Voracious Internet fuels an I/O revolution

Voracious Internet fuels an I/O revolution
Though Infiniband builds on technologies and concepts inherited from its immediate precursors, the Future I/O and Next Generation I/O standards, and borrows from earlier bus topologies such as PCI and PCI-X, many of the pieces that make up this server interconnect architecture are totally new. This combination of an evolutionary path coupled with essentially new throughput potential is proving to be a compelling argument for network designers.

"Developing a product is never easy and doing so in a fast-paced market operating on 'Internet time' is even more difficult," says Kevin Deierling, vice president of marketing for Mellanox Technologies Inc. (Santa Clara, Calif.), in this week's Focus section on Infiniband. "Before Infiniband began to surface as the best-of-class solution for the modern data center, Mellanox had committed to Next Generation I/O technology. Changing the company's focus to Infiniband would surely mean delays in getting product out."

The transition proved to be smoother than Deierling expected. And Mellanox's experience is proving to be typical, generating a rapid shift toward the high-bandwidth storage, clustering and communications specification.

Eyal Waldman, CEO, and Kevin Deierling, VP of marketing at Mellanox, tested their expertise in I/O technology with a fast rollout of Infiniband products.

In two weeks, engineers and software developers from dozens of OEMs, corporate end users and silicon providers will meet in Milpitas, Calif., at the Infiniband Trade Association's first industry-wide "plugfest." There, more than 45 companies will gather to test the compliance and interoperability of the building blocks they have spent almost two years developing, with the hope that full deployment of the new technology can be done across the industry within two years.

The reason for the pressure to upgrade the I/O capabilities of servers and clusters of servers in Internet data centers is the anticipated demand that is expected as more people and devices access the World Wide Web.

Also changing are the kinds of applications and the way they are affecting how Web sites are accessed and used. For example, the increased use of downloadable Java applets to add dynamism and interactivity to the desktop and increased functionality to Java-based handheld information appliances and embedded net-centric devices is fueling a trend to add Java capabilities on the server side, rather than down on the client.

Technologies such as "servlets," Java server pages and Enterprise Java Beans resident on servers do not have GUI issues and rarely need to be restarted. But they pose other performance issues, only some of which can be ameliorated with clever programming and good system software design. Problems such as waiting to read or write to the network, waiting for responses or processing results from a database query can be solved only by improving the underlying I/O structure of the server.

As the articles in this report on the server hardware/software infrastructure indicate, the new switched-fabric specification is perhaps the most comprehensive, complex and ambitious effort that the computer industry has undertaken. As a point-to-point switched-fabric topology with serial linkages, Infiniband is a radical break with the shared-bus topologies such as PCI and PCI-X that now provide connectivity. Then, too, the rates at which users of Internet data centers will be able to move data within the system and out onto the external intranets, virtual private networks and the Internet backbone will range from 1 to 10 Gbits/second, far outpacing even PCI-X, which is expected to top out just above 1 Gbit/s.

"Even those segments based on previously well-understood technologies are being pushed to their limits to provide the performance that will be needed over the next 10 years," said Perry Keller, project manager for computer-bus analysis and design validation at Agilent Technologies (Colorado Springs, Colo.), the author of one of the articles that follow. "We are pushing into new ground on so many fronts we have to be exceedingly careful that we do this just right, with a minimum of glitches."

Because so much is riding on getting Infiniband right the first time, some skeptics are arguing that the pace toward deployment should not be dictated by the need. Because Infiniband must work as soon as it is deployed with few hassles, the rate of deployment should rest on factors such as interoperability, compatibility and compliance to the Infiniband specifications, this camp argues.

According to potential system integrators, such as Hank Brineen, a Berkeley, Calif.-based engineering consultant, such issues should not be settled by things like plugfests, common during the development of such PC-based specifications such as PCI and the Universal Serial Bus. "From an engineering point of view the IBTA Infiniband Trade Association has put much of the legacy of PCI and USB behind us and adopted new methodologies and techniques," Brineen said. "But we haven't gotten rid of the old habits of doing things."

The best route is to develop as quickly as possible more-formalized and truly independent mechanisms for ensuring compliance and interoperability, said Michael Krause, interconnect architect at Hewlett-Packard Co. in Cupertino, Calif. Such techniques should be similar to those used on many IEEE standards and many of the more common network and communications standards, he said.

Krause said the industry needs an independent third-party entity to evaluate compliance and interoperability. "Just setting up a group within a trade association or a special-interest group to define, track and promote compliance and interoperability-or hiring an outside company to do the testing such as is done in much of the desktop software industry-is not enough," he said. "With Infiniband there is just too much riding on getting a truly trustworthy evaluation without the hint of politics or corporate influences, no matter how indirect or unintentional."

A good model for the IBTA to follow, Krause said, might be the route that many of the companies and organizations involved in a variety of communications technologies-such as asymmetric digital subscriber lines, Fast and Gigabit Ethernet, Fibre Channel and wireless 802.11 LANs-have taken: Go to, or establish, a fully independent test facility or organization at a university.

Many computer-communications companies, Krause said, work within a cooperative structure called the Interoperability Lab at the University of New Hampshire. There, more than 100 companies verify the interoperability or conformance of their products. Independent, focused interest groups, or consortiums, are formed by cooperative agreement among interested parties, usually companies that build the products.

Technically, moving toward an approach like the UNH laboratory would make a lot of sense, said Krause, " . . . because, Infiniband really has much more in common with the various communications protocols."