def test_load_balancing():
check_optimality(60, np.array([0.25, 0.25, 0.25, 0.25]), [15, 15, 15, 15])
check_optimality(1024, np.array([0.3, 0.5, 0.9]), [0, 255, 769])
check_optimality(60, np.array([0.44, 0.33, 0.22]), [42, 18, 0])
check_optimality(60, np.array([0.55, 0.44, 0.40]), [35, 16, 9])
check_optimality(1024 * 1024, np.array([0.3, 0.5, 0.9, 0.6]), [0, 169327, 602629, 276620])
check_optimality(1024 * 1024, np.array([0.0, 0.5, 0.0, 0.6]), [0, 428963, 0, 619613])
assert load_balance_peers(60, np.array([0.55, 0.44, 0.40]), min_size=10) == (41, 19, 0)
assert load_balance_peers(60, np.array([0.32, 0.55, 0.44]), min_size=10) == (0, 40, 20)
assert load_balance_peers(2, np.array([0.55, 0.20, 0.44]), min_size=10) == (1, 0, 1)
assert load_balance_peers(1, np.array([0.55, 0.20, 0.44]), min_size=10) == (1, 0, 0)
assert load_balance_peers(100, (None, None)) == (50, 50)
assert load_balance_peers(100, (None, None, None, None, None)) == (20, 20, 20, 20, 20)
assert load_balance_peers(100, (0, 0, 0, None, None)) == (0, 0, 0, 50, 50)
with pytest.raises(AssertionError):
load_balance_peers(100, (0, 0, 0))
for i in range(10):
vector_size = np.random.randint(1, 1024 ** 3)
num_peers = np.random.randint(1, 256)
scale = 1e-9 + np.random.rand() * 1e5
throughputs = np.random.rand(num_peers) * scale + 1e-6
min_size = np.random.choice([0, np.random.randint(0, vector_size // 10)])
assignment = load_balance_peers(vector_size, throughputs, min_size)
assert np.sum(assignment) == vector_size
assert np.min(assignment) >= 0
def test_partitioning():
for _ in range(100):
tensors = []
for _ in range(random.randint(1, 5)):
ndim = random.randint(0, 4)
shape = torch.Size([random.randint(0, 16) for _ in range(ndim)])
make_tensor = random.choice(
[torch.rand, torch.randn, torch.zeros, torch.ones])
tensors.append(make_tensor(shape))
total_size = sum(map(torch.Tensor.numel, tensors))
if total_size == 0:
continue
num_chunks = random.randint(1, min(100,
sum(x.numel() for x in tensors)))
part_sizes = load_balance_peers(total_size, [None] * num_chunks)
chunks = split_into_parts(tensors, part_sizes)
assert len(chunks) == num_chunks
shapes = [tensor.shape for tensor in tensors]
restored = restore_from_parts(chunks, shapes)
assert len(restored) == len(tensors)
assert all(new.shape == old.shape
for new, old in zip(restored, tensors))
assert all(
torch.allclose(new, old) for new, old in zip(restored, tensors))
async def leader_assemble_group(self) -> AllReduceRunner:
""" Form up all current followers into a group and prepare to _run_allreduce """
assert self.lock_looking_for_group.locked() and self.lock_request_join_group.locked()
assert not self.assembled_group.done()
group_id = DHTID.generate().to_bytes()
ordered_group_endpoints = list(self.current_followers)
ordered_group_endpoints.append(self.endpoint)
random.shuffle(ordered_group_endpoints)
throughputs, gathered = [], []
for endpoint in ordered_group_endpoints:
if endpoint == self.endpoint:
throughputs.append(self.throughput)
gathered.append(self.data_for_gather)
else:
follower_info = self.current_followers[endpoint]
throughputs.append(follower_info.throughput if follower_info.throughput >= 0 else None)
gathered.append(follower_info.gather if follower_info.gather else None)
part_sizes = load_balance_peers(self.total_size, throughputs, self.min_vector_size)
group_key_seed = random.randint(- 2 ** 31, 2 ** 31 - 1)
logger.debug(f"{self.endpoint} - leader started allreduce for {len(ordered_group_endpoints)} peers.")
allreduce_group = AllReduceRunner(group_id=group_id, tensors=self.averaged_tensors, endpoint=self.endpoint,
ordered_group_endpoints=ordered_group_endpoints, part_sizes=part_sizes,
gathered=gathered, group_key_seed=group_key_seed, **self.allreduce_kwargs)
await self.group_key_manager.update_key_on_group_assembled(allreduce_group, is_leader=True)
self.assembled_group.set_result(allreduce_group)
return allreduce_group
def check_optimality(vector_size, throughputs, ref_partitions):
partitions = list(load_balance_peers(vector_size, throughputs))
assert get_cost(vector_size, partitions, throughputs) <= get_cost(
vector_size, ref_partitions, throughputs)
