Simvia at the 32nd TELEMAC User Conference

The 32nd TELEMAC User Conference, organised by CERFACS, takes place from October 14 to 16, 2026 in Toulouse, France. This annual event brings together the community of users and developers of the TELEMAC system, the open-source software suite for modelling free-surface flows, sediment transport and waves.

Simvia will present a paper on the parallel scaling of TELEMAC-2D on the reference Malpasset dam-break case. Although it is a standard validation case, users lack quantitative guidance on two recurring questions: how many MPI processes to allocate for a given mesh, and at which resolution the standard numerical configuration stops being usable. To answer them, the study covers five mesh levels, from 26,000 to 6.66 million elements, run on 4 to 256 MPI processes with the NERD wave-equation scheme.

• Size your production runs: take away a practical rule on how many elements per core to target to use your computing resources efficiently.
• Avoid numerical dead ends: understand why mesh refinement applied on its own can make a computation fail, and how to anticipate it.
• Reproducible results: the entire benchmark relies on published case files and the TELEMAC validation framework.
• Connect with the TELEMAC community: three days of talks and discussions between users, developers and researchers.

Parallel Scaling of the Malpasset Dam-Break Benchmark in TELEMAC-2D

The benchmark protocol

• Five mesh levels built by successive element subdivision, from 26,000 to 6.66 million elements
• 4 to 256 MPI processes on 192-core AMD EPYC nodes (Oracle Cloud Infrastructure)
• Identical physical and numerical parameters across all meshes

Parallel performance

• Wall-clock time, speedup and parallel efficiency for each mesh
• Two distinct regimes: fine meshes scale strongly, coarse meshes saturate quickly
• Example: the 6.66-million-element mesh drops from 4,108 s on 16 cores to 644 s on 256 cores

A sizing rule

• Roughly 13,000 to 26,000 elements per core to maintain at least 70% parallel efficiency
• Mass conservation at round-off level (relative volume error of order 10⁻¹⁵) across all runs

The numerical feasibility limit

• At a fixed time step, each refinement level doubles the Courant number
• Beyond 6.66 million elements, the explicit NERD advection step no longer converges, even with 2,000 iterations, while the linear solver converges normally
• The remedy: reduce the time step in proportion to the refinement