Designing the Exascale Computer: Race Car or High-Speed Train?

Print Friendly, PDF & Email

In our last issue of The Exascale Report, we talked about the growing consensus that the global HPC community is, in fact, in a ‘race’ to exascale. And, as in most forms of racing, the stakes for coming in first are getting extremely high.While the enthusiasm around a global race is not all bad, the world won’t be much better off than we are today if the crowning achievement is reaching exascale-levels of computation with a stunt machine that fails to serve the bigger purpose of timely scientific discovery.

Not only will it take massive funding to win the race, it will take innovation and deep collaboration, enveloped in years of preparation, to build an environment that supports critical applications on exascale-class machines.

An exascale-class supercomputer that holds bragging rights based on benchmarks is a race car. Pure and simple. It will go from zero to 60 in the blink of an eye. Truly something to be admired. But if it has no practical applications, what good is it?

Launching an exascale-class supercomputer, with several key applications already ported and optimized, with dozens of scientists and researchers standing by to run their codes, well, that’s a high-speed train. And that’s exactly what we need, and how we should be targeting our research funding.

Can we get there? Of course. We can build that high-speed train and load it up with scientists and researchers who will take the journey together. But it will take a much more serious funding effort, and some innovative thinking to explore multiple paths – multiple tracks if you will.

In this issue, we talk about the Graph 500 – a good example of the type of innovative thinking necessary to lay down the tracks and prepare us for exascale applications.

Can the U.S. Compete?

This question has been raised in several forums. So, let’s explore this a bit more. Just recently, news came from Europe that exascale is being taken quite seriously. The European Commission stated it would double the investment in exascale research to the equivalent of $1.58 billion.

The U.S. is far behind this level of funding and is now finding itself limping along just trying to stay in the race. In fact, the U.S. is being outspent by pretty much all the other competitors in the race. China, Japan, Europe, Russia, and now India.

The American funding pool is not stagnant by any means. Pockets of research are underway. We’ve highlighted some of this in previous issues and will feature more of these programs in upcoming months. The folks at Argonne, PNNL, Oak Ridge, LLNL, and Sandia are all plugging away and making the best of the funds they can appropriate. And this month, we’re highlighting the new Computational Research and Theory center, designed to be an exascale facility, currently underway at the Lawrence Berkeley National Laboratory (LBNL).

And then there is the X-stack program that came out of the DOE ASCR program. This could be the most important program currently funded in the U.S. to affect exascale development. Funding levels will likely fall far short of what has been proposed. Thumbs up for this program, but we have concerns over the funding level. The U.S. approach seems to favor multiple, small projects, an approach that should not be taken lightly in terms of potential impact, but hardly gives the U.S. the overall momentum necessary to compete on a global scale. The X-stack program, which I believe is the first significant effort to come to fruition since Bill Harrod moved over to the U.S. Department of Energy (DOE), has the potential to define new programming models and future architectures, establishing the framework for future application environments for exascale-class systems.

Can the U.S. compete on a global scale? Yes, it can. The question is not if it can, but if it will.

Building the metaphorical high-speed train will take much more research (and much more money) than building the stunt-car racing machine.

While everyone wants to be the first, and set that new speed record, the harsh reality is that an exascale system capable of performing only benchmarks will not do the world much good. We need to take many scientists and researchers along for the ride. And we need to prepare for this now.

If we do this right, we can bring these new systems to market with applications ready to go.

Even without concrete definitions of what those first exascale systems will look like, there is much that can be done to ensure availability of applications soon after these systems arrive. We believe the Graph 500 is a good step in the right direction.

Exascale Funding Perspective

The White House is asking for just short of $90 million dollars for exascale in the next budget appropriation. Now contrast this to what the current political administration in the U.S. is looking to spend on passenger trains (actual, not metaphorical) and new and improved rail networks over the next six years: $53 billion.

Regarding the high-speed rail system, the administration’s goal is that 80 percent of Americans would have access to these high-speed commuter trains by 2035. Wouldn’t it be great if we had a goal for 80% of American scientists to have access to exascale class systems by 2035?

For related stories, visit The Exascale Report Archives.

Comments

  1. March-2012-guest says