def testPrepareSpilled(self):
from mars.worker.spill import write_spill_file
pool_address = '127.0.0.1:%d' % get_next_port()
session_id = str(uuid.uuid4())
mock_data = np.array([1, 2, 3, 4])
options.worker.spill_directory = tempfile.mkdtemp(prefix='mars_worker_prep_spilled-')
with create_actor_pool(n_process=1, backend='gevent', address=pool_address) as pool:
self.create_standard_actors(pool, pool_address, with_daemon=False, with_status=False)
pool.create_actor(SpillActor)
pool.create_actor(CpuCalcActor)
cluster_info_ref = pool.actor_ref(WorkerClusterInfoActor.default_uid())
chunk_meta_client = ChunkMetaClient(pool, cluster_info_ref)
pool.actor_ref(ChunkHolderActor.default_uid())
import mars.tensor as mt
from mars.tensor.fetch import TensorFetch
arr = mt.ones((4,), chunk_size=4)
arr_add = mt.array(mock_data)
result_tensor = arr + arr_add
graph = result_tensor.build_graph(compose=False, tiled=True)
modified_chunk = arr_add.chunks[0]
arr_add.chunks[0]._op = TensorFetch(
dtype=modified_chunk.dtype, _outputs=[weakref.ref(o) for o in modified_chunk.op.outputs],
_key=modified_chunk.op.key)
# test meta missing
with self.run_actor_test(pool) as test_actor:
graph_key = str(uuid.uuid4())
execution_ref = test_actor.promise_ref(ExecutionActor.default_uid())
execution_ref.execute_graph(session_id, graph_key, serialize_graph(graph),
dict(chunks=[result_tensor.chunks[0].key]), None, _promise=True) \
.then(lambda *_: test_actor.set_result(None)) \
.catch(lambda *exc: test_actor.set_result(exc, False))
with self.assertRaises(DependencyMissing):
self.get_result()
chunk_meta_client.set_chunk_meta(session_id, modified_chunk.key, size=mock_data.nbytes,
shape=mock_data.shape, workers=('0.0.0.0:1234', pool_address))
write_spill_file(modified_chunk.key, mock_data)
# test read from spilled file
with self.run_actor_test(pool) as test_actor:
def _validate(_):
data = test_actor._chunk_store.get(session_id, result_tensor.chunks[0].key)
assert_array_equal(data, mock_data + np.ones((4,)))
graph_key = str(uuid.uuid4())
execution_ref = test_actor.promise_ref(ExecutionActor.default_uid())
execution_ref.execute_graph(session_id, graph_key, serialize_graph(graph),
dict(chunks=[result_tensor.chunks[0].key]), None, _promise=True) \
.then(_validate) \
.then(lambda *_: test_actor.set_result(None)) \
.catch(lambda *exc: test_actor.set_result(exc, False))
self.get_result()
def testFailoverMessage(self):
mock_session_id = str(uuid.uuid4())
mock_graph_key = str(uuid.uuid4())
mock_chunk_key = str(uuid.uuid4())
addr = '127.0.0.1:%d' % get_next_port()
mock_worker_addr = '127.0.0.1:54132'
options.scheduler.worker_blacklist_time = 0.5
with create_actor_pool(n_process=1, backend='gevent',
address=addr) as pool:
cluster_info_ref = pool.create_actor(
SchedulerClusterInfoActor, [pool.cluster_info.address],
uid=SchedulerClusterInfoActor.default_uid())
session_manager_ref = pool.create_actor(
SessionManagerActor, uid=SessionManagerActor.default_uid())
resource_ref = pool.create_actor(ResourceActor,
uid=ResourceActor.default_uid())
pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_uid())
session_ref = pool.actor_ref(
session_manager_ref.create_session(mock_session_id))
chunk_meta_client = ChunkMetaClient(pool, cluster_info_ref)
chunk_meta_client.set_chunk_meta(mock_session_id,
mock_chunk_key,
size=80,
shape=(10, ),
workers=(mock_worker_addr, ))
with mock.patch(GraphActor.__module__ + '.' + GraphActor.__name__,
new=MockGraphActor):
session_ref.submit_tileable_graph(None, mock_graph_key)
graph_ref = pool.actor_ref(
GraphActor.gen_uid(mock_session_id, mock_graph_key))
expire_time = time.time(
) - options.scheduler.status_timeout - 1
resource_ref.set_worker_meta(mock_worker_addr,
dict(update_time=expire_time))
resource_ref.detect_dead_workers(_tell=True)
pool.sleep(0.2)
_, removes, lost_chunks = graph_ref.get_worker_change_args()
self.assertListEqual(removes, [mock_worker_addr])
self.assertListEqual(lost_chunks, [mock_chunk_key])
self.assertNotIn(mock_worker_addr,
resource_ref.get_workers_meta())
resource_ref.set_worker_meta(mock_worker_addr,
dict(update_time=time.time()))
self.assertNotIn(mock_worker_addr,
resource_ref.get_workers_meta())
pool.sleep(0.4)
resource_ref.set_worker_meta(mock_worker_addr,
dict(update_time=time.time()))
self.assertIn(mock_worker_addr,
resource_ref.get_workers_meta())
def testPrepareQuota(self, *_):
pinned = [True]
def _mock_pin(_graph_key, chunk_keys):
from mars.errors import PinChunkFailed
if pinned[0]:
raise PinChunkFailed
return chunk_keys
ChunkHolderActor.pin_chunks.side_effect = _mock_pin
pool_address = '127.0.0.1:%d' % get_next_port()
session_id = str(uuid.uuid4())
mock_data = np.array([1, 2, 3, 4])
with create_actor_pool(n_process=1, backend='gevent', address=pool_address) as pool:
self.create_standard_actors(pool, pool_address, with_daemon=False, with_status=False)
pool.create_actor(MockSenderActor, mock_data, 'in', uid='w:mock_sender')
pool.create_actor(CpuCalcActor)
cluster_info_ref = pool.actor_ref(WorkerClusterInfoActor.default_uid())
chunk_meta_client = ChunkMetaClient(pool, cluster_info_ref)
import mars.tensor as mt
from mars.tensor.fetch import TensorFetch
arr = mt.ones((4,), chunk_size=4)
arr_add = mt.array(mock_data)
result_tensor = arr + arr_add
graph = result_tensor.build_graph(compose=False, tiled=True)
modified_chunk = arr_add.chunks[0]
arr_add.chunks[0]._op = TensorFetch(
dtype=modified_chunk.dtype, _outputs=[weakref.ref(o) for o in modified_chunk.op.outputs],
_key=modified_chunk.op.key)
chunk_meta_client.set_chunk_meta(session_id, modified_chunk.key, size=mock_data.nbytes,
shape=mock_data.shape, workers=('0.0.0.0:1234', pool_address))
with self.run_actor_test(pool) as test_actor:
graph_key = str(uuid.uuid4())
execution_ref = test_actor.promise_ref(ExecutionActor.default_uid())
start_time = time.time()
execution_ref.execute_graph(
session_id, graph_key, serialize_graph(graph),
dict(chunks=[result_tensor.chunks[0].key]), None, _tell=True)
execution_ref.add_finish_callback(session_id, graph_key, _promise=True) \
.then(lambda *_: test_actor.set_result(time.time())) \
.catch(lambda *exc: test_actor.set_result(exc, False))
def _delay_fun():
time.sleep(1)
pinned[0] = False
threading.Thread(target=_delay_fun).start()
finish_time = self.get_result()
self.assertGreaterEqual(finish_time, start_time + 1)
def testReadyState(self, *_):
session_id = str(uuid.uuid4())
graph_key = str(uuid.uuid4())
mock_workers = ['localhost:12345', 'localhost:23456']
def _mock_get_workers_meta(*_, **__):
return dict((w, dict(hardware=dict(cpu_total=1, memory=1024**3)))
for w in mock_workers)
with patch_method(ResourceActor.get_workers_meta, new=_mock_get_workers_meta) as _, \
self._prepare_test_graph(session_id, graph_key, mock_workers) as (pool, graph_ref):
input_op_keys, mid_op_key, output_op_keys = self._filter_graph_level_op_keys(
graph_ref)
meta_client = ChunkMetaClient(
pool, pool.actor_ref(SchedulerClusterInfoActor.default_uid()))
op_ref = pool.actor_ref(
OperandActor.gen_uid(session_id, mid_op_key))
resource_ref = pool.actor_ref(ResourceActor.default_uid())
input_refs = [
pool.actor_ref(OperandActor.gen_uid(session_id, k))
for k in input_op_keys
]
def test_entering_state(target):
for key in input_op_keys:
op_ref.remove_finished_predecessor(key)
op_ref.start_operand(OperandState.UNSCHEDULED)
for ref in input_refs:
ref.start_operand(OperandState.UNSCHEDULED)
for ref in input_refs:
self.assertEqual(op_ref.get_state(),
OperandState.UNSCHEDULED)
ref.start_operand(OperandState.FINISHED)
pool.sleep(1)
self.assertEqual(target, op_ref.get_state())
for w in mock_workers:
resource_ref.deallocate_resource(session_id, mid_op_key, w)
# test entering state with no input meta
test_entering_state(OperandState.UNSCHEDULED)
# fill meta
input_chunk_keys, _, _ = self._filter_graph_level_chunk_keys(
graph_ref)
for ck in input_chunk_keys:
meta_client.set_chunk_meta(session_id,
ck,
workers=('localhost:12345', ),
size=800)
# test successful entering state
test_entering_state(OperandState.READY)
def testReadyState(self, *_):
session_id = str(uuid.uuid4())
graph_key = str(uuid.uuid4())
mock_workers = ['localhost:12345', 'localhost:23456']
with self._prepare_test_graph(session_id, graph_key,
mock_workers) as (pool, graph_ref):
input_op_keys, mid_op_key, output_op_keys = self._filter_graph_level_op_keys(
graph_ref)
meta_client = ChunkMetaClient(
pool, pool.actor_ref(SchedulerClusterInfoActor.default_name()))
op_ref = pool.actor_ref(
OperandActor.gen_uid(session_id, mid_op_key))
input_refs = [
pool.actor_ref(OperandActor.gen_uid(session_id, k))
for k in input_op_keys
]
def test_entering_state(target):
for key in input_op_keys:
op_ref.remove_finished_predecessor(key)
op_ref.start_operand(OperandState.UNSCHEDULED)
for ref in input_refs:
ref.start_operand(OperandState.UNSCHEDULED)
for ref in input_refs:
self.assertEqual(op_ref.get_state(),
OperandState.UNSCHEDULED)
ref.start_operand(OperandState.FINISHED)
pool.sleep(0.5)
self.assertEqual(target, op_ref.get_state())
# test entering state with no input meta
test_entering_state(OperandState.UNSCHEDULED)
# fill meta
input_chunk_keys, _, _ = self._filter_graph_level_chunk_keys(
graph_ref)
for ck in input_chunk_keys:
meta_client.set_chunk_meta(session_id,
ck,
workers=('localhost:12345', ),
size=800)
# test entering state with failure in fetching sizes
with patch_method(ChunkMetaClient.batch_get_chunk_size,
new=lambda *_: [None, None]):
test_entering_state(OperandState.UNSCHEDULED)
# test successful entering state
test_entering_state(OperandState.READY)
def testAssignerActor(self):
mock_scheduler_addr = '127.0.0.1:%d' % get_next_port()
with create_actor_pool(n_process=1, backend='gevent', address=mock_scheduler_addr) as pool:
cluster_info_ref = pool.create_actor(SchedulerClusterInfoActor, [pool.cluster_info.address],
uid=SchedulerClusterInfoActor.default_uid())
resource_ref = pool.create_actor(ResourceActor, uid=ResourceActor.default_uid())
pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_uid())
endpoint1 = 'localhost:12345'
endpoint2 = 'localhost:23456'
res = dict(hardware=dict(cpu=4, memory=4096))
def write_mock_meta():
resource_ref.set_worker_meta(endpoint1, res)
resource_ref.set_worker_meta(endpoint2, res)
g = gevent.spawn(write_mock_meta)
g.join()
assigner_ref = pool.create_actor(AssignerActor, uid=AssignerActor.default_uid())
session_id = str(uuid.uuid4())
op_key = str(uuid.uuid4())
chunk_key1 = str(uuid.uuid4())
chunk_key2 = str(uuid.uuid4())
chunk_key3 = str(uuid.uuid4())
op_info = {
'op_name': 'test_op',
'io_meta': dict(input_chunks=[chunk_key1, chunk_key2, chunk_key3]),
'retries': 0,
'optimize': {
'depth': 0,
'demand_depths': (),
'successor_size': 1,
'descendant_size': 0
}
}
chunk_meta_client = ChunkMetaClient(pool, cluster_info_ref)
chunk_meta_client.set_chunk_meta(session_id, chunk_key1, size=512, workers=(endpoint1,))
chunk_meta_client.set_chunk_meta(session_id, chunk_key2, size=512, workers=(endpoint1,))
chunk_meta_client.set_chunk_meta(session_id, chunk_key3, size=512, workers=(endpoint2,))
reply_ref = pool.create_actor(PromiseReplyTestActor)
reply_callback = ((reply_ref.uid, reply_ref.address), 'reply')
assigner_ref.apply_for_resource(session_id, op_key, op_info, callback=reply_callback)
while not reply_ref.get_reply():
gevent.sleep(0.1)
_, ret_value = reply_ref.get_reply()
self.assertEqual(ret_value[0], endpoint1)
def testFetchRemoteData(self):
pool_address = '127.0.0.1:%d' % get_next_port()
session_id = str(uuid.uuid4())
mock_data = np.array([1, 2, 3, 4])
with create_actor_pool(n_process=1, backend='gevent',
address=pool_address, distributor=MarsDistributor(2, 'w:0:')) as pool:
self.create_standard_actors(pool, pool_address, with_daemon=False, with_status=False,
with_resource=True)
pool.create_actor(CpuCalcActor)
pool.create_actor(MockSenderActor, mock_data, 'in', uid='w:mock_sender')
cluster_info_ref = pool.actor_ref(WorkerClusterInfoActor.default_uid())
chunk_meta_client = ChunkMetaClient(pool, cluster_info_ref)
import mars.tensor as mt
from mars.tensor.fetch import TensorFetch
arr = mt.ones((4,), chunk_size=4)
arr_add = mt.array(mock_data)
result_tensor = arr + arr_add
graph = result_tensor.build_graph(compose=False, tiled=True)
modified_chunk = arr_add.chunks[0]
arr_add.chunks[0]._op = TensorFetch(
dtype=modified_chunk.dtype, _outputs=[weakref.ref(o) for o in modified_chunk.op.outputs],
_key=modified_chunk.op.key)
with self.run_actor_test(pool) as test_actor:
graph_key = str(uuid.uuid4())
execution_ref = test_actor.promise_ref(ExecutionActor.default_uid())
execution_ref.execute_graph(session_id, graph_key, serialize_graph(graph),
dict(chunks=[result_tensor.chunks[0].key]), None, _tell=True)
execution_ref.add_finish_callback(session_id, graph_key, _promise=True) \
.then(lambda *_: test_actor.set_result(None)) \
.catch(lambda *exc: test_actor.set_result(exc, False))
with self.assertRaises(DependencyMissing):
self.get_result()
chunk_meta_client.set_chunk_meta(session_id, modified_chunk.key, size=mock_data.nbytes,
shape=mock_data.shape, workers=('0.0.0.0:1234',))
with self.run_actor_test(pool) as test_actor:
graph_key = str(uuid.uuid4())
execution_ref = test_actor.promise_ref(ExecutionActor.default_uid())
execution_ref.execute_graph(session_id, graph_key, serialize_graph(graph),
dict(chunks=[result_tensor.chunks[0].key]), None, _tell=True)
execution_ref.add_finish_callback(session_id, graph_key, _promise=True) \
.then(lambda *_: test_actor.set_result(None)) \
.catch(lambda *exc: test_actor.set_result(exc, False))
with self.assertRaises(DependencyMissing):
self.get_result()
chunk_meta_client.set_chunk_meta(session_id, modified_chunk.key, size=mock_data.nbytes,
shape=mock_data.shape, workers=('0.0.0.0:1234', pool_address))
with self.run_actor_test(pool) as test_actor:
def _validate(_):
data = test_actor._chunk_store.get(session_id, result_tensor.chunks[0].key)
assert_array_equal(data, mock_data + np.ones((4,)))
graph_key = str(uuid.uuid4())
execution_ref = test_actor.promise_ref(ExecutionActor.default_uid())
execution_ref.execute_graph(session_id, graph_key, serialize_graph(graph),
dict(chunks=[result_tensor.chunks[0].key]), None, _tell=True)
execution_ref.add_finish_callback(session_id, graph_key, _promise=True) \
.then(_validate) \
.then(lambda *_: test_actor.set_result(None)) \
.catch(lambda *exc: test_actor.set_result(exc, False))
self.get_result()
def testAssignerActor(self, *_):
mock_scheduler_addr = f'127.0.0.1:{get_next_port()}'
with create_actor_pool(n_process=1,
backend='gevent',
address=mock_scheduler_addr) as pool:
cluster_info_ref = pool.create_actor(
SchedulerClusterInfoActor, [pool.cluster_info.address],
uid=SchedulerClusterInfoActor.default_uid())
resource_ref = pool.create_actor(ResourceActor,
uid=ResourceActor.default_uid())
pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_uid())
endpoint1 = 'localhost:12345'
endpoint2 = 'localhost:23456'
res = dict(hardware=dict(cpu=4, mem_quota=4096))
resource_ref.set_worker_meta(endpoint1, res)
resource_ref.set_worker_meta(endpoint2, res)
assigner_ref = pool.create_actor(AssignerActor,
uid=AssignerActor.default_uid())
session_id = str(uuid.uuid4())
op_key = str(uuid.uuid4())
chunk_key1 = str(uuid.uuid4())
chunk_key2 = str(uuid.uuid4())
chunk_key3 = str(uuid.uuid4())
op_info = {
'op_name': 'test_op',
'io_meta':
dict(input_chunks=[chunk_key1, chunk_key2, chunk_key3]),
'retries': 0,
'optimize': {
'depth': 0,
'demand_depths': (),
'successor_size': 1,
'descendant_size': 0
}
}
chunk_meta_client = ChunkMetaClient(pool, cluster_info_ref)
chunk_meta_client.set_chunk_meta(session_id,
chunk_key1,
size=512,
workers=(endpoint1, ))
chunk_meta_client.set_chunk_meta(session_id,
chunk_key2,
size=512,
workers=(endpoint1, ))
chunk_meta_client.set_chunk_meta(session_id,
chunk_key3,
size=512,
workers=(endpoint2, ))
uid = OperandActor.gen_uid(session_id, op_key)
reply_ref = pool.create_actor(MockOperandActor, uid=uid)
assigner_ref.apply_for_resource(session_id, op_key, op_info)
while not reply_ref.get_worker_ep():
gevent.sleep(0.1)
self.assertEqual(reply_ref.get_worker_ep(), endpoint1)
with self.run_actor_test(pool) as test_actor, self.assertRaises(
TimeoutError):
assigner_p_ref = test_actor.promise_ref(assigner_ref)
try:
options.scheduler.assign_timeout = 1
res = dict(hardware=dict(cpu=4, mem_quota=0))
resource_ref.set_worker_meta(endpoint1, res)
resource_ref.set_worker_meta(endpoint2, res)
self.waitp(
assigner_p_ref.apply_for_resource(session_id,
op_key,
op_info,
_promise=True))
finally:
options.scheduler.assign_timeout = 600