Routing Protocol Concepts

Generally speaking, most routing protocols used today are based on one of two types of distributed routing algorithms: link-state or distance vector. In the next few sections, we'll discuss the different properties of distance vector and link-state routing algorithms.

Distance Vector Routing Protocols

Distance vector protocols are sometimes referred to as Bellman-Ford protocols, named after the person who invented the algorithm used for calculating the shortest paths^[2] and for the people who first described a distributed use of the algorithm^[3]. The term distance vector. is derived from the fact that the protocol includes a vector (list) of distances (hop counts or other metrics) associated with each destination prefix routing message.

Distance vector routing protocols, such as Routing Information Protocol (RIP), utilize a distributed computation approach to calculating the route to each destination prefix. In other words, distance vector protocols require that each node separately calculate the best path (output link) to each destination prefix.

After selecting the best path, a router then sends distance vectors to its neighbors, notifying them of the reachability of each destination prefix and of the corresponding metrics associated with the path it has selected to reach the prefix. In parallel, its neighbors also calculate the best path to each available destination and then notify their neighbors of the available path (and associated metrics) they've selected to reach the destination. Upon the receipt of messages from neighbors detailing the destination and associated metrics that the neighbor has selected, the router might determine that a better path exists via an alternative neighbor. The router will again notify its neighbors of its selected paths (and associated metrics) to reach each destination. This cycle continues until all the routers have converged upon a common understanding of the best paths to reach each destination prefix.

Initial specifications of distance vector routing protocols such as RIP Version 1 (RIP-1) had several drawbacks. For example, hop count was the only metric RIP-1 used to select a path. This imposed several limitations. Consider, for example, the RTA routing tables shown in Figure 4-1. One table represents routing information considered when using RIP, and the other when using OSPF. (OSPF is a link-state routing protocol that will be discussed in more detail in the following sections.)

When using RIP-1, RTA would select the direct link between RTA and RTB to reach network 192.10.5.0. RTA prefers this link because the direct path requires just one hop via the RTB path versus two hops via the RTC-RTB path. However, RTA has no knowledge that the RTA-RTB link is actually a very low-capacity, high-latency connection and that using the RTC-RTB path would provide a better level of service.

On the other hand, when using OSPF and metrics other than hop count alone for path selection, RTA will realize that the path to RTB via RTC (cost: 60 + 60 = 120; 2 hops) is actually more optimal than the direct path (cost: 2000; 1 hop).

Another issue with hop counts is the count to infinity restriction. Traditional distance vector protocols (for example, RIP-1) have a finite limit of hops, often 15, after which a route is considered unreachable. This would restrict the propagation of routing updates and would cause problems for large networks (those with more than 15 nodes in a given path). The reliance on hop counts is one deficiency of distance vector protocols, although newer distance vector protocols (that is, RIP-2 and EIGRP) are not constrained as such.

Another deficiency is the way that the routing information is exchanged. Traditional distance vector protocols work on the concept that routers exchange all the IP network numbers they can reach via periodic exchange of distance vector broadcasts—broadcasts that are sent when a "refresh timer" associated with the message exchange expires. Because of this, if the refresh timer expires and a fresh set of routing information is broadcast to your neighbors, the timer is reset, and no new information is sent until the timer expires again. Now, consider what would happen if a link or path became unavailable just after a refresh occurred. Propagation of the path failure would be suppressed until the refresh timer expired, thereby slowing convergence considerably.

Fortunately, newer distance vector protocols, such as EIGRP and RIP-2, introduce the concept of triggered updates. Triggered updates propagate failures as soon as they occur, speeding convergence considerably.

As you might have realized, in large networks, or even small networks with a large number of destination prefixes, periodic exchange of the routing table between neighbors might become very large and very difficult to maintain, contributing to slower convergence. Also, the amount of CPU and link overhead consumed by periodic advertisement of routing information can become quite large. Another property that newer distance vector protocols have adopted is to introduce reliability to the transmission of the distance vectors between neighbors, eliminating the need to periodically readvertise the entire routing table.

Convergence refers to the point in time at which the entire network becomes updated to the fact that a particular route has appeared, disappeared, or changed. Traditional distance vector protocols worked on the basis of periodic updates and hold-down timers: If a route is not received in a certain amount of time, the route goes into a hold-down state and gets aged out of the routing table. The hold-down and aging process translates into minutes in convergence time before the whole network detects that a route has disappeared. The delay between a route's becoming unavailable and its aging out of the routing tables can result in temporary forwarding loops or black holes.

Another issue in some distance vector protocols (for example, RIP) is that when an active route disappears, but the same route reappears with a higher metric (presumably emanating from another router, indicating a possible "good" alternative path), the route is still put into a hold-down state. Thus, the amount of time for the entire network to converge is still increased.

Another major drawback of first-generation distance vector protocols is their classful nature and their lack of support for VLSM or CIDR. These distance vector protocols do not exchange mask information in their routing updates and are therefore incapable of supporting these technologies. In RIP-1, a router that receives a routing update on a certain interface will apply to this update its locally defined subnet mask. IGRP does the same thing as RIP-1 but falls back to Class A, B, and C network masks if a portion of the transmitted network address does not match the local network address. This would lead to confusion (in case the interface belongs to a network that is variably subnetted) and a misinterpretation of the received routing update. Newer distance vector protocols, such as RIP Version 2 (RIP-2) and EIGRP, overcome the aforementioned shortcomings.

Several modifications have been made that alleviate deficiencies associated with traditional distance vector routing protocol behaviors. For example, RIP-2 and EIGRP support VLSM and CIDR. Also, IGRP and EIGRP have the capability to factor in composite metrics used to represent link characteristics along a path (such as bandwidth, utilization, delay, MTU, and so forth), which allows them to calculate more optimal paths than using a hop count alone.

The simplicity and maturity of distance vector protocols has led to their popularity. The primary drawback of traditional implementation of distance vector protocols is slow convergence, a property that can be a catalyst for introducing forwarding loops and/or black-holing traffic during topological changes. However, newer distance vector protocols—most notably, EIGRP—actually converge quite well.

This section wouldn't be complete without mentioning that BGP falls into the distance vector category. In addition to the standard distance vector properties, BGP employs an additional mechanism referred to as the path vector, used to avoid the count to infinity problem previously discussed. Essentially, the path vector contains a list of routing domains (AS numbers) through which the route has traversed. If a domain receives a route for which its domain identifier is already listed in the path, the route is ignored. This path information provides a mechanism that allows routing loops to be pruned. It can also be used to apply domain-based policies. This path attribute, and many other path attributes, are discussed in detail in the following chapters.

Link-State Routing Protocols

Link-state routing protocols, such as Open Shortest Path First (OSPF)^[4] and Intermediate System-to-Intermediate System (IS-IS)^[5], utilize a replicated distributed database model and are considered to be more-complex routing protocols. Link-state protocols work on the basis that routers exchange information elements, called link states, which carry information about links and nodes in the routing domain. This means that routers running link-state protocols do not exchange routing tables as distance vector protocols do. Rather, they exchange information about adjacent neighbors and networks and include metric information associated with the connection.

One way to view link-state routing protocols is as a jigsaw puzzle. Each router in the network generates a piece of the puzzle (link state) that describes itself and where it connects to adjacent puzzle pieces. It also provides a list of the metrics corresponding to the connection with each piece of the puzzle. The local router's piece of the puzzle is then reliably distributed throughout the network, router by router, via a flooding mechanism, until all nodes in the domain have received a copy of the puzzle piece. When distribution is complete, every router in the network has a copy of every piece of the puzzle and stores the puzzle pieces in what's referred to as a link-state database. Each router then autonomously constructs the entire puzzle, the result of which is an identical copy of the entire puzzle on each router in the network.

Then, by applying the SPF (shortest path first) algorithm (most commonly, the Dijkstra Algorithm) to the puzzle, each router calculates a tree of shortest paths to each destination, placing itself at the root.

Following are some of the benefits that link-state protocols provide:

• No hop count— There are no limits on the number of hops a route can take. Link-state protocols work on the basis of link metrics rather than hop counts.

  As an example of a link-state protocol's reliance on metrics rather than hop count, turn again to the RTA routing tables shown in Figure 4-1. In the OSPF case, RTA has picked the optimal path to reach RTB by factoring in the cost of the links. Its routing table lists the next hop of 192.10.3.2 (RTC) to reach 192.10.5.0 (RTB). This is in contrast to the RIP scenario, which resulted in a suboptimal path.

• Bandwidth representation— Link bandwidth and delays may be (manually or dynamically) factored in when calculating the shortest path to a certain destination. This leads to better load balancing based on actual link cost rather than hop count.

• Better convergence— Link and node changes are immediately flooded into the domain via link-state updates. All routers in the domain will instantly update their routing tables (some similar to triggered updates).

• Support for VLSM and CIDR— Link-state protocols exchange mask information as part of the information elements that are flooded into the domain. As a result, networks with variable-length subnet masks can be easily identified.

• Better hierarchy— Whereas distance vector networks are flat networks, link-state protocols provide mechanisms to divide the domain into different levels or areas. This hierarchical approach better scopes network instabilities within areas.

Although link-state algorithms have traditionally provided better routing scalability, which allows them to be used in bigger and more complex topologies, they still should be restricted to interior routing. Link-state protocols by themselves cannot provide a global connectivity solution required for Internet interdomain routing. In very large networks and in case of route oscillation caused by link instabilities, link-state retransmission and recomputation will become too large for any single router to handle.

Although a more detailed discussion of IGPs is beyond the scope of this book, two excellent references that discuss the different link-state and distance vector routing protocols are Interconnections, Second Edition:Bridges, Routers, Switches and Internetworking Protocols^[6] by Radia Perlman and OSPF: Anatomy of an Internet Routing Protocol^[7] by John T. Moy.

Most large service providers today use link-state routing protocols for intra-AS routing, primarily because of its fast convergence capabilities. The two most common protocols deployed in this space are OSPF and IS-IS.

Many older service providers have selected IS-IS as their IGP, and some newer providers select OSPF or IS-IS. Initially, it might seem that older networks use IS-IS rather than OSPF because the U.S. Government required support of ISO CLNP by networks in order for the networks to be awarded federal contracts. (Note that IS-IS is capable of carrying both CLNP and IP Network layer information, while OSPF is capable of carrying only IP information.) However, Internet folklore suggests that the driving factor was that IS-IS implementations were much more stable than OSPF implementations when early providers were selecting which routing protocol to use. This stability obviously had a significant impact on which IGP service providers selected.

Today, both IS-IS and OSPF are widely deployed in ISP networks. The maturity and stability of IS-IS has resulted in its remaining deployed in large networks, as well as its being the IGP of choice for some more recently deployed networks.