Source code for pymor.models.mpi

# This file is part of the pyMOR project (http://www.pymor.org).
# Copyright 2013-2020 pyMOR developers and contributors. All rights reserved.
# License: BSD 2-Clause License (http://opensource.org/licenses/BSD-2-Clause)

from collections import namedtuple

from pymor.core.base import ImmutableObject
from pymor.models.interface import Model
from pymor.operators.interface import Operator
from pymor.operators.mpi import mpi_wrap_operator
from pymor.tools import mpi
from pymor.vectorarrays.mpi import MPIVectorSpace


[docs]class MPIModel: """Wrapper class mixin for MPI distributed |Models|. Given a single-rank implementation of a |Model|, this wrapper class uses the event loop from :mod:`pymor.tools.mpi` to allow an MPI distributed usage of the |Model|. The underlying implementation needs to be MPI aware. In particular, the model's :meth:`~pymor.models.interface.Model.solve` method has to perform an MPI parallel solve of the model. Note that this class is not intended to be instantiated directly. Instead, you should use :func:`mpi_wrap_model`. """ def __init__(self, obj_id, *args, **kwargs): super().__init__(*args, **kwargs) self.obj_id = obj_id m = mpi.get_object(obj_id) self.parameters = m.parameters self.parameters_own = m.parameters_own self.parameters_internal = m.parameters_internal self.visualizer = MPIVisualizer(obj_id) def _compute_solution(self, mu=None, **kwargs): return self.solution_space.make_array( mpi.call(mpi.method_call_manage, self.obj_id, '_compute_solution', mu=mu, **kwargs) ) def visualize(self, U, **kwargs): self.visualizer.visualize(U, self, **kwargs) def __del__(self): mpi.call(mpi.remove_object, self.obj_id)
[docs]class MPIVisualizer(ImmutableObject): def __init__(self, m_obj_id): self.m_obj_id = m_obj_id def visualize(self, U, m, **kwargs): if isinstance(U, tuple): U = tuple(u.obj_id for u in U) else: U = U.obj_id mpi.call(_MPIVisualizer_visualize, self.m_obj_id, U, **kwargs)
def _MPIVisualizer_visualize(m, U, **kwargs): m = mpi.get_object(m) if isinstance(U, tuple): U = tuple(mpi.get_object(u) for u in U) else: U = mpi.get_object(U) m.visualize(U, **kwargs)
[docs]def mpi_wrap_model(local_models, mpi_spaces=('STATE',), use_with=True, with_apply2=False, pickle_local_spaces=True, space_type=MPIVectorSpace, base_type=None): """Wrap MPI distributed local |Models| to a global |Model| on rank 0. Given MPI distributed local |Models| referred to by the :class:`~pymor.tools.mpi.ObjectId` `local_models`, return a new |Model| which manages these distributed models from rank 0. This is done by first wrapping all |Operators| of the |Model| using :func:`~pymor.operators.mpi.mpi_wrap_operator`. Alternatively, `local_models` can be a callable (with no arguments) which is then called on each rank to instantiate the local |Models|. When `use_with` is `False`, an :class:`MPIModel` is instantiated with the wrapped operators. A call to :meth:`~pymor.models.interface.Model.solve` will then use an MPI parallel call to the :meth:`~pymor.models.interface.Model.solve` methods of the wrapped local |Models| to obtain the solution. This is usually what you want when the actual solve is performed by an implementation in the external solver. When `use_with` is `True`, :meth:`~pymor.core.base.ImmutableObject.with_` is called on the local |Model| on rank 0, to obtain a new |Model| with the wrapped MPI |Operators|. This is mainly useful when the local models are generic |Models| as in :mod:`pymor.models.basic` and :meth:`~pymor.models.interface.Model.solve` is implemented directly in pyMOR via operations on the contained |Operators|. Parameters ---------- local_models :class:`~pymor.tools.mpi.ObjectId` of the local |Models| on each rank or a callable generating the |Models|. mpi_spaces List of types or ids of |VectorSpaces| which are MPI distributed and need to be wrapped. use_with See above. with_apply2 See :class:`~pymor.operators.mpi.MPIOperator`. pickle_local_spaces See :class:`~pymor.operators.mpi.MPIOperator`. space_type See :class:`~pymor.operators.mpi.MPIOperator`. """ assert use_with or isinstance(base_type, Model) if not isinstance(local_models, mpi.ObjectId): local_models = mpi.call(mpi.function_call_manage, local_models) attributes = mpi.call(_mpi_wrap_model_manage_operators, local_models, mpi_spaces, use_with, base_type) wrapped_attributes = { k: _map_children(lambda v: mpi_wrap_operator(*v, with_apply2=with_apply2, pickle_local_spaces=pickle_local_spaces, space_type=space_type) if isinstance(v, _OperatorToWrap) else v, v) for k, v in attributes.items() } if use_with: m = mpi.get_object(local_models) if m.visualizer: wrapped_attributes['visualizer'] = MPIVisualizer(local_models) m = m.with_(**wrapped_attributes) m.disable_caching() return m else: class MPIWrappedModel(MPIModel, base_type): pass return MPIWrappedModel(local_models, **wrapped_attributes)
_OperatorToWrap = namedtuple('_OperatorToWrap', 'operator mpi_range mpi_source') def _mpi_wrap_model_manage_operators(obj_id, mpi_spaces, use_with, base_type): m = mpi.get_object(obj_id) attributes_to_consider = m._init_arguments if use_with else base_type._init_arguments attributes = {k: getattr(m, k) for k in attributes_to_consider} def process_attribute(v): if isinstance(v, Operator): mpi_range = type(v.range) in mpi_spaces or v.range.id in mpi_spaces mpi_source = type(v.source) in mpi_spaces or v.source.id in mpi_spaces if mpi_range or mpi_source: return _OperatorToWrap(mpi.manage_object(v), mpi_range, mpi_source) else: return v else: return v managed_attributes = {k: _map_children(process_attribute, v) for k, v in sorted(attributes.items()) if k not in {'cache_region', 'visualizer'}} if mpi.rank0: return managed_attributes def _map_children(f, obj): if isinstance(obj, dict): return {k: f(v) for k, v in sorted(obj.items())} elif isinstance(obj, (list, tuple, set)) and not isinstance(obj, _OperatorToWrap): return type(obj)(f(v) for v in obj) else: return f(obj)