Designing for the Future
Dr. John Gustafson, senior scientist and principal investigator for Sun's HPCS Program talks about the Defense Advanced Research Projects Agency (DARPA) research grant recently awarded to Sun to design supercomputers for the future.
Q: What kind of computing solutions is DARPA looking for?
A: Gustafson: The pitch is that they want High Productivity Computing
Solutions, and they make quite a point of the "P" standing for productivity
and not performance. Because this isn't just a peak operations per second
thing, this is the total cost of owning a system and the total performance
you get out of it including the time to write code. They want high-productivity giant computer systems that can deal with not only the type of needs that the Department of Defense (DOD) has, but really the whole United States industry demand for technical computing.
Q: How is this different from other Sun projects to develop complex systems?
A: Gustafson: In order to get to a breakaway solution, you really need
something like the DARPA program so you can take some risks and think
outside the box. It's a cliché, but it really fits here. We have to do
some very daring things, and not simply keep along the same compatible
strategies of evolving the existing product line. And DARPA gave us carte
blanche to do that. DARPA realizes that companies have to be fairly
conservative about their product plans, so they absorb the risk of the more
daring efforts in the hope of helping companies discover some incredibly
valuable breakthroughs. Sun simply could not do this R&D effort without
DARPA's help.
Q: Where is Sun now in the Phase Process for the DARPA HPC evolution?
A: Gustafson: Phase II is the research and development phase for creating a
product that has the potential to be the world's largest and most productive supercomputer. Phase I was exploration of possible technologies, and Phase II is about risk reduction and really clearing the path to make this a real product. That
is, we put our effort into removing the unknowns, the stuff that's scary
because no one has ever done it before. It's Phase III that is designed to
create a product. Phase III will last about five years and may go to only
one or two vendors, and we hope to be among those as well.
Q: Why is this an important thing for anyone other than DARPA?
A: Gustafson: Whatever gets developed for HPC, High Performance Computing, it
trickles down into enterprise computing and everything else. These technologies are critical for our future.
Q: What would be some examples that DARPA would use supercomputers in?
A: Gustafson: It would cover the exact same range as all of Sun's technical
market. Life Sciences, crash testing, design of vehicles, structural
analysis, circuit design. It's hard to find something that Sun doesn't
already do that doesn't have a DOD or DARPA based application to go along
with it. They are obviously interested in cryptography and analysis, and
they need to think of things like surveillance and reconnaissance, but we
are not just designing for DARPA and they don't want us to. They want us to
design for the overall technical market and that's exactly what we are doing.
They want these to be commercially viable systems that we create. We made
it such that this would conceivably address the entire scientific and technical computing market.
Q: What makes your design solution different from other computers available now?
A: Gustafson: The main difference is that it's not a cluster of computers like
Japan's Earth Simulator, or the ASCI machines or Linux clusters; it's a
single system architecture. It looks like a single image or a single
computer to the user, which is a lot easier to maintain. It's a lot more
secure and easier to debug, even though it's made of multiple processors.
What people have been doing for the last ten years is lashing together
smaller computers and asking programmers to figure out how to get them to
cooperate. Our design for DARPA presents a very big flat address space,
flat memory, and you don't have to keep track of a complicated topology of
what is connected to what.
Q: What's the difficulty in making a machine like that?
A: Gustafson: The main burden is even now we're running into physical limitations. Heat and cooling and the speed of light. All these things are limiting how fast a machine you can build. But we have some breakthrough technology that we think is very important to Sun's future that is going to be applied first in this design, and tried out here, that gets past some of the worst and nastiest problems in computer design.
Q: How has this DARPA grant changed Sun Lab's research approach?
A: Gustafson: We set this up as a new way of developing products at Sun, where
technologies we develop transition from things in the labs to being in the product groups. We've created a cross company team of maybe 100 people using Sun's best engineers to partner with us. We'll be meeting and gradually transferring what we learn in Phase II such that at the beginning of Phase III we'll be all set to go in the product groups.
Q: What is the most important thing that Sun's customers and partners should know about this project for DARPA?
A: Gustafson: The Federal Government has identified Sun as a thought leader in
this area and we're going to be leading the technology development for high
performance computing in the next ten years. That says quite a bit. It
gives us a road map beyond our current three-year time frame that we can
point people to and that shows we are aggressively going after the high end
of the market.
 Back to top
| 
