Buying Big Computers

Climate Supercomputing: In the previous blog I started an analysis of some of the arguments that are made by scientists in looking for resources, that is, money to support climate science. I took as a starting point the World Modeling Summit for Climate Prediction at the European Center for Medium-range Weather Forecasts. (See also article in Nature).

I argued that the potential impact on society that follows from climate change research is one of the reasons that climate change research motivates both broad and deep argument about funding climate change research. This distinguishes climate science from, for example, astronomy and high-energy physics.

Focusing specifically on the modeling aspects of climate science, expenditures on high-performance computers come to the forefront in the discussion. In fact, in the Statement from the Modeling Summit it was stated that in a facility to accelerate climate prediction, “The central component of this world facility will be one or more dedicated highend computing facilities …” There is no doubt that high-performance computers are essential to climate science and that climate modeling is an application that drives the requirements for high-performance computing. To some, expenditures on high-performance computers are among the highest priorities of the field. To others this is not so obvious.

More generally, expenditures in the United States on all aspects of high-performance computing are controversial. Like funding for climate change research, expenditures by the U.S. government on high performance computing has been examined and justified by one document after another. A recent document published is one by the National Academies Press, The Potential Impact of High-End Capability Computing on Four Illustrative Fields of Science and Engineering. At the core of this study was an examination of whether or not there were investigations that would benefit in the here and now from more high-performance computers, or would it be better to defer expenditures.

As I mentioned in the previous blog, high-performance computers have a relatively short life span. After about five years the computers are obsolete, and there will be the request for a new computer. Expenditure in high performance computing is not likely to lead to the generation of a definitive answer; it will lead to more questions and more complex algorithms. A high-performance computing center is a multiple tens of millions dollars per year expenditure. These computers are essential to climate research and a sustained, stable funding stream is required. This requires a different strategy, a different business model, than comes from arguing for “a center,” or a sequence of computational centers. This sustained commitment does demand accountability and metrics to indicate progress and best use of resources.

High-performance computers are not simple to use. Therefore, the expenditure on high-performance computers demands related expenditures in software and personnel. In fact, the field struggles to keep computer models viable --- able to use the computational systems. While a high-performance computer might be advertised as able to deliver several trillions of calculations per second, climate models, and many other applications, are only able to realize a small percentage of this potential. Again, this leads to scrutiny about the efficacy of expenditures in high-performance computers. To me, this suggests that a balance of expenditures is needed, and that priority attention needs to be given to the software and the viability of the software. There needs to be an anticipation of new computational systems, placing the software and algorithms out in front. If hardware, big computer systems, big iron, big silicon are out in front, each new computer system leads to an exercise in urgency of molding the models to use the computer system. This leads to an apparent inefficient use of resources. From a funding point of view, it seems much easier to advocate for large computers than it is to advocate for focused expenditures in software and algorithms.

There is one more aspect of expenditures in computing that I will mention here. An effective computational system needs to have a comparable data system associated with it. That data system needs to have an interface with an exotic and quixotic computational platform. The data system must also support the analysis of the simulations generated by the computer; therefore, it must also serve an exotic and quixotic user community. So the expenditure in high-performance computing begets expenditure in high-performance data systems.

Where does this leave us?

High-performance computers, supercomputers, are only a piece of the system that is needed to support climate modeling. Climate modeling is only a piece of the activities associated with climate science. Balanced, reasoned expenditures across all of these elements of the climate research enterprise are needed.

There are many aspects to developing a strategic, sustained managed enterprise in climate research. Some of the issues raised above demonstrate the point of view that is taken by program managers, agency leads, and elected officials when there is argument by the science community for more funds for more resources. There are legitimate questions of how the resources are managed, how priorities are determined, and the payoff of that research. The argument that climate research is important does not stand as a convincing argument. There are many important areas of research, and virtually every field feels that their research is important --- and underfunded.

It was also stated in the initial blog of this series that the community of scientists never uniformly agree that expenditures on major facilities is the best way forward. This is especially true when the expenditures are focused towards high-performance computers. These are expensive, difficult to use systems. While it is often argued that high-performance computing systems serve the community, the vast majority of the community is not served, directly, by expenditures in high-performance computing. Funds are concentrated for the direct use by what is, in fact, a small sub-culture of the community. It is an important and essential part of the climate community, but the climate modeling activities should not be confused with the totality of the community, or even the majority of the community. It is natural therefore, for others in the community to advocate for their research, because they will argue that their research is important for the future of the field. They will argue that discovery comes more from observations than computational experimentation. They will argue that discovery follows from intellect and creativity, that more minds, not more computations are the best way forward. Their arguments are substantive.

The point of this series of blogs is to address some of the strategic challenges that the field of climate research faces going forward. There is no doubt that improved understanding and prediction of the climate is important to society. No doubt that the field needs computing facilities, data systems, and observing systems. No doubt that there is much to be done now. But perpetuation of the fragmented approach to developing the facilities and the infrastructure to support the field is inefficient and, in fact, damaging. The intellectual capacity that needs to be focused on climate change is immense, and there are communities of thinkers that are able to contribute to the climate change problem today. The engagement of this intellectual community, the extension and scaling across minds, is as or more important than the extension and scaling of models across silicon chips. The climate community must move away from the strategies to secure funding that worked when the research was held primarily in the community of scientists. It must move away from the strategies that worked when agency budgets were growing. Climate science needs to recognize that its importance stands in relation to other important fields of research, like energy security, whose urgency and impact will trump the needs of climate change if it is just a matter of competing importance.


Importance of Justification



Figure 1: Picture of the ENIAC computer, which was the first computer used in weather forecasting. This image is taken from the Computer Science Department at Virgina Tech and the article The ENIAC by Kevin W. Richey. An evolving history of general circulation models and the introduction of computers to forecasting can be found here American Institute of Physics Narrative History of General Circulation Modeling. Here is a link to the University of Pennsylvania ENIAC Museum.