Beauty of supercomputing

Font Size

Print

Feedback

Email

Discuss

: Recently, there has been a lot of buzz surrounding the Eka supercomputer built by Computational Research Labs (CR Labs), a subsidiary of Tata Sons. It was ranked No 4 on a table ranking the world’s supercomputers released earlier this month at the Supercomputing conference in Reno, USA. We must all congratulate the team at CR Labs for this achievement. But there was one more achievement which very few have heard of. I have to set the stage before I disclose it.

What makes a computer a “supercomputer”? The Top 500 benchmark ranks computers by the rate at which they perform arithmetic operations on a specific mathematical problem. Anyone can download this benchmark, run it on a computer and send in the results. The top 500 computers then make it to the list. Supercomputers are constructed for a variety of reasons—to solve computational problems, for example, or for political motivations and national pride, among others. This says nothing about the computer’s capability to solve real world problems.

There are two general ways to build a supercomputer. One way is to assemble one with a very large number of generic components. Here’s how: buy lots of servers with Intel or AMD chips, and connect them with a fast network. These are commonly referred to as clusters. The Eka cluster has 2,000 such servers, each containing eight processors. The other way is to buy a specially designed supercomputer such as SGI Altix, Cray XT4, or IBM Blue Gene. But clusters are relatively economical, offering high performance per unit cost, especially on the Top 500 benchmark. Their drawback? They are hard to programme. The costs of software development are often large, and the resulting performance levels are very low. And it is software that realises the potential of a computer. It is software that sets Apple’s iPod and iPhone apart. India’s prowess in software services is well known. What do supercomputing and the Indian software industry have in common? It turns out that the skills required to programme the next generation of general purpose computers will be similar to the skills required to programme supercomputers.