From the birth of HPC, climate research has had a voracious appetite for computing resources. John Drake, chief computational scientist for the Climate End Station at Oak Ridge National Laboratory, explains what petascale computing will do to help feed this hunger and how the lab’s work supports the mission of the Intergovernmental Panel on Climate Change.
HPCwire: First, what is the Intergovernmental Panel on Climate Change and what is its main goal?
John Drake: The Intergovernmental Panel on Climate Change, or IPCC, was established in 1988 by the World Meteorological Organization and the United Nations Environment Programme to scientifically assess the global risk of climate change, its potential impact, and options for mitigation. Since that time the IPCC has published four climate change assessments, and a fifth assessment is scheduled for publication in 2013.
HPCwire: What does Oak Ridge National Laboratory have to do with the IPCC and climate research?
Drake: As the Department of Energy’s largest science and energy laboratory and a significant contributor to the fourth IPCC assessment, ORNL has been selected by the DOE and the National Science Foundation’s National Center for Atmospheric Research (NCAR) as a Climate End Station (CES) to help with the fifth IPCC assessment. Warren Washington of NCAR is chief scientist for the CES and I’m chief computational scientist.
HPCwire: Can you provide a brief overview of the Climate End Station?
Drake: Sure. The CES is a vehicle for engaging the broader science interests of the climate community with focus on the contribution of high performance computing. Specifically the CES helps to organize and coordinate the computational efforts — from the perspectives of both development of scalable climate modeling software to prioritization and allocation of computer time. It includes a mechanism to manage computer allocations awarded by DOE through their INCITE program. Imagine everything that has to happen to develop projects and to do science along the way. That requires lots of computer time and lots of decisions about where to go next. The primary software that is developed and used by the climate community is called the Community Climate System Model (CCSM) and it is also a joint effort involving a broad collaborative community. DOE allocates computing resources for climate studies, and CES was put in place as a value-added organization, to give structure on top of individual allocations, to choose priorities, help to schedule projects, and determine what development and goals need to be met to keep to the schedule. The CES chooses priorities for our 20-million node-hour allocation and makes sure that the development for the IPCC assessment is on schedule.
As I mentioned, Warren Washington is the science PI on this. There’s an executive board that advises Warren on who’s ready to do what, when. Through INCITE we applied for a very large computer allocation across systems at ORNL, NERSC and ANL and were fortunate to receive an award commensurate with the needs and goals of the climate community. We’re targeting things related to climate change and the role of carbon, versus broad climate science.
HPCwire: What is the process of creating an IPCC assessment?
Drake: For the first three years of the five-year assessment cycle, we’re primarily concerned with determining what areas to study, tracking HPC advances so we know how detailed our models can be, deploying the most current HPC technology, and creating and testing the actual models. Then we freeze the models and spend the fourth year running simulations and gathering data for the different scenarios. In the fifth year, we write the papers that report our results and have those papers reviewed by different working groups within the IPCC community and by a variety of government agencies. The result is an assessment that includes thousands of pages of peer-reviewed results of climate change models.
HPCwire: Regarding the 5th IPCC assessment, you’ve discussed the need to include additional factors in the planning phase, such as biogeochemical factors and higher resolution of ice sheets. Are these new factors or were they in the 4th assessment but needing more computing power?
Drake: It’s a little bit of both. New scientific results and increased compute performance are making more comprehensive models possible. Take sea level rise, for instance. In the 4th assessment the modeling community didn’t have good information, so there wasn’t much to say on this topic. Current models are based on phenomena that don’t include the land ice sheet dynamics. We do know that thermal expansion of the ocean causes some of the rise. But very little modeling has been done to look at what’s happening to ice sheets, for example in Greenland or the Antarctic. We have no predictive ability at this point to say what will happen with the Greenland ice sheet. Field research, conversely, indicates that if we remain on our current course, the melting of these ice sheets could raise sea levels quite a bit more than we currently predict (30 to 40 cm in the year 2100) by factoring in only the thermal expansion of the ocean. The modeling community didn’t have answers for the 4th assessment. We’re working on that now.
HPCwire: They didn’t have answers because the compute power wasn’t big enough then?
Drake: It was insufficient compute power and other things. People didn’t think the ice sheet melting was changing that much. We needed more comprehensive models, and to get to more realistic models and simulation we need more compute power.
HPCwire: So, what goes into the 5th assessment is a combination of factors. You have to balance computing capability over the next five years with the number and kinds of elements you include in the model.
Drake: Right. In our process, the scientific steering committee tries to lay out the target architectures and model configurations, and then we go to a prototype to see how well we do. It doesn’t help if the model is so complex no one can run it. We try to include as much as possible until the productivity level takes a real hit. Below five simulated years per day, climate scientists can’t get much new work done. We design the model with as many physical effects as we can, then address scalability. With more processors we’ll be able to increase the complexity and intricacy of the models significantly.
HPCwire: Could the 4th assessment have been done without HPC?
Drake: Probably, but not with the same detail. The bigger problem was that despite important new results, the science wasn’t cooked enough yet to go even further.
HPCwire: How would you characterize the main focus of the 5th assessment?
Drake: This is an ongoing conversation. The 4th assessment is broadly viewed as really nailing down the question of whether global warming is occurring and the primary causes of global warming. The 5th will move past this and ask more detailed questions about regional climate change. It will aim to go beyond the global view of whether it’s occurring and look at how, where, and how fast it’s occurring, and in which specific geographic areas, along with the signs of global warming. The melting of the Arctic ice cap will be a larger focus and the 5th assessment will put more emphasis on high resolution to capture regional details and extreme events and their relation to global warming. We’ve also been emphasizing the relationship with the carbon cycle. That’s biogeochemistry and explores how agriculture and the ecosystem are going to react under climate change scenarios. For this we develop global carbon models. There will also be much more detail and foundational science about emissions. The IPCC work has always tried to define scenarios based on the atmospheric concentration of greenhouse gases. The results in the 4th assessment are turning out to be overly optimistic. The rate of increase of CO2 in the atmosphere is larger than any of the scenarios being studied. The upper-range scenarios laid out in the 4th assessment are now considered mid-range.
HPCwire: Is the biogeochemical factor a major addition in the 5th assessment?
Drake: It’s one of the major additions. So is the ice sheet model, having more regional detail, and the aerosol indirect effect. We haven’t had good observational data on this to date.
HPCwire: What’s the status of the new factors that will be in the 5th assessment?
Drake: We have alpha versions of the CCSM4. It’s ongoing work.
HPCwire: What are your main workhorse computers?
Drake: We currently have the Cray XT4 system that is being upgraded to 250 teraflops and a Cray X1E vector system. We expect to have the Cray petaflops system later this year.
HPCwire: Do you work closely with vendors like Cray to know what’s coming down the pike?
Drake: We try to. We look at software life cycle models based on about 10 to 20 computer systems that we have swapped in and out. It’s tough to adapt to new architectures, but we try to look into the future to begin adapting. Multicore is a case in point. We’ve been getting ready for this for several years. We’re in the process now of making adaptations to CCSM for the NCCS 250 teraflops Cray XT4 system, which is based upon the AMD quad-core processor.
HPCwire: Where are you on the road toward petascale computing?
Drake: We’re upgrading our Cray XT4 system to 250 teraflops already. That machine will stay put and a petaflops Cray system will be brought in by the end of the year. This machine is in line to be the first petascale system available to the climate modeling community in the world.
HPCwire: How important is a petaflops system in terms of what it will enable you to do? Does it have real meaning in terms of advancing the science?
Drake: It starts to get really interesting. There is talk of simulating the atmosphere at ½ degree or maybe even ¼ degree resolution. We could go 360 degrees around the equator, chopping up the earth into ½ degree or ¼ degree blocks. In the ocean, we are experimenting with models that go to 1/10 of a degree. At resolutions less than that, you’re not representing eddies and that’s important because this is where ocean heat transport takes place. The types of phenomena we’re going to see at these resolutions will be very interesting. We can look in much more detail at the Arctic ice cap, in particular.
HPCwire: What about microclimates?
Drake: These will become more detailed as well. We’ll be able to look at the amount of rain falling in the atmosphere and chop the land up into more refined areas to see land use, including agriculture use, and to see forests are and track deforestation, etc. We’re looking at 1 km resolution for land use that will be much more realistic.
HPCwire: What about biofuels research?
Drake: The Computational End Station will host a lot of research on biofuels. People are looking at whether agriculture will start to change. If we plant enough biofuel crops, how will this affect the climate? There could be significant effects. We’re able to do those kinds of studies now. The chemistry coupling is becoming available.
HPCwire: What do you need from hardware, software and vendors to advance climate research as we move toward petascale and exascale hardware?
Drake: There’s an increased need for these communities to work together. Applications have to deal with an increasingly complex memory hierarchy and heterogeneous processors. These are large changes, not incremental changes. We’d like software to solve the problems while we all stay in our corners, but we need to come together. There’s no silver bullet. Identifying parallelism at the granularity you need is a real challenge with high numbers of processors. We need tool developers, language developers and hardware vendors to balance these things correctly. You can see some of these interactions developing in the tools and performance monitoring efforts. High performance languages and optimized libraries that better fit the emerging architectures offer other challenges and opportunities for scaling application codes to thousands or millions of processors.
HPCwire: What’s coming down the pike in HPC that you’re excited about?
Drake: Special purpose processors being used for science, streaming processors in particular. There is a lot of potential around graphics boards and accelerator processors. A common, standard programming language is holding them up. There’s no standard HPC programming language on these systems yet. As this gets solved, and this looks to be coming along in the next year or so, there could be a big leap forward.