def _distribute_try(computation_graph: ComputationGraph,
agents: Iterable[AgentDef],
hints: DistributionHints = None,
computation_memory=None,
communication_load=None,
attempt=0):
agents_capa = {a.name: a.capacity for a in agents}
# The distribution methods depends on the order used to process the node,
# we shuffle them to test a new configuration when retry a distribution
# after a failure
nodes = list(computation_graph.nodes)
shuffle(nodes)
mapping = defaultdict(set)
var_hosted = {}
# Distribute owned computation variable on the corresponding agent.
# For dcop build from an secp, this is the same thing as deploying the
# light variable on the light devices, as we were doing before.
for a in agents_capa:
for c in hints.must_host(a):
mapping[a].add(c)
var_hosted.update({c: a})
agents_capa[a] -= computation_memory(
computation_graph.computation(c))
# First mimic original secp adhoc behavior
for n in nodes:
if n.name in var_hosted:
continue
hostwith = hints.host_with(n.name)
# secp models have a constraint that should be hosted on the same
# agent than the variable of the model
if len(hostwith) == 1 and n.type == 'FactorComputation' and \
computation_graph.computation(hostwith[0]).type \
== 'VariableComputation':
dependent_var = [v.name for v in n.factor.dimensions]
candidates = [
a for a in agents_capa
if len(set(mapping[a]).intersection(dependent_var)) > 0
]
candidates.sort(key=lambda x: len(mapping[a]))
if candidates:
selected = candidates[0]
else:
selected = choice(list(agents_capa.keys()))
mapping[selected].update({n.name, hostwith[0]})
var_hosted[n.name] = selected
var_hosted[hostwith[0]] = selected
agents_capa[selected] -= computation_memory(n)
for n in nodes:
if n.name in var_hosted:
continue
footprint = computation_memory(n)
# Candidates : hints only with enough capacity
candidates = [(agents_capa[a], a) for a in hints.host_with(n.name)
if agents_capa[a] > footprint]
# If no hinted agents has enough capacity, fall back to all agents
if not candidates:
candidates = [(c, a) for a, c in agents_capa.items()
if c > footprint]
# Select the candidate that is already hosting the highest
# number of computations sharing a link with this one.
scores = []
for capacity, a in candidates:
count = 0
for l in computation_graph.links_for_node(n.name):
count += len([None for l_n in l.nodes if l_n in mapping[a]])
# The tuple is in this order so that we sort by score first,
# and then by available capacity.
scores.append((count, capacity, a))
scores.sort(reverse=True)
if scores:
selected = scores[0][2]
agents_capa[selected] -= footprint
else:
# Retry 3 times in case of failure, the nodes will be shuffled
# every time, increasing the probability to find a feasible
# distribution.
if attempt > 2:
raise ImpossibleDistributionException(
'Could not find feasible distribution after {} '
'attempts'.format(attempt))
else:
_distribute_try(computation_graph, agents, hints,
computation_memory, computation_graph,
attempt + 1)
mapping[selected].update({n.name})
var_hosted[n.name] = selected
return Distribution({a: list(mapping[a]) for a in mapping})
def test_must_host(self):
dh = DistributionHints(must_host={'a1': ['v1']})
self.assertIn('v1', dh.must_host('a1'))
def test_must_host_return_empty_when_not_specified(self):
dh = DistributionHints(must_host={'a1': ['v1']})
self.assertEqual(len(dh.must_host('a2')), 0)
def factor_graph_lp_model(cg: ComputationsFactorGraph,
agents: List[AgentDef],
hints: DistributionHints=None,
computation_memory=None,
communication_load=None):
"""
To distribute we need:
* com : the communication cost of an edge between a var and a fact
* mem_var : the memory footprint of a variable computation
* mem_fac : the memory footprint of a factor computation
These function depends on the algorithm.
Here
* mem_var and mem_fac are given by the computation_memory method.
* com is given by computation_memory
:return:
"""
variables = [n for n in cg.nodes if n.type == 'VariableComputation']
factors = [n for n in cg.nodes if n.type == 'FactorComputation']
agents = list(agents)
agents_names = [a.name for a in agents]
fixed_dist = Distribution({a.name: hints.must_host(a.name)
for a in agents})
# Only keep computations for which we actually need to find an agent.
vars_to_host = [v.name for v in variables
if not fixed_dist.has_computation(v.name)]
facs_to_host = [f.name for f in factors
if not fixed_dist.has_computation(f.name)]
# x_i^k : binary variable indicating if var x_i is hosted on agent a_k.
xs = _build_xs_binvar(vars_to_host, agents_names)
# f_j^k : binary variable indicating if factor f_j is hosted on agent a_k.
fs = _build_fs_binvar(facs_to_host, agents_names)
# alpha_ijk : binary variable indicating if x_i and f_j are both on a_k.
alphas = _build_alphaijk_binvars(cg, agents_names)
# LP problem with objective function (total communication cost).
pb = LpProblem('distribution', LpMinimize)
pb += _objective_function(cg, communication_load, alphas,
agents_names), 'Communication costs'
# Constraints.
# All variable computations must be hosted:
for i in vars_to_host:
pb += lpSum([xs[(i, k)] for k in agents_names]) == 1, \
'var {} is hosted'.format(i)
# All factor computations must be hosted:
for j in facs_to_host:
pb += lpSum([fs[(j, k)] for k in agents_names]) == 1, \
'factor {} is hosted'.format(j)
# Each agent must host at least one computation:
# We only need this constraints for agents that do not already host a
# computation:
empty_agents = [a for a in agents_names if not hints.must_host(a)]
for k in empty_agents:
pb += lpSum([xs[(i, k)] for i in vars_to_host]) + \
lpSum([fs[(j, k)] for j in facs_to_host]) >= 1, \
'atleastone {}'.format(k)
# Memory capacity constraint for agents
for a in agents:
# Decrease capacity for already hosted computations
capacity = a.capacity - \
sum([_computation_memory_in_cg(c, cg, computation_memory)
for c in hints.must_host(a.name)])
pb += lpSum([_computation_memory_in_cg(i, cg, computation_memory) *
xs[(i, a.name)] for i in vars_to_host]) \
+ lpSum([_computation_memory_in_cg(j, cg, computation_memory) *
fs[(j, a.name)] for j in facs_to_host]) <= capacity, \
'memory {}'.format(a.name)
# Linearization constraints for alpha_ijk.
for link in cg.links:
i, j = link.variable_node, link.factor_node
for k in agents_names:
if i in vars_to_host and j in facs_to_host:
pb += alphas[((i, j), k)] <= xs[(i, k)], \
'lin1 {}{}{}'.format(i, j, k)
pb += alphas[((i, j), k)] <= fs[(j, k)], \
'lin2 {}{}{}'.format(i, j, k)
pb += alphas[((i, j), k)] >= xs[(i, k)] + fs[(j, k)] - 1, \
'lin3 {}{}{}'.format(i, j, k)
elif i in vars_to_host and j not in facs_to_host:
# Var is free, factor is already hosted
if fixed_dist.agent_for(j) == k:
pb += alphas[((i, j), k)] == xs[(i, k)]
else:
pb += alphas[((i, j), k)] == 0
elif i not in vars_to_host and j in facs_to_host:
# if i is not in vars_vars_to_host, it means that it's a
# computation that is already hosted (from hints)
if fixed_dist.agent_for(i) == k:
pb += alphas[((i, j), k)] == fs[(j, k)]
else:
pb += alphas[((i, j), k)] == 0
else:
# i and j are both alredy hosted
if fixed_dist.agent_for(i) == k and fixed_dist.agent_for(j) \
== k:
pb += alphas[((i, j), k)] == 1
else:
pb += alphas[((i, j), k)] == 0
# Now solve our LP
# status = pb.solve(GLPK_CMD())
# status = pb.solve(GLPK_CMD(mip=1))
# status = pb.solve(GLPK_CMD(mip=0, keepFiles=1,
# options=['--simplex', '--interior']))
status = pb.solve(GLPK_CMD(keepFiles=0, msg=False, options=['--pcost']))
if status != LpStatusOptimal:
raise ImpossibleDistributionException("No possible optimal"
" distribution ")
else:
logger.debug('GLPK cost : %s', value(pb.objective))
comp_dist = fixed_dist
for k in agents_names:
agt_vars = [i for i, ka in xs
if ka == k and value(xs[(i, ka)]) == 1]
comp_dist.host_on_agent(k, agt_vars)
agt_rels = [j for j, ka in fs
if ka == k and value(fs[(j, ka)]) == 1]
comp_dist.host_on_agent(k, agt_rels)
return comp_dist
