Skip to content

Draft: Fix distributed data rendering with empty nodes

  • Surface representation uses PVGeometryFilter, that does MPI communication (AllReduce). As stated in #19963 (closed), some nodes may be not executed, leading to MPI lock.

    Now we create a subcontroller for PVGeometryFilter in surface representation, so only relevant nodes will communicate.

    A first fix was introduced in !4283 (merged) (and 6c156732). But it relies on updating the geometry filter MTime. We do not want that anymore to be able to do some optimization when filter does not change between execution (see DataObjectMeshCache)

Merge request reports