IT TRENDS
Towards a peta era
|
Today’s fastest computer operates at 280 teraflops. By next year, it is slated to be tripled
|
— Photo: by special Arrangement
Fuelling petaflops: Practical platforms for petascale computing may require between 150,000 and 500,000 processor cores.
The semi-annual list of the world’s ‘Top 500’ super computers, released last week in Dresden, Germany, has once again placed an IBM machine — the BlueGene/L System installed at the U.S. Department of Energy’s Lawrence Livermore National Laboratory — in the No. 1 spot. The leviathan recorded a benchmark of 280.6 teraflops or a trillion calculations per second.
Second place is taken by a Jaguar-Cray XT4/XT3 computer housed in another US government installation, the Oak Ridge National Laboratory. It clocked 101.7 T.Flops, only marginally faster than another Cray machine at Sandia National Labs, a sister institution ( 101.4 T.Flops).
There are eight India-based computers in the Top 500, a list compiled jointly every half-year by the Universities of Mannheim in Germany and Tennessee in the U.S. All of them are imported platforms from IBM and Hewlett Packard, that are working in un-named technology, biotechnology and geophysical exploration companies in this country.
The full list and a searchable data base — country-wise and machine-wise — is available at www.top500.org
Scalable architecture
Sun Microsystems has announced at the Dresden International supercomputing conference that it will deliver a 421 T.Flop machine later this year, based on a new scalable architecture called ‘Constellation,’ the result of a collaboration with the University of Texas’ Advanced Computing Centre.
This includes a newly developed Infiniband switch with 3456 ports which will significantly quicken the exchange of information between servers, memory and mass storage.
Indigenously developed machines — from the Centre for Development of Advanced Computing (C-DAC), Pune and the Institute of Mathematical Sciences (IMS), Chennai — which entered the lists a few years ago, with performances touching one teraflop, are no longer in the Top 500. This is a game of constantly shifting goal posts and the entry point is now 4 teraflops.
Revived efforts
India, whose supercomputing plans (except for unpublicised in-house strategic developments) appeared to be on a low key for some years now, seems to have revived its efforts to create indigenous general-purpose, high performance computers (HPCs)... fuelled by the emerging opportunities in pharma and genetic research, drug discovery, multimedia and geophysical exploration.
C-DAC has announced that it will escalate its flagship Param supercomputing programme to exceed 10 T.Flops in the third quarter of 2007. By 2010, C-DAC’s roadmap is to achieve 1 petaflop, senior scientists at its Pune headquarters told this correspondent recently.
A peta flop is 1000 tera flops or one quadrillion calculations per second — more than three times faster than the world’s no. 1 number cruncher today, the BlueGene/L.
Interestingly, IBM which has headed the list of the world’s fastest, most powerful computers for four years now, announced, that it was readying to deliver the world’s first petaflop computer — the next generation of its BlueGene machine to be called BG/P — in 2008.
Clearly the tera-era is set to fade away, giving way to a peta-era in computational power.. and while the personal computer is getting more powerful every day, it would still take about around 75,000 to 100,000 of them to match one of these petaflop supercomputers.
Peta scale ceiling
But the tail may end up wagging the dog: Hundreds of off-the-shelf PCs may end up as powerful as any of the dedicated supercomputers now being developed to break the peta scale ceiling. Researchers at Maryland University led by Prof Uzi Vishkin, have developed the prototype of a single-chip supercomputer with 64 parallel processing cores, which they are confident of scale to 1000.
Mass market appeal
And Microsoft has just joined with Hewlett Packard to create supercomputers with ‘mass market appeal.’ They hope to offer clusters of 64 or more nodes that can be up and running, in two hours. It would seem there is more than one path for those who aspire to attain number-crunching nivana on a peta scale.
How many cores needed?
Gabbar Singh’s menacing catch phrase, ‘kitne aadmi they’ may find its cyber age equivalent in the stiff competition between high performance computer researchers and their feverish concern: How many computing cores will it take to deliver a peta flop of performance?
In the last decade, microprocessor technology and architecture have driven the performance development of supercomputers, as well as the way in which we program such systems.
The linear development of microprocessors was suddenly interrupted and went into new directions with the advent in recent years, of multicore technology and application-specific accelerators.
The HPC community will enter the era of petascale computing most probably late in 2008. We might assume a petaflop system will be available in 2010 with about 150,000-500,000 cores.
However, users of petascale systems will benefit in terms of sustained performance only if system software and scalable algorithms can be developed.— A.P
ANAND PARTHASARATHY
Printer friendly
page
Send this article to Friends by
E-Mail
Sci Tech
