void Partitioner::partition_unpartitioned_elements (MeshBase & mesh,
const unsigned int n_subdomains)
{
MeshBase::element_iterator it = mesh.unpartitioned_elements_begin();
const MeshBase::element_iterator end = mesh.unpartitioned_elements_end();
const dof_id_type n_unpartitioned_elements = MeshTools::n_elem (it, end);
// the unpartitioned elements must exist on all processors. If the range is empty on one
// it is empty on all, and we can quit right here.
if (!n_unpartitioned_elements) return;
// find the target subdomain sizes
std::vector<dof_id_type> subdomain_bounds(mesh.n_processors());
for (processor_id_type pid=0; pid<mesh.n_processors(); pid++)
{
dof_id_type tgt_subdomain_size = 0;
// watch out for the case that n_subdomains < n_processors
if (pid < n_subdomains)
{
tgt_subdomain_size = n_unpartitioned_elements/n_subdomains;
if (pid < n_unpartitioned_elements%n_subdomains)
tgt_subdomain_size++;
}
//libMesh::out << "pid, #= " << pid << ", " << tgt_subdomain_size << std::endl;
if (pid == 0)
subdomain_bounds[0] = tgt_subdomain_size;
else
subdomain_bounds[pid] = subdomain_bounds[pid-1] + tgt_subdomain_size;
}
libmesh_assert_equal_to (subdomain_bounds.back(), n_unpartitioned_elements);
// create the unique mapping for all unpartitioned elements independent of partitioning
// determine the global indexing for all the unpartitoned elements
std::vector<dof_id_type> global_indices;
// Calling this on all processors a unique range in [0,n_unpartitioned_elements) is constructed.
// Only the indices for the elements we pass in are returned in the array.
MeshCommunication().find_global_indices (mesh.comm(),
MeshTools::bounding_box(mesh), it, end,
global_indices);
for (dof_id_type cnt=0; it != end; ++it)
{
Elem * elem = *it;
libmesh_assert_less (cnt, global_indices.size());
const dof_id_type global_index =
global_indices[cnt++];
libmesh_assert_less (global_index, subdomain_bounds.back());
libmesh_assert_less (global_index, n_unpartitioned_elements);
const processor_id_type subdomain_id =
cast_int<processor_id_type>
(std::distance(subdomain_bounds.begin(),
std::upper_bound(subdomain_bounds.begin(),
subdomain_bounds.end(),
global_index)));
libmesh_assert_less (subdomain_id, n_subdomains);
elem->processor_id() = subdomain_id;
//libMesh::out << "assigning " << global_index << " to " << subdomain_id << std::endl;
}
}
void ParmetisPartitioner::initialize (const MeshBase & mesh,
const unsigned int n_sbdmns)
{
const dof_id_type n_active_local_elem = mesh.n_active_local_elem();
// Set parameters.
_pmetis->wgtflag = 2; // weights on vertices only
_pmetis->ncon = 1; // one weight per vertex
_pmetis->numflag = 0; // C-style 0-based numbering
_pmetis->nparts = static_cast<Parmetis::idx_t>(n_sbdmns); // number of subdomains to create
_pmetis->edgecut = 0; // the numbers of edges cut by the
// partition
// Initialize data structures for ParMETIS
_pmetis->vtxdist.resize (mesh.n_processors()+1); std::fill (_pmetis->vtxdist.begin(), _pmetis->vtxdist.end(), 0);
_pmetis->tpwgts.resize (_pmetis->nparts); std::fill (_pmetis->tpwgts.begin(), _pmetis->tpwgts.end(), 1./_pmetis->nparts);
_pmetis->ubvec.resize (_pmetis->ncon); std::fill (_pmetis->ubvec.begin(), _pmetis->ubvec.end(), 1.05);
_pmetis->part.resize (n_active_local_elem); std::fill (_pmetis->part.begin(), _pmetis->part.end(), 0);
_pmetis->options.resize (5);
_pmetis->vwgt.resize (n_active_local_elem);
// Set the options
_pmetis->options[0] = 1; // don't use default options
_pmetis->options[1] = 0; // default (level of timing)
_pmetis->options[2] = 15; // random seed (default)
_pmetis->options[3] = 2; // processor distribution and subdomain distribution are decoupled
// Find the number of active elements on each processor. We cannot use
// mesh.n_active_elem_on_proc(pid) since that only returns the number of
// elements assigned to pid which are currently stored on the calling
// processor. This will not in general be correct for parallel meshes
// when (pid!=mesh.processor_id()).
_n_active_elem_on_proc.resize(mesh.n_processors());
mesh.comm().allgather(n_active_local_elem, _n_active_elem_on_proc);
// count the total number of active elements in the mesh. Note we cannot
// use mesh.n_active_elem() in general since this only returns the number
// of active elements which are stored on the calling processor.
// We should not use n_active_elem for any allocation because that will
// be inheritly unscalable, but it can be useful for libmesh_assertions.
dof_id_type n_active_elem=0;
// Set up the vtxdist array. This will be the same on each processor.
// ***** Consult the Parmetis documentation. *****
libmesh_assert_equal_to (_pmetis->vtxdist.size(),
cast_int<std::size_t>(mesh.n_processors()+1));
libmesh_assert_equal_to (_pmetis->vtxdist[0], 0);
for (processor_id_type pid=0; pid<mesh.n_processors(); pid++)
{
_pmetis->vtxdist[pid+1] = _pmetis->vtxdist[pid] + _n_active_elem_on_proc[pid];
n_active_elem += _n_active_elem_on_proc[pid];
}
libmesh_assert_equal_to (_pmetis->vtxdist.back(), static_cast<Parmetis::idx_t>(n_active_elem));
// ParMetis expects the elements to be numbered in contiguous blocks
// by processor, i.e. [0, ne0), [ne0, ne0+ne1), ...
// Since we only partition active elements we should have no expectation
// that we currently have such a distribution. So we need to create it.
// Also, at the same time we are going to map all the active elements into a globally
// unique range [0,n_active_elem) which is *independent* of the current partitioning.
// This can be fed to ParMetis as the initial partitioning of the subdomains (decoupled
// from the partitioning of the objects themselves). This allows us to get the same
// resultant partitioning independed of the input partitioning.
MeshTools::BoundingBox bbox =
MeshTools::bounding_box(mesh);
_global_index_by_pid_map.clear();
// Maps active element ids into a contiguous range independent of partitioning.
// (only needs local scope)
vectormap<dof_id_type, dof_id_type> global_index_map;
{
std::vector<dof_id_type> global_index;
// create the mapping which is contiguous by processor
dof_id_type pid_offset=0;
for (processor_id_type pid=0; pid<mesh.n_processors(); pid++)
{
MeshBase::const_element_iterator it = mesh.active_pid_elements_begin(pid);
const MeshBase::const_element_iterator end = mesh.active_pid_elements_end(pid);
// note that we may not have all (or any!) the active elements which belong on this processor,
// but by calling this on all processors a unique range in [0,_n_active_elem_on_proc[pid])
// is constructed. Only the indices for the elements we pass in are returned in the array.
MeshCommunication().find_global_indices (mesh.comm(),
bbox, it, end,
global_index);
for (dof_id_type cnt=0; it != end; ++it)
{
const Elem * elem = *it;
libmesh_assert (!_global_index_by_pid_map.count(elem->id()));
libmesh_assert_less (cnt, global_index.size());
libmesh_assert_less (global_index[cnt], _n_active_elem_on_proc[pid]);
_global_index_by_pid_map.insert(std::make_pair(elem->id(), global_index[cnt++] + pid_offset));
}
pid_offset += _n_active_elem_on_proc[pid];
}
// create the unique mapping for all active elements independent of partitioning
{
MeshBase::const_element_iterator it = mesh.active_elements_begin();
const MeshBase::const_element_iterator end = mesh.active_elements_end();
// Calling this on all processors a unique range in [0,n_active_elem) is constructed.
// Only the indices for the elements we pass in are returned in the array.
MeshCommunication().find_global_indices (mesh.comm(),
bbox, it, end,
global_index);
for (dof_id_type cnt=0; it != end; ++it)
{
const Elem * elem = *it;
libmesh_assert (!global_index_map.count(elem->id()));
libmesh_assert_less (cnt, global_index.size());
libmesh_assert_less (global_index[cnt], n_active_elem);
global_index_map.insert(std::make_pair(elem->id(), global_index[cnt++]));
}
}
// really, shouldn't be close!
libmesh_assert_less_equal (global_index_map.size(), n_active_elem);
libmesh_assert_less_equal (_global_index_by_pid_map.size(), n_active_elem);
// At this point the two maps should be the same size. If they are not
// then the number of active elements is not the same as the sum over all
// processors of the number of active elements per processor, which means
// there must be some unpartitioned objects out there.
if (global_index_map.size() != _global_index_by_pid_map.size())
libmesh_error_msg("ERROR: ParmetisPartitioner cannot handle unpartitioned objects!");
}
// Finally, we need to initialize the vertex (partition) weights and the initial subdomain
// mapping. The subdomain mapping will be independent of the processor mapping, and is
// defined by a simple mapping of the global indices we just found.
{
std::vector<dof_id_type> subdomain_bounds(mesh.n_processors());
const dof_id_type first_local_elem = _pmetis->vtxdist[mesh.processor_id()];
for (processor_id_type pid=0; pid<mesh.n_processors(); pid++)
{
dof_id_type tgt_subdomain_size = 0;
// watch out for the case that n_subdomains < n_processors
if (pid < static_cast<unsigned int>(_pmetis->nparts))
{
tgt_subdomain_size = n_active_elem/std::min
(cast_int<Parmetis::idx_t>(mesh.n_processors()), _pmetis->nparts);
if (pid < n_active_elem%_pmetis->nparts)
tgt_subdomain_size++;
}
if (pid == 0)
subdomain_bounds[0] = tgt_subdomain_size;
else
subdomain_bounds[pid] = subdomain_bounds[pid-1] + tgt_subdomain_size;
}
libmesh_assert_equal_to (subdomain_bounds.back(), n_active_elem);
MeshBase::const_element_iterator elem_it = mesh.active_local_elements_begin();
const MeshBase::const_element_iterator elem_end = mesh.active_local_elements_end();
for (; elem_it != elem_end; ++elem_it)
{
const Elem * elem = *elem_it;
libmesh_assert (_global_index_by_pid_map.count(elem->id()));
const dof_id_type global_index_by_pid =
_global_index_by_pid_map[elem->id()];
libmesh_assert_less (global_index_by_pid, n_active_elem);
const dof_id_type local_index =
global_index_by_pid - first_local_elem;
libmesh_assert_less (local_index, n_active_local_elem);
libmesh_assert_less (local_index, _pmetis->vwgt.size());
// TODO:[BSK] maybe there is a better weight?
_pmetis->vwgt[local_index] = elem->n_nodes();
// find the subdomain this element belongs in
libmesh_assert (global_index_map.count(elem->id()));
const dof_id_type global_index =
global_index_map[elem->id()];
libmesh_assert_less (global_index, subdomain_bounds.back());
const unsigned int subdomain_id =
std::distance(subdomain_bounds.begin(),
std::lower_bound(subdomain_bounds.begin(),
subdomain_bounds.end(),
global_index));
libmesh_assert_less (subdomain_id, static_cast<unsigned int>(_pmetis->nparts));
libmesh_assert_less (local_index, _pmetis->part.size());
_pmetis->part[local_index] = subdomain_id;
}
}
}