The rapid development of the fastest computers
Thomas Sterling is the kind of scientist the world could wish for. Not only is he brilliant: it is thanks to his work that ordinary individual PC parts, plugged together to form so-called Beowulf clusters, soar to unimagined peak computing performance.
He is also creative in dealing sparingly with the always scarce resources: Because he was not granted the requested computing power, he cannibalized decommissioned computers and built the first Beowulf clusters from the laboriously revived scrap. How has the picture changed since then! What is currently the fastest computer in the world consists – what a miracle – of new parts specially designed for this purpose; but at heart, in an evolution of Sterling's concept, it is a cluster.
Sterling doesn't need to hide any shortcomings behind inflated jargon either. His keynote speech at the Summer 2004 International Supercomputer Conference, a technical conference after all, was of such crystalline clarity that we were able to expand it into a "spectrum" article with a modest amount of additional explanation.
And he is of exemplary modesty (which is also noticeable by the fact that he patiently answers questions from the little "Spektrum" editor until he has finally grasped the essentials). Instead of basking in the fame that the phenomenal success of his concept has brought him, he is the first to pronounce the child of his own mind dead.
High-performance computing is a very special kind of business. It involves a lot of money - you have to fork out several hundred million for a really big device - and the customers are extraordinarily capricious: ministries, research agencies and parliaments change their minds, whether science needs a new supercomputer, often very suddenly.(Only the intelligence services have a fairly constant need, but they don't like to talk about it.) Even for a large company, it is therefore quite risky to develop completely new ideas at great expense in anticipation of future business. As a result, this industry, which is actually highly innovative, has a very strange tendency towards conservatives: People are sticking to a once successful concept, although there have long been better ones. The better, however, has no chance of getting up because it is still immature. Fortran, the Message Passing Interface and, more recently, Beowulf are among the dinosaurs that still dominate the scene today, although they've long outlived themselves.
Our "Supercomputing" dossier, which has been on the market since March 9th, summarizes, as is usual with dossiers, our articles on a topic from the last five years. But that in this short period of time the whole life of a Idea - birth, flowering and obsolescence - expires and is described by us is unusual. Thomas Sterling, one of the heroes of this story, is anything but a tragic hero because he has little time to cry for his Beowulf because of all the new ideas.
What else has happened in the supercomputing scene? A lot. There was an exceptional amount of detail to be updated on the articles. One computing performance record follows the other. Some of the products announced at the time are now on the market, strangely enough the websites for others have not been changed for a year.
But mind you: Those were just statements about market success. The presentation of the scientific basis is as fresh as the first day and is likely to remain so for quite some time.
Christoph Pöppe
