Sea change in supercomputing

The days of tennis court-sized number crunchers, may be over. New trends in 'cluster computing' hold out the hope that millions of networked PCs will create a mammoth Virtual supercomputer. Anand Parthasarathy explores the next step in high performance computing.

LIKE EVERY new technology, the computing-and-communication combo has two cultures: the traditional main stream of multinational IT companies and their high profile corporate customers and the murky byways of the `alter-net' society. The divide was underlined earlier this month by some significant computational `happenings': In Heidelberg, Germany, on July 1, the International Supercomputer Conference held its 16th meet and went through a semi annual exercise: a game of ``Mirror, mirror on the wall / Who's the fastest of them all?' The website www.top500.org, put up its latest list of the world's 500 fastest supercomputers. IBM continued to straddle the list with 201 of the 500 positions and six of the top ten. (Sun is the second with 81 systems, SGI 3rd with 63). IBM was number 1 with ``ASCI White'' ( for Accelerated Strategic Computing Initiative) currently crunching numbers for the US Department of Energy's Lawrence Livermore Laboratories, an institution which hopes the machine will simulate the performance of nuclear warheads with sufficient fidelity to making live testing unnecessary.The machine which takes up space equal to a tennis court, weighs 106 tonnes, uses 3000 km of wiring, consists of 8192 processors in a massively parallel processing (MPP) configuration. ``Top 500'' rates ASCI White the world's fastest, at 7.2 tera flops ( trillion floating point operations per second), though IBM claims it has achieved 12.3 T flops. There are another 11 machines worldwide, which are 1 tera flop or faster some of the Japanese machines from NEC. Hitachi are classic vector-based systems ( parallel processing was a later development).But the list includes a few surprises straws in the wind which tell of coming trends in high performance computing: A surprise entry ( at no. 439) is a machine the PRESTO-III built by the Tokyo Institute of Technology using 78 off-the-shelf AMD Athlon processors, chips identical to the ones powering many high end PCs. Another revealing trend was that some of the supercomputers in the list eschewed proprietary operating systems for open source alternatives IBM built Linux-based machines for the petroleum company, Shell and the National Centre for Super Computing Applications of the University of Illinois. The increasing preference for a `free' OS like Linux and the use of readily available Commodity-off-the-shelf (COTS) components are two facets of what is being perceived today as an exciting new evolutionary step in supercomputing launching a new era of `Virtual Supercomputing for all'. It has not happened overnight indeed the trend has come creeping in over almost a decade. Cluster computing In 1994, the Centre of Excellence in Space Data and Information Sciences (CESDIS), part of the Goddard Space Flight Centre in Greenbelt, Maryland, US, was give the job of creating a supercomputer for the National Aeronautics and Space Administration (NASA), on a tight budget. The key researchers, Thomas Sterling and Don Becker, decided to go in for 16 DX-4 processor-based PCs working in a cluster and linked by 10 MBPS ethernet. They called their system ``Beowulf'' after a medieval Scandinavian warrior and subject of the earliest known epic poem in the English language. This was similar in concept but not identical to the NOW (Network of workstations) project of the University of California at Berkeley, where the unused cycle time of PCs doing other jobs was harnessed. In the Beowulf cluster, the PC nodes are dedicated to the task at hand. Today ``Beowulf'' is the accepted term for using several small commercially available computers in a cluster, to provide the computing power of one large computer. The operating system is typically an Open variant of Unix like Linux. In 1997, the Utah-based company Linux Networkx (linuxnetworkx.com) created what was claimed to be the first commercial cluster computer based on Linux, for Brookhaven National Laboratories. The group that created that first ``Beowulf'' for NASA, is today a private company, ``Scyld Computing Corporation'' ( Scyld was Beowulf's father in the legend). And it is not alone: literally hundreds of small companies have sprung up worldwide to provide affordable ``Beowulf'' supercomputing solutions to small businesses or academic institutions which need the number crunching capability but not the stiff price of a conventional supercomputer. Students in institutes (like Caltech in the illustration alongside) routinely link up 24 - 48 Pentium- type PCs to achieve about 200 megaflops per node. Indian initiativesIndian supercomputing initiatives which centered around half a dozen central government institutions, were initially all based on the dedicated parallel processor concept, often going in ( like the DRDO's `PACE' computer) for custom-built processor chips). However both Param, the supercomputer from the Centre for Development of Advanced Computing (CDAC) and Anupam from the Bhabha Atomic Research Centre (BARC), have cannily recast their systems around off-the shelf processors. The `OpenFrame' Param 10000 based on UltraSparc workstations as processor nodes, has found acceptance in 12 Indian and quite a few foreign academic institutions. The current top-of-the-line model claims a peak computing power of 6.4 gigaflops. The newest Param 10000 will be commissioned next month at IIT Kharagpur. Meanwhile, BARC, announced on July 4, that it had achieved 15 gigaflops (probably peak ratings) with an 84-node version of Anupam created from Pentium III 600-MHz-based PCs, interconnected by fast ethernet and gigabit networks. Thus the two Indian supercomputing initiatives are also heading down the globally preferred path of cluster computing. As a nation we would then be ideally poised to ride the imminent next supercomputing wave the Virtual Distributed Supercomputer.Virtual Supercomputing This had its origins in 1999, when a project of the University of California at Berkeley, called `SETI@HOME' motivated 3 million volunteers with their own PCs to pitch in with their spare computing power to jointly search for extra terrestrial intelligence. A more down to earth goal is being addressed by an Intel-sponsored Philanthropic Peer-to-Peer programme which is helping the (US) National Foundation for Cancer Research to try and identify drugs that can block the four key proteins that promote cancer in humans.The programme hopes to motivate 6 million PC owners to download a 1.8 MB programme that will work on the problem in the background, or in a screen saver mode, whenever the owner boots the PC, without disturbing his or her work. It uploads the results periodically to the main research website.Another ambitious project launched in February this year, by an American internet service provider, Juno Online, hopes that its 4 million subscribers will `donate' some PC time, to create a Virtual Supercomputer for biotechnology research. If indeed 4 million PCs share the work, this would theoretically create a virtual machine that could crash the `petahertz barrier' ie it would be faster than 1 petahertz 1 billion megahertz. That would make today's `Top 500' supercomputers look like bullock carts in line up of racing cars. Industry watchers feel this is very much within the bounds of possibility. Indeed the models being mooted including paying millions of PC users to use their idle PC time, rather depending on volunteer services.

The Internet itself has evolved within a decade, beyond our wildest imagination. There is no reason why supercomputing should not emerge in the next decade as just another mundane housekeeping task for your PC and minesomething with which we join hands with millions, in a massive cyberchain around the globe.50 years of commercial computing A landmark in the history of computers passed virtually unheralded by a industry, too preoccupied with the problems of today, to remember its glorious yesterdays:

Fifty years ago on June 14, 1951, the world's first commercially available computer, the UNIVAC -1 ( short for UNIVersal Automatic Computer) was invented. It was made by Unisys Corporation, a company that is still in the business of computers.It was sold to the US Census Bureau. It was built as a huge cube, with a hollow core, with a desk inside for an engineer. It had a memory of 9 Kilobits; and a single processor working at 0.008 megahertz. It weighed 13 tonnes; drew 125 KW of power and was cooled with water to keep the vacuum tubes from heating.

The machine captured public imagination when it was used to predict that Gen. Eisenhower would win the US presidential race. The latest computer from Unisys, the ES 7000 enterprise server has 1.6 million times more memory that the UNIVAC1, is 1,12,500 times faster and weighs only 1/24 of the original. In a tongue in cheek `apology' for the many inconveniences resulting from its 1951 invention,

Unisys, last week said it is sorry for: making it impossible for anyone to do more than five minutes work without being interrupted by an emailed joke or chain letter ending the great morning tradition of newspaper with your coffee because by the time your coffee is hot, the ``news'' in the paper is already two generations behind the online edition. giving government, business and the average 12-year old the means of finding out more about you than you yourself knew. forcing you to go through a five minute start-up routine every time your PC crashes, while you are creating a three minute memo. eliminating the concept of regular working hours the dot.com bubble the bursting of the dotcom bubble.

Send this article to Friends by E-Mail

Section  : Science & Tech
Previous : James Franck: A versatile scientist
Next     : Integrated management of rootknot nematode on
           brinjal