
Route advertisements include several other attributes. The next hop attribute indicates the IP address of the router
associated with next hop along the path to the destination. The origin type identifies how the origin AS learned about
the route—within the AS (e.g., static configuration), EGP (a now-defunct distance-vector protocol), or injection from
another routing protocol. A neighbor AS may include a multiple exit discriminator (MED) in the route advertisement to