Posted by : bpburnwal on Oct 14, 2003 - 10:03 AM General

AUSTIN, Texas _ A few years ago, it would have been inconceivable to see a Dell Inc. server inside the University of Texas' supercomputing lab.
The Texas Advanced Computing Center does the heavy number-crunching for UT researchers and runs experiments with other university research labs around the world. It has helped researchers better understand black holes and how the ocean floor changes during earthquakes. Sitting in a chilly, closely-guarded room deep inside the center at the J.J. Pickle Research Campus in North Austin, three hulking supercomputers hum 24 hours a day, calculating billions of operations a second.

During the summer, they were joined by a sprawling collection of refrigerator-size racks filled with Dell servers _ about 300 pizza box-size machines, to be exact. Linked through a maze of cables and switches, the servers make up UT's newest supercomputer _ the Lonestar.

Officially launched Oct. 3, Lonestar is a Linux cluster, a bunch of servers linked together to work as one computer. UT is among a growing number of research institutions and large corporations to embrace clustered computers running the free Linux operating system.

"We needed a larger, more powerful system than we had," said Jay Boisseau, director of UT's computing center.

And cheaper, too. A Linux cluster costs about 90 percent less than a comparable Unix supercomputer, Dell says. That means a customer could spend about $200,000 on an Intel cluster instead of $2 million on a Unix server. Competitors argue that Linux clusters and bigsupercomputers are completely different, akin to comparing a Yugo to a Ferrari. But that hasn't stopped IBM Corp. and Hewlett-Packard Co. from also embracing the systems.

"It's built on the idea that if I can take these standard building blocks and connect them together, I can make a much more powerful, larger computer," said Pete Morowski, vice president of software in Dell's enterprise computing division.

The Round Rock, Texas, computer maker is pushing the benefits of clustering for one reason: It helps Dell sell thousands of additional low-end 2-processor Intel servers. Because servers are more profitable than desktops, that helps Dell keep revenue and profits up.

The company is pursuing a similar strategy in the corporate market with help from partners such as database software maker Oracle Corp., which provides software needed to make clustered systems more stable.

Dell's efforts are making it a small, but slowly growing, player in the supercomputer market, which is dominated by IBM, H-P and Sun Microsystems Inc.

Dell led the Intel-cluster market with 36 percent share of total revenue in the second quarter, according to tech research firm IDC. But that's a small fraction of the supercomputing market, which was led by H-P's 37 percent share of revenue.

IBM was second with 27.6 percent. Dell ranked fourth with 6.8 percent.




Two years ago, Dell was nowhere on the list of the top 500 fastest supercomputers in the world. During the summer, 15 Dell clusters made the list.

That's a tiny fraction compared with H-P and IBM, whose systems each accounted for about a third of the list. But it's an achievement for a company known more for selling low-cost PCs than making advances in computer research.

Clustering and grid computing is so important to the computer maker that it recently opened a multimillion-dollar lab at its Parmer Lane complex to test clusters and help develop solutions to problems encountered by its customers.

'`It's now become a fairly large business,'' chief executive Michael Dell said.

The company got into clusters because it discovered research institutions were buying hundreds of Dell servers and assembling the supercomputers themselves, he said.

``The initial pull came from customers,'' Dell said. ``Now it's become kind of a global phenomenon."

Clustering has been around since the 1990s but gained popularity in the past three years as technology budgets tightened and Linux, which is used on most clusters, began to move beyond university research departments.

``You can have a very substantial amount of computing for less money,'' said Sara Murphy, H-P's manager for grid and high-performance computing.

H-P and IBM long have dominated the supercomputing market with highly reliable, big-box servers running their own customized software.

But they, too, embraced low-cost Linux clusters, spending millions to develop software and services to improve their performance.

Clusters have some disadvantages. They can take far longer to install than a proprietary Unix system, which generally can be out of the box and running in a day or two. At the UT lab, the Dell cluster has been under construction by technicians from Cray Inc., a Dell partner, since July.

``The last thing (customers) want to do is spend two months putting the thing together after they've acquired the capital equipment," said David Turek, vice president of supercomputing at IBM, which sends technicians out to install and test Linux systems for its customers.

Dell farms out software, services and some installation to other companies.

Physically, clusters also can take up lots of floor space. They generate a tremendous amount of heat, even when the servers are sitting idle. And the sheer number of moving parts require-- more oversight and maintenance.

``They're a pain in the butt sometimes to manage. There's a lack of (software) management tools,'' said Bill Claybrook, an analyst with Aberdeen Group, a Boston research firm. ``There are a lot of things that people don't like, but (clusters) are so overwhelmingly inexpensive compared to the big boxes that people are willing to put up with it.''

Affordability and expansion possibilities were the deciding factors in UT's decision to devote a significant amount of precious floor space for the sprawling new Dell system, Boisseau said.

At UT, seven racks of two-processor Dell servers sitting in the middle of the computing center's supercomputing lab will be joined by another 200 servers by the end of the year, and UT plans to double the system's size next year.

At its peak, the $3.5 million system will be capable of 3.6 trillion operations a second, about three times as many as the lab's older IBM Unix-based system. The system features $1.9 million in hardware and $1.6 million in services and support.

Clustered computers aren't good for all experiments, but the new system will be a faster alternative for some researchers who have programs that can be broken into smaller-size bites, said Jim Foster, who helps oversee the lab.

``Now we have both systems for our scientists, whatever their needs might be,'' he said.