Climate Modelers Have Insatiable Appetite for HPC
Since the dawn of high performance computing, climate modeling has been one of its most demanding domains. The hunger for computational capability is unending, as researchers work to incorporate more of nature’s complexity into their models at higher resolutions. HPCwire talked with NOAA/GFDL Deputy Director Brian Gross and Venkatramani Balaji, head of the lab’s Modeling Systems Group.
HPCwire: How important have HPC-based modeling and simulation been in increasing human understanding of climate behavior and climate change?
Brian Gross: The climate system is inherently complex, measured by the number of processes and feedbacks between climate variables. It has interactions at all time and space scales, from minutes to millennia, and from millimeters to planet-scale. The role of HPC in addressing these inherent computational challenges to achieve the tremendous advances in our understanding of the Earth System cannot be overstated.
Venkatramani Balaji: In fact, Nature listed the first ocean-atmosphere coupled model — achieved by Suki Manabe, Kirk Bryan, and their collaborators at NOAA/GFDL in 1969 — as a milestone in scientific computing. That model, run on the HPC of the ’60s, was the first to show that adding CO2 to the atmosphere changes the radiative balance so as to increase surface temperatures. HPC-based modeling is the only science-based method to project future climate change.
HPCwire: In the 1990s, US climate researchers published a paper lamenting the lack of access to the most powerful supercomputers for climate modeling, which at that time were vector systems. Has anything been lost in the transition to non-vector supercomputers?
Gross: It turns out, no. On a scientific level, US labs without vector supercomputers kept pace with European and Japanese labs with vector machines. There is no evidence in hindsight that being denied access to vector machines hurt the US labs, whether measured in terms of scientific breakthroughs, or publications, or metrics of model skill.
Balaji: This is not to say that we went through the transition with no pain! The switch from vector to distributed memory machines was certainly disruptive and required a thorough technology refresh of the models. Labs had to expend a lot of effort recoding and then verifying that the new codes were capable of reproducing proven results.
Gross: We used the occasion also to instill better software engineering practices, and I think most people will agree that we’re the better for it. The models today are more agile and more configurable. We can build more complexity into our models than we were able to in the ’90s because of component-based design. We are now able to include atmospheric chemistry, aerosols, dynamic ecosystems on land and ocean, and we can study the complete Earth system. We couldn’t have done this very easily with models of the 1990 vintage.
HPCwire: What are the biggest challenges facing the climate modeling community today?
Gross: The principal challenges we face in climate modeling today remain the same as they have for decades: our limited understanding of the way the Earth System works, how accurately we can translate what we do know into computational algorithms and numerical models, quantifying uncertainty, and efficiently running our increasingly computationally intensive climate models on the largest HPC systems in the world.
It is worth pointing out that the direction of technology today, using more processors rather than faster processors, greatly favors weak scaling over strong scaling. The consequence is that we can often execute more complex, higher-resolution models at a fixed rate, as measured by, say, model years per day.
Balaji: But it’s much more difficult to execute a given model at a faster rate. This can often impede our scientific progress, given the very long time scales associated with some climate processes, such as the global ocean circulation and long-lived greenhouse gases like carbon dioxide. We’ll return to these challenges in a minute.
HPCwire: In the next few years, what are the goals for increased resolution of coupled earth system models?
Gross: The question of anthropogenic climate change on the scale of the planet is settled from a purely scientific viewpoint. However, understanding the details of climate change on a regional scale is harder. We’re not yet at a point where we can attribute local or regional climate change to human actions with the same confidence.
The goal for the current generation of IPCC-class models is to see if higher resolution yields better skill on regional scales. This is not a given. As processes that are currently unresolved become resolved, their representation in models changes from “parameterized” to “simulated.”
Balaji: There are key processes — for instance mesoscale eddies in the ocean, and deep convection in the atmosphere — that will undergo this transition over the next 5-10 years. Some current problems, such as cloud-climate feedback and ocean mixing, will be solved, but new ones might emerge. But certainly cloud-resolving and ocean-eddy-resolving coupled models promise to yield qualitatively new and exciting science.
HPCwire: What are the biggest barriers to greater scalability? Is it the codes, the models, or the limitations of the known science?
Balaji: All of these are barriers, but this list is incomplete. Why are hardware and system software not on your list? Our main difficulty is the speed of a single operation has not got faster for a while and is likely to become slower on the many-core and GPU cluster type systems. Compilers have not got any better for a long time at interpreting our codes, and are even more immature on the novel architectures.
Gross: The expectation had been that a given model at a given resolution would get faster over time just by advances in technology. We’ve just had a rude awakening.
Balaji: As an aside, I’d focus on time-to-solution rather than scalability per se. We all know tricks that make models run on more processors, yet take longer to reach the same solution. We class our models as 1 year/day models, 10 year/day models, and so on. Each can be used for a different class of scientific problem.
HPCwire: It seems that generational advances in computing power reduce uncertainty by enabling greater resolution, but adding new components to coupled models, such as for the carbon cycle, can offset these gains by increasing the complexity of the models? How do you balance these choices?
Gross: Good question. Our feeling is that the complexity comes first. When we feel we’ve reached a level of understanding of some process — say aerosol-cloud interactions, or dynamic vegetation — they get added to the models, and a new realm of scientific problem opens up. We then look at what hardware we can get with our computing budget, and that tells what resolutions we can use while achieving the target model years/day pace necessary for useful science.
HPCwire: How well do the atmospheric, oceanic and other components of coupled models and ensemble models “talk to” each other? How compatible are the physics and the scales in these models?
Balaji: We typically change one component at a time, so that you can do careful comparisons with previous results and trace differences back to a single component. But resolutions stay close, usually within a factor of two or so.
Not to say that the grids are the same. Atmosphere and ocean modelers have taken different routes to avoiding grid singularities and other numerical issues. Coupling technology is stable and mature. There are good, efficient, scalable, conservative coupling and regridding methods, but there’s always an open question as to whether they’ll keep scaling as we add resolution. Also, we’re not well situated to take advantage of AMR [adaptive mesh refinement], and so on. These methods are not much in use in the climate field today.
HPCwire: The goal is for the “Gaea” Cray XT6 supercomputer at ORNL to grow to a 720-teraflop Cray XE6 system in mid-2011. The plan is for “Gaea” to expand to 1.1 petaflops later on. What will these increases make possible?
Gross: Gaea puts within reach the eddy-resolving ocean models and cloud-resolving models we just spoke about. Separately, we’re already there. We believe we’ll be doing useful science with these models in coupled mode shortly after we get the full petaflop machine. Okay, maybe not cloud-resolving, but tropical storm-resolving.
Balaji: Additionally, we’re exploring predictability issues with our models. How sensitive are predictions to initial conditions? These studies explore probability distributions across ensembles of runs initialized with an advanced coupled data assimilation system. These will also stress the capacity of the machine.
Putting these two together, for predictability changes as a function of resolution, we could use up these cycles many times over. And I haven’t even mentioned the Earth System models, which apply this unique resource to substantially increase complexity,adding in atmospheric chemistry, fully interactive land-based ecosystem dynamics and carbon, nitrogen, and other biogeochemical cycles.
HPCwire: What elements of this supercomputer are especially important for weather and climate modeling?
Gross: We hope we’ve made it clear that we can now envision an unprecedented set of exciting science that was out of our reach before. The Cray SeaStar interconnect allows extraordinary levels of scaling, and we’re looking forward to seeing results on the Gemini upgrade, which should be even better.
HPCwire: How much of NOAA’s focus is on modeling weather and climate phenomena in the US, versus other areas of the world?
Balaji: All of our models are global, and the processes and feedbacks are linked on the planetary scale. It’s generally found that to get the climate right over the US, you do need to worry about clouds off the coast of Peru, or you need to get North Atlantic sea surface temperatures right to simulate drought in the Sahel, to take some prominent examples of global linkages. Some short runs are undertaken with regional models, but the fundamental basis of all research and operations is global models.
Gross: We are now configuring some variable resolution models such as the stretched cubed sphere, where resolution can be focused on the US, for instance.
HPCwire: There’s considerable pressure to reduce federal spending in every area possible. Why should strong funding for weather and climate modeling continue?
Gross: Just check out NOAA’s Next-Generation Strategic Plan. Climate change has already had profound implications for society, and climate model predictions and projections foretell a host of additional significant impacts both nationally and internationally.
We need the best possible science-based information on future climate so that decision-makers can develop and evaluate options that mitigate the human causes of climate change and allow society to adapt to foreseeable climate impacts. This information can only be obtained through state-of-the-science climate models. The cost of the associated HPC is trivial compared to the social gains from mitigation and adaptation.