Optimizing Geospatial Cyberinfrastructure to Improve the Computing Capability for Climate Studies

Climate simulation has significant uncertainties due to our current limited understanding of the processes and interactions between different components of the Earth. Model sensitivity analysis, which tests the sensitivity of model output to the input parameter values, is a standard practice for determining the model uncertainties and improving model accuracy. A common approach for climate model sensitivity analysis is to run a model many times by sweeping a large number of adjustable parameters. However, this approach is hampered by three computational challenges: computing intensity, data intensity, and procedure complexity. This dissertation proposes three optimization methodologies to address these challenges respectively, including 1) tackling the computing intensity challenge posed by climate simulation using Model as a Service, a new service model in the context of cloud computing; 2) managing and processing the big model output---"data intensity"---using a scalable big spatiotemporal data analytics framework; 3) solving the procedure complexity issue using a service-oriented cloud-based scientific workflow framework.;The feasibility and efficiency of these approaches is demonstrated by a case study---ModelE sensitivity analysis. Experiment result shows that 1) Model as a Service reduced the time spent on 300 model-runs (with forty machines) from 38 days to 5 days; 2) the data analytic framework archived 6 times speedup when processing 100 model-run outputs with 6 machines; and 3) the workflow framework transformed the complex analysis by dragging and connecting steps in a visually intuitive diagram. By integrating the three optimization methodologies seamlessly, the time spent on ModelE sensitivity analysis (including setting up model, running model and analyzing model output) is reduced from approximate two months to five days.;This research offers a computational solution to efficiently sweep a large number of adjustable climate model input parameters for identifying model uncertainty. It helps scientists to find answers in a more efficient way to the questions like "which model parameter is more sensitive to simulated climate changes?". This research also provides a valuable guideline on bridging the computing infrastructure and computing requirements for climate studies. Since the challenges of computing intensity, data intensity and procedure complexity are quite common in geosciences, the proposed optimization methodologies can be broaden from the climate science to broader geoscience domains. In addition, this research provides a potential solution to the uncertainty quantification (UQ) problems for general geoscience applications.