Internet Exchange BGP Route Server

Internet exchange points (IXPs) provide IP data interconnection facilities for their participants, typically using shared Layer-2 networking media such as Ethernet. The Border Gateway Protocol (BGP) , an inter-Autonomous System routing protocol, is commonly used to facilitate exchange of network reachability information over such media. While bilateral external BGP sessions between exchange participants were previously the most common means of exchanging reachability information, the overhead associated with dense interconnection can cause substantial operational scaling problems for partipants of larger Internet exchange points. Multilateral interconnection is a method of interconnecting BGP speaking routers using a third party brokering system, commonly referred to as a route server and typically managed by the IXP operator. Each of the multilateral interconnection participants (usually referred to as route server clients) announces network reachability information to the route server using external BGP, and the route server in turn forwards this information to each other route server client connected to it, according to its configuration. Although a route server uses BGP to exchange reachability information with each of its clients, it does not forward traffic itself and is therefore not a router. A route server can be viewed as similar in function to an route reflector, except that it operates using EBGP instead of iBGP. Certain adaptions to are required to enable an EBGP router to operate as a route server; these are outlined in of this document. Route server functionality is not mandatory in BGP implementations.`` The term "route server" is often in a different context used to describe a BGP node whose purpose is to accept BGP feeds from multiple clients for the purpose of operational analysis and troubleshooting. A system of this form may alternatively be known as a "route collector" or a "route-views server". This document uses the term "route server" exclusively to describe multilateral peering brokerage systems.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in .

A route server uses the Border Gateway Protocol to broker network reachability information between its clients. There are some differences between the behaviour of a BGP route server and a BGP implementation which is strictly compliant with . These differences are described as follows.

A route server MUST accept all UPDATE messages received from each of its clients for inclusion in its Adj-RIB-In. These UPDATE messages MAY be omitted from the route server's Loc-RIB or Loc-RIBs, due to filters configured for the purposes of implementing routing policy. The route server SHOULD perform one or more BGP Decision Processes to select routes for subsequent advertisement to its clients, taking into account possible configuration to provide multiple NLRI paths to a particular client as described in or multiple Loc-RIBs as described in . The route server SHOULD forward UPDATE messages from its Loc-RIB or Loc-RIBs to its clients as determined by local policy.

As a route server primarily performs a brokering service, modification of attributes could cause route server clients to alter their BGP Decision Process for received prefix reachability information, thereby changing the intended routing policies of exchange participants. Therefore, contrary to what is specified in section 5. of , route servers SHOULD NOT by default (unless explicitly configured) update well-known BGP attributes received from route server clients before redistributing them to their other route server clients. Optional recognized and unrecognized BGP attributes, whether transitive or non-transitive, SHOULD NOT be updated by the route server (unless enforced by local IXP operator configuration) and SHOULD be passed on to other route server clients.

The NEXT_HOP is a well-known mandatory BGP attribute which defines the IP address of the router used as the next hop to the destinations listed in the Network Layer Reachability Information field of the UPDATE message. As the route server does not participate in the actual routing of traffic, the NEXT_HOP attribute MUST be passed unmodified to the route server clients, similar to the "third party" next hop feature described in section 5.1.3. of .

AS_PATH is a well-known mandatory attribute which identifies the autonomous systems through which routing information carried in the UPDATE message has passed. As a route server does not participate in the process of forwarding data between client routers, and because modification of the AS_PATH attribute could affect route server client BGP Decision Process, the route server SHOULD NOT prepend its own AS number to the AS_PATH segment nor modify the AS_PATH segment in any other way. This differs from the behaviour specified in section 5.1.2 of , which requires that the BGP speaker prepends its own AS number as the last element of the AS_PATH segment. In contrast to what is recommended in section 6.3 of , route server clients need to be able to accept UPDATE messages where the leftmost AS in the AS_PATH attribute is not equal to the AS number of the route server that sent the UPDATE message. If the route server client BGP system has implemented a check for this, the BGP implementation MUST allow this check to be disabled and SHOULD allow the check to be disabled on a per-peer basis.

MULTI_EXIT_DISC is an optional non-transitive attribute intended to be used on external (inter-AS) links to discriminate among multiple exit or entry points to the same neighboring AS. Contrary to section 5.1.4 of , if applied to an NLRI UPDATE sent to a route server, this attribute SHOULD be propagated to other route server clients and the route server SHOULD NOT modify its value.

The BGP COMMUNITIES () and Extended Communities () attributes are attributes intended for labeling information carried in BGP UPDATE messages. Transitive as well as non-transitive Communities attributes applied to an NLRI UPDATE sent to a route server SHOULD NOT be modified, processed or removed, except as defined by local policy. If a Communities Attribute is intended for processing by the route server itself, as determined by local policy, it MAY be modified or removed.

While IXP participants often use route servers with the intention of interconnecting with as many other route server participants as possible, there are circumstances where control of path distribution on a per-client basis is important to ensure that desired interconnection policies are met. The control of path distribution on a per-client basis can lead to a path being hidden from the route server client. We refer to this as "path hiding". Neither nor its subsections form part of the normative specification of this document, but are included for information purposes only.

___ ___ / \ / \ ..| AS1 |..| AS2 |.. : \___/ \___/ : : \ / | : : \ / | : : IXP \/ | : : /\ | : : / \ | : : ___/____\_|_ : : / \ / \ : ..| AS3 |..| AS4 |.. \___/ \___/ Using the example in , AS1 does not directly exchange prefix information with either AS2 or AS3 at the IXP, but only interconnects with AS4. In the traditional bilateral interconnection model, per-client policy control to a third party exchange participant is accomplished either by not engaging in a bilateral interconnection with that participant or else by implementing outbound filtering on the BGP session towards that participant. However, in a multilateral interconnection environment, only the route server can perform outbound filtering in the direction of the route server client; route server clients depend on the route server to perform their outbound filtering for them. Assuming the BGP Decision Process is used, when the same prefix is advertised to a route server from multiple route server clients, the route server will select a single path for propagation to all connected clients. If, however, the route server has been configured to filter the calculated best path from reaching a particular route server client, then that client will not receive a path for that prefix, although alternate paths received by the route server might have been policy compliant for that client. This phenomenon is referred to as "path hiding". For example, in , if the same prefix were sent to the route server via AS2 and AS4, and the route via AS2 was preferred according to the BGP Decision Process on the route server, but AS2's policy prevented the route server from sending the path to AS1, then AS1 would never receive a path to this prefix, even though the route server had previously received a valid alternative path via AS4. This happens because the BGP Decision Process is performed only once on the route server for all clients. Path hiding will only occur on route servers which employ per-client policy control; if an IXP operator deploys a route server without implementing a per-client routing policy control system, then path hiding does not occur as all paths are considered equally valid from the point of view of the route server.

There are several approaches which can be taken to mitigate against path hiding.

The most portable method to allow for per-client policy control without the occurrence of path hiding, is by using a route server BGP implementation which performs the per-client best path calculation for each set of paths to a prefix, which results after the route server's client policies have been taken into consideration. This can be implemented by using per-client Loc-RIBs, with path filtering implemented between the Adj-RIB-In and the per-client Loc-RIB. Implementations can optimize this by maintaining paths not subject to filtering policies in a global Loc-RIB, with per-client Loc-RIBs stored as deltas. This implementation is highly portable, as it makes no assumptions about the feature capabilities of the route server clients.

The path distribution model described above assumes standard BGP session encoding where the route server sends a single path to its client for any given prefix. This path is selected using the BGP path selection decision process described in . If, however, it were possible for the route server to send more than a single path to a route server client, then route server clients would no longer depend on receiving a single path to a particular prefix; consequently, the path hiding problem described in would disappear. We present two methods which describe how such increased path diversity could be implemented.

The Diverse BGP Path proposal as defined in is a simple way to distribute multiple prefix paths from a route server to a route server client by using a separate BGP session from the route server to a client for each different path. The number of paths which may be distributed to a client is constrained by the number of BGP sessions which the server and the client are willing to establish with each other. The distributed paths may be established from the global BGP Loc-RIB on the route server in addition to any per-client Loc-RIB. As there may be more potential paths to a given prefix than configured BGP sessions, this method is not guaranteed to eliminate the path hiding problem in all situations. Furthermore, this method may significantly increase the number of BGP sessions handled by the route server, which may negatively impact its performance.

The Internet draft proposes a different approach to multiple path propagation, by allowing a BGP speaker to forward multiple paths for the same prefix on a single BGP session. As specifies that a BGP listener must implement an implicit withdraw when it receives an UPDATE message for a prefix which already exists in its Adj-RIB-In, this approach requires explicit support for the feature both on the route server and on its clients. If the ADD-PATH capability is negotiated bidirectionally between the route server and a route server client, and the route server client propagates multiple paths for the same prefix to the route server, then this could potentially cause the propagation of inactive, invalid or suboptimal paths to the route server, thereby causing loss of reachability to other route server clients. For this reason, ADD-PATH implementations on a route server should enforce send-only mode with the route server clients, which would result in negotiating receive-only mode from the client to the route server.

Route server implementation authors may wish to consider one of the methods described in to allow per-client route server policy control without "path hiding". Recommendations for route server operations are described separately in .