Network Working Group D. L. Mills Request for Comments: 975 M/A-COM Linkabit February 1986
Status of This Memo
This RFC proposes certain enhancements of the Exterior Gateway Protocol (EGP) to support a simple, multiple-level routing capability while preserving the robustness features of the current EGP model. It requests discussion and suggestions for improvements. Distribution of this memo is unlimited.
The enhancements, which do not require retrofits in existing implementations in order to interoperate with enhanced implementations, in effect generalize the concept of core system to include multiple communities of autonomous systems, called autonomous confederations. Autonomous confederations maintain a higher degree of mutual trust than that assumed between autonomous systems in general, including reasonable protection against routing loops between the member systems, but allow the routing restrictions of the current EGP model to be relaxed.
The enhancements involve the "hop count" or distance field of the EGP Update message, the interpretation of which is not covered by the current EGP model. This field is given a special interpretation within each autonomous confederation to support up to three levels of routing, one within the autonomous system, a second within the autonomous confederation and an optional third within the universe of confederations.
The historical development of Internet exterior-gateway routing algorithms began with a rather rigid and restricted topological model which emphasized robustness and stability at the expense of routing dynamics and flexibility. Evolution of robust and dynamic routing algorithms has since proved extraordinarily difficult, probably due more to varying perceptions of service requirements than to engineering problems.
The original exterior-gateway model suggested in RFC-827  and subsequently refined in RFC-888  severely restricted the Internet topology essentially to a tree structure with root represented by the BBN-developed "core" gateway system. The most important characteristic of the model was that debilitating resource-consuming routing loops between clusters of gateways (called autonomous
systems) could not occur in a tree-structured topology. However, the administrative and enforcement difficulties involved, not to mention the performance liabilities, made widespread implementation impractical.
Requirements for near-term interoperability between the BBN core gateways and the remainder of the gateway population implemented by other organizations required that an interim protocol be developed with the capability of exchanging reachability information, but not necessarily the capability to function as a true routing algorithm. This protocol is called the Exterior Gateway Protocol (EGP) and is documented in RFC-904 .
EGP was not designed as a routing algorithm, since no agreement could be reached on a trusted, common metric. However, EGP was designed to provide high-quality reachability information, both about neighbor gateways and about routes to non-neighbor gateways. At the present state of development, dynamic routes are computed only by the core system and provided to non-core gateways using EGP only as an interface mechanism. Non-core gateways can provide routes to the core system and even to other non-core gateways, but cannot pass on "third-party" routes computed using data received from other gateways.
As operational experience with EGP has accumulated, it has become clear that a more decentralized dynamic routing capability is needed in order to avoid resource-consuming suboptimal routes. In addition, there has long been resistance to the a-priori assumption of a single core system, with implications of suboptimal performance, administrative problems, impossible enforcement and possible subversion. Whether or not this resistance is real or justified, the important technical question remains whether a more dynamic, distributed approach is possible without significantly diluting stability and robustness.
This document proposes certain enhancements of EGP which generalize the concept of core system to include multiple communities of autonomous systems, called autonomous confederations. Autonomous confederations maintain a higher degree of mutual trust than that assumed between autonomous systems in general, including reasonable protection against routing loops between the member systems. The enhancements involve the "hop count" or distance field of the EGP Update
message, which is given a special interpretation as described later. Note that the interpretation of this field is not specified in RFC-904, but is left as a matter for further study.
The interpretation of the distance field involves three levels of metrics, in which the lowest level is available to the interior gateway protocol (IGP) of the autonomous system itself to extend the interior routes to the autonomous system boundary. The next higher level selects preferred routes within the autonomous system to those outside, while the third and highest selects preferred routes within the autonomous confederation to those outside.
The proposed model is believed compatible with the current specifications and practices used in the Internet. In fact, the entire present conglomeration of autonomous systems, including the core system, can be represented as a single autonomous confederation, with new confederations being formed from existing and new systems as necessary.
It was the intent in RFC-904 that the stipulated routing restrictions superceded all previous documents, including RFC-827 and RFC-888. The notion that a non-core gateway must not pass on third-party information was suggested in planning meetings that occured after the previous documents had been published and before RFC-904 was finalized. This effectively obsoletes prior notions of "stub" or any other asymmetry other than the third-party rule.
Thus, the only restrictions placed on a non-core gateway is that in its EGP messages (a) a gateway can be listed only if it belongs to the same autonomous system (internal neighbor) and (b) a net can be listed only if it is reachable via gateways belonging to that system. There are no other restrictions, overt or implied. The specification does not address the design of the core system or its gateways.
The restrictions imply that, to insure full connectivity, every non-core gateway must run EGP with a core gateway. Since the present core-gateway implementation disallows other gateways on EGP-neighbor paths, this further implies that every non-core gateway must share a net in common with at least one core gateway.
Note that there is no a-priori prohibition on using EGP as an IGP, or even on using EGP with a gateway of another non-core system,
providing that the third-party rule is observed. If a gateway in each system ran EGP with a gateway in every other system, the notion of core system would be unneccessary and superflous.
At one time during the evolution of the EGP model a strict hierarchical topology (tree structure) of autonomous systems was required, but this is not the case now. At one time it was forbidden for two nets to be connected by gateways of two or more systems, but this is not the case now. Autonomous systems are sets of gateways, not nets or hosts, so that a given net or host can be reachable via more than one system; however, every gateway belongs to exactly one system.
Consider the common case of two local-area nets A and B connected to the ARPANET by gateways of different systems. Now assume A and B are connected to each other by a gateway A-B belonging to the same system as the A-ARPANET gateway, which could then list itself and both the A and B nets in EGP messages sent to any other gateway, since both are now reachable in its system. However, the B-ARPANET gateway could list itself and only the B net, since the A-B gateway is not in its system.
In principle, we could assume the existence of a second gateway B-A belonging to the same system as the B-ARPANET gateway, which would entitle it to list the A net as well; however, it may be easier for both systems to sign a treaty and consider the A-B gateway under joint administration. The implementation of the treaty may not be trivial, however, since the joint gateway must appear to other gateways as two distinct gateways, each with its own autonomous-system number.
Another case occurs when for some reason or other a system has no path to a core gateway other than via another non-core system. Consider a third local-are net C, together with gateway C-A belonging to a system other than the A-ARPANET and B-ARPANET gateways. According to the restrictions above, gateway C-A could list net C in EGP messages sent to A-ARPANET, while A-ARPANET could list ARPANET in messages sent to C-A, but not other nets which it may learn about from the core. Thus, gateway C-A cannot acquire full routing information unless it runs EGP directly with a core gateway.
The second example above illustrates the need for a mechanism in which arbitrary routing information can be exchanged between non-core gateways without degrading the degree of robustness relative to a mutually agreed security model. One way of doing this is is to extend the existing single-core autonomous-system model to include multiple core systems. This requires both a topological model which can be used to define the scope of these systems together with a global, trusted metric that can be used to drive the routing computations. An appropriate topological model is described in the next section, while an appropriate metric is suggested in the following section.
An "autonomous system" consists of a set of gateways, each of which can reach any other gateway in the same system using paths via gateways only in that system. The gateways of a system cooperatively maintain a routing data base using an interior gateway protocol (IGP) and a intra-system trusted routing mechanism of no further concern here. The IGP is expected to include security mechanisms to insure that only gateways of the same system can acquire each other as neighbors.
One or more gateways in an autonomous system can run EGP with one or more gateways in a neighboring system. There is no restriction on the number or configuration of EGP neighbor paths, other than the requirement that each path involve only gateways of one system or the other and not intrude on a third system. It is specifically not required that EGP neighbors share a common network, although most probably will.
An "autonomous confederation" consists of a set of autonomous systems sharing a common security model; that is, they trust each other to compute routes to other systems in the same confederation. Each gateway in a confederation can reach any other gateway in the same confederation using paths only in that confederation. Although there is no restriction on the number or configuration of EGP paths other than the above, it is expected that some mechanism be available so that potential EGP neighbors can discover whether they are in the same confederation. This could be done by access-control lists, for example, or by partitioning the set of system numbers.
A network is "directly reachable" from an autonomous system if a gateway in that system has an interface to it. Every gateway in
that system is entitled to list all directly reachable networks in EGP messages sent to any other system. In general, it may happen that a particular network is directly reachable from more than one system.
A network is "reachable" from an autonomous system if it is directly reachable from an autonomous system belonging to the same confederation. A directly reachable net is always reachable from the same system. Every gateway in that confederation is entitled to list all reachable nets in EGP messages sent to any other system. It may happen that a particular net is either directly reachable or reachable from different confederations.
In order to preserve global routing stability in the Internet, it is explicitly assumed that routes within an autonomous system to a directly reachable net are always preferred over routes outside that system and that routes within an autonomous confederation are always preferred over routes outside that confederation. The mechanism by which this is assured is described in the next section.
In general, EGP Update messages can include two lists of gateways, one for those gateways belonging to the same system (internal neighbors) and the other for gateways belonging to different systems (external neighbors). Directly reachable nets must always be associated with gateways of the same system, that is, with internal neighbors, while non-directly reachable nets can be associated with either internal or external neighbors. Nets that are reachable, but not directly reachable, must always be associated with gateways of the same confederation.
There seems to be a general principle which characterizes distributed systems: The "nearer" a thing is the more dynamic and trustable it is, while the "farther" a thing is the more static and suspicious it is. For instance, the concept of network is intrinsic to the Internet model, as is the concept of gateways which bind them together. A cluster of gateways "near" each other (e.g. within an autonomous system) typically exchange routing information using a high-performance routing algorithm capable of sensitive monitoring of, and rapid adaptation to, changing performance indicators such as queueing delays and link loading.
However, clusters of gateways "far" from each other (e.g. widely separated autonomous systems) usually need only coarse routing information, possibly only "hints" on the best likely next hop to
the general destination area. On the other hand, mutual suspicion increases with distance, so these clusters may need elaborate security considerations, including peer authentication, confidentiality, secrecy and signature verification. In addition, considerations of efficiency usually dictate that the allowable network bandidth consumed by the routing protocol itself decreases with distance. The price paid for both of these things typically is in responsiveness, with the effect that the more distant clusters are from each other, the less dynamic is the routing algorithm.
The above observations suggest a starting point for the evolution of a globally acceptable routing metric. Assume the metric is represented by an integer, with low values representing finer distinctions "nearer" the gateway and high values coarser distinctions "farther" from it. Values less than a globally agreed constant X are associated with paths confined to the same autonomous system as the sender, values greater than X but less than another constant Y with paths confined to the autonomous confederation of the sender and values greater than Y associated with the remaining paths.
At each of these three levels - autonomous system, autonomous confederation and universe of confederations - multiple routing algorithms could be operated simultaneously, with each producing for each destination net a possibly different subtree and metric in the ranges specified above. However, within each system the metric must have the same interpretation, so that other systems can mitigate routes between multiple gateways in that system. Likewise, within each confederation the metric must have the same interpretation, so that other confederations can mitigate routes to gateways in that confederation. Although all confederations must agree on a common universe-of-confederations algorithm, not all confederations need to use the same confederation-level algorithm and not all systems in the same confederation need to use the same system-level algorithm.
The manner in which the eight-bit "hop count" or distance field in the EGP Update to be used is not specified in RFC-904, but left as a matter for further study. The above model provides both an interpretation of this field, as well as hints on how to design appropriate routing algorithms.
For the sake of illustration, assume the values of X and Y above are 128 and 192 respectively. This means that the gateways in a
particular system will assign distance values less than 128 for directly-reachable nets and that exterior gateways can compare these values freely in order to select among these gateways. It also means that the gateways in all systems of a particular confederation will assign distance values between 128 and 192 for those nets not directly reachable in the system but reachable in the confederation. In the following it will be assumed that the various confederations can be distinguished by some feature of the 16-bit system-number field, perhaps by reserving a subfield.
The following implementation model may clarify the above issues, as well as present at least one way to organize the gateway data base. The data base is organized as a routing table, the entries of which include a net number together with a list of items, where each item consists of (a) the gateway address, system number and distance provided by an EGP neighbor, (b) a time-to-live counter, local routing information and other information as necessary to manage the data base.
The routing table is updated each time an EGP Update message is received from a neighbor and possibly by other means, such as the system IGP. The message is first decoded into a list of quads consisting of a network number, gateway address, system number and distance. If the gateway address is internal to the neighbor system, as determined from the EGP message, the system number of the quad is set to that system; while, if not, the system number is set to zero, indicating "external."
Next, a new value of distance is computed from the old value provided in the message and subject to the following constraints: If the system number matches the local system number, the new value is determined by the rules for the system IGP but must be less than 128. If not and either the system number belongs to the same confederation or the system number is zero and the old distance is less than 192, the value is determined by the rules for the confederation EGP, but must be at least 128 and less than 192. Otherwise, the value is determined by the rules for the (global) universe-of-federations EGP, but must be at least 192.
For each quad in the list the routing table is first searched for matching net number and a new entry made if not already there. Next, the list of items for that net number is searched for matching gateway address and system number and a new entry made if not already there. Finally, the distance field is recomputed, the time-to-live field reset and local routing information inserted.
The time-to-live fields of all items in each list are incremented on a regular basis. If a field exceeds a preset maximum, the item is discarded; while, if all items on a list are discarded, the entire entry including net number is discarded.
When a gateway sends an EGP Update message to a neighbor, it must invert the data base in order by gateway address, rather than net number. As part of this process the routing table is scanned and the gateway with minimum distance selected for each net number. The resulting list is sorted by gateway address and partitioned on the basis of internal/external system number.
A gateway encountering a datagram (service unit) searches the routing table for matching destination net number and then selects the gateway on that list with minimum distance. As the result of the value assignments above, it should be clear that routes at a higher level will never be chosen if routes at a lower level exist. It should also be clear that route selection within a system cannot affect route selection outside that system, except through the intervention of the intra-confederation routing algorithm. If a simple min-system-hop algorithm is used for the confederation EGP, the IGP of each system can influence it only to the extent of reachability.
The proposed interpretation is backwards-compatibile with known EGP implementations which do not interpret the distance field and with several known EGP implementations that take private liberties with this field. Perhaps the simplest way to evolve the present system is to collect the existing implementations that do not interpet the distance field at all as a single confederation with the present core system and routing restrictions. All distances provided by this confederation would be assumed equal to 192, which would provide at least a rudimentary capability for routing within the universe of confederations.
One or more existing or proposed systems in which the distance field has a uniform interpretation throughout the system can be organized as autonomous confederations. This might include the Butterfly gateways now now being deployed, as well as clones elsewhere. These systems provide the capability to select routes into the system based on the distance fields for the different gateways. It is anticipated that the distance fields for the Butterfly system can be set to at least 128 if the routing
information comes from another Butterfly system and to at least 192 if from a non-Butterfly system presumed outside the confederation.
New systems using an implmentation model such as suggested above can select routes into a confederation based on the distance field. For this to work properly, however, it is necessary that all systems and confederations adopt a consistent interpretation of distance values exceeding 192.
Taken at face value, this document represents a proposal for an interpretation of the distance field of the EGP Update message, which has previously been assigned no architected interpretation, but has been often used informally. The proposal amounts to ordering the autonomous systems in a hierarchy of systems and confederations, together with an interpretation of the distance field as a three-level metric. The result is to create a corresponding three-level routing community, one prefering routes inside a system, a second preferring routes inside a confederation and the third with no preference.
While the proposed three-level hierarchy can readily be extended to any number of levels, this would create strain on the distance field, which is limited to eight bits in the current EGP model.
The concept of distance can easily be generalized to "administrative distance" as suggested by John Nagle and others.