UCAR > Communications > Staff Notes > December 1996 Search

A new C-90 arrives for climate modeling

NCAR's climate system model (CSM) and other integrated models have an spacious new home. Over Thanksgiving week, a CRAY Y-MP8I supercomputer leased by NCAR since 1991 was replaced by a CRAY C-90. The new machine--dubbed antero, like the one it replaced--has 256 million words of memory and 16 processors, twice as many as its predecessor. It can produce up to 5 billion floating-point operations per second.

"We believe it'll give us a factor of three to four increase in speed over the previous antero," says Bill Buzbee, SCD director. As the linchpin of the Climate System Laboratory (CSL), the new Cray will be dedicated to extensive climate simulations.

The C-90 has arrived at SCD through an 18-month extension of the previous lease arrangement for the Y-MP8I with Cray Research, Inc. The old lease had been due to expire in mid-1997.

The arrival of the C-90 comes after a stalled process for longer-term acquisition of a new supercomputer for climate modeling. In May, NCAR announced the selection of the Japan-based NEC Corp. to provide four large vector supercomputers over five years. However, the acquisition has been put on hold pending a formal complaint by Cray. Investigations are now under way by the International Trade Commission and the U.S. Commerce Department. Decisions are not expected until well into 1997.

"We spent the better part of two years developing requirements for supercomputing support for the CSM," says Jeff Reaves, UCAR associate vice president for finance and administration. "The extension of our lease with Cray enables us to provide the additional computing power we need for the Climate Modeling, Analysis, and Prediction program and other modeling projects."

The C-90 is in its testing-and-acceptance phase through December. Shortly after the two-day hardware installation and checking process, software was delivered and installed. The next step was to ask "friendly" users (primarily NCAR-based modelers) to begin running familiar programs and verify that the algorithms worked as expected. "Typically on a big mainframe, the acceptance tests take one or two weeks." says Bo Connell, head of SCD's Computer Production Group (formerly the Operations Group). •BH

[All photos by Carlye Calvin.]

Above and below: Workmen lower the C-90 into SCD's staging area.

Before the C-90's bright blue "skin" was attached, the computer's innards were in full view.

The C-90 weighs in at 256 million words of memory with 5 gigaflops. On the left is the machine's solid-state storage device; at right is its central processing unit.

Next step: a J-9

Hot on the heels of the C-90, another powerful new computer is joining the ranks in SCD, this one for the overall user community. Just before press time, NSF gave its approval for acquisition of a CRAY J-9 to be delivered this month.

NCAR already has two other J-9s, one with 20 processors devoted to the CSL and the other with 16 processors for users at large. The new machine also has 20 processors, but it boasts more random-access memory than any other NCAR machine to date: a billion words. "The large memory will offer community users the opportunity to run much larger jobs than they can now run on the [community] Y-MP," says Bill Buzbee.

This is one of two NCAR's current J-9 models, soon to be joined by a third.
The new J-9 earned a strong recommendation from the SCD advisory panel. Purchased outright rather than leased, it should be on hand for two or three years, said Buzbee. "If we can keep the old community Y-MP running as well, then the net effect will be a 50 percent increase in community computing power." •BH

Fred gets a mate

When SCD got its first robotic silo from StorageTek in 1989, the machine was dubbed "Fred," as in Flintstone. (For better or worse, the name didn't stick.) The silo's job was to store up to 6000 data-stuffed cartridges and provide access to them for users of NCAR supercomputers. Visitors were soon able to watch Fred at work via an internal camera and a lobby-mounted monitor.

Now Fred has company. A second silo arrived at SCD in late August and is currently going through its final stages of testing. From the outside, it looks identical to its predecessor. Inside, though, it's a silo on steroids. Its cartridges hold up to 50 gigabytes of data, compared to 800 megabytes for the older cartridges. Thus, the new machine eventually could hold up to 60 times more data than the old one does in its present configuration.

"The robotics are identical. The only difference is in the cartridges and in the tape drives attached to the silo," says Gene Harano, head of the SCD Mass Storage Systems group. The old silo has 16 drives, while the new one has 8 drives configured to accept the new, higher-capacity cartridges. If budgets allow, 8 additional drives could be added to the new silo later on.

Silos old (left) and new (right) sit side by side in the SCD machine room.
The difference between the old and new cartridges is roughly analogous to the difference between an audio cassette tape and a videocassette tape. Data are stored on an audio tape in straightforward fashion, perpendicular to the tape's length, while a videocassette uses what's called a helical-scan approach in which data are recorded in staggered, angled tracks.

The helical-scan technology means that more information can be packed onto a given stretch of tape, but storing or accessing short bits of data is relatively time-consuming (just as it seems to take forever to stop and start a videotape several times over). Because of this, "the new drives don't handle small files of less than 50 megabytes very well," says Gene. "It takes a long time to load the tape and get it positioned. If you write a small file, you're wasting a lot of time. But you come out way ahead for big files because of the much higher transfer rate."

SCD has opened the new silo with a starter set of about 600 cartridges on hand, representing 30 terabytes of data-storage potential. About 100 more of the new cartridges are being used as a back-end, manually mounted archive. They represent a potential solution to the mushrooming of long-term storage in the machine room. "We're slowly chewing up the room space with cartridges that have to stay on hand," says Gene. In time, many of the older cartridges in SCD's permanent archive will be copied onto the new format.

The old and new silos are joined at the hip, as it were. A pass-through connection allows one silo to transfer a cartridge to the other for temporary storage or use, although the dissimilar tape drives and cartridge formats prohibit the latter option for now. Eventually, the older silo could be retrofitted to share formats with the new one. •BH

In this issue...
Other issues of Staff Notes


Prepared by Jacque Marshall, jacque@ucar.edu, 303-497-8616