IBMs scaled-down version of its Blue Gene/L doesnt look like much of a supercomputer. About the size of a dishwasher, the Linux-based prototype is now the 73rd-fastest computer in the world and is just a sample of whats ahead, according to its chief architect.
This version of Blue Gene/ L is about 1/128th the size of the full Blue Gene/L supercomputer that IBM is building for the Lawrence Livermore National Laboratory in Livermore, Calif., and the first part of a two-pronged IBM project to create the fastest computers in the world.
The Blue Gene/L system being built for the Livermore lab—run by the University of California for the U.S. Department of Energy—is expected to be completed by the end of next year or in early 2005, said Al Gara, chief architect for the project, being run by IBM Research, in Yorktown Heights, N.Y.
Gara said he was surprised the Blue Gene/L prototype made it onto the list compiled by the Top 500 Supercomputer project and released at the SC2003 show in November but said it should give the industry an idea of how IBMs ambitious project is going. “Its a great precursor of whats to come,” Gara said. “This is just the tip of the iceberg.”
Eventually, what will come is Blue Gene/P, a supercomputer with a peak performance of 1 petaflop—or 1 quadrillion calculations per second. Blue Gene/L is the first of two programs within the project that IBM researchers will use to create the final product.
When completed, the $100 million Blue Gene/L is expected to have a peak performance of 360 teraflops, or 360 trillion operations per second, and will take over the top spot on the Top 500 list. The Earth Simulator in Japan, built by NEC Corp. last year, has a peak of about 40 teraflops, Gara said. The prototype, with 256 processors in a 21U (36.75-inch) rack, has a peak of about 2 teraflops.
Blue Gene/L will have more than 64,000 dual-processor compute nodes housed in 64 racks, according to Gara. The computer will be powered by IBMs PowerPC 440GX processors and will give users a familiar feel. Key to the project is the number of functions and features that will be housed directly on the silicon, including two processors and the interconnect technology. One processor will perform calculations, while the other is dedicated to communications among the nodes.
Each piece of silicon will also hold 4MB of Level 3 cache, as well as Levels 1 and 2 cache, and several network interfaces for such tasks as supporting the interconnect that runs the Message Passing Interface protocol used in parallel supercomputing clusters.
The density of the system is one of the challenges for any company building a supercomputer, said Gordon Haff, an analyst with Illuminata Inc., of Nashua, N.H. The density drives much of the design work being done by such companies as Silicon Graphics Inc. and Cray Inc., Haff said.
“If you can make it denser, you can reduce the latency significantly [and] ramp up the communication [among the processors],” Haff said. “Density in the data center environment means that you can free up more floor space and save money. But this [supercomputing environment] is different. This is [where density means] the closer you can keep the processors, the faster they can communicate and the faster the entire computer can possibly run.”
IBMs challenge now is ensuring that the density seen in the Blue Gene/L prototype can be scaled up for the full computer and that having all those processors so close together doesnt result in heat problems, Haff said.
Blue Gene/L is one of two prongs in IBMs overall Blue Gene project. Researchers are working on the design for Blue Gene/C, which Gara—who doesnt work on that part of the project—said probably will be more cost-effective and aimed at more specialized applications. Eventually, IBM researchers will look at the two designs and decide on the direction for Blue Gene/P.
“Its always about price,” Haff said. “I dont care if you have a budget like the Livermore lab; if you can get performance for a tenth of the price, that lets you do things you couldnt do before.”