 |
 |
||||
 |
|||||
|
Default | NEWS
11/6/2008 Hot is cool in future of computing Research into advanced computer systems is hot stuff. What makes Habanero particularly smoking will be its ability to help computers of the future live up to their potential. Vivek Sarkar, for years a parallel software guru at IBM, started the Habanero Software Research Project when he joined the Rice faculty as the E.D. Butcher Chair in Engineering and professor of computer science last fall. The project has backing from the National Science Foundation, IBM, Microsoft, Intel, BHP Billiton and the Defense Advanced Research Projects Agency, or DARPA.
At the same time, computer science students at Rice learning to use these new tools become highly employable, said Sarkar. And so will students learning remotely through tools and course notes Rice plans to make freely available on the Web. Traditionally, computers have become more powerful by increasing the number of transistors on the central processing unit and raising the clock speed by which they operate. But we are now in an era when increasing the clock speed needs more energy and releases more heat. So a transition is taking place. Instead of raising the clock speed, manufacturers make computers more powerful by increasing the number of processor cores. A computer you buy today typically sports a dual- or quad-core processor, which splits tasks to process them in a more energy-efficient manner -- but only if the program and/or operating system know how to take advantage of it. Building programs that share tasks among available cores is important now -- and will be critical when computers come with 1,000-core chips. That increase in parallelism seems outrageous, but with 128-core processors already being tested, Sarkar said such machines aren’t far off. To make computers faster, there’s no alternative to multicore processors, said Sarkar. “In the 2015 to 2020 time frame, we can expect exascale supercomputers that run 1,000 times faster than today’s petascale systems,” he said, referring to a state-of-the-art IBM machine at Los Alamos National Laboratory that handles 1015 instructions per second. “If you tried to make a single processor run 1,000 times faster by increasing its clock speed, there would be a number of limitations with potentially a billionfold increase in the amount of power required and heat generated. It’s just not practical,” he said. Enter multicore processors, which don’t necessarily run faster than today’s machines but, like laborers building the pyramids, achieve monumental tasks by sharing the load. And they won’t only be for supercomputers. Sarkar expects they’ll find their way onto desktops and into handhelds running all kinds of programs. Sarkar intends Habanero to be an ecosystem by which programmers can access a common set of tools -- languages, compilers and runtime systems -- to enable their programs to run in parallel on multicore computers. He said the retraining required for these “many-core” chips will need to go far beyond simply learning new tricks; this is a paradigm shift. “There’s been a classical view of teaching undergraduate students what a computer looks like,” said Sarkar. “That’s been largely unchanged for the last 20 or 30 years. All that changed were the numbers -- how fast the clock goes or how big the memory is. “But there’s a big issue when you’re talking about getting software ready for these many-core computers. People have to be trained to write this new kind of software. They have to be trained that computers look different now.” Habanero will start with a version for Java, one of the most common programming languages and one that runs across many types of computers, from mainframes to desktops to cell phones. “Habanero-Java is quite far along already,” said Sarkar, who said knowledge of Java will be a prerequisite for juniors taking a new course, Principles of Parallel Programming, he’ll teach next fall. “People use Java for all kinds of things, and it always comes up high on the list as a dominant programming language in different surveys,” he said. “While Java is supported on current multicore processors, it will take time to become available on new, bleeding-edge many-core architectures, so we’re also working on Habanero-C. H-J will be available for the course next fall, and H-C might be available 18 to 24 months after that.” He expects other programming languages will also feel the heat of Habanero in time. Sarkar said teaching materials developed at Rice will be freely available online, along with the necessary software to teach and use Habanero-Java. “If anyone else wants to learn the language by following the lectures, they can. If they want to teach it anywhere in the world, they can just download the software and the lecture material and so on." So what does Sarkar want from a many-core computer? “First of all, I want it to be power-efficient, so it never runs out of charge during my working day,” he said, reflecting the wishes of many a laptop user looking for a power outlet. “In terms of computational capability for the things that are physically around us, there’s opportunity for customization, where they could sense your normal routines. For instance, I’d like my iPhone to be my personal assistant -- to anticipate everything I need. My phone would work differently for me than yours would work for you. “All our personal devices will be backed by data centers that store our profiles and provide virtually unlimited computation and storage capabilities to millions of customers,” he said. “In contrast, the more our personal devices can be tailored to us, the more interesting they will become. They can provide things the data center cannot provide as easily.” |
 |
 |
 |
 |
 |
 |
 |
