The Defense Advanced Research Projects Agency’s multiyear, $650 million mission to develop high-performance supercomputers is more than just a hardware project, said DARPA officials. The contractors must ensure that the systems are productive and easy to use for both national security and commercial applications.
Therefore, the DARPA effort also involves building programming languages, development tools and methods for scaling applications across tens of thousands of processors.
“High-performance computing is at a critical juncture,” said William Harrod, manager of DARPA’s High Productivity Computing Systems program.
Cray Inc. and IBM in November won contracts valued at $500 million to develop “economically viable high-productivity” supercomputers by the end of 2010, he noted.
The agency’s development effort began in 2002 with five vendors. The two remaining suppliers have been tasked with delivering a line of systems whose high end is capable of sustaining petaflop speeds and securely running national defense and commercial applications.
Building such systems requires a programming environment that Harrod said “is easier to use and that has less of a learning curve than the environments on today’s HPCs,” or high-performance computers.
The new systems will also need an architecture and an operating system that enables efficient execution, he said.
A Single Language
Current programming languages require an enormous amount of effort from coders to transform ideas into algorithms and then turn that work into something that can run on a supercomputer, said Stanley Ahalt, executive director at the Ohio Supercomputer Center in Columbus.
Timeline | |
|
“Right now, we have a collection of codes written in older languages that are very difficult to modify,” he said, citing languages like Fortran.
Under the contract with DARPA, Cray is developing a programming language called Chapel and IBM is developing what it calls x10 to address these problems. Only one language will emerge from the research, officials said.
“There will be one language at the end of the day, and the government, multiple companies and HPC user communities are going to have to put in some effort to adopt [it],” said Steve Scott, chief technology officer at Cray.
Tony Befi, a vice president in IBM’s HPC group, said engineers are working to make the systems more productive so that they will be more cost-effective and attractive for broad commercial uses.
“There is money that could be spent if these assets were usable and productive enough to warrant the investment,” he said.