IT TRENDS
`Home made' solutions for high performance computing
|
Last month, a Linux cluster assembled in Chennai, made it to the world's `Top 500' supercomputer club, a rare example of a non-commercial machine vying with industry heavyweights.
|
Project leader N.D. Hari Dass and the final configuration of the Institute of Mathematical Sciences,Chennai `Kabru' cluster supercomputer...now in a global league.
THE SEMI annual list compiled jointly by German and American Academics, at the Universities of Mannheim and Tennessee, ranking the world's `Top 500' supercomputers
(www.top500.org/list/2004/06) is one of the most closely watched indices of international heavy number crunching capability. For long, the list was the prerogative of half a dozen major computer companies. Indeed even today, 45 percent of the 500 machines are made by IBM and more than half use Intel processors.
And in the latest list (and 23rd) announced last month, the number one rank remained unchanged since 2002 — the Earth Simulator made by Japanese company, NEC, for modelling the world's climate and clocking a maximum 35.86 trillion floating point operations per second or teraflops. The new number 2 is the `Thunder' made by a relatively unknown assembler, California Digital Corporation and supplied to the Lawrence Livermore Laboratories. Its top speed is about half that of the NEC machine at 19.94 teraflops.
But one of the surprises in the June 2004 Top 500, is `Kabru', the do-it-yourself cluster Linux supercomputer assembled at Chennai's Institute of Mathematical Sciences (IMSc), just two months ago. It has been ranked no. 257 with a maximum computational speed of 959 Giga FLOPS (or billion floating point operations per second) and a peak speed of 1382.4 GFLOPS. The peak speed is achieved when the machine is fully stretched, but this may not be sustainable for long. The peak speed makes this machine, a `teraflop' (trillions of operations per second) supercomputer.
There are 6 more India-based supercomputers in the Top 500 but unlike the IMSc machine, they have not been indigenously assembled, but are commercial machines imported from the U.S., and deployed by industry here. Beside Tech Pacific Exports and PCS Trading which have each, a HP Superdome and Geoscience which has deployed two IBM Blade Centers and an IBM xSeries, the other supercomputer being used in India is another IBM xSeries, credited to a "Semiconductor Company", and thought to be with Intel, Bangalore. The last is slightly faster than the `desi' machine at 1196.41 GFLOPS maximum.
The machine, which put India in the supercomputing league last year, the indigenously developed Param Padma from the Centre for Development in Advanced Computing (C-DAC), does not make the list this year. Other machines have overtaken it and the datum has moved up. There is more than mere `desi' pride in the IMSc achievement. The teraflop performance was achieved only in April this year, by stringing together 144 separate computers based on the Intel Xeon processor and running the Redhat 8.0 version of Linux.
Institute faculty members were partnered by two Indian companies — Netweb Technologies from Delhi and Summation Enterprises from Mumbai — as well as two hardware suppliers, Supermicro and Dolphinics. But the real achievement lay in the very concept that `home made' machines can today achieve supercomputing speeds.
A cluster computer is a currently popular way of achieving supercomputing speeds, by putting together dozens of smaller, standard configuration computers, to create a cluster, which packs more power.
Indeed, the Top 500 site's statistics section only underlines that the days of traditional and proprietary number crunching behemoths are almost over: Of the 500 machines ranked last month, 58.2 per cent are cluster architectures. N.D. Haridass who led the team at IMSc, that assembled the Kabru supercomputer, says the machine is now being regularly used to perform large-scale simulations in Lattice Gauge Theory.
He adds: " It is believed that a string-like object is formed between quarks which are the constituents of protons and neutrons. Investigating the properties of this string numerically is a very big challenge and we carrying this out with Dr Pushan Majumdar of the University of Graz. A few years ago this task would have been impossible even on the biggest supercomputers but with important progress in algorithms this can now be done on a cluster machine like Kabru".
And the guys who are putting them together are getting younger. In April this year, over 1000 students converged on the University of San Francisco to see if they could achieve performance of Top 500 class, by stringing together their laptop PCs.
There were networking problems, which reduced a theoretically possible 404 billion Flops to something like 180 BFlops.
They have gone home to rejig their scheme but next year don't be surprised if the latest entrant in the Top 500 is an ad hoc machine put together by guys who live in jeans and are not too crazy about trimming their facial fungus.
They've already given this a name: Flashmob computing — teenage jargon for a sudden gathering summoned by the Internet. And it may yet end up as the most exciting development in collaborative, egoless, high performance computing in this century.
Anand Parthasarathy
Printer friendly
page
Send this article to Friends by
E-Mail
Sci Tech
