InformationWeek

Another user of Sequent's NUMA Q system is more interested in hotel reservation systems and sea cruises then smoke and mirrors. "We need NUMA Q because we 've grown so rapidly and continue to expect growth," says Scott Heintzeman, VP of knowledge technologies for travel and hospitality company Carlson Cos. in Minneapolis. "We don't want more capacity than we need, but we also don't want to fall short."

Trust Adjuster
Heintzeman enjoys the modularity of the NUMA architecture because it lets him add to each of three Sequent systems. "We can increase capacity on a day's notice," says Heintzeman. "With other systems, I don't have such confidence."

When someone in Carlson's marketing group makes a change, Heintzeman has to update hotel software in 34 countries. He's not worried about reliability: "NUMA Q will meet our needs for the foreseeable future." But Heintzeman won't keep his eyes closed to other vendors. "We'll always have some new technology to consider," he says, "but if I have a technology that allows me to grow 11- or 12-fold, that's enough."

Such decision-making is a driving force behind Sequent's NUMA strategy. "It used to be 15% of the people in the company made decisions," says Casey Powell, chairman and CEO of Sequent. "Now decisions are made at all levels of the company, so we have to install a system to support the decision-making, and hook up the other 85% of the people."

Such requirements aren't the domain of Sequent customers only. Larry Mohr, VP of IS for F.W. Webb in Burlington, Mass., doesn't think the largest heating and ventilation company in New England needs him to be its technical guru. "Moving to Data General's NUMA is a move to performance and growth," says Mohr.

Webb runs a Data General AV 20000 NUMA system with Intel inside for order entry, inventory, EDI, and bar coding for nearly 2 million records. Mohr says backups take just 60% of the time they did with his older Data General AV 9500, which was based on Motorola chips. "We're also in great shape with large records," he adds. "I can go in and select in the middle of the day, in a half-hour. That used to be a four-hour job."

Even with such NUMA scalability, Data General isn't expecting a mainframe-killer anytime soon. "Some NUMA systems will be bigger than some mainframes, but they won't supplant the mainframes because the customers are stuck on all that legacy software they won't rewrite," says Ron Skates, chairman and CEO of Data General, in Westboro, Mass.

Another NUMA vendor, Silicon Graphics Inc., is more concerned about competing in the high-end Unix server market than looking ahead to mainframe competition. The company, struggling to be recognized as a commercial server vendor, is betting the bank on the NUMA architecture, having built its entire server technology infrastructure on NUMA. Silicon Graphics says it has already sold thousands of NUMA servers and expects the investment will pay off.

Ihab Abu-Hakima, VP and general manager of Silicon Graphics' enterprise server division, says SMP is severely limited by the bus architecture. "In five years, SMP won't be sold," he predicts. "Within four years, nothing over four processors will rely on a bus."

Unix vendors disagree. "SMP will keep growing, and the broader market will never need NUMA," insists Dan Glessner, HP's marketing director for Enterprise 9000 servers. HP has invested in NUMA, but only for the technical and high-end data warehousing markets.

Similarly, Sun says it achieves all the SMP scalability needed-up to 64 processors-with a cross-bar interconnect that attaches every board to every other board. "The 'NU' in NUMA is a bug, not a desirable feature," says Shahin Khan, Sun's director of marketing for high-performance computing. Yet Sun is developing a NUMA-like technology, called COMA (cache-only memory architecture), that differs from NUMA only in how it deals with cache coherency.

Hollywood Entertainment's Hnanicek conducted benchmarks on the best of HP and Sun last year before choosing NUMA Q from Sequent. "A 12-processor SMP system isn't three times faster than a four-way, but with NUMA we get linear scalability," he says. "We can't detect any overhead ."