Approaches to distributed execution of hydrologic models: methods for ensemble Monte Carlo risk modelling with and without workflows

Abstract

Recent water reforms in Australia and the release of the Murray-Darling Basin Plan have been supported by climate models and detailed hydrological modelling including river system models. Sensitivity analysis of these river system models provides valuable insights into the often complex and non-linear relationships between uncertainty in input variables and parameters and model outputs. An understanding of these relationships is an important component of assessing the risks in the planning process. However, a comprehensive sensitivity analysis is computationally intensive, requiring many thousands of simulations to examine a few parameters and may require months of computer time to complete. In this paper we consider a sensitivity analysis of a river system model using new and emergent technologies and discuss the merits of four methodologies for undertaking this analysis. In each case some new tools and techniques have been developed and these are applicable to sensitivity, uncertainty and error analysis of other simulation models. The Murray-Darling basin is represented by a range of regional river models that are connected together to describe the entire basin. CSIRO recently calibrated regional Source models that, when combined, describe all of the Murray-Darling Basin. The Murrumbidgee regional model was selected from this project and subsequently simplified to reduce the runtime while still being representative of the system's behaviour. As part of a risk assessment, the sensitivity of this model was explored. The sensitivity analysis examined uncertainty in inflows, rainfall, evaporation and groundwater/surface water interaction, via 100,000 simulations and the results can be found in Peeters et. al. (2013), submitted to this conference. The four methodologies considered to support this work are: 1. Running all 100,000 simulations on a single computer; 2. Running the simulations using several dedicated machines; 3. Running the simulations using ad-hoc computing resources; and 4. Multi-core execution, where runs are executed on a cluster. Method 1 was the simplest, but requires the most computer time. Method 2 improved total runtime, but required dedicated computer resources. Method 3 gave reasonable runtimes, did not require dedicated resources, but did require constant monitoring and input. Method 4 was the most complex to configure, but provided very fast runtimes and automated input and output marshalling and cluster job creation and submission. Methods 1 through 3 used Source's command line interface, while for method 4 the Source model was imported as a workflow activity into Project Trident via 'the Hydrologists Workbench'. Project Trident is a scientific workflow system developed by Microsoft Research and the Hydrologists Workbench is a suite of add-on tools for Trident developed by CSIRO's Water for a Healthy Country Flagship. Using Trident and the Hydrologists Workbench for sensitivity analysis allows the modeler to easily leverage available resources without requiring extensive or complex coding.