Community Opinion: Does the U.S. Need an Exascale Reality Check?

Print Friendly, PDF & Email

Many people believe that the current levels of U.S. investment and commitment to exascale research fall far short of what is necessary to get the country to exascale-class systems by the end of this decade.

Yet, thousands of very smart HPC community participants have accepted the position, reinforced by hundreds of presentations, that the nation will achieve exascale by 2018.

While other countries look at unified approaches to this daunting technology journey, the U.S. appears to be in jeopardy of losing momentum without a solid, coordinated plan to bring the necessary resources together in support of a national exascale initiative.

The Exascale Report Asks

  • What do you think is missing when it comes to an exascale reality check?
  • Do you think the U.S. is showing a strong enough commitment to long-term, sustained exascale funding?
  • What would be the result of the U.S. not driving a coordinated exascale initiative – and not making the necessary funding commitments to a sustained level of R&D for the remainder of the decade?
  • Do you think exascale-class systems could be a turning point toward decreasing or even ending U.S. dependence on fossil fuel?
  • If you had an exascale system available to you today – what would you do with it?

Q: What do you think we are missing when it comes to an exascale reality check?

What aren’t we missing?  The many constraints (e.g., little power per op and enormous concurrency requirements) mean that the many helpers that we take for granted today won’t be available.  We need to invest more in algorithms and in the analysis that will let us evaluate approaches, as well in innovative architecture (particularly for the part of Exascale that won’t be just commodity parts).  We should be building simulators for slices of an Exascale system and analysis techniques that can help us answer the outstanding questions, at a level of detail that will permit confidence in setting program directions.

Q: Do you think the U.S. is showing a strong enough commitment to long-term, sustained exascale funding?

No.  Just look at the trouble with funding of the UHPC program and the lack of a significant research program, at either NSF or DOE, in Exascale and trans-exascale hardware, software, and algorithms.

Q: What would be the result of the U.S. not driving a coordinated exascale initiative – and not making the necessary funding commitments to a sustained level of R&D for the remainder of the decade?

The obvious – we won’t get to Exascale this decade.  But more serious is the consequence of this – for decades, computing power has been getting faster exponentially.  This will stop and it will be a shock – many computational scientists have come to depend on ever increasing computing power.  (Yes, even with research investment, we won’t be able to continue the increase in computing power – but we can better manage approaching the limit, ensuring a smooth transition to a time when computing power does not increase as rapidly.)

Q: Do you think exascale-class systems could be a turning point toward decreasing or even ending U.S. dependence on fossil fuel?

Maybe (e.g., fusion, more practical fission, more efficient solar) but the case still needs to be made.  The various exascale workshops did begin to look at this question, but an answer convincing enough to build a billion dollar project on will require more details.  A first step would be something like the point designs that NSF funded back when the community was trying to figure out how to achieve petaflops.  The DOE co-design centers might help, once they begin in earnest.

Q: If you had an exascale system available to you today – what would you do with it?

Do what we (and NSF) are doing with Blue Waters, and what DOE does with INCITE – make it available for research that can’t be done in any other way (note that this means a thousand petascale runs for an ensemble calculation is not what I’d do with an exascale system).


Q: What do you think we are missing when it comes to an Exascale reality check?

To reach Exascale computing, without just throwing hardware at the problem, we need significant advances in: single node performance, interconnect between nodes, and software development tools. And of course a power budget that is reasonable.

All the above can be achieved if we have a well thought out program, managed by a group of highly motivated people. This program must have milestones that must be met along the road to Exascale computing.

Additionally, more and more there are discussions about application specific Exascale compute. One size does not fit all, especially when we consider programming models and power consumption. In July, there is a workshop titled” “MyAnton” in Park City (sponsored by the DOE and Argonne (ICiS)) that focuses on application-specific computing.

Q: Do you think the U.S. is showing a strong enough commitment to long-term, sustained Exascale funding?

NO. There is lots of talk and workshops, but I have yet to see a strong commitment.

Q: What would be the result of the U.S. not driving a coordinated Exascale initiative – and not making the necessary funding commitments to a sustained level of R&D for the remainder of this decade?

The Mainland Chinese will get there first. They have demonstrated this for the Linpack in the latest Top 500 report. Whatever it takes, including a new power plant. I believe they are less concerned about elegance (note: their efficiency of sustained vs. peak for Linpack) and more concerned as to getting there first.

Q: Do you think Exascale-class systems could be a turning point toward decreasing or even ending U.S. dependence on fossil fuel?

The better the simulation capability we have, the finer the fidelity of the results, the faster we get to achieving some of the major national priorities. Renewable energy is one of these priorities. Increasing the efficiency of ANY form of fuel use is beneficial to society.

Q: If you had an Exascale system available to you today – what would you do with it?

I believe we are on the cusp of significant medical breakthroughs in the use of genomic sequencing and bio informatics and their use in Pharma. Major simulations can take hours and/or days. And these are still approximations and/or reduced data sets.

Understanding the role of DNA and its changes over time (how and why), and how to then use this information to directly impact Pharma will have a dramatic impact on society as a whole and its beneficial effects, both in health care and reduced medical costs. This will make an investment Exascale look like chicken feed from a return on investment perspective.


Q: What do you think we are missing when it comes to an Exascale reality check?

The biggest challenges facing Exascale systems is energy efficiency and programmability. On the energy efficiency front, just relying on expected semiconductor scaling through the rest of the decade would result in Exascale systems in excess of 200 megawatts. Substantial research and development will be required in the areas of low-power circuits, energy efficient architectures, and energy-aware programming systems and applications. On the programmability front, no one is prepared for the billion-way parallelism that will be required for Exascale systems. Today’s supercomputing programming systems are unlikely to easily evolve to match the requirements of Exascale.

Q: Do you think the U.S. is showing a strong enough commitment to long-term, sustained exascale funding?

While there have been many U.S. sponsored Exascale studies over the last few years, actual work and funding support for the R&D for Exascale has been slow to start. Given the design cycles of components (such as GPUS) that will go into Exascale systems, the window in which Exascale critical application studies can affect hardware designs will close within the next 2-3 years. A more concerted and urgent effort that is presently visible will be required. Other countries in Europe and Asia are investing in high performance computing systems and infrastructure and the U.S. is at risk of losing its leading position in this field.

Q: What would be the result of the U.S. not driving a coordinated exascale initiative – and not making the necessary funding commitments to a sustained level of R&D for the remainder of this decade?

The result may be two-fold: (1) Exascale systems will be delayed well into the next decade, delaying the science that will benefit from Exascale systems, and (2) Other countries will seize the opportunity to take a leadership position in high-performance computing, which can then be used to enable much better scientific discovery and engineering design. Investments by other countries in high-performance computing is being made as a means to become more economically competitive. The U.S. risks losing its edge in these areas and being less competitive.


Q: What do you think we are missing when it comes to an exascale reality check?

Exascale technologies are probably a decade in the future. With enough money we can put sufficient equipment in a room with power and skilled people to run an Exascale benchmark. It would be a completely impractical machine today and even the most optimistic estimates have current technologies still pushing the bounds of practicality in ten years. Computing technology goes through periodic revolutionary change: pipelining to achieve Megascale, vectors to achieve Gigascale, message passing to achieve Terascale. These changes can be disruptive to customers invested in the old technology; to lead in these transitions requires sustained strategic investment in what is initially a commercially unattractive market.

Q: Do you think the U.S. is showing a strong enough commitment to long-term, sustained exascale funding?

The funding so far is commendable, particularly programs such as UHPC which we are proud to be a part of. Without sustained research investment, the U.S. will risk falling behind. We don’t want to simply reach exascale just to hit an arbitrary benchmark. We want to make Exascale systems practical and productive. These will be truly massive systems by any measure and we need research to uncover new opportunities that will make them far more practical, especially in terms of power consumption, programmability, and overall cost.

Q: What would be the result of the U.S. not driving a coordinated exascale initiative – and not making the necessary funding commitments to a sustained level of R&D for the remainder of the decade?

It is the same risk you take in the business world if you don’t invest in R&D. Someone else will get there first, and someone else will eat your lunch. Exascale systems are massive information engines 1000 times more capable than we have today. These systems will, without a doubt, uncover new physics, chemistry, biology and lead to new understanding, products and technologies. They will run missions that are critical to national competitiveness and defense. It requires a well-funded and well-coordinated initiative to remain in first place.

Q: Do you think exascale-class systems could be a turning point toward decreasing or even ending U.S. dependence on fossil fuel?

With Exascale systems we expect to uncover concepts in physics that could lead to new energy sources and more efficient ways to store or transport energy. We see the foundations of this work at the National Ignition Facility and in “Smart Grid” research. We know that they will have a tremendous impact in areas such as drug discovery, transportation, financial analytics and product design.

Q: If you had an exascale system available to you today – what would you do with it?

I might want to do a whole host of interesting things, but I would likely find it extremely difficult to take advantage of such a machine today. Applications, and application designers, are not yet ready to take advantage of the degree of parallel processing we expect. We’re talking millions or perhaps billions of threads in a single application. We’re learning and innovating every day and already developing techniques to allow software to scale to this degree, but, we’re not there yet. A coordinated research initiative is needed to bring the applications and tools in line with the capabilities of an Exascale system.

For related stories, visit The Exascale Report Archives.