Looking beyond PCI to switched fabric I/O
While disk drive capacities keep doubling and CPUs continue to churn out more and more MHz, server bus I/O has not kept pace. As a result, much of the discussion at last month`s Server I/O `99 conference in Monterey, CA, centered on the successor to the PCI bus standard.
PCI-X, a spec developed by Compaq, Hewlett-Packard, and IBM last fall, appears to be the interim successor to PCI. At the conference, Ken Jansen, director of advanced server architecture and design at Compaq, said the PCI-X spec addresses users` needs for faster, more efficient I/O while maintaining backward compatibility with a large installed base. "PCI has been a very successful standard, and PCI-X will give it new life and new performance capabilities," says Jansen.
The proposed 32/64-bit PCI-X standard boosts speed from 33/66MHz to 133MHz (or multiple slots at 100MHz per) and provides backward compatibility with PCI, as well as easy integration with next-generation switched-fabric interfaces. Compaq officials, confident that the spec will be approved by summer, anticipate PCI-X-compliant products to be delivered late this year or early in 2000.
Despite its merits, PCI-X is clearly seen as an interim solution. Recently, two vendor camps led by Compaq and Intel divulged plans for a switched fabric architecture to replace buses as the next server I/O structure. The Compaq-led plan, called Future I/O, is backed by IBM and HP. Intel`s proposal, Next Generation I/O (NGIO), is supported by Dell, Hitachi, NEC, Siemens, and Sun.
Both Future I/O and NGIO propose using a switched fabric architecture, which differs significantly from bus-based systems. Instead of data requests moving from the server CPU to a bus and then to an I/O subsystem, a switched fabric moves requests from the CPU to a high-speed switch. The switch then takes the request, similar to a controller, and routs it to a subsystem or another server. Supporters say that by usurping the bus, arbitration issues and latency can be eliminated and servers are no longer limited to a finite number of PCI slots.
Mitch Shults, director of NGIO marketing for Intel, says NGIO is built not only to improve speed and reliability, but to remove the physical limitations of PCI slots.
"End-users will see a fundamental improvement in the level of system reliability and scalability, with a minimum of incremental effort on their part," says Schults.
By essentially eliminating the bus, NGIO is designed in part to eliminate I/O delays caused by device arbitration and latency. The inherent nature of a switched fabric also removes limits caused by a limited number of PCI slots and extends the physical distance between server memory and peripheral controllers. The NGIO and Future I/O camps believe they can deliver compliant products in the next couple of years.
For adapter/controller vendors like Costa Mesa, CA-based Emulex, supporting more than one standard is not great business, but at times a necessary evil. "We`re monitoring both proposals," says Mike Kane, Emulex`s director of marketing. "I don`t expect there`s going to be any material effect on things shipping for at least three to four years."
"In the next few years, I think Fibre Channel will be pretty well entrenched in the enterprise server space," says Kane. "Beyond that, there might be an opportunity for a new standard, but I would be very hesitant to guess what that would be."
Technically, supporting divergent standards is "more of an irritant than anything else," to Eric Herzog, a vice president at Mylex Corp, in Milpitas, CA. He says that training staff to support two different standards and dealing with dual inventory issues could be greater concerns. Both Herzog and Kane expect Future I/O and NGIO to meet somewhere in the middle.
Others, such as Adaptec CEO Larry Boucher, believe the leverage of industry-leading system vendors backing switched fabric initiatives could hypothetically replace Fibre Channel. "There is no question that what they are specifying is a modern alternative to what was defined 10 years ago by the Fibre Channel committee," contends Boucher. "Future I/O and NGIO are defined as both a switched fabric system and a subsystem interconnect architecture. So how many switched fabrics do we need to connect systems and subsystems?" (Adaptec recently sold off its Fibre Channel business.)
Intel`s Schults and Compaq`s Jensen say that eliminating Fibre Channel is not a motivating factor and that merging fabrics is still a long way off. "That`s not something we`re trying to do here," says Schults. "What NGIO is more about is breaking the paradigm of just stuffing in more I/O cards."
"I think over time you`re going to see a merging of these (switched fabrics)," says Jensen. "Are you going to build one fabric that handles everything for everybody? No. You`re going to see a migration."