MPLS and Traffic Engineering in IP Networks

0163-6804/99/$10.00 © 1999 IEEE nternet growth in recent times has been very impressive. A report from the U.S. Department of Commerce [1] suggests that the rate at which the Internet has been adopted has surpassed all other technologies preceding it, including radio, television, and the personal computer. Today, the Internet has become a convenient and cost-effective medium for collaboration, education, electronic commerce, and entertainment. A common consensus is that the Internet will metamorphose into a medium for the convergence of voice, video, and data communications. Although the long-term market behavior of the Internet is difficult to forecast, Internet traffic is clearly growing in a geometric progression. Reported compounded traffic growth rates range from two to ten times per annum. Large Internet service providers (ISPs) have responded to the challenge of Internet growth by employing three complementary technical instruments: • Network architecture • Capacity expansion • Traffic engineering Network architecture deals with the abstract structure of networks, the components or object classes of the network, their functions, and the relationships between them. A good, scalable network architecture, premised on sound architectural principles, is imperative in the quickly evolving Internet environment. The second instrument employed by large ISPs to respond to traffic growth is rapid expansion of capacity and network infrastructure. In 1996 most large ISPs in the United States operated backbones with DS3 (44.736 Mb/s) links. In 1997 and 1998, OC12c (622 Mb/s) links became pervasive. In 1999 a number of major ISPs upgraded to OC-48c (2.488 Gb/s) links. By the year 2000, some ISPs expect to begin deployment of IP backbones with OC-192c (9.953 Gb/s) links, provisioned directly over dense wavelength-division multiplexing (DWDM) facilities. The third instrument employed by service providers to address the Internet growth challenge is traffic engineering. This subject has attracted significant attention in recent times [2–9]. A motivation for Internet traffic engineering is the realization that architectural paradigms and simple capacity expansion are necessary, but not sufficient, to deliver highquality Internet service under all circumstances. Internet traffic engineering is the aspect of Internet network engineering that addresses the issue of performance optimization of operational networks. It encompasses the application of technology and scientific principles to the measurement, modeling, characterization, and control of Internet traffic [2]. It also includes the application of knowledge and techniques to achieve specific performance objectives, including reliable and expeditious movement of traffic through the network, efficient utilization of network resources, and planning of network capacity. Ultimately, good traffic engineering increases the value of a network to both the service provider and the Internet user community. Historically, effective traffic engineering has been difficult to achieve in public IP networks. The reason for this is the limited functional capabilities of conventional IP technologies. One particular shortcoming of conventional IP systems is the inadequacy of measurement functions. For example, a traffic matrix, which is a basic data set needed for traffic engineering, is difficult to estimate from interface statistics on IP routers. The limitations of intradomain routing control functions are another issue with conventional IP systems. Interior gateway protocols (IGPs), such as Intermediate System–Intermediate System (IS-IS) and Open Shortest Path First (OSPF), commonly used to route traffic within autonomous systems in the Internet, are topology-driven and employ per-packet progressive connection control. Each router makes independent routing decisions using a local instantiation of a synchronized routing area link state database. Route selection is based on shortest path computations using simple additive link metrics. This approach is highly distributed and scalable, but flawed. The flaw is that these protocols do not consider the characteristics of offered traffic and network capacity constraints when making routing decisions. This results in subsets of network resources becoming congested, while other resources along alternate paths remain underutilized [2]. This type of congestion problem is a symptom of poor resource allocation, and is an issue that traffic engineering specifically attempts to redress. Recent developments in multiprotocol label switching (MPLS) [2–8] open new possibilities to address some of the limitations of IP systems concerning traffic engineering. A framework for MPLS is presented in [5] and an architecture for it described in [8]. The requirements for traffic engineering over MPLS were articulated in [2]. Although MPLS is a relatively simple technology (based on the classical label swapping paradigm), it enables the introduction of sophisticated control capabilities that advance the traffic engineering function in IP networks [2–4, 6, 7]. A particularly interesting aspect of MPLS is that it efficiently supports origination connection control through explicit label-switched paths. When MPLS is comDaniel O. Awduche, UUNET (MCI Worldcom)