Entry No.125

IT Writers Awards

The world's best-known supercomputer is probably Deep Blue, the IBM machine that defeated world chess champion Garry Kasparov in a tournament in 1997. When they're not beating chess grandmasters, supercomputers are generally employed performing more mundane tasks such as nuclear simulations, economic analysis or modelling vehicle crashes.

A machine such as Deep Blue has about as much in common with your desktop PC as pond sludge does with a Bengal tiger. Supercomputers are highly evolved machines out of reach of all but the most wealthy corporations, governments and research institutions. Which is why a project such as Bunyip, the budget super computer at the Australian National University, is of extreme interest.

Bunyip falls into a special class of supercomputers, known as Beowulf, which link common, off-the-shelf components (COTS), such as Intel processors, mass-market hard drives and network equipment, using the Linux operating system.

"The first Beowulf machines were built back in 1994 by a group associated with the earth and space sciences project at NASA's Godard Space Flight Centre," says Chris Johnson, head of ANU's department of computer science. "Although the name was originally applied just to the Godard machine, it has since become an overall classification for computers built along the same principles."

Bunyip cost $250,000 to put together and uses 192 Intel Pentium III processors each running at 550Mhz in addition to 36 gigabytes of random access memory (RAM) and 1.3 terabytes of hard-drive storage

Performance is a tricky thing to get a handle on, which is why many publications benchmark things such as frame rates - the speed a computer can draw pictures, such as those in complex 3D games like Unreal - or how fast it can calculate a spreadsheet.

The problem is that Bunyip isn't a PC and so its performance has to be measured in terms of the number of floating point operations (a type of mathematical operation) per second that it can perform. In this case, Bunyip is able to execute 158.4 trillion floating point operations per second (also called gigaflops, or gflops). By way of comparison, the average desktop computer can manage about 1 gigaflop, although this can be misleading because, in much the same way that horsepower doesnít reflect how fast a car can go, gigaflops donít really reflect the real-world performance of a computer.

Suffice to say, however, that Bunyip is reasonably fast. ìWe expect to get at least five years out of this machine,î says Johnson. ìAfter that, its performance will become painfully slow due to the increase in computer performance. It will still be useful, however, for doing research on parallel processing.î

Parallel processing is the taking of a problem and dividing it into manageable chunks. The chunks are processed by individual processors and the interim results are then returned and assimilated into a final result. Part of Bunyipís charter is as a tool for developing better parallel-processing techniques, the other part being research into optimising communication between individual parts of the machine. This is important for Beowulf systems, which use fast networks and switches to connect their processors. Other supercomputers are integrated much more tightly to minimise the amount of time it takes for a piece of data to move through the system.

Cost has been one of the main drivers behind the development of the Beowulf cluster, says Johnson. "[Bunyip] delivers a huge amount of computing power at an acceptable cost, one thatís far less than the cost for a specialised supercomputer.î The cost of specialised machines, such as those used in nuclear research laboratories in the United States, can run to tens of millions of dollars - and thatís just for starters.

The use of the Linux operating system assists the meeting of cost goals because it is freely available, as is its source code. Access to source code is the digital equivalent of being able to get under the hood of a car and play around with the greasy bits, an attribute prized highly by computer researchers and the digerati.

ìAny changes or alterations that we make to [Linux] are also made available to the community at large,î says Johnson. ìThat is because of the way Linux is licensed under the GNU Public License. But itís not the software thatís the most important outcome from this project. Itís the intellectual property and the know-how that is generated that will generate the real value.î

Itís unlikely computers such as Bunyip will replace the hyper expensive supercomputers that live in the rarefied atmosphere at the apex of the computing tree. ìThese machines slot in just underneath those hyper expensive supercomputers,î Johnson says. ìThere will always be a market for the highest performance available but Bunyip, and machines like it, provide another option that's both cost effective and within the reach of a greater pool of people."

Back to Voting Form
Back to Best Features  
Top of page

Content Copyright © the author/publisher listed above

Design Copyright © Consensus Pty Ltd

This web-site uses frames, click here for the full picture