By Phil Jones
Sometime next quarter, according to Rob Zimmer, director of strategy for IBM’s Network Hardware Division, IBM will start shipping a product that could hold the key to the renaissance of Big Blue’s network credentials. Rainier, a microprocessor switch adapter, is unlikely to become a household name. But if it fulfils its promise of enabling wire-speed policy enforcement in Gigabit network switches, it could bring a new lease of life to an IBM division which has struggled to make any impression on the market since the death of SNA.
According to Zimmer policy enforcement, the process of dynamically allocating and prioritizing network resources according to predefined application and user specific requirements, will be the key issue facing public and private network owners in the internet age. The e-commerce phenomena, after all, is turning all business processes into network transactions. Policing network transactions in the internet age, is therefore the same as running the business – except, of course, that business policies must now be delivered at wire speed.
Today’s layer 2 network switches, such as IBM’s own Prizma chip, create the potential for wire speed business management by driving packet streams at speeds of 500Gbps and upwards. But this throughput is painfully constrained the moment policy management enters the equation. Here, at layers 3 and above, says Chuck Sannipoli, manager of systems enablement for wired communications at IBM’s Research Triangle Park facility in Raleigh, North Carolina, the best efforts of conventional RISC processors (typically found doing the look-up functions in layer 3 routers) and their real-time operating systems don’t cut the mustard. It was possible says Sannipoli, for conventional hybrid switch/routers to offer some elements of QoS packet processing at speeds of 100Mbps but it is not possible at 1Gbps, and it certainly is not possible at 10Gbps.
Rainier is IBM’s response to the quality of service (QoS) bottleneck problem. The switch adapter complements IBM’s Prizma switch fabric device, relieving the switch engine of the burden of identifying and prioritizing packet streams, and matching them against policy profiles. It is not likely to be the only in-silicon answer to the QoS bottleneck, but it if does come to market in the next quarter and if, of course, it works as IBM claims it will, it could prove a clear differentiator for IBM’s own enterprise switch products, and for the products of any other manufacturer which takes up IBM’s offer to place Rainer on the merchant silicon market.
Its pre-release credentials certainly look impressive. The 0.25 micron device carries two million gates and 3Mbits of memory. Indeed, half the chips circuits are dedicated to memory and a specially built interconnect is included which, Sannipoli says will put the device in touch with a ton of memory which should ensure that it will spend minimal time going off-board to look up policy information. This policy information is constantly updated by a neighboring control processor which acts as a client to the local LDAP (lightweight directory access protocol), and should reduce the business of keeping policy information up to date to background task with direct bearing on the performance of the switch.
If this were all Rainer could do it would be a boon to most switch makers, but in typical IBM fashion, the device has been engineered to cover almost any eventually. It comes, for instance, with its own 4Gbps switch fabric on-board, giving it the potential to act as a standalone switch device with enough capacity to handle 40 100Mbps Ethernet ports, or 80 100Mbps Ethernet ports if two Rainier are run back-to-back.
In practice Rainier is unlikely to be used as switch fabric. Instead Sannipoli expects multiple Rainier devices to provide a front-end process and aggregation role in front of the much faster switch fabrics which will emerge in the next generation Gigabit and 10 Gigabit switches. And the device won’t be constrained to handling particular traffic streams. Unlike Prizma, the device is programmable (at the factory floor, but Sannipoli said it may become field programmable in time) and will handle Ethernet, and Frame Relay traffic as well as ATM packets. Any frame to cell conversion that is required will also be assigned to Rainer and executed, IBM claims, at wire speed.
In themselves, these features of Rainier are not exceptional. Other chip designers are addressing wire speed switching and packet conversion issues, and some are also focussing on higher level policy management issues. Where Rainier appears to differ is in the comprehensiveness of IBM’s approach. As well as basic core functions, such as loads, stores, tests and branches, Rainier has also been designed to handle a variety of more complex tasks, such as tree searches for longest prefix matching – a key capability for traffic prioritization, and the kind of task which quickly congests a conventional Risc processor architecture.
On the face of it, Rainier’s abilities appear impressive, and an deal complement to the raw switching performance of Prizma which has been cited as industry leading by independent organizations such as the Tolly Group. But is it really the right product, and if it is, has IBM got the network technology credibility needed to build a market lead for Rainier before its rivals come to market with their own solutions?
From IBM’s perspective it is essential that the answer to both these questions is yes. During the past decade IBM’s dominance of the corporate WAN market evaporate along with its Systems Network Architecture user base, and its decision to back ATM in the enterprise has proven a costly mistake. In the LAN Token Ring has been overcome by the Ethernet community, while in the increasingly important public network infrastructure, IBM has still to make any significant impression. Just in terms of its own future, its seems unlikely that IBM’s network division can afford to make another costly technology error.
