[Distributed] Support multiple servers (#1886)

* client init graph on the backup servers.

* fix.

* test multi-server.

* fix anonymous dist tensors.

* check #parts.

* fix init_data

* add multi-server multi-client tests.

* update tests in kvstore.

* fix.

* verify the loaded partition.

* fix a bug.

* fix lint.

* fix.

* fix example.

* fix rpc.

* fix pull/push handler for backup kvstore

* fix example readme.

* change ip.

* update docstring.
Co-authored-by: Ubuntu <ubuntu@ip-172-31-19-1.us-west-2.compute.internal>

examples/pytorch/graphsage/experimental/README.md

@@ -55,7 +55,7 @@ To run unsupervised training:
 python3 ~/dgl/tools/launch.py \
 --workspace ~/dgl/examples/pytorch/graphsage/experimental \
 --num_client 4 \
---conf_path data/ogb-product.json \
+--part_config data/ogb-product.json \
 --ip_config ip_config.txt \
 "python3 train_dist_unsupervised.py --graph-name ogb-product --ip_config ip_config.txt --num-epochs 3 --batch-size 1000 --num-client 4"
 ```
@@ -76,13 +76,13 @@ python3 partition_graph.py --dataset ogb-product --num_parts 1
 To run supervised training:
 
 ```bash
-python3 train_dist.py --graph-name ogb-product --ip_config ip_config.txt --num-epochs 3 --batch-size 1000 --conf_path data/ogb-product.json --standalone
+python3 train_dist.py --graph-name ogb-product --ip_config ip_config.txt --num-epochs 3 --batch-size 1000 --part_config data/ogb-product.json --standalone
 ```
 
 To run unsupervised training:
 
 ```bash
-python3 train_dist_unsupervised.py --graph-name ogb-product --ip_config ip_config.txt --num-epochs 3 --batch-size 1000 --conf_path data/ogb-product.json --standalone
+python3 train_dist_unsupervised.py --graph-name ogb-product --ip_config ip_config.txt --num-epochs 3 --batch-size 1000 --part_config data/ogb-product.json --standalone
 ```
 
 Note: please ensure that all environment variables shown above are unset if they were set for testing distributed training.

examples/pytorch/graphsage/experimental/ip_config.txt

-172.31.16.250 5555 1
-172.31.30.135 5555 1
-172.31.27.41 5555 1
-172.31.30.149 5555 1
+172.31.19.1 5555 2
+172.31.23.205 5555 2
+172.31.29.175 5555 2
+172.31.16.98 5555 2

examples/pytorch/graphsage/experimental/train_dist_unsupervised.py

@@ -349,7 +349,7 @@ def run(args, device, data):
 def main(args):
     if not args.standalone:
         th.distributed.init_process_group(backend='gloo')
-    g = dgl.distributed.DistGraph(args.ip_config, args.graph_name, conf_file=args.conf_path)
+    g = dgl.distributed.DistGraph(args.ip_config, args.graph_name, part_config=args.conf_path)
     print('rank:', g.rank())
     print('number of edges', g.number_of_edges())
 

python/dgl/distributed/dist_graph.py

@@ -12,7 +12,7 @@ from .kvstore import KVServer, KVClient
 from .standalone_kvstore import KVClient as SA_KVClient
 from .._ffi.ndarray import empty_shared_mem
 from ..frame import infer_scheme
-from .partition import load_partition
+from .partition import load_partition, load_partition_book
 from .graph_partition_book import PartitionPolicy, get_shared_mem_partition_book
 from .graph_partition_book import NODE_PART_POLICY, EDGE_PART_POLICY
 from .shared_mem_utils import _to_shared_mem, _get_ndata_path, _get_edata_path, DTYPE_DICT
@@ -23,6 +23,41 @@ from .rpc_server import start_server
 from .graph_services import find_edges as dist_find_edges
 from .dist_tensor import DistTensor, _get_data_name
 
+INIT_GRAPH = 800001
+
+class InitGraphRequest(rpc.Request):
+    """ Init graph on the backup servers.
+
+    When the backup server starts, they don't load the graph structure.
+    This request tells the backup servers that they can map to the graph structure
+    with shared memory.
+    """
+    def __init__(self, graph_name):
+        self._graph_name = graph_name
+
+    def __getstate__(self):
+        return self._graph_name
+
+    def __setstate__(self, state):
+        self._graph_name = state
+
+    def process_request(self, server_state):
+        if server_state.graph is None:
+            server_state.graph = _get_graph_from_shared_mem(self._graph_name)
+        return InitGraphResponse(self._graph_name)
+
+class InitGraphResponse(rpc.Response):
+    """ Ack the init graph request
+    """
+    def __init__(self, graph_name):
+        self._graph_name = graph_name
+
+    def __getstate__(self):
+        return self._graph_name
+
+    def __setstate__(self, state):
+        self._graph_name = state
+
 def _copy_graph_to_shared_mem(g, graph_name):
     new_g = g.shared_memory(graph_name, formats='csc')
     # We should share the node/edge data to the client explicitly instead of putting them
@@ -218,16 +253,20 @@ class DistGraphServer(KVServer):
                                               num_clients=num_clients)
         self.ip_config = ip_config
         # Load graph partition data.
-        self.client_g, node_feats, edge_feats, self.gpb, graph_name = load_partition(part_config,
-                                                                                     server_id)
-        print('load ' + graph_name)
-        if not disable_shared_mem:
-            self.client_g = _copy_graph_to_shared_mem(self.client_g, graph_name)
+        if self.is_backup_server():
+            # The backup server doesn't load the graph partition. It'll initialized afterwards.
+            self.gpb, graph_name = load_partition_book(part_config, self.part_id)
+            self.client_g = None
+        else:
+            self.client_g, node_feats, edge_feats, self.gpb, \
+                    graph_name = load_partition(part_config, self.part_id)
+            print('load ' + graph_name)
+            if not disable_shared_mem:
+                self.client_g = _copy_graph_to_shared_mem(self.client_g, graph_name)
 
-        # Init kvstore.
         if not disable_shared_mem:
             self.gpb.shared_memory(graph_name)
-        assert self.gpb.partid == server_id
+        assert self.gpb.partid == self.part_id
         self.add_part_policy(PartitionPolicy(NODE_PART_POLICY, self.gpb))
         self.add_part_policy(PartitionPolicy(EDGE_PART_POLICY, self.gpb))
 
@@ -240,20 +279,13 @@ class DistGraphServer(KVServer):
                 self.init_data(name=_get_data_name(name, EDGE_PART_POLICY),
                                policy_str=EDGE_PART_POLICY,
                                data_tensor=edge_feats[name])
-        else:
-            for name in node_feats:
-                self.init_data(name=_get_data_name(name, NODE_PART_POLICY),
-                               policy_str=NODE_PART_POLICY)
-            for name in edge_feats:
-                self.init_data(name=_get_data_name(name, EDGE_PART_POLICY),
-                               policy_str=EDGE_PART_POLICY)
 
     def start(self):
         """ Start graph store server.
         """
         # start server
         server_state = ServerState(kv_store=self, local_g=self.client_g, partition_book=self.gpb)
-        print('start graph service on server ' + str(self.server_id))
+        print('start graph service on server {} for part {}'.format(self.server_id, self.part_id))
         start_server(server_id=self.server_id, ip_config=self.ip_config,
                      num_clients=self.num_clients, server_state=server_state)
 
@@ -314,6 +346,12 @@ class DistGraph:
             self._gpb = get_shared_mem_partition_book(graph_name, self._g)
             if self._gpb is None:
                 self._gpb = gpb
+
+            # Tell the backup servers to load the graph structure from shared memory.
+            for server_id in range(self._client.num_servers):
+                rpc.send_request(server_id, InitGraphRequest(graph_name))
+            for server_id in range(self._client.num_servers):
+                rpc.recv_response()
             self._client.barrier()
             self._client.map_shared_data(self._gpb)
 
@@ -692,3 +730,5 @@ def edge_split(edges, partition_book=None, rank=None, force_even=True):
         # Get all edges that belong to the rank.
         local_eids = partition_book.partid2eids(partition_book.partid)
         return _split_local(partition_book, rank, edges, local_eids)
+
+rpc.register_service(INIT_GRAPH, InitGraphRequest, InitGraphResponse)

python/dgl/distributed/dist_tensor.py

 """Define distributed tensor."""
 
 import os
-import uuid
 
 from .graph_partition_book import PartitionPolicy, NODE_PART_POLICY, EDGE_PART_POLICY
 from .rpc_client import is_initialized
@@ -71,13 +70,15 @@ class DistTensor:
 
         if init_func is None:
             init_func = _default_init_data
+        exist_names = g._client.data_name_list()
         # If a user doesn't provide a name, we generate a name ourselves.
+        # We need to generate the name in a deterministic way.
         if name is None:
             assert not persistent, 'We cannot generate anonymous persistent distributed tensors'
-            name = uuid.uuid4().hex[:10]
+            name = 'anonymous-' + str(len(exist_names) + 1)
         self._name = _get_data_name(name, part_policy.policy_str)
         self._persistent = persistent
-        if self._name not in g._client.data_name_list():
+        if self._name not in exist_names:
             g._client.init_data(self._name, shape, dtype, part_policy, init_func)
             self._owner = True
         else:

python/dgl/distributed/kvstore.py

@@ -162,8 +162,7 @@ class InitDataRequest(rpc.Request):
                                policy_str=self.policy_str,
                                data_tensor=data_tensor)
         else:
-            kv_store.init_data(name=self.name,
-                               policy_str=self.policy_str)
+            kv_store.init_data(name=self.name, policy_str=self.policy_str)
         res = InitDataResponse(INIT_MSG)
         return res
 
@@ -439,18 +438,26 @@ class SendMetaToBackupRequest(rpc.Request):
         data shape
     policy_str : str
         partition-policy string, e.g., 'edge' or 'node'.
+    pull_handler : callable
+        The callback function when data is pulled from kvstore.
+    push_handler : callable
+        The callback function when data is pushed to kvstore.
     """
-    def __init__(self, name, dtype, shape, policy_str):
+    def __init__(self, name, dtype, shape, policy_str, pull_handler, push_handler):
         self.name = name
         self.dtype = dtype
         self.shape = shape
         self.policy_str = policy_str
+        self.pull_handler = pull_handler
+        self.push_handler = push_handler
 
     def __getstate__(self):
-        return self.name, self.dtype, self.shape, self.policy_str
+        return self.name, self.dtype, self.shape, self.policy_str, self.pull_handler, \
+                self.push_handler
 
     def __setstate__(self, state):
-        self.name, self.dtype, self.shape, self.policy_str = state
+        self.name, self.dtype, self.shape, self.policy_str, self.pull_handler, \
+                self.push_handler = state
 
     def process_request(self, server_state):
         kv_store = server_state.kv_store
@@ -460,6 +467,8 @@ class SendMetaToBackupRequest(rpc.Request):
             dlpack = shared_data.to_dlpack()
             kv_store.data_store[self.name] = F.zerocopy_from_dlpack(dlpack)
             kv_store.part_policy[self.name] = kv_store.find_policy(self.policy_str)
+            kv_store.pull_handlers[self.name] = self.pull_handler
+            kv_store.push_handlers[self.name] = self.push_handler
         res = SendMetaToBackupResponse(SEND_META_TO_BACKUP_MSG)
         return res
 
@@ -895,6 +904,11 @@ class KVClient(object):
         """Get machine ID"""
         return self._machine_id
 
+    @property
+    def num_servers(self):
+        """Get the number of servers"""
+        return self._server_count
+
     def barrier(self):
         """Barrier for all client nodes.
 
@@ -1032,6 +1046,21 @@ class KVClient(object):
         self._full_data_shape[name] = tuple(shape)
         self._pull_handlers[name] = default_pull_handler
         self._push_handlers[name] = default_push_handler
+
+        # Now we need to tell the backup server the new tensor.
+        if self._client_id % num_clients_per_part == 0:
+            request = SendMetaToBackupRequest(name, F.reverse_data_type_dict[dtype],
+                                              part_shape, part_policy.policy_str,
+                                              self._pull_handlers[name],
+                                              self._push_handlers[name])
+            # send request to all the backup server nodes
+            for i in range(self._group_count-1):
+                server_id = self._machine_id * self._group_count + i + 1
+                rpc.send_request(server_id, request)
+            # recv response from all the backup server nodes
+            for _ in range(self._group_count-1):
+                response = rpc.recv_response()
+                assert response.msg == SEND_META_TO_BACKUP_MSG
         self.barrier()
 
     def delete_data(self, name):
@@ -1047,7 +1076,7 @@ class KVClient(object):
         self.barrier()
         part_policy = self._part_policy[name]
         num_partitions = part_policy.partition_book.num_partitions()
-        num_clients_per_part = rpc.get_num_client() / num_partitions
+        num_clients_per_part = rpc.get_num_client() // num_partitions
         if self._client_id % num_clients_per_part == 0:
             # send request to every server nodes
             request = DeleteDataRequest(name)
@@ -1108,7 +1137,9 @@ class KVClient(object):
         # Send meta data to backup servers
         for name, meta in response.meta.items():
             shape, dtype, policy_str = meta
-            request = SendMetaToBackupRequest(name, dtype, shape, policy_str)
+            request = SendMetaToBackupRequest(name, dtype, shape, policy_str,
+                                              self._pull_handlers[name],
+                                              self._push_handlers[name])
             # send request to all the backup server nodes
             for i in range(self._group_count-1):
                 server_id = self._machine_id * self._group_count + i + 1

python/dgl/distributed/partition.py

@@ -138,6 +138,9 @@ def load_partition(conf_file, part_id):
     assert EID in graph.edata, "the partition graph should contain edge mapping to global edge Id"
 
     gpb, graph_name = load_partition_book(conf_file, part_id, graph)
+    nids = F.boolean_mask(graph.ndata[NID], graph.ndata['inner_node'])
+    partids = gpb.nid2partid(nids)
+    assert np.all(F.asnumpy(partids == part_id)), 'load a wrong partition'
     return graph, node_feats, edge_feats, gpb, graph_name
 
 def load_partition_book(conf_file, part_id, graph=None):
@@ -162,6 +165,8 @@ def load_partition_book(conf_file, part_id, graph=None):
     with open(conf_file) as conf_f:
         part_metadata = json.load(conf_f)
     assert 'num_parts' in part_metadata, 'num_parts does not exist.'
+    assert part_metadata['num_parts'] > part_id, \
+            'part {} is out of range (#parts: {})'.format(part_id, part_metadata['num_parts'])
     num_parts = part_metadata['num_parts']
     assert 'num_nodes' in part_metadata, "cannot get the number of nodes of the global graph."
     assert 'num_edges' in part_metadata, "cannot get the number of edges of the global graph."

python/dgl/distributed/rpc.py

@@ -766,7 +766,9 @@ def send_requests_to_machine(target_and_requests):
         service_id = request.service_id
         msg_seq = incr_msg_seq()
         client_id = get_rank()
-        server_id = target
+
+        server_id = random.randint(target*get_num_server_per_machine(),
+                                   (target+1)*get_num_server_per_machine()-1)
         data, tensors = serialize_to_payload(request)
         msg = RPCMessage(service_id, msg_seq, client_id, server_id, data, tensors)
         send_rpc_message(msg, server_id)

python/dgl/distributed/server_state.py

@@ -42,7 +42,7 @@ class ServerState:
 
     def __init__(self, kv_store, local_g, partition_book):
         self._kv_store = kv_store
-        self.graph = local_g
+        self._graph = local_g
         self.partition_book = partition_book
 
     @property

tests/distributed/test_dist_graph_store.py

@@ -66,14 +66,14 @@ def emb_init(shape, dtype):
 def rand_init(shape, dtype):
     return F.tensor(np.random.normal(size=shape), F.float32)
 
-def run_client(graph_name, part_id, num_nodes, num_edges):
+def run_client(graph_name, part_id, num_clients, num_nodes, num_edges):
     time.sleep(5)
     gpb, graph_name = load_partition_book('/tmp/dist_graph/{}.json'.format(graph_name),
                                           part_id, None)
     g = DistGraph("kv_ip_config.txt", graph_name, gpb=gpb)
-    check_dist_graph(g, num_nodes, num_edges)
+    check_dist_graph(g, num_clients, num_nodes, num_edges)
 
-def check_dist_graph(g, num_nodes, num_edges):
+def check_dist_graph(g, num_clients, num_nodes, num_edges):
     # Test API
     assert g.number_of_nodes() == num_nodes
     assert g.number_of_edges() == num_edges
@@ -129,7 +129,8 @@ def check_dist_graph(g, num_nodes, num_edges):
         loss.backward()
         optimizer.step()
         feats = emb(nids)
-        assert_almost_equal(F.asnumpy(feats), np.ones((len(nids), 1)) * -lr)
+        if num_clients == 1:
+            assert_almost_equal(F.asnumpy(feats), np.ones((len(nids), 1)) * -lr)
         rest = np.setdiff1d(np.arange(g.number_of_nodes()), F.asnumpy(nids))
         feats1 = emb(rest)
         assert np.all(F.asnumpy(feats1) == np.zeros((len(rest), 1)))
@@ -137,7 +138,7 @@ def check_dist_graph(g, num_nodes, num_edges):
         policy = dgl.distributed.PartitionPolicy('node', g.get_partition_book())
         grad_sum = dgl.distributed.DistTensor(g, (g.number_of_nodes(),), F.float32,
                                               'emb1_sum', policy)
-        assert np.all(F.asnumpy(grad_sum[nids]) == np.ones((len(nids), 1)))
+        assert np.all(F.asnumpy(grad_sum[nids]) == np.ones((len(nids), 1)) * num_clients)
         assert np.all(F.asnumpy(grad_sum[rest]) == np.zeros((len(rest), 1)))
 
         emb = DistEmbedding(g, g.number_of_nodes(), 1, 'emb2', emb_init)
@@ -156,7 +157,8 @@ def check_dist_graph(g, num_nodes, num_edges):
         optimizer.step()
         with F.no_grad():
             feats = emb(nids)
-        assert_almost_equal(F.asnumpy(feats), np.ones((len(nids), 1)) * math.sqrt(2) * -lr)
+        if num_clients == 1:
+            assert_almost_equal(F.asnumpy(feats), np.ones((len(nids), 1)) * math.sqrt(2) * -lr)
         rest = np.setdiff1d(np.arange(g.number_of_nodes()), F.asnumpy(nids))
         feats1 = emb(rest)
         assert np.all(F.asnumpy(feats1) == np.zeros((len(rest), 1)))
@@ -188,8 +190,8 @@ def check_dist_graph(g, num_nodes, num_edges):
 
     print('end')
 
-def check_server_client(shared_mem):
-    prepare_dist()
+def check_server_client(shared_mem, num_servers, num_clients):
+    prepare_dist(num_servers)
     g = create_random_graph(10000)
 
     # Partition the graph
@@ -203,15 +205,16 @@ def check_server_client(shared_mem):
     # We cannot run multiple servers and clients on the same machine.
     serv_ps = []
     ctx = mp.get_context('spawn')
-    for serv_id in range(1):
-        p = ctx.Process(target=run_server, args=(graph_name, serv_id, 1, shared_mem))
+    for serv_id in range(num_servers):
+        p = ctx.Process(target=run_server, args=(graph_name, serv_id,
+                                                 num_clients, shared_mem))
         serv_ps.append(p)
         p.start()
 
     cli_ps = []
-    for cli_id in range(1):
+    for cli_id in range(num_clients):
         print('start client', cli_id)
-        p = ctx.Process(target=run_client, args=(graph_name, cli_id, g.number_of_nodes(),
+        p = ctx.Process(target=run_client, args=(graph_name, 0, num_clients, g.number_of_nodes(),
                                                  g.number_of_edges()))
         p.start()
         cli_ps.append(p)
@@ -227,8 +230,10 @@ def check_server_client(shared_mem):
 @unittest.skipIf(dgl.backend.backend_name == "tensorflow", reason="TF doesn't support some of operations in DistGraph")
 def test_server_client():
     os.environ['DGL_DIST_MODE'] = 'distributed'
-    check_server_client(True)
-    check_server_client(False)
+    check_server_client(True, 1, 1)
+    check_server_client(False, 1, 1)
+    check_server_client(True, 2, 2)
+    check_server_client(False, 2, 2)
 
 @unittest.skipIf(dgl.backend.backend_name == "tensorflow", reason="TF doesn't support some of operations in DistGraph")
 def test_standalone():
@@ -242,10 +247,10 @@ def test_standalone():
     partition_graph(g, graph_name, num_parts, '/tmp/dist_graph')
     dist_g = DistGraph("kv_ip_config.txt", graph_name,
                        part_config='/tmp/dist_graph/{}.json'.format(graph_name))
-    check_dist_graph(dist_g, g.number_of_nodes(), g.number_of_edges())
+    check_dist_graph(dist_g, 1, g.number_of_nodes(), g.number_of_edges())
 
 def test_split():
-    prepare_dist()
+    #prepare_dist()
     g = create_random_graph(10000)
     num_parts = 4
     num_hops = 2
@@ -290,7 +295,7 @@ def test_split():
         assert np.all(np.sort(edges1) == np.sort(F.asnumpy(edges5)))
 
 def test_split_even():
-    prepare_dist()
+    #prepare_dist(1)
     g = create_random_graph(10000)
     num_parts = 4
     num_hops = 2
@@ -348,10 +353,10 @@ def test_split_even():
     assert np.all(all_nodes == F.asnumpy(all_nodes2))
     assert np.all(all_edges == F.asnumpy(all_edges2))
 
-def prepare_dist():
+def prepare_dist(num_servers):
     ip_config = open("kv_ip_config.txt", "w")
     ip_addr = get_local_usable_addr()
-    ip_config.write('%s 1\n' % ip_addr)
+    ip_config.write('{} {}\n'.format(ip_addr, num_servers))
     ip_config.close()
 
 if __name__ == '__main__':

tests/distributed/test_new_kvstore.py

@@ -151,6 +151,7 @@ def start_client(num_clients):
     dgl.distributed.connect_to_server(ip_config='kv_ip_config.txt')
     # Init kvclient
     kvclient = dgl.distributed.KVClient(ip_config='kv_ip_config.txt')
+    kvclient.map_shared_data(partition_book=gpb)
     assert dgl.distributed.get_num_client() == num_clients
     kvclient.init_data(name='data_1', 
                        shape=F.shape(data_1), 
@@ -162,8 +163,6 @@ def start_client(num_clients):
                        dtype=F.dtype(data_2), 
                        part_policy=node_policy,
                        init_func=init_zero_func)
-
-    kvclient.map_shared_data(partition_book=gpb)
     
     # Test data_name_list
     name_list = kvclient.data_name_list()
@@ -264,7 +263,6 @@ def start_client(num_clients):
                        part_policy=node_policy,
                        init_func=init_zero_func)
     kvclient.register_push_handler('data_3', add_push)
-    kvclient.map_shared_data(partition_book=gpb)
     data_tensor = F.tensor([[6.,6.],[6.,6.],[6.,6.]], F.float32)
     kvclient.barrier()
     time.sleep(kvclient.client_id + 1)