BACKGROUND AND CHARGE
The U.S. research community is a world leader in the study and understanding of climate and climate variability. Modeling capabilities are being applied to the study of anthropogenic impacts on the climate system, such as those resulting from additions of radiatively active constituents to the atmosphere, and from slow natural variations of climate. Short-term climatic variations, such as those occurring with the El Niño/Southern Oscillation, are becoming better understood and may be increasingly predictable as a result of the observations and modeling by this community.
This skill of predicting short-term climate variations and the information gained to better understand natural variability and the response to natural and anthropogenic perturbations is of great societal, ecological, and economic value for future planning. Recently a key use of climate models has been the production of legally mandated climate assessments (the U.S. National Assessment) and assessments required by international agreement (the assessments of long-term climate change performed by the Intergovernmental Panel on Climate Change and the Ozone Assessments called for by the Montreal protocols). Regional assessments to characterize climate impacts on a more local scale are increasingly in use, as they become valuable for planning purposes in both the public and private sectors.
Recognizing the societal importance of climate modeling, the National Oceanic and Atmospheric Administration (NOAA) and the National Science Foundation (NSF) requested the Climate Research Committee (CRC) of the National Research Council (NRC) to investigate the current state of
U.S. climate modeling and its ability to meet these assessment demands. In response to the request, the NRC produced a report entitled Capacity of U.S. Climate Modeling to Support Climate Assessment Activities (NRC, 1998a). This report evaluated allocation of resources to high-end modeling and whether these resources were being used effectively. The CRC found that “insufficient human and computational resources are being devoted to high-end, computer intensive, comprehensive modeling, perhaps, in part, because of the absence of a nationally coordinated modeling strategy.” This present study focuses on the challenges posed in the 1998 report and as specified in the statement of task given to the panel.
The purpose of this study is to provide relevant federal agencies and the scientific community with an assessment of the nation's technical modeling needs and a vision of how government, interacting with the rest of the scientific community, can optimize the use of modeling talents in the United States. This study addresses the challenges posed in the Climate Research Committee's 1998 report, Capacity of U.S. Climate Modeling to Support Climate Change Assessment Activities. In pursuit of these objectives, the panel:
Examines the major types of climate modeling, paying particular attention to both the similarities (e.g., potential synergisms) and unique characteristics of each. Specific issues to be addressed include model construction and testing, data input and archival, ensemble simulation, interrogation and diagnostics, evaluation, and operational utilization.
Describes the computational and human resources required to effectively conduct climate modeling in the United States to meet the needs of the climate applications, policy, and scientific communities. This evaluation will include consideration of shifts in computational architectures and potential for, and cost of, improvements in model codes. It will also consider the utilization of common climate modeling tools, protocols, and data, and the availability of cooperative opportunities between different scales of modeling effort and institutions.
Quantitatively assesses the computational and human resources that are presently directed toward climate modeling in the United States.
Describes ways in which the efficacy of the U.S. climate modeling enterprise might be improved, given the current needs and resources. The report will define a set of issues that are fundamental to the enhancement and sustenance of climate modeling in the United States.
CLIMATE MODELS AND OBSERVATIONS
Climate models are mathematical representations of the major systems (atmosphere, ocean, land, snow, and ice) whose interactions determine climatic means and climate variability. For the most complex mod-
els, components of the climate system are linked, or coupled, using algorithms describing the connections between these systems. Due to limitations in resolution such components as radiation, clouds, and turbulent processes are generally unresolved in climate models and therefore require separate numerical representation. Although the climate components in these models can be separately built and evaluated, the nature of their coupling determines the behavior of the climate model. To evaluate model realism, model outputs are compared to each other and to environmental observations. The results of these comparisons form the basis for changes to model code, which improve the mathematical representation of physical processes.
Model simulations can be used for short-term environmental prediction, climate prediction, and assessment of future climatic responses to anthropogenic forcing. They can be combined with observations to produce model-assimilated data sets, and to design and improve climate observing systems. Atmospheric analysis through the assimilation of weather data into weather forecast models, and the need to downscale and interpret the output of climate data to the local region, provides a continual and necessary interaction between climate modeling and weather modeling. Because models are tested and improved through comparison to observational data, progress in modeling and observations are interdependent. An effective and integrated system for producing and delivering climate information needs to be supported by data collected from a dedicated climate observing system. Because the present atmospheric observing system was built primarily for weather prediction, and as such is subjected to major changes in time, it is inadequate to unambiguously detect and monitor climate change. Without regular and systematic analysis of parameters controlling the climate system, it is impossible to clearly document climatic variability and long-term climate trends. The panel therefore notes that the lack of a suitable sustained observing system for climate limits progress in climate modeling.
The building of parameterizations of individual model elements; the running of uncoupled atmosphere, land, and ocean models; and the diagnoses and analyses of coupled climate model outputs can be accomplished at the workstation level. The integration of the components into comprehensive coupled climate models, the running of these coupled models, and the integration of global data with models can be attained only by using the very highest end of supercomputers.
The panel concludes that sustained computational capabilities of 10-
100 Tflops 1would meet the needs of the different types of climate modeling; this capability is almost attainable using present technology. The potential for using this capability to achieve adequate throughput is determined by the efficiency of a given model code on the available supercomputer architecture (parallel vector processors and massively parallel commodity processors). The panel concludes that parallel vector computers provide superior processor speeds, greater usability, and lower human resource requirements; however, the massively parallel commodity processor machines are currently the only ones that can be purchased in the United States. The primary drawback to a massively parallel architecture is that the speed at which climate modeling code will run on a large number of parallel processors does not linearly increase with the number of processors but is controlled by Amdahl 's law. This law essentially states that incomplete parallelization of model code creates significant computational inefficiencies and reduces the speed at which that code is run on a large number of processors. Even perfectly written code must deal with the irreducibly sequential underlying dynamics so that there is an absolute theoretical limit with massively parallel machines to the speedup factor possible over the speed of a single processor. This limit is far less than the theoretical maximum based on the number of processing elements.
As part of this study a survey was conducted to quantitatively assess the computational and human resources presently directed toward climate modeling in the United States. The survey responses indicated that access to increased computational power is desired across all modeling scales but is most apparent at the highest end. Smaller and intermediate-size modeling groups are able to accomplish modeling undreamed of a generation ago, but they expressed the desire for increased access to supercomputing facilities. A recurrent theme in the survey results was the difficulty of hiring and retaining computer technologists because of extraordinary competition from the information technology sector.
The survey also provided information on supercomputing capabilities for climate modeling in the United States. With the exception of a few centers devoted to prediction, most of the computational load in existing centers is devoted to modeling for research purposes. Computing capabilities at large modeling centers have sustained speeds between 10 and 100 Gflops (with most being toward the lower end of the range) on actual model codes using massively parallel processing systems. These systems enable present coupled climate models to be run for hundreds of model years at resolutions of 300 km in the atmosphere and approximately 100
Computer speed is frequently measured in units of “floating operations per second”, or flops. Megaflops (Mflops) indicates a speed of a million operations per second. Gigaflops (Gflops) equal 1000 Mflops, Teraflops (Tflops) equal 1000 Gflops.
km in the ocean. One of the major problems faced by large modeling centers is the conversion of existing model codes, previously optimized for machines using small numbers of vector processors, to those that can efficiently run on parallel architectures. The scarcity of human resources in information technology further compounds this problem.
RESEARCH AND OPERATIONS
During the panel's examination of measures to improve the effectiveness of U.S. climate modeling, a distinction between modeling in response to societal need (“operational” modeling) and modeling for research arose from the necessarily differing roles played by the research community in each. Societal requirements for the regular delivery of useful climate information products places demands on the research community that are difficult to meet because of insufficient resources, the lack of research organization capable of concentrating the resources needed to respond to these demands, and an inappropriate management structure to carry out the regular and systematic production of products. Although operational modeling depends on research for its success, by itself, it is not a research activity and cannot be well addressed in a research culture. The panel concluded that the present research infrastructure spread among many research agencies, each operating in its own interests according to its own culture, is not capable of responding to the modeling demands of regular assessment and prediction, nor is the management structure of the U.S. Global Change Research Program (USGCRP) able to instill such a culture or otherwise provide the focus required for regular climate product production.
Analogous to operational weather forecasting, centralized climate modeling activities and the maintenance of a climate observing system should exist outside the research domain but should have a close interaction with the research community. The beneficial interaction between research and operational climate modeling communities requires that different groups be able to run a variety of models, interchange model components and parameterizations, and compare output and observational data in common formats. This “Common Modeling Infrastructure” is defined as a set of standards, protocols, and associated tools for physical parameters, model codes, file formats, diagnostics, visualization, and data storage, as well as an ‘exchange infrastructure' that fosters efficient collaboration among modeling groups.
The state of the climate system can only be defined using sustained observations of critical components of the climate system. Observational data assimilated into a comprehensive coupled climate model will enable the verification and enhancement of model code to produce accurate and
continual climate analysis. From panel deliberations and acknowledgments of the robust linkages between research, observations, and climate modeling, the panel endorses previous NRC reports (NRC, 1998b, 1999a, 1999b, 1999c, 2000a, 2000b) that called for the development of a sustained climate observing system.
Based on panel expertise, input from a one-day workshop held as part of this study, and from a survey distributed to large, intermediate-size, and small climate modeling centers, the panel makes the following recommendations.
The Need for Centralized Operations
Recommendation 1: In order to augment and improve the effectiveness of the U.S. climate modeling effort so that it can respond to societal needs, the panel recommends that enhanced and stable resources be focused on dedicated and centralized operational activities capable of addressing each of the following societally important activities:
short-term climate prediction on scales of months to years;
study of climate variability and predictability on decadal-to-centennial time scales;
national and international assessments of anthropogenic climate change;
national and international ozone assessments;
assessment of the regional impacts of climatic change.
The Need for Open Access to the Most Appropriate Computer Architecture
Recommendation 2: The panel recommends the adoption of a scientific computing policy ensuring open access to systems best suited to the needs of the climate modeling community.
Recommendation 3: Researchers should have improved access to modern, high-end computing facilities connected with the centralized operational activities discussed in Recommendation 1. These facilities should be sufficiently capable to enable comprehensive study of the climate system and help develop models and techniques to address relevant high-end climate modeling problems.
The Need for a Common Modeling Infrastructure
Recommendation 4: In order to maximize the effectiveness of different operational climate modeling efforts, these efforts should be linked to each other and to the research community by a common modeling and data infrastructure. Furthermore, operational modeling should maintain links to the latest advances in computer science and information technology.
Human Resource Needs in Support of Climate Modeling Activities
The climate modeling community faces a severe shortage of qualified technical and scientific staff members, who, because of high salaries and incentives, find the high-tech industry more desirable than the research and operational modeling centers. (Some overseas groups, such as the European Centre for Medium-range Weather Forecasts (ECMWF), have overcome this difficulty by offering lucrative salary packages that U.S. modeling groups have been unable to match.) A further complication is the dependence of university-based modeling groups on the vagaries of short-term funding for employee salaries. The state of affairs also affects graduate programs, wherein universities see many students accepting attractive offers from private industry prior to completion of their degrees. This strain on human resources has resulted in declining graduate enrollments in all areas of the climate sciences and in the growing disparity in the quality of life of scientists —especially young ones—and their private sector counterparts.
The shortage of highly skilled technical workers is, however, not unique to the climate modeling community; it is part of a larger shortage affecting nearly all areas of science and engineering except those with strong linkages to the private sector. The complexity of the problem and the lack of expertise on the panel to address this issue precludes this panel from making any specific recommendations related to human resources.
Institutional Arrangements for Delivery of Climate Services
Recommendation 5: Research studies on the socio-economic aspects of climate and climate modeling should be undertaken at appropriate institutions to design the institutional and governmental structures required to provide effective climate services. This assessment should include:
an examination of present and future societal needs for climate information;
a diagnosis of existing institutional capabilities for providing climate services;
an analysis of institutional and governmental constraints for sustaining a climate observing system, modeling the climate system, com-
municating with the research community, and delivering useful climate information;
an analysis of the human resources available and needed to accomplish the above tasks;
an analysis of costs and required solutions to remove the constraints in accomplishing the above tasks;
recommendations on the most effective form of institutional and governmental organization to produce and deliver climate information for the public and private sectors.
VISION FOR THE FUTURE
The panel envisions an operational entity or entities that would create and deliver climate information products of benefit to society. This entity would not only concentrate resources on the needed modeling activities but would also establish and maintain a climate observing system to develop and test climate models and make available climate information for public and private use. Researchers would interact with this group to develop, diagnose, and improve models and observational systems. This interaction would provide the research community with otherwise unobtainable resources and result in enormous benefits and a sound foundation for future improvements.