T-HySEA (Tsunami-HySEA) | ChEESE - Centre of Excellence for Exascale Supercomputing in the area of the Solid Earth

Code name	T-HySEA (Tsunami-HySEA)
Developer(s)	Manuel J. Castro Díaz Jorge Macías Sánchez Marc de la Asunción Contact person: Jorge Macías Sánchez
Link	https://edanya.uma.es/hysea/
Short description	T-HySEA solves the 2D shallow water equations on hydrostatic and dispersive versions. T-HySEA is based on a high-order Finite Volume (FV) discretization (hydrostatic) with Finite Differences (FD) for the dispersive version on two-way structured nested meshes in spherical coordinates. Initial conditions from the Okada model or initial deformation, synchronous and asynchronous multi-Okada, rectangular and triangular faults.
Original code level	3
Current code level	8-9
Pilot(s) involved	PD2, PD7,PD8
Main results and References	MAIN RESULTS: Code Audit driven improvements. Better load balancing. Resuming a stored simulation. Asynchronous NetCDF file writing. Asynchronous CPU-GPU memory transfers. Direct GPU to GPU data transfers using GPUDirect in architectures that support it. Added compression of the NetCDF output files. Support for bigger meshes by using int64_t datatype. Developed a Monte-Carlo version capable of performing many simulations in a single execution, one on each GPU. REFERENCES: Macías, J., Castro, M.J., Ortega, S., Escalante, C., González-Vida, J.M. (2017). Performance benchmarking of Tsunami-HySEA model for NTHMP's inundation mapping activities. Pure and Applied Geophysics , 1-37. [doi: 10.1007/s00024-017-1583-1] Macías, J., Castro, M.J., Escalante, C. (2020). Performance assessment of Tsunami-HySEA model for NTHMP tsunami currents benchmarking. Laboratory data. Coastal Engineering, 158, 103667, ISSN 0378-3839, [doi: 10.1016/j.coastaleng.2020.103667]. Macías, J., Castro, M.J. Ortega, S., and González-Vida, J.M., (2020). Performance assessment of Tsunami-HySEA model for NTHMP tsunami currents benchmarking. Field cases. Ocean Modeling, 152, 101645, ISSN 1463-5003, [doi: 10.1016/j.ocemod.2020.101645].
Performance results	Strong scaling reaches up to 87 % efficiency using 32 GPUs with direct GPU to GPU data transfers. Use of direct GPU to GPU data transfers improves runtimes and scaling up to a 20 %. Asynchronous file writing reduces the runtimes up to a 50 % for high frequency saving.