def testEmptyGraph(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(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()) pool.create_actor(AssignerActor, uid=AssignerActor.gen_uid(session_id)) 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_key = str(uuid.uuid4()) serialized_graph = serialize_graph(DAG()) graph_ref = pool.create_actor(GraphActor, session_id, graph_key, serialized_graph, uid=GraphActor.gen_uid( session_id, graph_key)) graph_ref.execute_graph() self.assertEqual(graph_ref.get_state(), GraphState.SUCCEEDED)
def _start_calc_pool(self): mock_addr = f'127.0.0.1:{get_next_port()}' with self.create_pool(n_process=1, backend='gevent', address=mock_addr) as pool: pool.create_actor(SchedulerClusterInfoActor, [mock_addr], uid=SchedulerClusterInfoActor.default_uid()) pool.create_actor(WorkerClusterInfoActor, [mock_addr], uid=WorkerClusterInfoActor.default_uid()) pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_uid()) pool.create_actor(StatusActor, mock_addr, uid=StatusActor.default_uid()) pool.create_actor(PlasmaKeyMapActor, uid=PlasmaKeyMapActor.default_uid()) pool.create_actor(WorkerDaemonActor, uid=WorkerDaemonActor.default_uid()) pool.create_actor(DispatchActor, uid=DispatchActor.default_uid()) pool.create_actor(StorageManagerActor, uid=StorageManagerActor.default_uid()) pool.create_actor(IORunnerActor) pool.create_actor(QuotaActor, 1024**2, uid=MemQuotaActor.default_uid()) shared_holder_ref = pool.create_actor( SharedHolderActor, uid=SharedHolderActor.default_uid()) pool.create_actor(InProcHolderActor) pool.create_actor(CpuCalcActor, uid=CpuCalcActor.default_uid()) with self.run_actor_test(pool) as test_actor: try: yield pool, test_actor finally: shared_holder_ref.destroy()
def _prepare_test_graph(self, session_id, graph_key, mock_workers): addr = f'127.0.0.1:{get_next_port()}' a1 = mt.random.random((100,)) a2 = mt.random.random((100,)) s = a1 + a2 v1, v2 = mt.split(s, 2) graph = TileableGraph([v1.data, v2.data]) builder = TileableGraphBuilder(graph) next(iter(builder.build())) with create_actor_pool(n_process=1, backend='gevent', address=addr) as pool: 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()) pool.create_actor(AssignerActor, uid=AssignerActor.gen_uid(session_id)) graph_ref = pool.create_actor(GraphActor, session_id, graph_key, serialize_graph(graph), uid=GraphActor.gen_uid(session_id, graph_key)) for w in mock_workers: resource_ref.set_worker_meta(w, dict(hardware=dict(cpu=4, cpu_total=4, memory=1600))) graph_ref.prepare_graph() graph_ref.analyze_graph() graph_ref.create_operand_actors(_start=False) yield pool, graph_ref
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(SchedulerClusterInfoActor, [address], uid=SchedulerClusterInfoActor.default_uid()) pool.create_actor(WorkerClusterInfoActor, [address], uid=WorkerClusterInfoActor.default_uid()) pool.create_actor(PlasmaKeyMapActor, uid=PlasmaKeyMapActor.default_uid()) pool.create_actor(StorageManagerActor, uid=StorageManagerActor.default_uid()) pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_uid()) pool.create_actor(DispatchActor, uid=DispatchActor.default_uid()) pool.create_actor(QuotaActor, 1024 * 1024 * 20, uid=MemQuotaActor.default_uid()) shared_holder_ref = pool.create_actor(SharedHolderActor, plasma_size, uid=SharedHolderActor.default_uid()) pool.create_actor(StatusActor, address, uid=StatusActor.default_uid()) pool.create_actor(IORunnerActor) pool.create_actor(StorageClientActor, uid=StorageClientActor.default_uid()) pool.create_actor(InProcHolderActor) pool.create_actor(ReceiverManagerActor, uid=ReceiverManagerActor.default_uid()) try: yield pool finally: shared_holder_ref.destroy()
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(SchedulerClusterInfoActor, schedulers=[pool_address], uid=SchedulerClusterInfoActor.default_uid()) pool.create_actor(WorkerClusterInfoActor, schedulers=[pool_address], uid=WorkerClusterInfoActor.default_uid()) resource_ref = pool.create_actor(ResourceActor, uid=ResourceActor.default_uid()) pool.create_actor(ChunkHolderActor, self.plasma_storage_size, uid=ChunkHolderActor.default_uid()) status_ref = pool.create_actor(StatusActor, pool_address, uid=StatusActor.default_uid()) status_ref.enable_status_upload() 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_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(SchedulerClusterInfoActor, schedulers=[address], uid=SchedulerClusterInfoActor.default_uid()) pool.create_actor(WorkerClusterInfoActor, schedulers=[address], uid=WorkerClusterInfoActor.default_uid()) pool.create_actor(PlasmaKeyMapActor, uid=PlasmaKeyMapActor.default_uid()) pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_uid()) pool.create_actor(DispatchActor, uid=DispatchActor.default_uid()) pool.create_actor(QuotaActor, 1024 * 1024 * 20, uid=MemQuotaActor.default_uid()) chunk_holder_ref = pool.create_actor( ChunkHolderActor, plasma_size, uid=ChunkHolderActor.default_uid()) pool.create_actor(SpillActor) pool.create_actor(StatusActor, address, uid=StatusActor.default_uid()) yield pool chunk_holder_ref.destroy()
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(SchedulerClusterInfoActor.default_uid()))) self.assertTrue( pool.has_actor( pool.actor_ref(SessionManagerActor.default_uid()))) self.assertTrue( pool.has_actor(pool.actor_ref(DispatchActor.default_uid()))) with new_session(endpoint) as session: api = session._api t = mt.ones((3, 3), chunk_size=2) result = session.run(t, timeout=_exec_timeout) np.testing.assert_array_equal(result, np.ones((3, 3))) self.assertNotIn(session._session_id, api.session_manager.get_sessions())
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 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(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()) pool.create_actor(AssignerActor, uid=AssignerActor.default_uid()) 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_uid(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_uid(session_id, graph_key)) def _mock_cancels(*_, **__): graph_meta_ref = pool.actor_ref(GraphMetaActor.gen_uid(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_uid(session_id, graph_key)) def _mock_cancels(*_, **__): graph_meta_ref = pool.actor_ref(GraphMetaActor.gen_uid(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 testOperandActorWithCancel(self, *_): 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(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()) pool.create_actor(AssignerActor, uid=AssignerActor.gen_uid(session_id)) graph_ref = pool.create_actor(GraphActor, session_id, graph_key, serialize_graph(graph), uid=GraphActor.gen_uid(session_id, graph_key)) def _build_mock_ref(uid=None, address=None): try: return pool.create_actor( FakeExecutionActor, exec_delay=0.2, uid=FakeExecutionActor.gen_uid(address)) except ActorAlreadyExist: return pool.actor_ref(FakeExecutionActor.gen_uid(address)) # 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_uid(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)
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(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()) pool.create_actor(AssignerActor, uid=AssignerActor.gen_uid(session_id)) graph_ref = pool.create_actor(GraphActor, session_id, graph_key, serialize_graph(graph), uid=GraphActor.gen_uid( session_id, graph_key)) def _build_mock_ref(uid=None, address=None): try: return execution_creator( pool, FakeExecutionActor.gen_uid(address)) except ActorAlreadyExist: return pool.actor_ref(FakeExecutionActor.gen_uid(address)) # 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_uid(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 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 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 _start_worker_process(self, cuda=False, cuda_device=None, extra_env=None, modules=None, check_timeout=None): mock_scheduler_addr = f'127.0.0.1:{get_next_port()}' try: with create_actor_pool(n_process=1, backend='gevent', address=mock_scheduler_addr) as pool: pool.create_actor(SchedulerClusterInfoActor, [mock_scheduler_addr], uid=SchedulerClusterInfoActor.default_uid()) pool.create_actor(SessionManagerActor, uid=SessionManagerActor.default_uid()) pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_uid()) resource_ref = pool.create_actor( ResourceActor, uid=ResourceActor.default_uid()) args = [ 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, '--log-level', 'debug', '--log-format', '%(asctime)-15s %(message)s', '--ignore-avail-mem' ] if modules: args.extend(['--load-modules', ','.join(modules)]) env = os.environ.copy() env.update(extra_env or dict()) if cuda: env['CUDA_VISIBLE_DEVICES'] = cuda_device proc = subprocess.Popen(args, env=env) worker_endpoint = self._wait_worker_ready( proc, resource_ref, timeout=check_timeout) yield pool, worker_endpoint 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 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 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(SchedulerClusterInfoActor, schedulers=[mock_scheduler_addr], uid=SchedulerClusterInfoActor.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 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(SchedulerClusterInfoActor, schedulers=[mock_scheduler_addr], uid=SchedulerClusterInfoActor.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 _start_worker_process(self, no_cuda=True, cuda_device=None): 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(SchedulerClusterInfoActor, [mock_scheduler_addr], uid=SchedulerClusterInfoActor.default_uid()) pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_uid()) resource_ref = pool.create_actor( ResourceActor, uid=ResourceActor.default_uid()) args = [ 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' ] env = os.environ.copy() if no_cuda: args.append('--no-cuda') else: env['CUDA_VISIBLE_DEVICES'] = cuda_device proc = subprocess.Popen(args, env=env) worker_endpoint = self._wait_worker_ready(proc, resource_ref) yield pool, worker_endpoint 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 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(SchedulerClusterInfoActor, [address], uid=SchedulerClusterInfoActor.default_uid()) pool.create_actor(WorkerClusterInfoActor, [address], uid=WorkerClusterInfoActor.default_uid()) pool.create_actor(PlasmaKeyMapActor, uid=PlasmaKeyMapActor.default_uid()) if with_resource: pool.create_actor(ResourceActor, uid=ResourceActor.default_uid()) if with_daemon: pool.create_actor(WorkerDaemonActor, uid=WorkerDaemonActor.default_uid()) if with_status: pool.create_actor(StatusActor, address, uid=StatusActor.default_uid()) pool.create_actor( ChunkHolderActor, cls.plasma_storage_size, uid=ChunkHolderActor.default_uid()) pool.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_uid()) pool.create_actor(DispatchActor, uid=DispatchActor.default_uid()) pool.create_actor(QuotaActor, quota_size, uid=MemQuotaActor.default_uid()) pool.create_actor(ExecutionActor, uid=ExecutionActor.default_uid())
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(SchedulerClusterInfoActor, [pool.cluster_info.address], uid=SchedulerClusterInfoActor.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_uid( 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 start_processes(self, n_schedulers=2, n_workers=2, etcd=False, cuda=False, modules=None, log_scheduler=True, log_worker=True, env=None): 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 = [] append_args_scheduler = [] append_args_worker = [] 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_scheduler += ['-Dscheduler.dump_graph_data=true'] if not cuda: append_args_worker += ['--no-cuda'] proc_env = os.environ.copy() if env: proc_env.update(env) self.proc_schedulers = [ subprocess.Popen([ sys.executable, '-m', 'mars.scheduler', '-H', '127.0.0.1', '-p', p, '--log-level', 'debug' if log_scheduler else 'warning', '--log-format', 'SCH%d %%(asctime)-15s %%(message)s' % idx, '-Dscheduler.retry_delay=5', '-Dscheduler.default_cpu_usage=0', '-Dscheduler.status_timeout=10' ] + append_args + append_args_scheduler, env=proc_env) for idx, p in enumerate(scheduler_ports) ] cuda_count = resource.cuda_count() self.proc_workers = [ subprocess.Popen([ sys.executable, '-m', 'mars.worker', '-a', '127.0.0.1', '--cpu-procs', '1', '--log-level', 'debug' if log_worker else 'warning', '--log-format', 'WOR%d %%(asctime)-15s %%(message)s' % idx, '--cache-mem', '16m', '--ignore-avail-mem', '--cuda-device', str(idx % cuda_count) if cuda_count else '0', '-Dworker.prepare_data_timeout=30' ] + append_args + append_args_worker, env=proc_env) for idx in range(n_workers) ] actor_client = new_client() self.cluster_info = actor_client.actor_ref( SchedulerClusterInfoActor.default_uid(), 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 ProcessRequirementUnmetError( 'Schedulers does not met requirement: %d < %d.' % (len(started_schedulers), n_schedulers)) actor_address = self.cluster_info.get_scheduler( SessionManagerActor.default_uid()) self.session_manager_ref = actor_client.actor_ref( SessionManagerActor.default_uid(), address=actor_address) actor_address = self.cluster_info.get_scheduler( ResourceActor.default_uid()) resource_ref = actor_client.actor_ref( ResourceActor.default_uid(), address=actor_address) if resource_ref.get_worker_count() < n_workers: raise ProcessRequirementUnmetError( '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 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
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)] ChunkMetaClient.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: cluster_info1 = pool1.create_actor( SchedulerClusterInfoActor, endpoints, uid=SchedulerClusterInfoActor.default_name()) pool1.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name()) with create_actor_pool(n_process=1, backend='gevent', address=endpoints[1]) as pool2: cluster_info2 = pool2.create_actor( SchedulerClusterInfoActor, endpoints, uid=SchedulerClusterInfoActor.default_name()) pool2.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name()) actor_client = new_client() client1 = ChunkMetaClient( actor_client, actor_client.actor_ref(cluster_info1)) client2 = ChunkMetaClient( actor_client, actor_client.actor_ref(cluster_info2)) local_ref1 = actor_client.actor_ref( ChunkMetaActor.default_name(), address=endpoints[0]) local_ref2 = actor_client.actor_ref( ChunkMetaActor.default_name(), address=endpoints[1]) key1 = str(uuid.uuid4()) key2 = str(uuid.uuid4()) key3 = str(uuid.uuid4()) keys = [key1, key2, key3] client1.set_chunk_broadcasts(session_id, key1, [endpoints[1]]) client1.set_chunk_size(session_id, key1, 512) client1.set_chunk_shape(session_id, key1, (10, ) * 2) client1.add_worker(session_id, key1, 'abc') client2.set_chunk_broadcasts(session_id, key2, [endpoints[0]]) client2.set_chunk_size(session_id, key2, 512) client1.set_chunk_shape(session_id, key2, (10, ) * 2) client2.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]]) client1.batch_set_chunk_broadcasts(session_id, [key3], [[endpoints[1]]]) meta3 = WorkerMeta(chunk_size=512, chunk_shape=(10, ) * 2, workers=(endpoints[0], )) local_ref1.batch_set_chunk_meta(session_id, [key3], [meta3]) self.assertEqual( local_ref2.get_chunk_meta(session_id, key3).chunk_size, 512) self.assertEqual( local_ref2.get_chunk_meta(session_id, key3).chunk_shape, (10, ) * 2) client1.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 testChunkMetaActors(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)] ChunkMetaClient.get_scheduler.side_effect = _mock_get_scheduler session1 = str(uuid.uuid4()) session2 = str(uuid.uuid4()) with create_actor_pool(n_process=1, backend='gevent', address=endpoints[0]) as pool1: cluster_info1 = pool1.create_actor( SchedulerClusterInfoActor, endpoints, uid=SchedulerClusterInfoActor.default_name()) pool1.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name()) with create_actor_pool(n_process=1, backend='gevent', address=endpoints[1]) as pool2: cluster_info2 = pool2.create_actor( SchedulerClusterInfoActor, endpoints, uid=SchedulerClusterInfoActor.default_name()) pool2.create_actor(ChunkMetaActor, uid=ChunkMetaActor.default_name()) actor_client = new_client() client1 = ChunkMetaClient( actor_client, actor_client.actor_ref(cluster_info1)) client2 = ChunkMetaClient( actor_client, actor_client.actor_ref(cluster_info2)) loc_ref1 = actor_client.actor_ref( ChunkMetaActor.default_name(), address=endpoints[0]) loc_ref2 = actor_client.actor_ref( ChunkMetaActor.default_name(), address=endpoints[1]) key1 = (str(uuid.uuid4()), str(uuid.uuid4())) key2 = str(uuid.uuid4()) key3 = str(uuid.uuid4()) key4 = (str(uuid.uuid4()), str(uuid.uuid4())) key5 = str(uuid.uuid4()) key6 = str(uuid.uuid4()) keys = [key1, key2, key3, key4, key5, key6] client1.set_chunk_size(session1, key1, 512) client2.set_chunk_size(session1, key2, 1024) client2.set_chunk_size(session2, key3, 1024) self.assertEqual(client1.get_chunk_size(session1, key1), 512) self.assertEqual(client2.get_chunk_size(session1, key2), 1024) self.assertEqual(client1.get_chunk_size(session1, key2), 1024) self.assertEqual(client2.get_chunk_size(session1, key1), 512) self.assertListEqual( client1.batch_get_chunk_size(session1, [key1, key2]), [512, 1024]) self.assertListEqual( client2.batch_get_chunk_size(session1, [key1, key2]), [512, 1024]) client1.set_chunk_shape(session1, key1, (10, )) client2.set_chunk_shape(session1, key2, (10, ) * 2) client2.set_chunk_shape(session2, key3, (10, ) * 2) self.assertEqual(client1.get_chunk_shape(session1, key1), (10, )) self.assertEqual(client2.get_chunk_shape(session1, key2), (10, ) * 2) self.assertEqual(client1.get_chunk_shape(session1, key2), (10, ) * 2) self.assertEqual(client2.get_chunk_shape(session1, key1), (10, )) self.assertListEqual( client1.batch_get_chunk_shape(session1, [key1, key2]), [(10, ), (10, ) * 2]) self.assertListEqual( client2.batch_get_chunk_shape(session1, [key1, key2]), [(10, ), (10, ) * 2]) mock_endpoint = '127.0.0.1:%d' % get_next_port() with create_actor_pool(n_process=1, backend='gevent', address=mock_endpoint) as pool3: cluster_info3 = pool3.create_actor( SchedulerClusterInfoActor, endpoints, uid=SchedulerClusterInfoActor.default_name()) client3 = ChunkMetaClient( actor_client, actor_client.actor_ref(cluster_info3)) self.assertListEqual( client3.batch_get_chunk_shape(session1, [key1, key2]), [(10, ), (10, ) * 2]) client1.add_worker(session1, key1, 'abc') client1.add_worker(session1, key1, 'def') client2.add_worker(session1, key2, 'ghi') client1.add_worker(session2, key3, 'ghi') self.assertEqual(sorted(client1.get_workers(session1, key1)), sorted(('abc', 'def'))) self.assertEqual(sorted(client2.get_workers(session1, key2)), sorted(('ghi', ))) batch_result = client1.batch_get_workers( session1, [key1, key2]) self.assertEqual(sorted(batch_result[0]), sorted( ('abc', 'def'))) self.assertEqual(sorted(batch_result[1]), sorted(('ghi', ))) affected = [] for loc_ref in (loc_ref1, loc_ref2): affected.extend( loc_ref.remove_workers_in_session(session2, ['ghi'])) self.assertEqual(affected, [key3]) self.assertEqual(sorted(client1.get_workers(session1, key2)), sorted(('ghi', ))) self.assertIsNone(client1.get_workers(session2, key3)) client1.delete_meta(session1, key1) self.assertIsNone(client1.get_workers(session1, key1)) self.assertIsNone( client1.batch_get_chunk_size(session1, [key1, key2])[0]) self.assertIsNone( client1.batch_get_workers(session1, [key1, key2])[0]) client2.batch_delete_meta(session1, [key1, key2]) self.assertIsNone(client1.get_workers(session1, key2)) self.assertIsNone( client1.batch_get_chunk_size(session1, [key1, key2])[1]) self.assertIsNone( client1.batch_get_workers(session1, [key1, key2])[1]) meta4 = WorkerMeta(chunk_size=512, chunk_shape=(10, ) * 2, workers=(endpoints[0], )) loc_ref2.batch_set_chunk_meta(session1, [key4], [meta4]) self.assertEqual( loc_ref2.get_chunk_meta(session1, key4).chunk_size, 512) self.assertEqual( loc_ref2.get_chunk_meta(session1, key4).chunk_shape, (10, ) * 2) meta5 = WorkerMeta(chunk_size=512, chunk_shape=(10, ) * 2, workers=(endpoints[0], )) meta6 = WorkerMeta(chunk_size=512, chunk_shape=(10, ) * 2, workers=(endpoints[0], )) client1.batch_set_chunk_meta(session1, [key5, key6], [meta5, meta6]) self.assertEqual( loc_ref1.get_chunk_meta(session1, key5).chunk_size, 512) self.assertEqual( loc_ref2.get_chunk_meta(session1, key6).chunk_size, 512)
def prepare_graph_in_pool(self, expr, clean_io_meta=True, compose=False): session_id = str(uuid.uuid4()) graph_key = str(uuid.uuid4()) graph = expr.build_graph(compose=compose) serialized_graph = serialize_graph(graph) chunked_graph = expr.build_graph(compose=compose, 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(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()) pool.create_actor(AssignerActor, uid=AssignerActor.gen_uid(session_id)) graph_ref = pool.create_actor(GraphActor, session_id, graph_key, serialized_graph, uid=GraphActor.gen_uid( session_id, graph_key)) graph_ref.prepare_graph(compose=compose) fetched_graph = graph_ref.get_chunk_graph() self.assertIsNotNone(fetched_graph) self.assertEqual(len(chunked_graph), len(fetched_graph)) graph_ref.analyze_graph(do_placement=False) 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.analyze_graph() 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_info=clean_io_meta) op_infos = graph_ref.get_operand_info() if not clean_io_meta: orig_metas = dict() for n in fetched_graph: try: meta = orig_metas[n.op.key] except KeyError: meta = orig_metas[n.op.key] = dict(predecessors=set(), successors=set(), input_chunks=set(), chunks=set()) meta['predecessors'].update([ pn.op.key for pn in fetched_graph.iter_predecessors(n) ]) meta['successors'].update( [sn.op.key for sn in fetched_graph.iter_successors(n)]) meta['input_chunks'].update( [pn.key for pn in fetched_graph.iter_predecessors(n)]) meta['chunks'].update([c.key for c in n.op.outputs]) 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 = orig_metas[n.op.key] 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'])) yield pool, graph_ref
def _start_processes(self, n_schedulers=2, n_workers=2, etcd=False, cuda=False, modules=None, log_scheduler=True, log_worker=True, env=None, scheduler_args=None, worker_args=None, worker_cpu=1): 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 = [] append_args_scheduler = scheduler_args or [] append_args_worker = worker_args or [] 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 = f'etcd://127.0.0.1:{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_scheduler += ['-Dscheduler.dump_graph_data=true'] proc_env = os.environ.copy() if env: proc_env.update(env) self.proc_schedulers = [ subprocess.Popen([sys.executable, '-m', 'mars.scheduler', '-H', '127.0.0.1', '-p', p, '--log-level', 'debug' if log_scheduler else 'warning', '--log-format', f'SCH{idx} %(asctime)-15s %(message)s' '-Dscheduler.retry_delay=5', '-Dscheduler.default_cpu_usage=0', '-Dscheduler.status_timeout=10'] + append_args + append_args_scheduler, env=proc_env) for idx, p in enumerate(scheduler_ports)] cuda_count = resource.cuda_count() cuda_devices = [int(d) for d in os.environ['CUDA_VISIBLE_DEVICES'].split(',')] \ if os.environ.get('CUDA_VISIBLE_DEVICES') else list(range(cuda_count)) self.proc_workers = [ subprocess.Popen([sys.executable, '-m', 'mars.worker', '-a', '127.0.0.1', '--cpu-procs', str(worker_cpu), '--log-level', 'debug' if log_worker else 'warning', '--log-format', f'WOR{idx} %(asctime)-15s %(message)s', '--cache-mem', '16m', '--ignore-avail-mem', '--cuda-device', str(cuda_devices[idx % cuda_count]) if cuda_count else '', '-Dworker.prepare_data_timeout=30'] + append_args + append_args_worker, env=proc_env) for idx in range(n_workers) ] actor_client = new_client() self.cluster_info = actor_client.actor_ref( SchedulerClusterInfoActor.default_uid(), address=self.scheduler_endpoints[0]) check_time = time.time() while True: try: try: started_schedulers = self.cluster_info.get_schedulers() except Exception as e: raise ProcessRequirementUnmetError(f'Failed to get scheduler numbers, {e}') if len(started_schedulers) < n_schedulers: raise ProcessRequirementUnmetError( f'Schedulers does not met requirement: {len(started_schedulers)} < {n_schedulers}.') actor_address = self.cluster_info.get_scheduler(SessionManagerActor.default_uid()) self.session_manager_ref = actor_client.actor_ref( SessionManagerActor.default_uid(), address=actor_address) actor_address = self.cluster_info.get_scheduler(ResourceActor.default_uid()) resource_ref = actor_client.actor_ref(ResourceActor.default_uid(), address=actor_address) if not actor_client.has_actor(self.session_manager_ref) \ or resource_ref.get_worker_count() < n_workers: raise ProcessRequirementUnmetError( f'Workers does not met requirement: {resource_ref.get_worker_count()} < {n_workers}') break except: # noqa: E722 if time.time() - check_time > 20: raise time.sleep(0.1) gevent.hub.Hub.NOT_ERROR = old_not_errors