Around the world, efforts are ramping up to cross the next major computing threshold with machines that are 50-100x more performant than today’s fastest number crunchers. Earlier this year, the United States announced its goal to stand up two capable exascale machines by 2023 as part of the Exascale Computing Project and Distinguished Argonne Fellow Dr. Paul Messina is leading the charge.
Since the project launched last February ECP has awarded $122 million in funding with $39.8 million going toward 22 application development projects, $34 million for 35 software development proposals and $48 million for four co-design centers. At SC16, we spoke with Dr. Messina about the mission of the project, the progress made so far — including a review of these three funding rounds — and the possibility of an accelerated timeline.
Here are highlights from that discussion (the full interview is included at the end of the article).
Why exascale matters
“In the history of computing as one gets the ability to do more calculations or deal with more data, we are able to tackle problems we couldn’t deal with otherwise. A lot of the problems that over the years we first could simulate and validate with an experiment in one-dimension, we’re now able to do it in two or three-dimensions. With exascale, we expect to be able to do things in much greater scale and with more fidelity. In some cases we hope to be able to do predictive simulations, not just to verify that something works the way we thought it would. An example of that would be discovering new materials that are better for batteries, for energy storage.
“Exascale is an arbitrary stepping stone along the way that will continue. Just as we had gigaflops and teraflops, peta- and so on, exascale is one along the way. But when you have an increase in compute power by a factor of one-hundred, chances are you will be able to tackle things that you cannot tackle now. Even at this conference you will hear about certain problems that exascale isn’t good enough for, so that indicates that it’s a stepping stone along the way. But we have identified dozens of applications that are important, problems that can’t be solved today and that we believe with exascale capability we will be able to solve. Precision medicine is one, additive manufacturing for very complex materials is another, climate science and carbon capture simulation, for example, are among the applications we are investing in.”
On the significance of ECP being a project as opposed to a program
“There have been research efforts and investigations into exascale since 2007, nine years ago. At the point that it became a project, it indicates that we really want to get going on it. The reason it is a project is that there are so many things that have to be done simultaneously and in concert with each other. The general outline of the project is that we invest in applications, we invest in the software stack, we invest in hardware technology with the vendor community — the people who develop the technologies so that those technologies will eventually land in products that will be in exascale systems and that will be better suited to our applications — and we also invest in the facilities from their knowledge of what works when they install systems. Those four big pieces have to work together and this is a holistic approach.
“The project will have milestones, some of which are shared between the applications and the software so if application A says ‘I need a programming language feature to express this kind of calculation more easily,’ then we want the compiler and programming models part of the software to try to address that but then they have to address it together — if it doesn’t work, try again. That’s why it’s a project, because we have to orchestrate the various pieces. It can’t be just invent a nice programming model, tackle a very exciting application. We have to work together to be successful at exascale; same thing goes with the hardware architecture, the node technology and the system technology.”
The mission of ECP
“The mission is to create an exascale ecosystem so that towards the end of the project there will be companies that will be able to bid exascale systems in response to an RFP by the facilities, not the project, but the typical DOE facilities at Livermore, Argonne, Berkeley, Oak Ridge and Los Alamos. There will be a software stack that we hope will not only meet the needs of the exascale applications, but will also be a good HPC software stack because one of our goals is also to help industry and the medium-sized HPC users more easily get into HPC. If the software stack is compatible at the middle end as well as the very highest end, it gives them an on-ramp. And a major goal is the applications we are funding to be ready on day one to use the systems when they are installed. These systems have a lifetime of four to five years. If it takes two years for the applications to get ready to use them productively, half the life of the system has gone by before they can start cranking out results, so part of the ecosystem is a large cadre of application teams that know how to use exascale, they’ve implemented exciting applications, and that will help spread the knowledge and expertise.”
The global exascale race
“The fact that these countries and world regions like the EU have announced major investments in exascale development is an indication that exascale matters. Those countries would not be investing heavily in exascale development if they didn’t think it was useful. The US currently has a goal to develop exascale capability with systems installed and accepted in a time range of seven to ten years. It is a range, and certainly the government is considering an acceleration of that — it might be six to seven years. Any acceleration comes at a price. This project is investing very heavily in applications and software, not just on buying the system from vendors — so it’s a big investment but one that I think is necessary to be able to get the benefits of exascale, to have the applications ready to use and exploit the systems.
“Could we be doing better? If this project had started two-three years ago we would be farther ahead, but that didn’t happen. We got going about a year ago — it isn’t clear that we would be the first country that has an exaflop system. But remember I haven’t used the word exaflop until now. I’ve talked about exascale. What we’re focusing on is having applications and a software stack that runs effectively in a ratio that would indicate that it’s exascale. It might take two exaflops, so who gets an exaflop first might not be as important as who gets the equivalent of exascale. We also have goals around energy usage, 20-30 MW, which is a lot but if we didn’t have a goal like that we might end up with 60 or 100 MW, which is very expensive.
“If we are asked as a project to accelerate, we will do our best to accelerate — it will require more money and more risk, but within reason we will certainly do that.”
“I often emphasize that for the technologies that we’re hoping the vendors will develop partly with our funding and the software stack that we’re developing in collaboration with universities and industry that that will create a sustainable ecosystem. It will not just be that we’ve gotten to exascale, systems can be anointed as exascale, we breath a sigh of relief and relax. It needs to be sustainable and that’s why we really want systems that are in the vendor’s product line — they’re not something they are building just for us one of a kind. It needs to be part of the business model that they want to follow, and software that is usable by many different applications, which will make it sustainable — open source almost exclusively, which again helps sustainability because many people can then contribute to it and help evolve it beyond exascale.”