Take the massively parallel computing technology of a company such as nCube, apply it to routers and come up with an innovative solution for the router performance problems which currently plague the backbones of internet service providers (ISPs). This is the idea of Vadim Antonov, the Russian founder and chief technology officer of Palo Alto, California-based Pluris Inc. Antonov believes massively parallel routers will be capable of routing 10,000 times the number of packets per second of traditional routers on the market today from companies such as Cisco Systems. Applying massively parallel architectures to input-output is really the breakthrough idea, says Pluris CEO David Bernstein. Traditional routers from Cisco or even Ascend use a single master processor to do all of the calculations against the routing tables. Our architecture allows for parallel processing of routing so that we can scale to much higher levels, he says.
Bernstein, a co-founder of Unix vendor Santa Cruz Operation, was recruited by venture backer Weiss, Peck & Greer to guide the company through its development phase. Antonov had his ‘Eureka moment’ while working for large system vendor nCube, but his background is in networking and Unix programming. At the beginning of the decade, he set up Russia’s first internet service provider before being head-hunted to help UUNet – now owned by WorldCom – build up its backbone network in the US. After UUNet, he was involved in designing and building SprintLink, a high speed data network for Sprint, a long distance communications carrier. The Pluris architecture can combine up to 64,000 microprocessors. And with each node in the massively parallel router capable of routing data at speeds equivalent to the OC-12 ATM (asynchronous transfer mode) standard, the company’s idea could potentially solve the internet’s routing bandwidth problems for the next 10 years.
Such a big leap in router performance is necessary, says Bernstein, because although semiconductor technology continues to double in power approximately every 18 months, in accordance with Moore’s law, internet demand is doubling every three or four months. While Antonov’s idea sounds simple, it took him a year to design and develop a reliable and sufficiently scalable interconnect, the mechanism used to link up the processors, and to develop the software for ensuring that data packets are delivered by the router in the correct order. The company has three patents relating to its interconnect currently pending and a prototype already up and running in the laboratory. This prototype uses off-the-shelf Intel Pentium processors – but the heat dissipated by a rack of up to 64,000 Pentiums, not to mention the cost, means the finished product will have to ship with cooler embedded processors from ARM, Hitachi or MIPS rather than a standard Intel processor. With the finished product, promises Pluris, customers will be able to start small, but add modules to the router as data throughput grows. Clearly, the Pluris product will be expensive but it is intended for the world’s biggest telecommunications companies and ISPs. Before the end of the year, the company will recruit a direct salesforce to sell to telcos such as British Telecom, France Telecom, Sprint, MCI and other big internet backbone operators.
Next Granite Systems
The company then plans to develop partnerships and a channel network to improve its market penetration. But the success of the salesforce will depend on how well the technology works in practice, and if customers are prepared to take a risk. Certainly, massively parallel processor (MPP) technology has never lived up to its original promise, as proved by the struggles of MPP systems vendors such as nCube and Meiko. MPP systems are often not as scalable as customers expect and can also prove difficult to program, raising manageability questions about Pluris’s massively parallel routers. One of the key reasons for Cisco’s success, for example, has been the manageability software it provides with its routers. But probably the biggest barrier facing Pluris is simply the reluctance of large telcos and ISPs to buy from a newcomer to the field. The biggest telcos require 24-hour global support that a small company like Pluris will struggle to provide. Should Pluris prove that its technology is workable – a public demonstration is due soon – the most likely scenario is that Pluris will be the next Granite Systems, rather than the next Cisco. Granite was bought by Cisco in the fall of 1996 for $220m, despite the fact that it had not begun trading. But for the most promising technology, the network elite – Bay Networks, Cisco and 3-Com – are prepared to pay a premium.