 
|
|||||||||||||||||||||
              
|
Pittsburgh supercomputer is complete, and scientists are champing at the bit to use it
Monday, October 01, 2001 By Byron Spice, Science Editor, Post-Gazette
Astrophysicist Mike Norman would love to witness star birth, to see exactly how a giant cloud of gas collapses under the force of gravity to form stars. He doubts he'll ever see it through a telescope, but he's pretty sure he will someday through the lens of a supercomputer.
LOCATION: Westinghouse Energy Center, Monroeville
MANUFACTURED BY: Compaq Computer Corp.
SYSTEM: Consists of 750 interlinked Compaq AlphaServer computers housed in 150 6-foot-high cabinets.
SIZE: The 150 cabinets, arrayed in nine rows, occupy 3,500 square feet. (A college basketball court is 4,200 square feet.)
PROCESSORS: Each AlphaServer has four Alpha EV68 processors. Each processor can perform up to 2 billion floating-point calculations per second, or 2 gigaflops.
SYSTEM CAPABILITY: The 3,000 processors can perform up to 6 teraflops, or 6 trillion calculations per second. Virtually every man, woman and child on earth would have to perform a calculation each second to keep pace.
RANDOM ACCESS MEMORY: 2.7 trillion bytes, or terabytes
SYSTEM STORAGE CAPACITY: 50 terabytes on hard disk, 300 terabytes on tape or disk storage.
CABLES: About 6,000. The longest interconnect cable is 33 meters.
POWER DEMAND: Half a megawatt.
COOLING: 500 gallons of chilled water circulates through the computer room each minute, supplying 100 tons of cooling -- equivalent to the energy released from melting 100 tons of ice in 24 hours.
SPONSOR: The National Science Foundation last year awarded a three-year, $45 million grant to the Pittsburgh Supercomputing Center to build and operate the new machine.
A related story ...
On the Web:
That someday may come soon, now that technicians have finished assembling the Terascale Computing System in the basement of the Westinghouse Energy Center in Monroeville. The sprawling computer, which takes up almost as much space as a basketball court, links 3,000 high-speed computer processors.
With the capability to perform up to 6 trillion calculations every second when running flat out, it will be the world's most powerful computer doing unclassified research. In benchmarking tests last Friday, the machine established itself as the second-most powerful in the world, second only to the ASCI White computer that simulates nuclear explosions at Lawrence Livermore National Laboratory.
Such rankings are transitory; with machines being upgraded and even faster machines planned, the Terascale could be obsolete within three years unless it is expanded. But for scientists such as Norman, a San Diego astrophysicist who has used supercomputers for 25 years, the added computing power available now with the Terascale machine means he can increase the resolution, or fine detail, of his simulations.
And that means new discoveries are at hand.
"It's just like a telescope," explained Norman, a professor at the University of California, San Diego. "You increase the resolving power and new things start popping into view."
Completing the assembly
First, the Terascale needs to become operational. Today was the deadline set by the National Science Foundation for getting all of the machine's pieces in place. But it will be available only sporadically for scientific computing over the next few months, as staff members at the Pittsburgh Supercomputing Center test and fine tune it.
The machine consists of 750 of Compaq Computer Corporation's AlphaServer computers, each of which contains four processors. Getting all of those pieces together by today looked iffy after the terrorist attacks on New York and Washington temporarily shut down air cargo traffic, including shipments from Compaq's plant in Ayr, Scotland. The final 180 AlphaServers arrived nine days ago and were connected a week ago.
"It's quite a piece of work," said Tony Woodman, technical program manager for Compaq. Since early August, Woodman has supervised crews working two, 10-hour shifts a day, seven days a week, to assemble the Terascale machine. "It's a cable nightmare," he added, with an estimated 6,000 cables -- together weighing more than four tons -- snaking below the floor or in racks above the cabinets to connect the machines.
The 150 cabinets are arranged in nine rows and kept as close together as possible so that the longest cable measures 33 meters. Designers were battling against the speed of light, which limits how fast signals can travel the cables. Every additional meter of cable would reduce the machine's throughput by 0.5 percent, said Lynn Layman, the supercomputing center's facility manager.
That's no big deal for some machines, but even half a percent can mean big losses for a $45 million machine that theoretically can perform 6 trillion calculations a second.
"It's a lot more complex than a standard cluster," Woodman said, noting the system includes a lot of custom equipment, such as the switching console at the center of the clustered boxes. "It's on the bleeding edge of technology."
Which is exactly where Norman, the astrophysicist, likes to be.
Formerly a staff member at the National Center for Supercomputing Applications in Urbana-Champaign, Ill., Norman is a steady customer at all of the NSF's supercomputing sites. He's one of a select number of "early users" who get first crack at Pittsburgh's new machines, including a prototype for the Terascale that began operating in January.
Using that smaller machine, which can perform up to 342 billion calculations per second, Norman's team early this year was able to simulate part of the star formation process. It was the most detailed such simulation ever run, modeling a gas cloud within a three-dimensional, 512-point grid.
What Norman and other astrophysicists would like to learn is why star formation is so inefficient. Only about 1 percent of the gas in these clouds is converted into stars, probably the consequence of magnetic fields counteracting the force of gravity.
"We've had hand-waving arguments about this for 30 years," Norman said, but the questions remain -- how does this occur and how much does it occur?
This year's run on the Terascale prototype took 450 hours and was both revealing and frustrating. It showed a cloud being compressed into a dozen, spinning disk-like cores and at roughly the same efficiency as the rate of star formation observed by astronomers. But then the simulation petered out. Star birth and the formation of protoplanetary disks -- the swirl of matter from which planets form -- require more detailed simulations.
"The fact that we're in the ballpark is encouraging," Norman said. But further work will require a model that is twice as detailed -- a 1,012-point grid. Computing that grid in three dimensions requires calculations at a billion grid points, compared to the 134 million points of the 512-point grid.
"The Pittsburgh machine is the only machine where I could attempt the bigger simulation," Norman noted.
Getting up to speed
Though it has the raw power to perform 6 trillion calculations per second, or "teraflops," that is only if all of the processors are running flat out. In actual operation, some processors will spend some of their time waiting -- waiting while other processors finish calculating a different part of a problem, or just waiting the 33 billionths of a second it takes signals to travel down one of those long cables.
The ultimate speed of the machine will depend on how well designers can minimize the delays in moving data around the machine and on the ability of programmers to write software that efficiently allocates computing tasks among those thousands of processors.
Compaq engineers on Friday put the machine through a series of benchmarking tests that are used to rank the Top 500 computers in the world. The top-ranked ASCI White computer, for instance, is theoretically able to perform up to 12 teraflops, but reaches 7.2 teraflops on the benchmark tests. In its initial benchmarking, the Terascale reached 2.8 teraflops -- besting Lawrence Berkeley Laboratory's 2.5-teraflop Seaborg computer -- but officials expect the Pittsburgh machine will top 3 teraflops as they continue to test and tweak the system this week.
The Terascale's performance likely will improve further as operators and users become acquainted with the new machine. This learning period, sometimes called "the six-month miasma," occurs after delivery of every new machine, said Michael Levine, scientific co-director of the Pittsburgh Supercomputing Center.
The NSF this summer announced an even faster unclassified machine, the Distributed Terascale Facility. That's a $53 million project, which by mid-2002 will tie together the existing NSF supercomputing centers in Illinois and San Diego. Also, a new 6.1-teraflop computer will be built at the Illinois site. By April 2003, the networked supercomputers will have a peak performance of 11.6 teraflops.
The Distributed Terascale Facility "is a bit of an experiment," said Bob Borchers, director of the NSF's advanced computer infrastructure division. The problem of communication delays between computers will be magnified by the long distances between the sites, which will make the distributed facility unsuitable for some types of problems.
The same high-speed links will be extended to Pittsburgh next year, so the Terascale computer could be tied in to boost the power of the distributed facility even further, he added.
But Borchers said that the days of single, massive computers such as the Terascale are hardly endangered. Already, officials have discussed the possibility of eventually upgrading the Pittsburgh computer so its performance can reach 20 or 25 teraflops.
"We will try to keep the Pittsburgh machine viable and in the high end," Borchers said.
|
||||||||||||||||||||
Pittsburgh Supercomputing Center,