Skip to content
Open
Changes from 1 commit
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 1 addition & 5 deletions content/docs/configuration/basics/configuration-mapping.md
Original file line number Diff line number Diff line change
Expand Up @@ -143,7 +143,7 @@ To use this feature, build preCICE from source with [Kokkos-Kernels or Ginkgo en
| Mapping tag | Dependencies | `cuda` | `hip` | `sycl` | `openmp` | `cpu` | MPI-parallelization |
|-------------------------|--------------------------------------------------------------------------------------------------------------------|:------:|:-----:|:------:|:--------:|:-----:|---------------------|
| `rbf-pum-direct` | (Eigen, [mandatory](https://precice.org/installation-source-dependencies.html#required-dependencies) for preCICE) | — | — | — | — | ✓ | distributed |
| `rbf-pum-direct` | Kokkos-Kernels and Kokkos (preCICE >= v3.4.1) | ✓ | ✓ | ✓ | ✓ | — | distributed |
| `rbf-pum-direct` | Kokkos-Kernels and Kokkos (preCICE >= v3.5.0) | ✓ | ✓ | ✓ | ✓ | — | distributed |
| `rbf-global-direct` | Ginkgo and Kokkos (preCICE >= v3.2) | ✓ | ✓ | — | — | — | gather-scatter |
| `rbf-global-direct` | (Eigen, [mandatory](https://precice.org/installation-source-dependencies.html#required-dependencies) for preCICE) | — | — | — | — | ✓ | gather-scatter |
| `rbf-global-iterative` | Ginkgo and Kokkos (preCICE >= v3.2) | ✓ | ✓ | — | ✓ | — | gather-scatter |
Expand Down Expand Up @@ -171,10 +171,6 @@ Mapping configurations that follow a gather-scatter approach are always computed

By contrast, distributed mappings solve their rank-local problem in parallel. In practice, this means that the configuration snippet above assigns multiple MPI ranks to a single GPU (with device ID 0). Since oversubscribing a device is typically undesired, use `gpu-device-id="auto"` to assign MPI ranks in a round-robin fashion to the available GPUs. Also note -- due to the different parallelization strategies -- `n-threads` configures the number of threads per executing rank, i.e., for a distributed parallelization, `n-threads=10` assigns 10 OpenMP threads to each MPI rank, whereas for a gather-scatter parallelization `n-threads=10` assigns 10 OpenMP threads to the primary rank only.

{% note %}
Note that as of preCICE v3.4 the `rbf-pum-direct` executor only supports `consistent` constraints. Supporting a `conservative` is [work in progress](https://github.com/precice/precice/pull/2536).
{% endnote %}

More details on the feature can be found in [Schneider et al. 2023](https://doi.org/10.23967/c.coupled.2023.016).

## Geometric multiscale mapping
Expand Down
Loading