A supercomputer designed and built by The Australian National University in Canberra, Australian Capital Territory, has won the prestigious year 2000 Gordon Bell Prize for Supercomputing in the best price/performance category. The prize was announced at the annual Supercomputing Conference SC2000, held in Dallas, Texas, USA with 5000 of the world's leading supercomputing and high-performance networking experts attending.
The ANU "Bunyip" supercomputer was designed and constructed by Mr Robert Edwards of the ANU Faculty of Engineering and Information Technology. The prize-winning software for the Bunyip was developed by Dr Jonathan Baxter and Mr Douglas Aberdeen of the Research School of Information Sciences and Engineering. The Australian branch of the US company LinuxCare, based in Canberra, provided support. The Bunyip team is one of several supercomputer projects at the ANU.
The Gordon Bell prize has been awarded annually since 1987 to mark the best performance for a useful program run on a supercomputer. The prize is awarded in open international competition but in practice is usually won by American entries. The prize carries prestige but only $5000 cash. Entries are judged by a panel of three respected figures in high performance computing.
The ANU team of Jonathan Baxter, Douglas Aberdeen and Robert Edwards were awarded the price/performance prize for the performance of a neural network training program run on a locally built commodity cluster supercomputer. Previous prize-winners include teams from the peak government nuclear physics and applied computing research laboratories in the USA, at Sandia National Labs, Argonne National Labs, and Lawrence Livermore National Labs.
The ANU Bunyip is a commodity cluster computer which combines 196 ordinary personal computer processors in 96 computer boxes. Each processor is rated at 550MHz, like many desktop processors. Faster individual processors can be bought, but cost more than their extra performance is worth. These computers are connected using a dedicated fast Ethernet network, like that used at a smaller scale in many offices and departments. Six network switches connect the 96 computers in 4 groups of 24. To keep costs down only one of the computers has a keyboard and screen. The freely available LINUX version of the UNIX operating system is used in all of the computers. This system makes it easier to develop and run very efficient scientific application programs.
The cost of the Bunyip's combined computers and interconnection network was just under $260,000, which is around one third of the price of commercial computers with similar processor performance. It was funded by a competitive internal grant within the Australian National University, with contributions from the university's Research School of Information Sciences and Engineering and the Faculty of Engineering and Information Technology.
The prize-winning application program is used to train a neural network that can recognise Japanese characters. A neural network is a powerful technique for a computer program to make decisions, recognising things that it has never seen before by their closeness to previously seen patterns. By analysing a large number of Japanese characters printed in different ways this program sets up a neural network to recognise fresh previously unseen characters. Once the analysis has been completed (referred to as "training"), the trained program may fit on a much smaller computer and recognise characters very quickly. But the training phase can be very time-consuming -almost 20,000 hours of single computer time is needed to train the neural network for Japanese characters, by analysing 6 million examples. This time is long enough to make the project impractical, unless a multi-processor computer is used -and this cuts the time to a more practical 100 hours.
The ANU team's specialised programming techniques allow this to be done with an average performance of 152 GFLOPS (billions of floating point operations per second) at a cost of US 92 cents per MFLOPS. This prize-winning performance compares with 1998 prize winners cost of over US $10.50 per MFLOPS. Ten years ago the cost was over $125.
Large-scale pattern recognition is a powerful method with a growing number of applications. It can be used by computers to analyse and predict movements of financial data, analyse geological data for the mining industry, recognise faces from camera images, help to make sense of human speech. It can also be used in optical character recognition. Japanese Optical Character Recognition is the process of automatically recognising machine-printed Japanese documents and converting them into electronic form. The most difficult aspect of the job is recognising the individual characters, since there are approximately 4000 different characters in common use.
The Bunyip computer will continue to be used at the ANU on a variety of problems in mathematics and science. This technology also has industrial applications in the new information economy - which is shown by the commercial interest in hiring some of the team members away overseas. The ANU would welcome initiatives from government and industry to help provide the resources needed to further develop the expertise within Australia, for Australia's benefit.
For more details see the following World Wide Web site
http://tux.anu.edu.au/Projects/Beowulfor
contact:
Dr Chris Johnson
Head, Department of Computer
Science
The Australian National University
Canberra, ACT
0200
Australia
phone +61
(02) 6249 4509
fax +61 (02) 6249 0010
email Chris.Johnson@anu.edu.au
Chris Johnson