'

Blue Gene and the PetaFLOP promise

Andrew Brockfield shies away from the term 'grudge match', but he concedes there's national pride involved as the race to build the world's fastest supercomputers continues to push the bar upwards at dizzying speed.

Andrew Brockfield shies away from the term 'grudge match', but he concedes there's national pride involved as the race to build the world's fastest supercomputers continues to push the bar upwards at dizzying speed.

One of the products of that race -- Blue Gene, a modular system that this year unseated Japan's long-reigning performance champ the Earth Simulator as the fastest computer in the world -- recently spent several weeks at IBM's Melbourne offices before being transferred over the weekend to APAC '05, the high-performance computing (HPC) conference that kicks off today on the Gold Coast.

Designed for portability, the Blue Gene system on display at the show -- a one-eighth filled rack with 128 computing nodes containing 256 700MHz IBM PowerPC 440 processors -- is just a shadow of the 65,536-processor Blue Gene/L, the world's fastest computer, whose 32 full racks provide 136.8 teraFLOPs (trillion floating point operations per second) of computing power to simulate nuclear explosions at Lawrence Livermore National Laboratories (LLNL) in the US. LLNL will soon take delivery of 32 more nodes, raising total computing power to around 270 teraFLOPs and further widening the gap with Earth Simulator.

The TOP500 has become a barometer of computing progress, and high rankings are seen as merit badges of a sort - particularly for the likes of Intel, upon whose chips 333 of the TOP500 relatively inexpensive HPC systems are now running, up from virtually none just three years ago. Topping the charts is so important that the two-and-a-half-year domination of Japan's Earth Simulator spurred US government enquiries as to why that country hadn't dethrone its trans-Pacific neighbours.

IBM, whose Deep Blue technology succeeded in battling world chess champion Gary Kasparov to a draw several years ago, used that specialised technology as the basis of the five-year, US$400 million R&D effort that begat Blue Gene and finally broke Earth Simulator's lock on the top spot.

A few years ago, just a few systems had reached 1 teraFLOP, but every current member of the official TOP500 list offers more than 1.1 teraFLOPs. Even the scaled-down Blue Gene system at APAC '05 offers 1.4 teraFLOPs. Ready availability of such immense computing power has transformed complex modelling in scientific research, such as biotechnology and astrophysics, as well as computationally intense industries such as petroleum exploration.

Scientists and researchers are finding Blue Gene's power so appealing, in fact, that IBM is now mass-producing the systems (in a relative sense; only 16 of the machines currently exist) from its Rochester, Minnesota server factory. For around US$2.5 million you can have one of your own -- and a growing number of companies are jumping at the chance.

"The range of applications Blue Gene is suited to is not as broad as [generic] Power PC, Intel or AMD based architectures," says Brockfield, an HPC specialist who has stewarded the Blue Gene system during its Australian tour. "Bear in mind that it was first designed as a pure research project, and when we first designed the machine we didn't think astrophysics would be an application domain for Blue Gene. We've been pleasantly surprised at the breadth of applications that Blue Gene is suited to."

Scientists must rewrite much of their code to work on Blue Gene, but that hasn't hurt its popularity. During its Australian tour, the demonstration system has been flat-out running tasks from local scientists involved with regional HPC groups including VPAC (Victorian PAC) and APAC (Australian Partnership for Advanced Computing), at whose conference the system is being displayed.

Just as significant as the systems' computing power, however, is its portability: drawing less than one-tenth of the power of a conventional Intel-based system, the demonstration unit is cool enough to run quite happily on the APAC '05 exhibition floor without any of the specialising cooling systems found in conventional data centres. Blue Gene is more than 60% cooler than power-hungry Intel-based servers, which run hot enough that cooling the systems has become more complicated than building them.

HPC giants aren't sitting on their laurels. With 32 additional Blue Gene nodes soon set to double Blue Gene/L's performance to 270 teraFLOPs, experts are now pondering HPC's next goal: the petaFLOP computer, offering 1000 teraFLOPs - enough to handle impossibly complex computations such as modelling of protein folding within the body. Brockfield can't say when Blue Gene/P, as it's now known, will be delivered -- but with Japan already planning a retaliatory strike on Blue Gene/L, odds are that it won't be long.