India’s biggest initiative to ramp up its supercomputing capability is taking shape as the country aims to bridge the widening gap with its global peers in a sector seen crucial to scientific output.
For the 12th Five-Year Plan, spanning 2012-2017, the ministry of science and technology is considering an ecosystem of supercomputers at an outlay of R6,000 crore, which would ensure that at least 25% of the scientific population of the country has access to a supercomputer. The sense of urgency has been amplified by the relatively flat rate of the capacity addition in India over the past few years, while a country like China has been doubling or quadrupling its capability every year.
Supercomputing is key for research and businesses in sectors, such as climate modelling, defence, aerospace and drug discovery, which require enormous amounts of number crunching. Areas, such as animation, finance and telecom, too, require different levels of high-performance computing.
“This was the worrisome thing for us, because the nation’s scientific productivity has been shown to be directly related to its facilities, both high-resolution instruments and computing power. That is one of the reasons why we felt we should have a large number of supercomputers,” says N Balakrishnan, associate director, Indian Institute of Science (IISc), and professor at the Institute's Supercomputer Education and Research Centre (SERC).
Balakrishnan is coordinating a national effort to map India’s capability in supercomputing, which has suggested building a four-tier ecosystem of machines and greater involvement of the private sector. The proposed roadmap follows a brainstorming of the scientific community in May this year. A final report of the proposals is currently being prepared.
The target for the 12th plan is to set up a top-tier of around six supercomputers with about 3-6 Petaflops computing power in different locations, followed by 12-20 zonal supercomputers of 200-500 Teraflops. The third rung would involve 20-50 machines with 10-Teraflop speed and, finally, 50-100 regional supercomputers of one Teraflops. A Teraflops or Tflops is one trillion floating point operations per second, while a Petaflops is thousand trillion (or quadrillion) floating point operations per second.
India’s fastest supercomputer currently is the ‘Eka’, built by Pune-based Computational Research Laboratories (CRL), a Tata Group subsidiary, which has a peak performance of 172.60 Teraflops. The combined performance of India’s top 16 supercomputers is 318.16 Teraflops according to an SERC ranking in June.
In May, the Indian Space Research Organisation (Isro) had commissioned a 220-TFlop machine, but its peak performance could not be evaluated for the June ranking.
The software development, for which an outlay of R600 crore is envisaged, would be equally crucial, says Balakrishnan. “On the software development side, both for programming on the multicore as well as for scientific software development, the private industries will play a significant role.”
It is relatively easy to build Petaflop supercomputers with the capability inherent in hardware, but building computer programmes to make use of the entire power is challenging, says Subram Natarajan, executive, Deep Computing, at the Systems and Technology Group of IBM India and South Asia.
The supercomputing demand from private industry is also increasing, but not to the extent it is in the US, China or Japan, says Vipin Chaudhary, CEO of CRL, Pune. “The push for these kinds of really high-end stuff has to come from government because the really large users, for example weather and climate modelling, homeland security, are typically government,” he says. “How do you use these systems, maintain them, write code for these systems? All that experience doesn't exist if you don't use such large systems. That’s why government expenditure in those programmes is crucial in this domain.”
The demand for mainstream supercomputing, involving 2-10 Tflops machines, has increased dramatically in India because of the need to solve more complex problems and with greater accuracy, says IBM's Natarajan. “In general, the awareness to create supercomputing to solve problems in a smart manner is now more prevalent than what it was three or four years ago.”
Most of India's supercomputing development seen in the 1990s was driven by national initiatives like missile development and the light combat aircraft project that required enormous computing capability. In recent years, greater investment into equipment in scientific institutions has resulted in greater demand for computer simulations, says Balakrishnan. “It is not a one-time investment over a plan period that you do. You will have to have a plan by which you double (capability) every year. You can't double in the same location, so you will have to phase it.”
Balakrishnan reckons that at the end of the 12th plan when the proposed ecosystem is in place, the country should be thinking about the exaflop initiative. An exaflop is one million trillion calculations per second and the global race to build the exaflop machine is aimed at a finish in 2018 or thereabouts.
— Ajay Sukumaran