DENVER – The Bureau of Reclamation announced the winners of the Divide and Conquer Challenge. This challenge seeks to significantly improve the execution speed of numerical models that simulate hydraulics, sediment erosion, transport, and deposition in rivers and reservoirs.
"Reclamation is facing an increasing demand to look at large river systems and how the hydrology is changing over time, which involves using computationally intensive simulation models," said Senior Advisor Research and Development Levi Brekke. "The results of this competition show promise that we can accelerate these models and provide results to decision-makers in a quicker, more cost-effective manner."
The winners are:
First Place: $115,000 Christophe Choquet, Vanves, Ile de France, France - The SRH-2D solver has been implemented using a GPU accelerated Parallel Smoothed Aggregation algorithm with the use of previous simulation time steps as the input of the finest multi-grid level. To speed up the overall process, most of the calculations between solver calls were implemented in OpenCL with a Fortran binding. By using a parallel solver, processing intermediate calculations on the GPU and thus reducing CPU to GPU transfers, speedup in the 30x to 60x range was obtained.
Second Place: $85,000 Xiaofeng Liu, State College, Pennsylvania - This solution provides a single unified code which can run in one of three modes: serial, MPI, and OpenMP. User experience of the parallel version of SRH-2D is the same as the original serial version, which enables painless transition. The solution leverages the high-performance linear equation system solvers in Hypre, which is designed for massively parallel computers.
Third place: $45,000 Mahdi Esmaily, Ithaca, NY, USA - This solution is using MPI technology and a graph partitioning algorithm that was developed from scratch in order to ensure that all components of SRH were parallelized. A dedicated library was developed for specialized calls
to the MPI library to simplify the parallelization process while ensuring its backward compatibility.
Fourth Place: $10,000 Zhi Jian Wang, Lawrence, KS, USA - This team used domain decomposition and MPI to parallelize the tool on multiple CPU cores. The linear solver is a preconditioned CGS approach by an incomplete LU decomposition.
These solutions will be used in Reclamation's sedimentation and river hydraulics model to demonstrate the increased execution speed against Reclamation’s baseline model.
Reclamation is partnering with the Federal Highways Administration Office of Innovation Implementation - Resource Center, NASA Tournament Lab, and Freelancer.com on this competition. Please visit https://www.usbr.gov/research/challenges/srh.html to learn more about this competition.
