Take the biggest Compaq server and the latest version of NT and SQL Server, and you might support a useful database in the tens of gigabytes. That's great for lots of applications--maybe most applications--but there's a class of applications that require more, a lot more. Hewlett-Packard, Sun, IBM and the like have maintained a very nice high-margin business serving these elite application needs. We're talking real glass-house stuff here, with big air conditioners and fancy power systems.

Corporate America has considered the glass house a reasonable cost of doing business. Of course, some have deemed the cost of a corporate jet reasonable, too--which, by the way, it isn't (until I get one). But the folks at Intel see this--and most other niches, I suppose--as a niche just waiting to be filled by them and their partners. Much as they did with PCI and I2O, they've proposed the VI (Virtual Interface) architecture and thrown it to an independent development group to market the technology. The VI architecture is intended to make cluster computing with commodity hardware realistic and inexpensive.

Fast, Efficient and Cheap VI caught my attention when a small company called Giganet visited me in mid-September. Giganet has a switch and a PCI card that represent the first embodiment of the VI standard. The switch can move tens of gigabits per second, and each PCI card runs at 1 Gbps. This sounds rather ho-hum given all the Gigabit Ethernet gear on the market, but there is an important difference, and Giganet has an impressive story to tell.

Just like PCI some years ago and I2O two years ago, VI will be a critical standard to watch and understand for server technology. The key difference between VI and typical networking protocols is that when you start moving packets at Gigabit Ethernet speeds, CPU utilization shoots through the roof.

I've previously cited the ability of a CPU to move data as roughly equal to its clock speed (at 200 MHz, a CPU can move data at about 200 Mbps). For most protocol stacks, this equation has been fairly accurate, but the VI architecture changes this drastically. The architecture, as detailed by Intel, is intended to move data quickly and efficiently. Giganet has shown that its hardware can move more than 800 Mbps over its gear with less than 10 percent processor utilization on high-end Pentium servers.

High Speed, Low Latency Since we're talking clustering, not only do you need to move lots of data very quickly, you also have to move small packets with minimal latency. Successful cluster computing requires that data in memory be shared easily between processors. At hundreds of MIPS, these transfers had better happen with latencies in times measured in a few microseconds or less, otherwise millions of cycles can be wasted waiting for data. Today, Gigabit Ethernet latencies are measured in hundreds of microseconds. VI hardware operates at latencies of a few microseconds. By contrast, PCI cards right on the processor bus exhibit latencies just less than 1 microsecond.

And Giganet's stuff isn't just fast, it's reasonably priced. The interface cards cost less than $800; the eight-port switch less than $6,500. For about $10,000, you can start building servers into VI clusters. Unfortunately, though, there are some roadblocks to cluster nirvana. Most significant, don't look for OS-level support for VI from Microsoft anytime soon. The best I could get from Microsoft was that it would support VI sometime after NT 5 shipped. Until then, applications themselves must be VI-aware.

The first demo of this technology was conducted recently with Giganet's hardware, Dell servers and a special version of DB-2 from IBM. A terabyte database of census data was created and used--and it all worked. So is VI in your future? Very possibly. I don't think your Amex billing statement will be printed from VI architecture-supported databases anytime soon, but data mining could find a home on such systems. VI is definitely a standard to watch.

Send your comments on this column to Art Wittmann at awittmann@nwc.com.