Cray, NVIDIA, the Swiss National Supercomputing Centre (CSCS) and MeteoSwiss recently announced the acceptance of MeteoSwiss’ new supercomputing platform for operational weather forecasting, a Cray CS-Storm GPU supercomputer system with NVIDIA® Tesla® K80 GPUs. It is the world’s first operational weather forecasting system using GPGPUs as the primary computational engine, and it represents a successful return on years of effort by MeteoSwiss, C2SM/ETH and CSCS in porting the COSMO weather model to GPUs.
This system is the latest in a long series of investments in Cray supercomputers by weather forecasting and climate research organizations, which over the past two years have included the United Kingdom’s Met Office; Danish Meteorological Institute; Australian Bureau of Meteorology; Korea Meteorological Administration; South African Weather Service; Finnish Meteorological Institute; and a number of systems for the United States’ National Oceanic and Atmospheric Administration (NOAA), including those supporting the National Weather Service.
What is driving this investment in weather forecasting and climate research?
I will discuss two interdependent areas that underpin these investments:
- The scientific and operational challenges of delivering higher-accuracy forecasts and providing a broader set of forecast products.
- The benefit to society generated by higher quality and more diverse weather and climate forecast products.
Scientific and operational challenges of increasing forecast accuracy
Meteorologists have been striving to increase the accuracy of weather forecasts since the first ones were issued in the mid-19th century by pioneers such as Robert FitzRoy. In the 1950s and 1960s, the development of computers and observational technologies enabled forecast generation using the numerical weather prediction theories of Vilhelm Bjerknes and Lewis Fry Richardson. Today the numerical models that support our weather and climate forecasts are true “grand challenge” problems, requiring on the order of ten thousand billion (1016) mathematical operations and the exchange of one hundred and fifty billion (1.5 x 1014) bytes of information to generate a contemporary global 10-day high resolution forecast. Additionally, forecasts must be completed and issued within strict time windows in order to provide value to end users, and each forecast simulation must be integrated into a multi-stage operational workflow that includes ingestion and pre-processing of observations, assimilation of data into the forecast model, the forecast itself, post-processing and forecast product generation. In addition to compute power, these complex workflows require both efficient and flexible scheduling and high-performance data movement.
Efforts to improve the accuracy of weather forecasts include:
- Developing more accurate “initial state” for the models, by increasing both the number of input sources used to generate the state and the quality of the assimilation methodologies.
- Increasing the resolution of forecast models, allowing smaller physical features and processes to be captured; this commonly also requires the time-step of a model (the amount of time the model is extrapolated forward before the next re-calculation) to be reduced to ensure numerical stability.
- The introduction of new physics and chemistry into the models, and the replacement of simple approximations with more realistic processes.
- The introduction of interactions with other components of the Earth system, such as solar radiation, the land and ocean surface (and sub-surface), atmospheric chemistry and vegetation. Using more complete earth system models becomes important when considering forecasts over multiple days and weeks, and essential when investigating changes in weather and climate over periods of decades and centuries.
- The use of many models with similar initial conditions and physics run in parallel (known as ensemble modelling), which allows for a better understanding of the likelihood of different weather and climate scenarios.
- Running a larger variety of simulations, such as extremely high resolution “city-scale” models, and on-demand models to forecast and track severe weather events such as hurricanes and tropical cyclones.
One common thread with all of these is that they significantly increase the amount of computational effort required to complete a forecast. For example, if the resolution of a forecast is doubled, computational intensity may increase by almost 6x. Because a forecast is only useful if it’s delivered within a fixed time window, a higher-resolution forecast must be run across a larger set of resources, challenging the scalability of forecast models. The data movement requirements (both within the models and between elements of the workflow) also scale with greater computational effort, requiring tightly integrated supercomputer systems and careful planning of the end-to-end workflow to ensure bottlenecks are not introduced.
Benefits of higher quality and more diverse weather and climate forecast products
Improving and expanding weather and climate forecasting activities is a challenging proposition, requiring investment in fundamental scientific research, development of more scalable numerical models, and enhanced supercomputer resources upon which those models can run. But the benefits are clear: In addition to the expansion of scientific knowledge, investment in weather and climate forecasting provides important socio-economic returns. The most obvious benefit of increasing the accuracy of weather forecasting is the protection of human life from severe weather events; for example, successfully forecasting the unusual track of hurricane/tropical storm Sandy from seven days prior to landfall. [A re-analysis of the Sandy event by ECMWF has also shown the impact of differing model resolutions on forecast accuracy, clearly demonstrating (at least for this singular event) the benefits of continuing to improve forecast models.] A huge range of human activities benefits from accurate weather forecasts. Airlines, for example, rely on forecasts to accurately fuel planes, maximize fuel efficiency and ensure safe operations. Energy providers use them to predict demand and optimize energy generation and distribution. Furthermore, climate forecasts can allow governments to plan long-lived infrastructure projects, such as flood defences that are resistant to future weather patterns, reducing the likelihood that expensive re-engineering work will be required within the project’s lifespan. NOAA estimated in their 2011 report, “Value of a Weather Ready Nation,” that the American public gains $31.5 billion in benefits from weather forecasts each year, while the annual cost of generating those forecasts, both public and private, is only $5.1 billion. More recently the Met Office calculated that the U.K. government’s investment of £97 million in a new generation of supercomputers would yield £2 billion in benefits.
Given the clear value of higher-accuracy and more diverse weather and climate forecasts, and the return on investment weather centers around the world are able to demonstrate to their funding agencies and users, I expect the investment in improving and expanding forecasting capabilities to continue over the coming years. I look forward to working with our weather and climate scientists around the world to help them achieve this goal.