Skip to content. | Skip to navigation

EUROfusion

Abstract of RFM3OPT Project

Finite-difference time-domain (FDTD) methods are one of the most popular numerical techniques used to simulate reflectometry. They offer a comprehensive description of the plasma phenomena. However, this method requires a fine spatial grid discretization to keep the error to a minimum, which in turns implies a high-resolution time discretization to comply with CFL numerical stability condition. Simulations in three-dimensions (3D) are therefore very demanding computationally, both in terms of floating point operations, as well as memory resources. They become only possible if the problem can be efficiently distributed over large numbers of resources. The REFMUL3 code was developed during 2016 within a collaboration between the Instituto de Plasmas e Fusão Nuclear (IPFN-IST) in Lisbon and the HLST, precisely to meet this goal [1]. It is a 3D full-wave code using a FDTD Yee scheme [2] with full polarisation that copes simultaneously with o- and x-modes, supports a general external magnetic field and dynamic plasma. REFMUL3 employs a hybrid MPI/OpenMP parallelisation using explicit 3D domain decomposition, which yields very good scaling properties up to a few thousands of cores. This has opened the door to the simulation of much larger domains (grid-counts), which, as expected, exposed new challenges regarding data input/output (I/O) operations. This challenge has been addressed within HLST-REFMULIO project (2018), by adding parallel I/O capabilities to the code using the HDF5 library [3]. Further enhancing the parallel I/O capabilities of REFMUL3 was the objective of REFMUL3+ project (2019), which introduced a check-point/restart file infrastructure, as well as a generalisation of the I/O infrastructure using the Portable Data Interface (PDI) library [4]. Finally, some effort was also dedicated to the performance tuning of the existing parallel HDF5 I/O infrastructure. The project RFM3OPT aims towards a direct continuation of the latter, profiting from the knowledge and experience acquired. Additionally, the inclusion of the native VTK file formats for the output of REFMUL3 is also proposed, as is, if time allows, complementary performance optimizations related to the parallel infrastructure of the code, related to the overlap of MPI communication with computations and the parallelisation of the impulse calibration response.