def _partition_by_atoms(
self, vertices, n_atoms, max_atoms_per_core, subgraph, graph,
graph_to_subgraph_mapper, resource_tracker):
""" Try to partition subvertices on how many atoms it can fit on\
each subvert
:param vertices: the vertexes that need to be partitioned at the same \
time
:type vertices: iterable list of\
:py:class:`pacman.model.partitionable_graph.abstract_partitionable_vertex.AbstractPartitionableVertex`
:param n_atoms: the atoms of the first vertex
:type n_atoms: int
:param max_atoms_per_core: the max atoms from all the vertexes\
considered that have max_atom constraints
:type max_atoms_per_core: int
:param subgraph: the partitioned_graph of the problem space to put\
subverts in
:type subgraph: :py:class:`pacman.model.subgraph.subgraph.Subgraph`
:param graph: the partitionable_graph object
:type graph:\
:py:class:`pacman.model.partitionable_graph.partitionable_graph.PartitionableGraph`
:param graph_to_subgraph_mapper: the mapper from\
partitionable_graph to partitioned_graph
:type graph_to_subgraph_mapper:\
py:class:'pacman.modelgraph_subgraph_mapper.graph_mapper.GraphMapper'
:param resource_tracker: A tracker of assigned resources
:type resource_tracker:\
:py:class:`pacman.utilities.resource_tracker.ResourceTracker`
:type no_machine_time_steps: int
"""
n_atoms_placed = 0
while n_atoms_placed < n_atoms:
lo_atom = n_atoms_placed
hi_atom = lo_atom + max_atoms_per_core - 1
if hi_atom >= n_atoms:
hi_atom = n_atoms - 1
# Scale down the number of atoms to fit the available resources
used_placements, hi_atom = self._scale_down_resources(
lo_atom, hi_atom, vertices, resource_tracker,
max_atoms_per_core, graph)
# Update where we are
n_atoms_placed = hi_atom + 1
# Create the subvertices
for (vertex, used_resources) in used_placements:
vertex_slice = Slice(lo_atom, hi_atom)
subvertex = vertex.create_subvertex(
vertex_slice, used_resources,
"{}:{}:{}".format(vertex.label, lo_atom, hi_atom),
partition_algorithm_utilities.get_remaining_constraints(
vertex))
# update objects
subgraph.add_subvertex(subvertex)
graph_to_subgraph_mapper.add_subvertex(
subvertex, vertex_slice, vertex)
def _reallocate_resources(
used_placements, resource_tracker, lo_atom, hi_atom, graph):
""" readjusts resource allocation and updates the placement list to\
take into account the new layout of the atoms
:param used_placements: the original list of tuples containing\
placement data
:type used_placements: iterable of tuples
:param resource_tracker: the tracker of resources
:type resource_tracker:\
:py:class:`pacman.utilities.resource_tracker.ResourceTracker`
:param lo_atom: the low atom of a slice to be considered
:type lo_atom: int
:param hi_atom: the high atom of a slice to be considered
:type hi_atom: int
:param graph: the partitionable graph used by the partitioner
:type graph:
:py:class:`pacman.model.partitionable_graph.partitionable_graph.PartitionableGraph`
:return: the new list of tuples containing placement data
:rtype: iterable of tuples
"""
new_used_placements = list()
for (placed_vertex, x, y, p, placed_resources,
ip_tags, reverse_ip_tags) in used_placements:
# Deallocate the existing resources
resource_tracker.unallocate_resources(
x, y, p, placed_resources, ip_tags, reverse_ip_tags)
# Get the new resource usage
vertex_slice = Slice(lo_atom, hi_atom)
new_resources = placed_vertex.get_resources_used_by_atoms(
vertex_slice, graph)
# Re-allocate the existing resources
(x, y, p, ip_tags, reverse_ip_tags) = \
resource_tracker.allocate_constrained_resources(
new_resources, placed_vertex.constraints)
new_used_placements.append(
(placed_vertex, x, y, p, new_resources, ip_tags,
reverse_ip_tags))
return new_used_placements
def test_immutability_as_slice(self):
s = Slice(0, 10)
with self.assertRaises(AttributeError):
s.as_slice = slice(2, 10)
def test_immutability_n_atoms(self):
s = Slice(0, 10)
with self.assertRaises(AttributeError):
s.n_atoms = 3
def test_immutability_hi_atom(self):
s = Slice(0, 10)
with self.assertRaises(AttributeError):
s.hi_atom = 3
def _scale_up_resource_usage(
self, used_resources, hi_atom, lo_atom, max_atoms_per_core, vertex,
resources, ratio, graph):
""" Try to push up the number of atoms in a subvertex to be as close\
to the available resources as possible
:param used_resources: the resources used by the machine so far
:type used_resources:\
:py:class:`pacman.model.resources.resource.Resource`
:param hi_atom: the total number of atoms to place for this vertex
:type hi_atom: int
:param lo_atom: the number of atoms already partitioned
:type lo_atom: int
:param max_atoms_per_core: the min max atoms from all the vertexes \
considered that have max_atom constraints
:type max_atoms_per_core: int
:param vertex: the vertexes to scale up the num atoms per core for
:type vertex:\
:py:class:`pacman.model.partitionable_graph.abstract_partitionable_vertex.AbstractPartitionableVertex`
:param resources: the resource estimate for the vertex for a given\
number of atoms
:type resources:\
:py:class:`pacman.model.resources.resource.Resource`
:param ratio: the ratio between max atoms and available resources
:type ratio: int
:return: the new resources used and the new hi_atom
:rtype: tuple of\
(:py:class:`pacman.model.resources.resource.Resource`,\
int)
"""
previous_used_resources = used_resources
previous_hi_atom = hi_atom
# Keep searching while the ratio is still in range,
# the next hi_atom value is still less than the number of atoms,
# and the number of atoms is less than the constrained number of atoms
while ((ratio < 1.0) and ((hi_atom + 1) < vertex.n_atoms) and
((hi_atom - lo_atom + 2) < max_atoms_per_core)):
# Update the hi_atom, keeping track of the last hi_atom which
# resulted in a ratio < 1.0
previous_hi_atom = hi_atom
hi_atom += 1
# Find the new resource usage, keeping track of the last usage
# which resulted in a ratio < 1.0
previous_used_resources = used_resources
vertex_slice = Slice(lo_atom, hi_atom)
used_resources = vertex.get_resources_used_by_atoms(
vertex_slice, graph)
ratio = self._find_max_ratio(used_resources, resources)
# If we have managed to fit everything exactly (unlikely but possible),
# return the matched resources and high atom count
if ratio == 1.0:
return used_resources, hi_atom
# At this point, the ratio > 1.0, so pick the last allocation of
# resources, which will be < 1.0
return previous_used_resources, previous_hi_atom
def _scale_down_resources(
self, lo_atom, hi_atom, vertices, resource_tracker,
max_atoms_per_core, graph):
""" Reduce the number of atoms on a core so that it fits within the
resources available.
:param lo_atom: the number of atoms already partitioned
:type lo_atom: int
:param hi_atom: the total number of atoms to place for this vertex
:type hi_atom: int
:param vertices: the vertexes that need to be partitioned at the same \
time
:type vertices: iterable of\
:py:class:`pacman.model.partitionable_graph.abstract_partitionable_vertex.AbstractPartitionableVertex`
:param max_atoms_per_core: the min max atoms from all the vertexes \
considered that have max_atom constraints
:type max_atoms_per_core: int
:param graph: the partitionable_graph object
:type graph:\
:py:class:`pacman.model.graph.partitionable_graph.PartitionableGraph`
:param resource_tracker: Tracker of used resources
:type resource_tracker: spinnmachine.machine.Machine object
:return: the list of placements made by this method and the new amount\
of atoms partitioned
:rtype: tuple of (iterable of tuples, int)
:raise PacmanPartitionException: when the vertex cannot be partitioned
"""
used_placements = list()
# Find the number of atoms that will fit in each vertex given the
# resources available
min_hi_atom = hi_atom
for i in range(len(vertices)):
vertex = vertices[i]
# get max resources available on machine
resources = \
resource_tracker.get_maximum_constrained_resources_available(
vertex.constraints)
# get resources used by vertex
vertex_slice = Slice(lo_atom, hi_atom)
used_resources = vertex.get_resources_used_by_atoms(
vertex_slice, graph)
# Work out the ratio of used to available resources
ratio = self._find_max_ratio(used_resources, resources)
while ratio > 1.0 and hi_atom >= lo_atom:
# Scale the resources by the ratio
old_n_atoms = (hi_atom - lo_atom) + 1
new_n_atoms = int(float(old_n_atoms) / (ratio * 1.1))
# Avoid infinite looping
if old_n_atoms == new_n_atoms:
new_n_atoms -= 1
# Find the new resource usage
hi_atom = lo_atom + new_n_atoms - 1
if hi_atom >= lo_atom:
vertex_slice = Slice(lo_atom, hi_atom)
used_resources = vertex.get_resources_used_by_atoms(
vertex_slice, graph)
ratio = self._find_max_ratio(used_resources, resources)
# If we couldn't partition, raise an exception
if hi_atom < lo_atom:
raise exceptions.PacmanPartitionException(
"No more of vertex {} would fit on the board:\n"
" Allocated so far: {} atoms\n"
" Request for SDRAM: {}\n"
" Largest SDRAM space: {}".format(
vertex, lo_atom - 1,
used_resources.sdram.get_value(),
resources.sdram.get_value()))
# Try to scale up until just below the resource usage
used_resources, hi_atom = self._scale_up_resource_usage(
used_resources, hi_atom, lo_atom, max_atoms_per_core, vertex,
resources, ratio, graph)
# If this hi_atom is smaller than the current minimum, update the
# other placements to use (hopefully) less resources
if hi_atom < min_hi_atom:
min_hi_atom = hi_atom
used_placements = self._reallocate_resources(
used_placements, resource_tracker, lo_atom, hi_atom, graph)
# Attempt to allocate the resources for this vertex on the machine
try:
(x, y, p, ip_tags, reverse_ip_tags) = \
resource_tracker.allocate_constrained_resources(
used_resources, vertex.constraints)
used_placements.append(
(vertex, x, y, p, used_resources,
ip_tags, reverse_ip_tags))
except exceptions.PacmanValueError as e:
raise exceptions.PacmanValueError(
"Unable to allocate requested resources to"
" vertex {}:\n{}".format(vertex, e))
# reduce data to what the parent requires
final_placements = list()
for (vertex, _, _, _, used_resources, _, _) in used_placements:
final_placements.append((vertex, used_resources))
return final_placements, min_hi_atom
