JASON and the Have-Nots
Intel and Microsoft opened up their funding competition to the top 25 computer science departments in the US. The two winners were joined by Stanford University and MIT in the last round of the contest. But in the end, the public universities won the hearts of the vendors over those wealthy elitists. (Stanford was apparently quite upset by this loss, as it seemed that someone from the university - Bill Dally, we're looking at you - leaked word of Berkeley's victory to a news outlet. An odd PR strategy to be sure, but Dally is a JASON, so he must know what he's doing.)
To help model future hardware, Berkeley will turn to its own RAMP systems, which combine hundreds of FPGAs (field programmable gate arrays) on a single board. Researchers can tweak the RAMP systems to mimic various processor architectures and test how their code will run across tons of cores. Ideally, the RAMP designs save on hardware costs and time by freeing researchers from pricey systems and by allowing them to speculate about future hardware rather than waiting for prototypes from hardware makers.
The funding will let 8 full-time faculty and more than 30 students on their way to PhDs research the multi-threaded software problems at Berkeley.
"There is a real sense of urgency that we need to figure this stuff out," Patterson told us.
Many of the researchers have been looking into these types of problems for years, and lots of work is already underway. For example, the Berkeley staff are exploring personalized medicine applications where a doctor or even patient could use a handheld device to simulate how various treatments might affect the blood. They're also examining systems that would project a sound field into a room "that recreates a concert in your living room as if you were at the event," said Krste Asanović, another professor at Berkeley. And there's work underway on a parallel browser that could make it possible to display and use much richer applications on your cellphone.
All of the work done at Berkeley will be released under open source software licenses, meaning that Microsoft or rivals could, say, pick up the browser code.
While the needs for this type of research appear obvious, some prominent technology players remain fond of yesteryear's methods.
'Putting the cart in front of the horse'
Linux figurehead Linus Torvalds, for example, recently attacked Berkeley's parallelism efforts. Torvalds relegated the concerns around multi-threaded software to the niche of researchers.
"Designing future hardware around the needs of scientific computing seems ass-backwards," he wrote on a message board. "It's putting the cart in front of the horse."
Torvalds' comments shocked some of the Berkeley group.
"I was kind of staggered by that comment - that one of the leaders of computing sees the future in linear time computing," said Kurt Keutzer, a professor at Berkeley.
As the Berkeley set views things, the most interesting future applications will require far more horsepower than Torvalds seems to imagine.
Quite frankly, it must be a bit disheartening for Linux developers to witness Torvalds speak in such terms. He seems married to a past that's disappearing at pace with every release of a new processor.
Other critics might argue that we're heading toward a world where software returns to the data center rather than running on personal clients. The so-called Cloud Computing scenario would have servers do the serious work for users who just need relatively dumb, low-powered devices.
Patterson countered this argument by saying that a lack of pervasive network connections along with improvements in things such as flash memory will continue to make powerful client devices very attractive.
The Microsoft and Intel funding should help the researchers tackle these kinds of questions for five years. ®