Devangshu Datta: Parallel vs vector

TECHNO BEAT

There are two basic approaches to a problem. One is to dedicate all the resources of a single mind to seeking a solution. The other is to split the problem up into segments and parcel it out to members of a team.

For example, it needed a single massive intellect to derive the relationship between the sides of a triangle. A dozen men would not have worked things out any better than Pythagoras.

But if the Pythagoras Formula is known, it becomes more efficient to put many men on the job of finding right-angled triangles.

The Manhattan Project would never have been successful without splitting things up. Richard Feynman organised a task force of "clerks" (some of whom were the wives of scientists), to work the millions of arithmetical calculations that went into the design of the first atom-bomb.

This method of splitting up problems is called parallel processing. For many years, computer scientists found a single-minded approach or "vector-computing" more practical. Computers were expensive and there were few of them around.

Supercomputers had a small market "" there were just two or three players and a handful of customers. Cray and NEC built multi-million machines and installed maybe three systems a year. The customers consisted of research institutions.

Parallel processing became practical with the advent of cheap chips in the 1980s. When the Internet took off, parallel processing exploded; it became possible to put together very large clusters of chips.

Some parallel projects work on donated time. These programs use CPU (central processing unit) resources whenever and wherever available. The Search for Extra-Terrestrial Intelligence project (Seti) and tests for large prime numbers depend on parallel processing.

In Seti, home computers test radio-telescope data for signs of intelligence. In prime number testing, home computers factor very large numbers seeking primes. These machines calculate offline, usually while the CPU is otherwise idle.

By the late 1990s, parallel processing looked set to replace vectors: cheap PC and Mac chips can be strung together in ever-larger clusters. But splitting up some tasks is difficult and parallel processing is inefficient at tackling certain problems. The bigger the cluster, the more the programming hassles.

Still, most modern supercomputers are clusters. The fastest is a vector, however. The $ 350 million NEC-constructed Earth Simulator in Tokyo runs 35.6 trillion calculations a second. It's the size of a Boeing 737.

The Earth Simulator is used to track global sea temperatures, rainfall and crustal movements to predict natural disasters over centuries. The Japanese government reckons it could save billions by giving early warnings about storms and ecological damage.

Speed isn't just an index of geek-machismo. Supercomputers can calculate the spread of pandemics like AIDS and bird flu, simulated bioterrorist attacks and forest fires in real-time.

They can also speed drug discovery and drug-testing. Most of all, they can model the effect of new weapons "" that's the big market. After nuclear tests became unfashionable, simulations are the norm.

India has a presence in the supercomputer market. Seven systems in the global Top 500 list are Indian-designed. All are parallels put together with commercially available chips. Mostly, these are used for weather modelling. But super-computing capability is one reason India's nuclear programmes are taken seriously.

IBM is building two parallel-processing supercomputers for its Blue Gene project, projected to beat the Earth Simulator in terms of speed. Cray is also back in the game, and concentrating on vectors. IBM is spending $ 290 million on Blue Gene while Cray has a $ 90-million contract to design a custom-machine for nuclear weapons testing.

Parallel or vector? The debate has been renewed. Parallels are often more cost-effective, but vectors may solve certain types of problems, especially those involving complex simulations, quicker and with fewer programming overheads.

And, these problems often crop up in military research, which is carried out regardless of cost. In this case, the spin-offs from the intellectual competition may be beneficial. The state of the art suggests that there will continue to be room for both vectors and clusters.