def testChunkBroadcast(self, *_):
proc_count = 2
endpoints = ['127.0.0.1:%d' % get_next_port() for _ in range(proc_count)]
keys = []
def _mock_get_scheduler(key):
return endpoints[keys.index(key[1]) % len(endpoints)]
ChunkMetaActor.get_scheduler.side_effect = _mock_get_scheduler
session_id = str(uuid.uuid4())
with create_actor_pool(n_process=1, backend='gevent', address=endpoints[0]) as pool1:
pool1.create_actor(ClusterInfoActor, endpoints, uid=ClusterInfoActor.default_name())
pool1.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name())
with create_actor_pool(n_process=1, backend='gevent', address=endpoints[1]) as pool2:
pool2.create_actor(ClusterInfoActor, endpoints, uid=ClusterInfoActor.default_name())
pool2.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name())
client = new_client()
ref1 = client.actor_ref(ChunkMetaActor.default_name(), address=endpoints[0])
ref2 = client.actor_ref(ChunkMetaActor.default_name(), address=endpoints[0])
local_ref1 = client.actor_ref(LocalChunkMetaActor.default_name(), address=endpoints[0])
local_ref2 = client.actor_ref(LocalChunkMetaActor.default_name(), address=endpoints[1])
key1 = str(uuid.uuid4())
key2 = str(uuid.uuid4())
keys = [key1, key2]
ref1.set_chunk_broadcasts(session_id, key1, [endpoints[1]])
ref1.set_chunk_size(session_id, key1, 512)
ref1.set_chunk_shape(session_id, key1, (10,) * 2)
ref1.add_worker(session_id, key1, 'abc')
ref2.set_chunk_broadcasts(session_id, key2, [endpoints[0]])
ref2.set_chunk_size(session_id, key2, 512)
ref1.set_chunk_shape(session_id, key2, (10,) * 2)
ref2.add_worker(session_id, key2, 'def')
pool2.sleep(0.1)
self.assertEqual(local_ref1.get_chunk_meta(session_id, key1).chunk_size, 512)
self.assertEqual(local_ref1.get_chunk_meta(session_id, key1).chunk_shape, (10,) * 2)
self.assertEqual(local_ref1.get_chunk_broadcasts(session_id, key1), [endpoints[1]])
self.assertEqual(local_ref2.get_chunk_meta(session_id, key1).chunk_size, 512)
self.assertEqual(local_ref2.get_chunk_meta(session_id, key1).chunk_shape, (10,) * 2)
self.assertEqual(local_ref2.get_chunk_broadcasts(session_id, key2), [endpoints[0]])
ref1.delete_meta(session_id, key1)
pool2.sleep(0.1)
self.assertIsNone(local_ref1.get_chunk_meta(session_id, key1))
self.assertIsNone(local_ref2.get_chunk_meta(session_id, key1))
self.assertIsNone(local_ref1.get_chunk_broadcasts(session_id, key1))
local_ref1.remove_workers_in_session(session_id, ['def'])
local_ref2.remove_workers_in_session(session_id, ['def'])
pool2.sleep(0.1)
self.assertIsNone(local_ref1.get_chunk_meta(session_id, key2))
self.assertIsNone(local_ref2.get_chunk_meta(session_id, key2))
self.assertIsNone(local_ref2.get_chunk_broadcasts(session_id, key2))
def create_standard_actors(cls,
pool,
address,
quota_size=None,
with_daemon=True,
with_status=True,
with_resource=False):
quota_size = quota_size or (1024 * 1024)
pool.create_actor(PlasmaKeyMapActor,
uid=PlasmaKeyMapActor.default_name())
pool.create_actor(ClusterInfoActor,
schedulers=[address],
uid=ClusterInfoActor.default_name())
if with_resource:
pool.create_actor(ResourceActor, uid=ResourceActor.default_name())
if with_daemon:
pool.create_actor(WorkerDaemonActor,
uid=WorkerDaemonActor.default_name())
if with_status:
pool.create_actor(StatusActor,
address,
uid=StatusActor.default_name())
pool.create_actor(ChunkHolderActor,
cls.plasma_storage_size,
uid=ChunkHolderActor.default_name())
pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name())
pool.create_actor(TaskQueueActor, uid=TaskQueueActor.default_name())
pool.create_actor(DispatchActor, uid=DispatchActor.default_name())
pool.create_actor(QuotaActor,
quota_size,
uid=MemQuotaActor.default_name())
pool.create_actor(ExecutionActor, uid=ExecutionActor.default_name())
def execute_case():
pool.create_actor(ClusterInfoActor,
[pool.cluster_info.address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(
ResourceActor, uid=ResourceActor.default_name())
pool.create_actor(ChunkMetaActor,
uid=ChunkMetaActor.default_name())
pool.create_actor(AssignerActor,
uid=AssignerActor.gen_name(session_id))
graph_ref = pool.create_actor(GraphActor,
session_id,
graph_key,
serialize_graph(graph),
uid=GraphActor.gen_name(
session_id, graph_key))
execution_ref = pool.create_actor(FakeExecutionActor, sleep=1)
# handle mock objects
OperandActor._get_raw_execution_ref.side_effect = lambda: execution_ref
mock_resource = dict(
hardware=dict(cpu=4, cpu_total=4, memory=512))
def write_mock_meta():
resource_ref.set_worker_meta('localhost:12345',
mock_resource)
resource_ref.set_worker_meta('localhost:23456',
mock_resource)
v = gevent.spawn(write_mock_meta)
v.join()
graph_ref.prepare_graph()
fetched_graph = graph_ref.get_chunk_graph()
graph_ref.scan_node()
graph_ref.place_initial_chunks()
final_keys = set()
for c in fetched_graph:
if fetched_graph.count_successors(c) == 0:
final_keys.add(c.op.key)
graph_ref.create_operand_actors()
graph_meta_ref = pool.actor_ref(
GraphMetaActor.gen_name(session_id, graph_key))
start_time = time.time()
cancel_called = False
while True:
gevent.sleep(0.1)
if not cancel_called and time.time() > start_time + 0.8:
cancel_called = True
graph_ref.stop_graph(_tell=True)
if time.time() - start_time > 30:
raise SystemError('Wait for execution finish timeout')
if graph_meta_ref.get_state() in (GraphState.SUCCEEDED,
GraphState.FAILED,
GraphState.CANCELLED):
break
def testHolder(self):
pool_address = '127.0.0.1:%d' % get_next_port()
with create_actor_pool(n_process=1,
backend='gevent',
address=pool_address) as pool:
pool.create_actor(ClusterInfoActor,
schedulers=[pool_address],
uid=ClusterInfoActor.default_name())
pool.create_actor(KVStoreActor, uid=KVStoreActor.default_name())
pool.create_actor(DispatchActor, uid=DispatchActor.default_name())
pool.create_actor(QuotaActor,
1024 * 1024 * 10,
uid=MemQuotaActor.default_name())
cache_ref = pool.create_actor(ChunkHolderActor,
self.plasma_storage_size,
uid=ChunkHolderActor.default_name())
pool.create_actor(SpillActor)
try:
test_ref = pool.create_actor(CacheTestActor)
test_ref.run_test_cache()
while not test_ref.get_exc_info()[0]:
gevent.sleep(0.1)
exc_info = test_ref.get_exc_info()[1]
if exc_info:
six.reraise(*exc_info)
finally:
pool.destroy_actor(cache_ref)
def _prepare_test_graph(self, session_id, graph_key, mock_workers):
addr = '127.0.0.1:%d' % get_next_port()
a1 = mt.random.random((100,))
a2 = mt.random.random((100,))
s = a1 + a2
v1, v2 = mt.split(s, 2)
graph = DAG()
v1.build_graph(graph=graph, compose=False)
v2.build_graph(graph=graph, compose=False)
with create_actor_pool(n_process=1, backend='gevent', address=addr) as pool:
pool.create_actor(ClusterInfoActor, [pool.cluster_info.address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(ResourceActor, uid=ResourceActor.default_name())
pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name())
pool.create_actor(AssignerActor, uid=AssignerActor.default_name())
graph_ref = pool.create_actor(GraphActor, session_id, graph_key, serialize_graph(graph),
uid=GraphActor.gen_name(session_id, graph_key))
for w in mock_workers:
resource_ref.set_worker_meta(w, dict(hardware=dict(cpu_total=4)))
graph_ref.prepare_graph()
graph_ref.analyze_graph()
graph_ref.create_operand_actors(_start=False)
yield pool, graph_ref
def testEnsureTimeout(self, *_):
from mars.errors import PromiseTimeout
pool_address = '127.0.0.1:%d' % get_next_port()
with create_actor_pool(n_process=1, backend='gevent', address=pool_address) as pool:
pool.create_actor(PlasmaKeyMapActor, uid=PlasmaKeyMapActor.default_name())
pool.create_actor(ClusterInfoActor, schedulers=[pool_address],
uid=ClusterInfoActor.default_name())
pool.create_actor(KVStoreActor, uid=KVStoreActor.default_name())
pool.create_actor(DispatchActor, uid=DispatchActor.default_name())
pool.create_actor(QuotaActor, 1024 * 1024 * 10, uid=MemQuotaActor.default_name())
pool.create_actor(SpillActor, uid=SpillActor.default_name())
cache_ref = pool.create_actor(ChunkHolderActor, self.plasma_storage_size,
uid=ChunkHolderActor.default_name())
try:
options.worker.prepare_data_timeout = 2
test_ref = pool.create_actor(CacheTestActor)
test_ref.run_test_ensure_timeout()
while not test_ref.get_exc_info()[0]:
pool.sleep(0.1)
exc_info = test_ref.get_exc_info()[1]
self.assertIsNotNone(exc_info)
self.assertIsInstance(exc_info[1], PromiseTimeout)
finally:
options.worker.prepare_data_timeout = 600
pool.destroy_actor(cache_ref)
def testExecute(self):
pool_address = '127.0.0.1:%d' % get_next_port()
with create_actor_pool(n_process=1,
backend='gevent',
address=pool_address) as pool:
pool.create_actor(ClusterInfoActor,
schedulers=[pool_address],
uid=ClusterInfoActor.default_name())
cache_ref = pool.create_actor(ChunkHolderActor,
self._plasma_helper._size,
uid='ChunkHolderActor')
pool.create_actor(KVStoreActor, uid=KVStoreActor.default_name())
pool.create_actor(DispatchActor, uid='DispatchActor')
pool.create_actor(QuotaActor, 1024 * 1024, uid='MemQuotaActor')
pool.create_actor(CpuCalcActor)
pool.create_actor(ExecutionActor, uid='ExecutionActor')
try:
test_ref = pool.create_actor(ExecuteTestActor)
test_ref.run_test()
while not test_ref.get_exc_info()[0]:
gevent.sleep(0.1)
exc_info = test_ref.get_exc_info()[1]
if exc_info:
six.reraise(*exc_info)
finally:
pool.destroy_actor(cache_ref)
def testStatus(self):
pool_address = '127.0.0.1:%d' % get_next_port()
with create_actor_pool(n_process=1,
backend='gevent',
address=pool_address) as pool:
pool.create_actor(ClusterInfoActor,
schedulers=[pool_address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(ResourceActor,
uid=ResourceActor.default_name())
pool.create_actor(ChunkHolderActor,
self.plasma_storage_size,
uid=ChunkHolderActor.default_name())
status_ref = pool.create_actor(StatusActor,
'127.0.0.1:1234',
uid=StatusActor.default_name())
def delay_read():
gevent.sleep(2)
return resource_ref.get_workers_meta()
gl = gevent.spawn(delay_read)
gl.join()
v = gl.value
self.assertIsNotNone(v)
pool.destroy_actor(status_ref)
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:
pool.create_actor(ClusterInfoActor, [pool.cluster_info.address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(ResourceActor,
uid=ResourceActor.default_name())
chunk_meta_ref = pool.create_actor(
ChunkMetaActor, uid=ChunkMetaActor.default_name())
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_name())
session_id = 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]),
'output_size': 512,
'retries': 0,
'optimize': {
'depth': 0,
'demand_depths': (),
'successor_size': 1,
'descendant_size': 0
}
}
chunk_meta_ref.set_chunk_meta(session_id,
chunk_key1,
size=512,
workers=(endpoint1, ))
chunk_meta_ref.set_chunk_meta(session_id,
chunk_key2,
size=512,
workers=(endpoint1, ))
chunk_meta_ref.set_chunk_meta(session_id,
chunk_key3,
size=512,
workers=(endpoint2, ))
workers = assigner_ref.get_worker_assignments(session_id, op_info)
self.assertEqual(workers[0], endpoint1)
def testLocalCluster(self):
endpoint = gen_endpoint('0.0.0.0')
with LocalDistributedCluster(endpoint,
scheduler_n_process=2,
worker_n_process=3) as cluster:
pool = cluster.pool
self.assertTrue(
pool.has_actor(pool.actor_ref(
ClusterInfoActor.default_name())))
self.assertTrue(
pool.has_actor(
pool.actor_ref(SessionManagerActor.default_name())))
self.assertTrue(
pool.has_actor(pool.actor_ref(DispatchActor.default_name())))
with new_session(endpoint) as session:
api = session._api
t = mt.ones((3, 3), chunk_size=2)
result = session.run(t)
np.testing.assert_array_equal(result, np.ones((3, 3)))
self.assertNotIn(session._session_id,
api.session_manager.get_sessions())
def testErrorOnPrepare(self, *_):
session_id = str(uuid.uuid4())
addr = '127.0.0.1:%d' % get_next_port()
with create_actor_pool(n_process=1, backend='gevent', address=addr) as pool:
pool.create_actor(ClusterInfoActor, [pool.cluster_info.address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(ResourceActor, uid=ResourceActor.default_name())
pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name())
pool.create_actor(AssignerActor, uid=AssignerActor.default_name())
resource_ref.set_worker_meta('localhost:12345', dict(hardware=dict(cpu_total=4)))
resource_ref.set_worker_meta('localhost:23456', dict(hardware=dict(cpu_total=4)))
# error occurred in create_operand_actors
graph_key = str(uuid.uuid4())
expr = mt.random.random((8, 2), chunk_size=2) + 1
graph = expr.build_graph(compose=False)
serialized_graph = serialize_graph(graph)
graph_ref = pool.create_actor(GraphActor, session_id, graph_key, serialized_graph,
uid=GraphActor.gen_name(session_id, graph_key))
def _mock_raises(*_, **__):
raise RuntimeError
with patch_method(GraphActor.create_operand_actors, new=_mock_raises):
with self.assertRaises(RuntimeError):
graph_ref.execute_graph()
self.assertEqual(graph_ref.get_state(), GraphState.FAILED)
graph_ref.destroy()
# interrupted during create_operand_actors
graph_key = str(uuid.uuid4())
graph_ref = pool.create_actor(GraphActor, session_id, graph_key, serialized_graph,
uid=GraphActor.gen_name(session_id, graph_key))
def _mock_cancels(*_, **__):
graph_meta_ref = pool.actor_ref(GraphMetaActor.gen_name(session_id, graph_key))
graph_meta_ref.set_state(GraphState.CANCELLING)
with patch_method(GraphActor.create_operand_actors, new=_mock_cancels):
graph_ref.execute_graph()
self.assertEqual(graph_ref.get_state(), GraphState.CANCELLED)
# interrupted during previous steps
graph_key = str(uuid.uuid4())
graph_ref = pool.create_actor(GraphActor, session_id, graph_key, serialized_graph,
uid=GraphActor.gen_name(session_id, graph_key))
def _mock_cancels(*_, **__):
graph_meta_ref = pool.actor_ref(GraphMetaActor.gen_name(session_id, graph_key))
graph_meta_ref.set_state(GraphState.CANCELLING)
return dict()
with patch_method(GraphAnalyzer.calc_operand_assignments, new=_mock_cancels):
graph_ref.execute_graph()
self.assertEqual(graph_ref.get_state(), GraphState.CANCELLED)
def testExecute(self):
pool_address = '127.0.0.1:%d' % get_next_port()
with create_actor_pool(n_process=1, backend='gevent', address=pool_address) as pool:
pool.create_actor(ClusterInfoActor, schedulers=[pool_address],
uid=ClusterInfoActor.default_name())
cache_ref = pool.create_actor(
ChunkHolderActor, self.plasma_storage_size, uid=ChunkHolderActor.default_name())
pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name())
pool.create_actor(TaskQueueActor, uid=TaskQueueActor.default_name())
pool.create_actor(DispatchActor, uid=DispatchActor.default_name())
pool.create_actor(QuotaActor, 1024 * 1024, uid=MemQuotaActor.default_name())
pool.create_actor(CpuCalcActor)
pool.create_actor(ExecutionActor, uid=ExecutionActor.default_name())
try:
with self.run_actor_test(pool) as test_actor:
import mars.tensor as mt
from mars.tensor.expressions.datasource import TensorOnes, TensorFetchChunk
arr = mt.ones((10, 8), chunk_size=10)
arr_add = mt.ones((10, 8), chunk_size=10)
arr2 = arr + arr_add
graph = arr2.build_graph(compose=False, tiled=True)
for chunk in graph:
if isinstance(chunk.op, TensorOnes):
chunk._op = TensorFetchChunk(
dtype=chunk.dtype, _outputs=[weakref.ref(o) for o in chunk.op.outputs],
_key=chunk.op.key)
session_id = str(uuid.uuid4())
chunk_holder_ref = test_actor.promise_ref(ChunkHolderActor.default_name())
refs = test_actor._chunk_store.put(session_id, arr.chunks[0].key,
np.ones((10, 8), dtype=np.int16))
chunk_holder_ref.register_chunk(session_id, arr.chunks[0].key)
del refs
refs = test_actor._chunk_store.put(session_id, arr_add.chunks[0].key,
np.ones((10, 8), dtype=np.int16))
chunk_holder_ref.register_chunk(session_id, arr_add.chunks[0].key)
del refs
executor_ref = test_actor.promise_ref(ExecutionActor.default_name())
def _validate(_):
data = test_actor._chunk_store.get(session_id, arr2.chunks[0].key)
assert_array_equal(data, 2 * np.ones((10, 8)))
executor_ref.enqueue_graph(session_id, str(id(graph)), serialize_graph(graph),
dict(chunks=[arr2.chunks[0].key]), None, _promise=True) \
.then(lambda *_: executor_ref.start_execution(session_id, str(id(graph)), _promise=True)) \
.then(_validate) \
.then(lambda *_: test_actor.set_result(None)) \
.catch(lambda *exc: test_actor.set_result(exc, False))
self.get_result()
finally:
pool.destroy_actor(cache_ref)
def testAssignerActor(self):
with create_actor_pool(backend='gevent') as pool:
pool.create_actor(ClusterInfoActor, [pool.cluster_info.address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(ResourceActor, uid=ResourceActor.default_name())
kv_store_ref = pool.create_actor(KVStoreActor, uid=KVStoreActor.default_name())
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')
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]),
'output_size': 512,
'retries': 0,
'optimize': {
'depth': 0,
'demand_depths': (),
'successor_size': 1,
'descendant_size': 0
}
}
kv_store_ref.write('/sessions/%s/chunks/%s/data_size' % (session_id, chunk_key1), 512)
kv_store_ref.write('/sessions/%s/chunks/%s/workers/%s' % (session_id, chunk_key1, endpoint1), '')
kv_store_ref.write('/sessions/%s/chunks/%s/data_size' % (session_id, chunk_key2), 512)
kv_store_ref.write('/sessions/%s/chunks/%s/workers/%s'
% (session_id, chunk_key2, endpoint1), '')
kv_store_ref.write('/sessions/%s/chunks/%s/data_size' % (session_id, chunk_key3), 512)
kv_store_ref.write('/sessions/%s/chunks/%s/workers/%s'
% (session_id, chunk_key3, 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, endpoint1)
def run_transfer_worker(pool_address, session_id, chunk_keys, spill_dir,
msg_queue):
from mars.config import options
options.worker.spill_directory = spill_dir
plasma_size = 1024 * 1024 * 10
# don't use multiple with-statement as we need the options be forked
with plasma.start_plasma_store(plasma_size) as store_args:
options.worker.plasma_socket = plasma_socket = store_args[0]
with create_actor_pool(n_process=2,
backend='gevent',
distributor=WorkerDistributor(2),
address=pool_address) as pool:
try:
pool.create_actor(ClusterInfoActor,
schedulers=[pool_address],
uid=ClusterInfoActor.default_name())
pool.create_actor(ChunkMetaActor,
uid=ChunkMetaActor.default_name())
pool.create_actor(DispatchActor,
uid=DispatchActor.default_name())
pool.create_actor(QuotaActor,
1024 * 1024 * 20,
uid=MemQuotaActor.default_name())
holder_ref = pool.create_actor(HolderActor, uid='HolderActor')
chunk_holder_ref = pool.create_actor(
ChunkHolderActor,
plasma_size,
uid=ChunkHolderActor.default_name())
pool.create_actor(SpillActor)
pool.create_actor(SenderActor, uid='%s' % str(uuid.uuid4()))
pool.create_actor(SenderActor, uid='%s' % str(uuid.uuid4()))
pool.create_actor(ReceiverActor, uid='%s' % str(uuid.uuid4()))
pool.create_actor(ReceiverActor, uid='%s' % str(uuid.uuid4()))
register_actor = pool.create_actor(WorkerRegistrationTestActor)
register_actor.register(session_id, chunk_keys)
check_time = time.time()
while not register_actor.get_finished():
gevent.sleep(0.5)
if time.time() - check_time > 60:
raise SystemError('Wait result timeout')
register_actor.destroy()
msg_queue.put(plasma_socket)
check_time = time.time()
while not holder_ref.obtain():
gevent.sleep(1)
if time.time() - check_time > 60:
raise SystemError('Wait result timeout')
finally:
pool.destroy_actor(chunk_holder_ref)
def _run_operand_case(session_id, graph_key, tensor, execution_creator):
graph = tensor.build_graph(compose=False)
with create_actor_pool(n_process=1, backend='gevent') as pool:
pool.create_actor(ClusterInfoActor, [pool.cluster_info.address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(ResourceActor,
uid=ResourceActor.default_name())
pool.create_actor(ChunkMetaActor,
uid=ChunkMetaActor.default_name())
pool.create_actor(AssignerActor, uid=AssignerActor.default_name())
graph_ref = pool.create_actor(GraphActor,
session_id,
graph_key,
serialize_graph(graph),
uid=GraphActor.gen_name(
session_id, graph_key))
addr_dict = dict()
def _build_mock_ref(uid=None, address=None):
if address in addr_dict:
return addr_dict[address]
else:
r = addr_dict[address] = execution_creator(pool)
return r
# handle mock objects
OperandActor._get_raw_execution_ref.side_effect = _build_mock_ref
mock_resource = dict(hardware=dict(cpu=4, cpu_total=4, memory=512))
resource_ref.set_worker_meta('localhost:12345', mock_resource)
resource_ref.set_worker_meta('localhost:23456', mock_resource)
graph_ref.prepare_graph()
fetched_graph = graph_ref.get_chunk_graph()
graph_ref.analyze_graph()
final_keys = set()
for c in fetched_graph:
if fetched_graph.count_successors(c) == 0:
final_keys.add(c.op.key)
graph_ref.create_operand_actors()
graph_meta_ref = pool.actor_ref(
GraphMetaActor.gen_name(session_id, graph_key))
start_time = time.time()
while True:
pool.sleep(0.1)
if time.time() - start_time > 30:
raise SystemError('Wait for execution finish timeout')
if graph_meta_ref.get_state() in (GraphState.SUCCEEDED,
GraphState.FAILED,
GraphState.CANCELLED):
break
def execute_case():
pool.create_actor(ClusterInfoActor,
[pool.cluster_info.address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(
ResourceActor, uid=ResourceActor.default_name())
kv_store_ref = pool.create_actor(
KVStoreActor, uid=KVStoreActor.default_name())
pool.create_actor(AssignerActor,
uid=AssignerActor.gen_name(session_id))
graph_ref = pool.create_actor(GraphActor,
session_id,
graph_key,
serialize_graph(graph),
uid=GraphActor.gen_name(
session_id, graph_key))
execution_ref = execution_creator(pool)
# handle mock objects
OperandActor._get_raw_execution_ref.side_effect = lambda: execution_ref
mock_resource = dict(
hardware=dict(cpu=4, cpu_total=4, memory=512))
def write_mock_meta():
resource_ref.set_worker_meta('localhost:12345',
mock_resource)
resource_ref.set_worker_meta('localhost:23456',
mock_resource)
v = gevent.spawn(write_mock_meta)
v.join()
graph_ref.prepare_graph()
graph_data = kv_store_ref.read(
'/sessions/%s/graphs/%s/chunk_graph' %
(session_id, graph_key)).value
fetched_graph = deserialize_graph(graph_data)
graph_ref.scan_node()
graph_ref.place_initial_chunks()
final_keys = set()
for c in fetched_graph:
if fetched_graph.count_successors(c) == 0:
final_keys.add(c.op.key)
graph_ref.create_operand_actors()
start_time = time.time()
while True:
gevent.sleep(0.1)
if time.time() - start_time > 30:
raise SystemError('Wait for execution finish timeout')
if kv_store_ref.read('/sessions/%s/graph/%s/state' % (session_id, graph_key)).value.lower() \
in ('succeeded', 'failed', 'cancelled'):
break
def testExecuteWorker(self):
mock_scheduler_addr = '127.0.0.1:%d' % get_next_port()
try:
with create_actor_pool(n_process=1, backend='gevent',
address=mock_scheduler_addr) as pool:
pool.create_actor(ClusterInfoActor, schedulers=[mock_scheduler_addr],
uid=ClusterInfoActor.default_name())
pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name())
resource_ref = pool.create_actor(ResourceActor, uid=ResourceActor.default_name())
proc = subprocess.Popen([sys.executable, '-m', 'mars.worker',
'-a', '127.0.0.1',
'--schedulers', mock_scheduler_addr,
'--cpu-procs', '1',
'--cache-mem', '10m',
'--ignore-avail-mem'])
worker_ips = []
def waiter():
check_time = time.time()
while True:
if not resource_ref.get_workers_meta():
gevent.sleep(0.5)
if proc.poll() is not None:
raise SystemError('Worker dead. exit code %s' % proc.poll())
if time.time() - check_time > 20:
raise SystemError('Check meta_timestamp timeout')
continue
else:
break
val = resource_ref.get_workers_meta()
worker_ips.extend(val.keys())
gl = gevent.spawn(waiter)
gl.join()
test_ref = pool.create_actor(WorkerProcessTestActor)
test_ref.run_test(worker_ips[0], _tell=True)
check_time = time.time()
while not test_ref.get_reply():
gevent.sleep(0.1)
if time.time() - check_time > 20:
raise SystemError('Check reply timeout')
finally:
if proc.poll() is None:
proc.send_signal(signal.SIGINT)
check_time = time.time()
while True:
time.sleep(1)
if proc.poll() is not None or time.time() - check_time >= 5:
break
if proc.poll() is None:
proc.kill()
if os.path.exists(options.worker.plasma_socket):
os.unlink(options.worker.plasma_socket)
def setUp(self):
endpoint = '127.0.0.1:%d' % get_next_port()
self.endpoint = endpoint
self.pool = create_actor_pool(n_process=1,
backend='gevent',
address=endpoint)
self.pool.create_actor(ClusterInfoActor, [endpoint],
uid=ClusterInfoActor.default_name())
self.pool.create_actor(SessionManagerActor,
uid=SessionManagerActor.default_name())
self.pool.create_actor(ResourceActor, uid=ResourceActor.default_name())
self.api = MarsAPI(endpoint)
def run_transfer_worker(pool_address, session_id, plasma_socket, chunk_keys,
spill_dir, msg_queue):
from mars.config import options
from mars.utils import PlasmaProcessHelper
options.worker.plasma_socket = plasma_socket
options.worker.spill_directory = spill_dir
plasma_helper = PlasmaProcessHelper(size=1024 * 1024 * 10, socket=options.worker.plasma_socket)
try:
plasma_helper.run()
with create_actor_pool(n_process=2, backend='gevent', distributor=WorkerDistributor(2),
address=pool_address) as pool:
try:
pool.create_actor(ClusterInfoActor, schedulers=[pool_address],
uid=ClusterInfoActor.default_name())
pool.create_actor(KVStoreActor, uid=KVStoreActor.default_name())
pool.create_actor(DispatchActor, uid=DispatchActor.default_name())
pool.create_actor(QuotaActor, 1024 * 1024 * 20, uid=MemQuotaActor.default_name())
holder_ref = pool.create_actor(HolderActor, uid='HolderActor')
chunk_holder_ref = pool.create_actor(ChunkHolderActor, plasma_helper._size,
uid=ChunkHolderActor.default_name())
pool.create_actor(SpillActor)
pool.create_actor(SenderActor, uid='%s' % str(uuid.uuid4()))
pool.create_actor(SenderActor, uid='%s' % str(uuid.uuid4()))
pool.create_actor(ReceiverActor, uid='%s' % str(uuid.uuid4()))
pool.create_actor(ReceiverActor, uid='%s' % str(uuid.uuid4()))
register_actor = pool.create_actor(WorkerRegistrationTestActor)
register_actor.register(session_id, chunk_keys)
check_time = time.time()
while not register_actor.get_finished():
gevent.sleep(0.5)
if time.time() - check_time > 60:
raise SystemError('Wait result timeout')
register_actor.destroy()
msg_queue.put(1)
check_time = time.time()
while not holder_ref.obtain():
gevent.sleep(1)
if time.time() - check_time > 60:
raise SystemError('Wait result timeout')
finally:
pool.destroy_actor(chunk_holder_ref)
finally:
plasma_helper.stop()
def testWorkerProcessRestart(self):
mock_scheduler_addr = '127.0.0.1:%d' % get_next_port()
try:
with create_actor_pool(n_process=1,
backend='gevent',
address=mock_scheduler_addr) as pool:
pool.create_actor(ClusterInfoActor,
schedulers=[mock_scheduler_addr],
uid=ClusterInfoActor.default_name())
pool.create_actor(ChunkMetaActor,
uid=ChunkMetaActor.default_name())
resource_ref = pool.create_actor(
ResourceActor, uid=ResourceActor.default_name())
proc = subprocess.Popen([
sys.executable, '-m', 'mars.worker', '-a', '127.0.0.1',
'--schedulers', mock_scheduler_addr, '--cpu-procs', '1',
'--cache-mem', '10m', '--spill-dir', self._spill_dir,
'--ignore-avail-mem'
])
worker_endpoint = self._wait_worker_ready(proc, resource_ref)
daemon_ref = pool.actor_ref(WorkerDaemonActor.default_name(),
address=worker_endpoint)
dispatch_ref = pool.actor_ref(DispatchActor.default_name(),
address=worker_endpoint)
cpu_slots = dispatch_ref.get_slots('cpu')
calc_ref = pool.actor_ref(cpu_slots[0],
address=worker_endpoint)
daemon_ref.kill_actor_process(calc_ref)
check_start = time.time()
while not daemon_ref.is_actor_process_alive(calc_ref):
gevent.sleep(0.1)
if time.time() - check_start > 10:
raise TimeoutError('Check process restart timeout')
finally:
if proc.poll() is None:
proc.send_signal(signal.SIGINT)
check_time = time.time()
while True:
time.sleep(0.1)
if proc.poll(
) is not None or time.time() - check_time >= 5:
break
if proc.poll() is None:
proc.kill()
if os.path.exists(options.worker.plasma_socket):
os.unlink(options.worker.plasma_socket)
def start_transfer_test_pool(**kwargs):
address = kwargs.pop('address')
plasma_size = kwargs.pop('plasma_size')
with create_actor_pool(n_process=1, backend='gevent', address=address, **kwargs) as pool:
pool.create_actor(PlasmaKeyMapActor, uid=PlasmaKeyMapActor.default_name())
pool.create_actor(ClusterInfoActor, schedulers=[address],
uid=ClusterInfoActor.default_name())
pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name())
pool.create_actor(DispatchActor, uid=DispatchActor.default_name())
pool.create_actor(QuotaActor, 1024 * 1024 * 20, uid=MemQuotaActor.default_name())
chunk_holder_ref = pool.create_actor(ChunkHolderActor,
plasma_size, uid=ChunkHolderActor.default_name())
pool.create_actor(SpillActor)
pool.create_actor(StatusActor, address, uid=StatusActor.default_name())
yield pool
chunk_holder_ref.destroy()
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:
pool.create_actor(ClusterInfoActor, [pool.cluster_info.address],
uid=ClusterInfoActor.default_name())
pool.create_actor(AssignerActor, uid=AssignerActor.default_name())
session_manager_ref = pool.create_actor(
SessionManagerActor, uid=SessionManagerActor.default_name())
resource_ref = pool.create_actor(ResourceActor, uid=ResourceActor.default_name())
chunk_meta_ref = pool.create_actor(
ChunkMetaActor, uid=ChunkMetaActor.default_name())
session_ref = pool.actor_ref(session_manager_ref.create_session(mock_session_id))
chunk_meta_ref.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_tensor_graph(None, mock_graph_key)
graph_ref = pool.actor_ref(GraphActor.gen_name(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 testExecuteWorker(self):
mock_scheduler_addr = '127.0.0.1:%d' % get_next_port()
try:
with create_actor_pool(n_process=1,
backend='gevent',
address=mock_scheduler_addr) as pool:
pool.create_actor(ClusterInfoActor,
schedulers=[mock_scheduler_addr],
uid=ClusterInfoActor.default_name())
pool.create_actor(ChunkMetaActor,
uid=ChunkMetaActor.default_name())
resource_ref = pool.create_actor(
ResourceActor, uid=ResourceActor.default_name())
proc = subprocess.Popen([
sys.executable, '-m', 'mars.worker', '-a', '127.0.0.1',
'--schedulers', mock_scheduler_addr, '--cpu-procs', '1',
'--cache-mem', '10m', '--spill-dir', self._spill_dir,
'--ignore-avail-mem'
])
worker_endpoint = self._wait_worker_ready(proc, resource_ref)
test_ref = pool.create_actor(WorkerProcessTestActor)
test_ref.run_test(worker_endpoint, _tell=True)
check_time = time.time()
while not test_ref.get_reply():
gevent.sleep(0.1)
if time.time() - check_time > 20:
raise TimeoutError('Check reply timeout')
finally:
if proc.poll() is None:
proc.send_signal(signal.SIGINT)
check_time = time.time()
while True:
time.sleep(0.1)
if proc.poll(
) is not None or time.time() - check_time >= 5:
break
if proc.poll() is None:
proc.kill()
if os.path.exists(options.worker.plasma_socket):
os.unlink(options.worker.plasma_socket)
def testStatus(self):
pool_address = '127.0.0.1:%d' % get_next_port()
with create_actor_pool(n_process=1,
backend='gevent',
address=pool_address) as pool:
pool.create_actor(ClusterInfoActor,
schedulers=[pool_address],
uid=ClusterInfoActor.default_name())
pool.create_actor(KVStoreActor, uid='KVStoreActor')
pool.create_actor(ChunkHolderActor,
self._plasma_helper._size,
uid='ChunkHolderActor')
pool.create_actor(StatusActor, '127.0.0.1:1234', uid='StatusActor')
def delay_read():
gevent.sleep(2)
return self._kv_store.read('/workers/meta', recursive=True)
gl = gevent.spawn(delay_read)
gl.join()
v = gl.value
print(v)
def testStatus(self):
pool_address = '127.0.0.1:%d' % get_next_port()
old_spill_dir = options.worker.spill_directory
dir_name = options.worker.spill_directory = tempfile.mkdtemp(
prefix='temp-mars-spill-')
try:
with create_actor_pool(n_process=1,
backend='gevent',
address=pool_address) as pool:
pool.create_actor(ClusterInfoActor,
schedulers=[pool_address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(
ResourceActor, uid=ResourceActor.default_name())
pool.create_actor(ChunkHolderActor,
self.plasma_storage_size,
uid=ChunkHolderActor.default_name())
status_ref = pool.create_actor(StatusActor,
pool_address,
uid=StatusActor.default_name())
status_ref.update_slots(dict(cpu=4))
status_ref.update_stats(dict(min_est_finish_time=10))
def delay_read():
gevent.sleep(1.5)
return resource_ref.get_workers_meta()
gl = gevent.spawn(delay_read)
gl.join()
v = gl.value
self.assertIsNotNone(v)
pool.destroy_actor(status_ref)
finally:
options.worker.spill_directory = old_spill_dir
shutil.rmtree(dir_name)
def start_processes(self,
n_schedulers=2,
n_workers=2,
etcd=False,
modules=None,
log_scheduler=True,
log_worker=True):
old_not_errors = gevent.hub.Hub.NOT_ERROR
gevent.hub.Hub.NOT_ERROR = (Exception, )
scheduler_ports = [str(get_next_port()) for _ in range(n_schedulers)]
self.scheduler_endpoints = ['127.0.0.1:' + p for p in scheduler_ports]
append_args = []
if modules:
append_args.extend(['--load-modules', ','.join(modules)])
if etcd:
etcd_port = get_next_port()
self.etcd_helper = EtcdProcessHelper(port_range_start=etcd_port)
self.etcd_helper.run()
options.kv_store = 'etcd://127.0.0.1:%s' % etcd_port
append_args.extend(['--kv-store', options.kv_store])
else:
append_args.extend(
['--schedulers', ','.join(self.scheduler_endpoints)])
if 'DUMP_GRAPH_DATA' in os.environ:
append_args += ['-Dscheduler.dump_graph_data=true']
self.proc_schedulers = [
subprocess.Popen([
sys.executable, '-m', 'mars.scheduler', '-H', '127.0.0.1',
'--level', 'debug' if log_scheduler else 'warning', '-p', p,
'--format', '%(asctime)-15s %(message)s',
'-Dscheduler.retry_delay=5'
] + append_args) for p in scheduler_ports
]
self.proc_workers = [
subprocess.Popen([
sys.executable, '-m', 'mars.worker', '-a', '127.0.0.1',
'--cpu-procs', '1', '--level',
'debug' if log_worker else 'warning', '--cache-mem', '16m',
'--ignore-avail-mem', '-Dworker.prepare_data_timeout=30'
] + append_args) for _ in range(n_workers)
]
actor_client = new_client()
self.cluster_info = actor_client.actor_ref(
ClusterInfoActor.default_name(),
address=self.scheduler_endpoints[0])
check_time = time.time()
while True:
try:
started_schedulers = self.cluster_info.get_schedulers()
if len(started_schedulers) < n_schedulers:
raise RuntimeError(
'Schedulers does not met requirement: %d < %d.' %
(len(started_schedulers), n_schedulers))
actor_address = self.cluster_info.get_scheduler(
SessionManagerActor.default_name())
self.session_manager_ref = actor_client.actor_ref(
SessionManagerActor.default_name(), address=actor_address)
actor_address = self.cluster_info.get_scheduler(
ResourceActor.default_name())
resource_ref = actor_client.actor_ref(
ResourceActor.default_name(), address=actor_address)
if resource_ref.get_worker_count() < n_workers:
raise RuntimeError(
'Workers does not met requirement: %d < %d.' %
(resource_ref.get_worker_count(), n_workers))
break
except:
if time.time() - check_time > 20:
raise
time.sleep(0.1)
gevent.hub.Hub.NOT_ERROR = old_not_errors
def testSimpleTransfer(self):
import tempfile
session_id = str(uuid.uuid4())
local_pool_addr = 'localhost:%d' % get_next_port()
remote_pool_addr = 'localhost:%d' % get_next_port()
remote_chunk_keys = [str(uuid.uuid4()) for _ in range(9)]
msg_queue = multiprocessing.Queue()
remote_spill_dir = os.path.join(
tempfile.gettempdir(),
'mars_spill_%d_%d' % (os.getpid(), id(run_transfer_worker)))
proc = multiprocessing.Process(target=run_transfer_worker,
args=(remote_pool_addr, session_id,
remote_chunk_keys,
remote_spill_dir, msg_queue))
proc.start()
try:
remote_plasma_socket = msg_queue.get(30)
except:
if proc.is_alive():
proc.terminate()
raise
with create_actor_pool(n_process=1,
distributor=WorkerDistributor(1),
backend='gevent',
address=local_pool_addr) as pool:
pool.create_actor(ClusterInfoActor,
schedulers=[local_pool_addr],
uid=ClusterInfoActor.default_name())
pool.create_actor(ChunkMetaActor,
uid=ChunkMetaActor.default_name())
pool.create_actor(DispatchActor, uid=DispatchActor.default_name())
pool.create_actor(QuotaActor,
1024 * 1024 * 20,
uid=MemQuotaActor.default_name())
cache_ref = pool.create_actor(ChunkHolderActor,
self.plasma_storage_size,
uid=ChunkHolderActor.default_name())
pool.create_actor(SpillActor)
sender_refs = [
pool.create_actor(SenderActor,
uid='w:1:%s' % str(uuid.uuid4())),
pool.create_actor(SenderActor,
uid='w:2:%s' % str(uuid.uuid4())),
]
receiver_refs = [
pool.create_actor(ReceiverActor,
uid='w:1:%s' % str(uuid.uuid4())),
pool.create_actor(ReceiverActor,
uid='w:1:%s' % str(uuid.uuid4())),
pool.create_actor(ReceiverActor,
uid='w:2:%s' % str(uuid.uuid4())),
pool.create_actor(ReceiverActor,
uid='w:2:%s' % str(uuid.uuid4())),
]
try:
for data_id in (-1, 1):
chunk_key = remote_chunk_keys[data_id]
with self.run_actor_test(pool) as test_actor:
from mars.worker.spill import build_spill_file_name
from mars.serialize import dataserializer
from numpy.testing import assert_array_equal
remote_dispatch_ref = test_actor.promise_ref(
DispatchActor.default_name(),
address=remote_pool_addr)
remote_plasma_client = plasma.connect(
remote_plasma_socket, '', 0)
remote_store = PlasmaChunkStore(remote_plasma_client)
def _call_send_data(sender_uid):
sender_ref = test_actor.promise_ref(
sender_uid, address=remote_pool_addr)
return sender_ref.send_data(session_id,
chunk_key,
local_pool_addr,
_promise=True)
def _test_data_exist(*_):
try:
local_data = test_actor._chunk_store.get(
session_id, chunk_key)
except KeyError:
with open(build_spill_file_name(chunk_key),
'rb') as spill_file:
local_data = dataserializer.load(
spill_file)
try:
remote_data = remote_store.get(
session_id, chunk_key)
except KeyError:
with open(
build_spill_file_name(
chunk_key, remote_spill_dir),
'rb') as spill_file:
remote_data = dataserializer.load(
spill_file)
assert_array_equal(local_data, remote_data)
del local_data, remote_data
remote_dispatch_ref.get_free_slot('sender', _promise=True) \
.then(_call_send_data) \
.then(_test_data_exist) \
.then(
lambda *_: test_actor.set_result(chunk_key),
lambda *exc: test_actor.set_result(exc, False),
)
self.assertEqual(self.get_result(60), chunk_key)
remote_holder_ref = pool.actor_ref('HolderActor',
address=remote_pool_addr)
remote_holder_ref.trigger()
finally:
for ref in sender_refs:
pool.destroy_actor(ref)
for ref in receiver_refs:
pool.destroy_actor(ref)
pool.destroy_actor(cache_ref)
os.unlink(remote_plasma_socket)
os.kill(proc.pid, signal.SIGINT)
t = time.time()
while proc.is_alive() and time.time() < t + 5:
time.sleep(1)
if proc.is_alive():
proc.terminate()
def testGraphActor(self):
session_id = str(uuid.uuid4())
graph_key = str(uuid.uuid4())
arr = mt.random.randint(10, size=(10, 8), chunk_size=4)
arr_add = mt.random.randint(10, size=(10, 8), chunk_size=4)
arr2 = arr + arr_add
graph = arr2.build_graph(compose=False)
serialized_graph = serialize_graph(graph)
chunked_graph = arr2.build_graph(compose=False, tiled=True)
addr = '127.0.0.1:%d' % get_next_port()
with create_actor_pool(n_process=1, backend='gevent',
address=addr) as pool:
pool.create_actor(ClusterInfoActor, [pool.cluster_info.address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(ResourceActor,
uid=ResourceActor.default_name())
pool.create_actor(ChunkMetaActor,
uid=ChunkMetaActor.default_name())
pool.create_actor(AssignerActor,
uid=AssignerActor.gen_name(session_id))
graph_ref = pool.create_actor(GraphActor,
session_id,
graph_key,
serialized_graph,
uid=GraphActor.gen_name(
session_id, graph_key))
graph_ref.prepare_graph(compose=False)
fetched_graph = graph_ref.get_chunk_graph()
self.assertIsNotNone(fetched_graph)
self.assertEqual(len(chunked_graph), len(fetched_graph))
graph_ref.scan_node()
op_infos = graph_ref.get_operand_info()
for n in fetched_graph:
depth = op_infos[n.op.key]['optimize']['depth']
self.assertIsNotNone(depth)
successor_size = op_infos[
n.op.key]['optimize']['successor_size']
self.assertIsNotNone(successor_size)
descendant_size = op_infos[
n.op.key]['optimize']['descendant_size']
self.assertIsNotNone(descendant_size)
resource_ref.set_worker_meta('localhost:12345',
dict(hardware=dict(cpu_total=4)))
resource_ref.set_worker_meta('localhost:23456',
dict(hardware=dict(cpu_total=4)))
graph_ref.place_initial_chunks()
op_infos = graph_ref.get_operand_info()
for n in fetched_graph:
if fetched_graph.count_predecessors(n) != 0:
continue
target_worker = op_infos[n.op.key]['target_worker']
self.assertIsNotNone(target_worker)
graph_ref.create_operand_actors(_clean_io_meta=False)
op_infos = graph_ref.get_operand_info()
for n in fetched_graph:
self.assertEqual(op_infos[n.op.key]['op_name'],
type(n.op).__name__)
io_meta = op_infos[n.op.key]['io_meta']
orig_io_meta = dict(
predecessors=list(
set(pn.op.key
for pn in fetched_graph.iter_predecessors(n))),
successors=list(
set(sn.op.key
for sn in fetched_graph.iter_successors(n))),
input_chunks=list(
set(pn.key
for pn in fetched_graph.iter_predecessors(n))),
chunks=list(c.key for c in n.op.outputs),
)
self.assertSetEqual(set(io_meta['predecessors']),
set(orig_io_meta['predecessors']))
self.assertSetEqual(set(io_meta['successors']),
set(orig_io_meta['successors']))
self.assertSetEqual(set(io_meta['input_chunks']),
set(orig_io_meta['input_chunks']))
self.assertSetEqual(set(io_meta['chunks']),
set(orig_io_meta['chunks']))
self.assertEqual(op_infos[n.op.key]['output_size'],
sum(ch.nbytes for ch in n.op.outputs))
def testSimpleTransfer(self):
import tempfile
session_id = str(uuid.uuid4())
local_pool_addr = 'localhost:%d' % get_next_port()
remote_pool_addr = 'localhost:%d' % get_next_port()
remote_chunk_keys = [str(uuid.uuid4()) for _ in range(9)]
msg_queue = multiprocessing.Queue()
remote_plasma_socket = '/tmp/plasma_%d_%d.sock' % (os.getpid(), id(run_transfer_worker))
remote_spill_dir = os.path.join(tempfile.gettempdir(),
'mars_spill_%d_%d' % (os.getpid(), id(run_transfer_worker)))
proc = multiprocessing.Process(
target=run_transfer_worker,
args=(remote_pool_addr, session_id, remote_plasma_socket,
remote_chunk_keys, remote_spill_dir, msg_queue)
)
proc.start()
try:
msg_queue.get(30)
except:
if proc.is_alive():
proc.terminate()
raise
with create_actor_pool(n_process=1, distributor=WorkerDistributor(3),
backend='gevent', address=local_pool_addr) as pool:
pool.create_actor(ClusterInfoActor, schedulers=[local_pool_addr],
uid=ClusterInfoActor.default_name())
pool.create_actor(KVStoreActor, uid=KVStoreActor.default_name())
pool.create_actor(DispatchActor, uid=DispatchActor.default_name())
pool.create_actor(QuotaActor, 1024 * 1024 * 20, uid=MemQuotaActor.default_name())
cache_ref = pool.create_actor(ChunkHolderActor, self.plasma_storage_size,
uid=ChunkHolderActor.default_name())
pool.create_actor(SpillActor)
sender_refs = [
pool.create_actor(SenderActor, uid='w:1:%s' % str(uuid.uuid4())),
pool.create_actor(SenderActor, uid='w:2:%s' % str(uuid.uuid4())),
]
receiver_refs = [
pool.create_actor(ReceiverActor, uid='w:1:%s' % str(uuid.uuid4())),
pool.create_actor(ReceiverActor, uid='w:1:%s' % str(uuid.uuid4())),
pool.create_actor(ReceiverActor, uid='w:2:%s' % str(uuid.uuid4())),
pool.create_actor(ReceiverActor, uid='w:2:%s' % str(uuid.uuid4())),
]
test_ref = pool.create_actor(TransferTestActor, local_pool_addr, remote_pool_addr,
remote_plasma_socket, remote_spill_dir)
try:
for data_id in (-1, 1):
chunk_key = remote_chunk_keys[data_id]
test_ref.do_transfer_test(session_id, chunk_key)
check_time = time.time()
while test_ref.get_results()[0] != chunk_key:
gevent.sleep(0.5)
if not proc.is_alive():
raise SystemError('Transfer worker dead. exit code %s' % proc.exitcode)
if time.time() - check_time > 60:
raise SystemError('Wait result timeout')
exc = test_ref.get_results()[1]
if exc:
six.reraise(*exc)
remote_holder_ref = pool.actor_ref('HolderActor', address=remote_pool_addr)
remote_holder_ref.trigger()
finally:
for ref in sender_refs:
pool.destroy_actor(ref)
for ref in receiver_refs:
pool.destroy_actor(ref)
pool.destroy_actor(cache_ref)
pool.destroy_actor(test_ref)
os.unlink(remote_plasma_socket)
if proc.is_alive():
proc.terminate()
def testOperandActorWithCancel(self, *_):
import logging
logging.basicConfig(level=logging.DEBUG)
arr = mt.random.randint(10, size=(10, 8), chunk_size=4)
arr_add = mt.random.randint(10, size=(10, 8), chunk_size=4)
arr2 = arr + arr_add
session_id = str(uuid.uuid4())
graph_key = str(uuid.uuid4())
graph = arr2.build_graph(compose=False)
with create_actor_pool(n_process=1, backend='gevent') as pool:
pool.create_actor(ClusterInfoActor, [pool.cluster_info.address],
uid=ClusterInfoActor.default_name())
resource_ref = pool.create_actor(ResourceActor, uid=ResourceActor.default_name())
pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name())
pool.create_actor(AssignerActor, uid=AssignerActor.default_name())
graph_ref = pool.create_actor(GraphActor, session_id, graph_key, serialize_graph(graph),
uid=GraphActor.gen_name(session_id, graph_key))
addr_dict = dict()
def _build_mock_ref(uid=None, address=None):
if address in addr_dict:
return addr_dict[address]
else:
r = addr_dict[address] = pool.create_actor(
FakeExecutionActor, exec_delay=0.2)
return r
# handle mock objects
OperandActor._get_raw_execution_ref.side_effect = _build_mock_ref
mock_resource = dict(hardware=dict(cpu=4, cpu_total=4, memory=512))
for idx in range(20):
resource_ref.set_worker_meta('localhost:%d' % (idx + 12345), mock_resource)
graph_ref.prepare_graph(compose=False)
fetched_graph = graph_ref.get_chunk_graph()
graph_ref.analyze_graph()
final_keys = set()
for c in fetched_graph:
if fetched_graph.count_successors(c) == 0:
final_keys.add(c.op.key)
graph_ref.create_operand_actors()
graph_meta_ref = pool.actor_ref(GraphMetaActor.gen_name(session_id, graph_key))
start_time = time.time()
cancel_called = False
while True:
pool.sleep(0.05)
if not cancel_called and time.time() > start_time + 0.3:
cancel_called = True
graph_ref.stop_graph(_tell=True)
if time.time() - start_time > 30:
raise SystemError('Wait for execution finish timeout')
if graph_meta_ref.get_state() in (GraphState.SUCCEEDED, GraphState.FAILED, GraphState.CANCELLED):
break
self.assertEqual(graph_meta_ref.get_state(), GraphState.CANCELLED)
