PALO ALTO, Calif. - Packet Design Inc. this week will take the ambitious step of supplementing the familiar Transmission Control Protocol and Border Gateway Protocol (BGP) on the transport layer with the proprietary BGP Scalable Transport, or BST, in an effort to provide more than a makeshift patch for BGP's well-known hierarchical-routing limitations.
While BST is not intended to replace TCP, Packet Design chief executive officer Judy Estrin doesn't expect it to be adopted without a fight. The protocol must be implemented directly within routers to be effective, she said, yet major router manufacturers are bound to show some reluctance to implement a protocol from an independent third party. Hence, Packet Design is trying to win over some Internet service providers to prod the equipment manufacturers, prior to bringing BST to the Internet Engineering Task Force as a proposed transport standard.
BGP's biggest problem, Estrin said, is not in its external version, which operates between router autonomous systems (known as eBGP), but in the internal, or iBGP, version that operates within an autonomous-system region.
BGP uses TCP, a point-to-point protocol, as its transport method. That's not a problem when two peer routers communicate with each other across autonomous-system boundaries, Estrin said, but it becomes a big limiting factor when a peer router must distribute BGP information to all routers inside an autonomous system. Scaling to hundreds of routers is very difficult, configuration is slow and misconfiguration leading to security holes or errors is commonplace.
Since each pair of peers among routers has to maintain a separate TCP connection, transport sizes quickly scale to unwieldy dimensions. If true peering is not maintained, the network must support very large routing tables that make it difficult for networks to reconverge. Packet Design's answer is to use a flooding methodology, familiar in Open Shortest Path First and Intermediate System-Intermediate System protocols, to form the basis of a new transport method. When information is "flooded" to a node's nearest neighbors, detailed topology of the network is not required, and information distribution is reliable regardless of the spatial diversity of the route network. The routers that are configured are BST neighbors only and not every router receiving new TCP state information.
The BST stack runs in parallel to TCP, so that no forklift replacements of TCP are required. TCP, in fact, is used at startup and to communicate with non-BST routers.
Estrin said that network administrators usually try to address TCP limitations along two dimensions. When the issue of sheer scaling of TCP is important, route reflectors can be installed in the network. These can remain in place when BST is used, although the new transport protocol alleviates the need for additional reflectors in the network.
To address issues of reliability, some vendors suggest total router mirroring, an expensive proposition that relies on fully redundant equipment; or users can adopt a software-based graceful restart, which is only in an IETF draft stage now. On the security front, a handful of vendors are suggesting a Secure-BGP extension to BGP, although Estrin said that this particular protocol does not handle scaling problems.
Packet Design developers realize that a BST universe still will need to communicate with routers based on TCP. A translation facility, Ptran, is part of the code and can be implemented in those routers that serve as peers to external autonomous systems. Other secondary routers within an AS also keep information on TCP connections and Internet Protocol Secure states, so that BST-to-TCP translations can be maintained if a primary peer router goes down.
One of the first places BST shows up may be inside a router, between the route processors within a single chassis. In such an embedded application, BST can significantly reduce the BGP processing carried out by hardware and firmware. BST will find its next applications in routers within a point of presence and in inter-POP topologies within a single AS.
In theory, Estrin said, BST can even be used between AS domains, with eBGP riding on top of the BST protocol. The performance advantages, though, clearly are greatest when the BST protocol is used inside a single autonomous system of routers.
Packet Design will begin offering the BST suite to developers in December, based on an initial license fee of $100,000 and a per-device royalty. The eventual goal is to have router OEMs serve as the BST customer, though ISPs and specialized service providers will help to push the protocols, and may want to develop reference implementations. Estrin said that she anticipates plenty of debate within the industry on BST's merits but that she expects "a fair assessment from most BGP developers, even taking into account the NIH factor we'll no doubt see."
