def testLocalCluster(self):
endpoint = gen_endpoint('0.0.0.0')
with LocalDistributedCluster(endpoint,
scheduler_n_process=2,
worker_n_process=3,
shared_memory='20M') 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 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(SchedulerClusterInfoActor, [endpoint],
uid=SchedulerClusterInfoActor.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 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 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