The IBM SP
SCD welcomed the newest member of its computing family with the arrival of an IBM RS/6000 SP at the Mesa Lab on 11 August. The $6.2 million machine arrived on schedule and without a hitch. According to SCD associate director Bernie O'Lear, "By 3:30 this afternoon [Wednesday], power was going to the machine and the operating system was successfully booted on all 160 nodes."
The scene on the mesa Wednesday morning was quite different than that for arrivals of supercomputers in years past. Unlike the gigantic mainframes once lowered into the SCD machine room through a hatch in the ceiling, the SP arrived in dozens of boxes wheeled from the loading dock through the basement to the machine room.
One of two trucks that brought the IBM SP to Boulder pulls up to the loading dock. (Photo by Greg McArthur.)
The distributed arrival is a clue to the new machine's architecture. The SP includes 160 nodes, each with two processors. Each node has its own memory, and message-passing protocols are used to exchange information between the nodes. This distinguishes the SP from SCD's vector computers, such as the Cray C-90, and from massively parallel machines like the SGI Origin 2000. The SP's peak speed is estimated at 256 gigaflops, more than tripling SCD's previous peak capacity.
Al Kellie at the press conference.
(Photo by Greg McArthur)
At a brief press conference on the mesa, several officials from UCAR, NCAR, and IBM lauded the SP and its promise for weather and climate modeling. "We're tickled pink that we can come out and visit with customers who are doing the kinds of things you're doing," said Lou Bifano, IBM vice president for the SP series. "We've had a lot of fun designing and building [the SP], and we hope you have just as much fun using it."|
SCD director Al Kellie noted that the SP is "part of the vanguard of new, robust supercomputing systems being produced by U.S. manufacturers." He added, "This is our first little baby step along the way to terascale computing."
Technicians from NCAR and IBM have been working around the clock to ready the new computer for users. "What's happening today and tomorrow is they're making the connections for the internal networks," said SCD network engineer Pete Siemsen. "One of them is an extremely high-speed proprietary technology that IBM makes for the SPs." This high-speed network allows for four connections from each node to every other node, according to Pete. It's used for the bulk of data processing. The other network is a fast Ethernet (100 megabits per second), used as a control network to guide the machine's operations.
Left: SCD's Wes Wildcat records the SP's arrival for posterity. Right: Two of the key people in the SP acquisition and delivery were George Fuentes, head of supercomputer systems for SCD's High Performance Systems Section, and George Maroulis, client representative for the Global Government Industry Sector at IBM.
(Photos by Greg McArthur.)
Network testing should continue into late Thursday or Friday. Software testing, again with on-site assistance from IBM staff, will take place next week. It will likely be the week of 23 August before users are able to log onto the SP. "Things are going better than expected--really smooth," says Pete. "There've been no surprises." --Bob Henson
Watch for more coverage of the SP transition in the upcoming issue of the SCDzine and through
the SCD home page. More details on the new machine and its impact will also
appear in the September issue of Staff Notes Monthly and the fall issue
of the UCAR Quarterly.
Left: Movers bring the IBM SP mainframe into its new home.
Right: IBM's Lou Bifano grants an interview at the Mesa Lab press conference.
(Photos by Carlye Calvin.)
Note: The rest of this issue will be posted in early September 1999.
Prepared for the Web by Jacque Marshall