Disable gloo by default (#29805)

* update, test=develop

python/paddle/distributed/fleet/base/role_maker.py

@@ -220,15 +220,8 @@ class Gloo(object):
             rank, nodes = self._get_rank_nodes(Role.WORKER)
             gloo = init(rank, nodes, "WORKER")
             self._worker_comm = gloo
-        else:
-            rank, nodes = self._get_rank_nodes(Role.SERVER)
-            gloo = init(rank, nodes, "SERVER")
-            self._server_comm = gloo
+        # TODO (sandyhouse): initialize gloo for server and all
 
-        if self._need_init_all:
-            rank, nodes = self._get_rank_nodes(Role.ALL)
-            gloo = init(rank, nodes, "ALL")
-            self._nodes_comm = gloo
         if start_http_server:
             http_server_d["running"] = False
             http_server.join()

python/paddle/distributed/fleet/launch.py

@@ -219,7 +219,7 @@ def launch_collective(args):
     global_envs = copy.copy(os.environ.copy())
     gloo_rendezvous_dir = tempfile.mkdtemp()
     # add gloo env
-    global_envs["PADDLE_WITH_GLOO"] = str(os.getenv("PADDLE_WITH_GLOO", "1"))
+    global_envs["PADDLE_WITH_GLOO"] = str(os.getenv("PADDLE_WITH_GLOO", "0"))
     global_envs["PADDLE_GLOO_RENDEZVOUS"] = "3"
     global_envs["PADDLE_GLOO_FS_PATH"] = gloo_rendezvous_dir
 

python/paddle/distributed/fleet/launch_utils.py

@@ -954,7 +954,7 @@ class ParameterServerLauncher(object):
                 "TRAINING_ROLE": "PSERVER",
                 "PADDLE_TRAINERS_NUM": str(self.worker_num),
                 "POD_IP": cur_server.endpoint.split(":")[0],
-                "PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "1")),
+                "PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "0")),
                 "PADDLE_GLOO_RENDEZVOUS": "3",
                 "PADDLE_GLOO_FS_PATH": self.gloo_rendezvous_dir,
                 "PADDLE_GLOO_HTTP_ENDPOINT": self.http_port
@@ -1018,7 +1018,7 @@ class ParameterServerLauncher(object):
                 self.heter_worker_endpoints,
                 "TRAINING_ROLE": "TRAINER",
                 "PADDLE_TRAINER_ID": str(cur_worker.rank),
-                "PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "1")),
+                "PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "0")),
                 "PADDLE_GLOO_RENDEZVOUS": "3",
                 "PADDLE_GLOO_FS_PATH": self.gloo_rendezvous_dir,
                 "FLAGS_selected_gpus": "0",
@@ -1088,7 +1088,7 @@ class ParameterServerLauncher(object):
                 "TRAINING_ROLE": "HETER_TRAINER",
                 "PADDLE_TRAINERS_NUM": str(self.worker_num),
                 "POD_IP": cur_heter_worker.endpoint.split(":")[0],
-                "PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "1")),
+                "PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "0")),
                 "PADDLE_GLOO_RENDEZVOUS": "3",
                 "PADDLE_GLOO_FS_PATH": self.gloo_rendezvous_dir,
                 "FLAGS_selected_gpus": "0",

python/paddle/distributed/parallel.py

@@ -142,6 +142,8 @@ def init_parallel_env():
     _check_var_exists("PADDLE_TRAINER_ENDPOINTS")
 
     # 3: init gloo context (step 1: httpsever start)
+    init_gloo = int(os.getenv("PADDLE_WITH_GLOO", "0"))
+    if init_gloo:
         ep_rank_0 = parallel_env.trainer_endpoints[0].split(":")
         ep_rank = parallel_env.trainer_endpoints[parallel_env.rank].split(":")
         manager = Manager()
@@ -185,6 +187,7 @@ def init_parallel_env():
     # dividing init_gloo into two part beacause nccl and gloo
     # are separately looking for free ports which sometimes
     # leads to port-conflict.
+    if init_gloo:
         wait_server_ready([parallel_env.trainer_endpoints[0]])
 
         gloo_strategy = core.GlooParallelStrategy()

python/paddle/fluid/tests/unittests/test_collective_api_base.py

@@ -178,6 +178,7 @@ class TestDistBase(unittest.TestCase):
             "LD_PRELOAD": os.getenv("LD_PRELOAD", ""),
             "GLOG_v": "0",
             "NCCL_P2P_DISABLE": "1",
+            "PADDLE_WITH_GLOO": "1",
             "BACKEND": backend,
             "PATH_ID": path_id
         }