Using serial Partitioner on multiple ranks leads to unexpected duplication of FunctionSpace #187
Description
What happened?
When attempting to set up a FunctionSpace whose grid points are all on one particular MPI task, I find the serial partitioner does not act the way I expect it to, and the full grid appears to be created on every rank.
See snippets and outputs below for code specifics.
Is this reflective of user error in setting up the serial partitioner, or is there a bug?
What are the steps to reproduce the bug?
When I run this code on 6 MPI tasks...
const atlas::Grid grid("F8");
eckit::LocalConfiguration conf{};
conf.set("partition", 0);
const atlas::grid::Partitioner part("serial", conf);
const atlas::functionspace::StructuredColumns cols1(grid, part);
std::cout << "on rank = " << eckit::mpi::comm().rank() << ", StructuredColumns 1 size = " << cols1.size() << std::endl;
I get the output...
on rank = 0, StructuredColumns 1 size = 512
on rank = 1, StructuredColumns 1 size = 512
on rank = 2, StructuredColumns 1 size = 512
on rank = 3, StructuredColumns 1 size = 512
on rank = 4, StructuredColumns 1 size = 512
on rank = 5, StructuredColumns 1 size = 512
Whereas with this code...
const atlas::Grid grid("F8");
std::vector<int> zeros(grid.size(), 0);
const atlas::grid::Distribution dist(eckit::mpi::comm().size(), grid.size(), zeros.data());
const atlas::functionspace::StructuredColumns cols2(grid, dist);
std::cout << "on rank = " << eckit::mpi::comm().rank() << ", StructuredColumns 2 size = " << cols2.size() << std::endl;
I get the expected all-on-rank-0 distribution...
on rank = 0, StructuredColumns 2 size = 512
on rank = 1, StructuredColumns 2 size = 0
on rank = 2, StructuredColumns 2 size = 0
on rank = 3, StructuredColumns 2 size = 0
on rank = 4, StructuredColumns 2 size = 0
on rank = 5, StructuredColumns 2 size = 0
Version
0.36
Platform (OS and architecture)
Linux x86_64
Relevant log output
No response
Accompanying data
No response
Organisation
JCSDA