A new parallel framework of distributed SWAT calibration

With the development of large-scale hydrologic modeling, computational efficiency is becoming more and more important. Rapid modeling and analysis are needed to deal with emergency environmental disasters. The Soil and Water Assessment Tool （SWAT） is a popular hydrologic model, which is less applied in large-scale watershed simulation because of its sequential characteristics. For improving the computational efficiency of the SWAT model, we present a new parallel processing solution for hydrologic cycle and calibration based on MPI （Message Passing Interface）. We partitioned sub-basins during the processes based on a load balancing method Then the calibration was parallelized using a master-slave scheme, in which different input parameters were allocated to different processes to run the hydrologic cycle and compute the function value. Because of the slow convergence and local optimization of the SCE-UA （Shuffled Complex Evolution-developed by University of Arizona） algorithm in SWAT calibration, a genetic algorithm （GA） is developed to optimize the calibration step. Then by dividing the default communicator into several sub-communicators, all the hydrologic cycles were par- allelized in their own sub-communicators to achieve further acceleration. In this paper the results show speedups for the hydrologic cycle calculations, as well as in the optimized calibration step. In the case study, we tested the parallel hydrologic cycle by four processes, and got a speedup of 3.06. In the calibration section, after applying the GA optimization, with 10 cores, we got a speed increase of 8.0 in our GA parallel framework compared with the GA sequential calibration, which is much better than the original SWAT calibration. After the sub-communicators added, this process was speeded up even further. The study demonstrated that the GA parallel framework with multi-sub-communicators is an effective and efficient solution for the hydrologists in large scale hydrology simulations.