fix nng perf hostname error leading deadlock

Author: SolenoidWGT

.coveragerc

@@ -1,5 +1,5 @@
 [run]
-concurrency = multiprocessing
+concurrency = multiprocessing,thread
 omit =
   ding/utils/slurm_helper.py
   ding/utils/file_helper.py

ding/framework/message_queue/perfs/tests/test_perf_nng.py

@@ -1,16 +1,21 @@
-from ding.framework.message_queue.perfs.perf_nng import nng_perf_main
 import multiprocessing as mp
 import pytest
+import socket
+import torch
+from ding.framework.message_queue.perfs.perf_nng import nng_perf_main
 
 
 @pytest.mark.benchmark
 # @pytest.mark.multiprocesstest
 def test_nng():
-    params = [
-        ("12960", None, "127.0.0.1", "learner", "0"), ("12961", "tcp://127.0.0.1:12960", "127.0.0.1", "collector", "1")
-    ]
-    ctx = mp.get_context("spawn")
-    with ctx.Pool(processes=2) as pool:
-        pool.starmap(nng_perf_main, params)
-        pool.close()
-        pool.join()
+    if torch.cuda.is_available() and torch.cuda.device_count() >= 2:
+        address = socket.gethostbyname(socket.gethostname())
+        params = [
+            ("12960", None, address, "learner", "0"),
+            ("12961", "tcp://{}:12960".format(address), "127.0.0.1", "collector", "1")
+        ]
+        ctx = mp.get_context("spawn")
+        with ctx.Pool(processes=2) as pool:
+            pool.starmap(nng_perf_main, params)
+            pool.close()
+            pool.join()

ding/framework/message_queue/perfs/tests/test_perf_torchrpc_nccl.py

@@ -1,18 +1,22 @@
-from ding.framework.message_queue.perfs.perf_torchrpc_nccl import rpc_model_exchanger
-from ding.compatibility import torch_ge_1121
 import multiprocessing as mp
 import pytest
 import torch
 import platform
+import socket
+from ding.utils.system_helper import find_free_port
+from ding.framework.message_queue.perfs.perf_torchrpc_nccl import rpc_model_exchanger
+from ding.compatibility import torch_ge_1121
 
 
 @pytest.mark.benchmark
 @pytest.mark.cudatest
 # @pytest.mark.multiprocesstest
 def test_perf_torchrpc_nccl():
+    address = socket.gethostbyname(socket.gethostname())
+    init_method = "tcp://{}:{}".format(address, find_free_port(address))
     if platform.system().lower() != 'windows' and torch.cuda.is_available():
         if torch_ge_1121() and torch.cuda.device_count() >= 2:
-            params = [(0, "tcp://127.0.0.1:12387", False, True), (1, "tcp://127.0.0.1:12387", False, True)]
+            params = [(0, init_method, False, True), (1, init_method, False, True)]
             ctx = mp.get_context("spawn")
             with ctx.Pool(processes=2) as pool:
                 pool.starmap(rpc_model_exchanger, params)

ding/framework/message_queue/tests/test_torch_rpc.py

@@ -1,13 +1,15 @@
+import pytest
+import torch
+import platform
+import time
+import socket
+
 from ding.framework.message_queue.torch_rpc import DeviceMap, TORCHRPCMQ, DEFAULT_DEVICE_MAP_NUMS
 from torch.distributed import rpc
 from multiprocessing import Pool, get_context
 from ding.compatibility import torch_ge_1121
 from ditk import logging
-
-import pytest
-import torch
-import platform
-import time
+from ding.utils.system_helper import find_free_port
 
 mq = None
 recv_tensor_list = [None, None, None, None]
@@ -22,6 +24,7 @@ def torchrpc(rank):
     global mq
     global recv_tensor_list
     mq = None
+    address = socket.gethostbyname(socket.gethostname())
     recv_tensor_list = [None, None, None, None]
     logging.getLogger().setLevel(logging.DEBUG)
     name_list = ["A", "B", "C", "D"]
@@ -34,7 +37,7 @@ def torchrpc(rank):
     mq = TORCHRPCMQ(
         rpc_name=name_list[rank],
         global_rank=rank,
-        init_method="tcp://127.0.0.1:12398",
+        init_method="tcp://{}:12398".format(address),
         remote_parallel_entrance=remote_mq_entrance,
         attach_to=attach_to,
         async_rpc=False,
@@ -81,6 +84,7 @@ def torchrpc_cuda(rank):
     mq = None
     recv_tensor_list = [None, None, None, None]
     name_list = ["A", "B"]
+    address = socket.gethostbyname(socket.gethostname())
     logging.getLogger().setLevel(logging.DEBUG)
 
     if rank == 0:
@@ -96,7 +100,7 @@ def torchrpc_cuda(rank):
     mq = TORCHRPCMQ(
         rpc_name=name_list[rank],
         global_rank=rank,
-        init_method="tcp://127.0.0.1:12390",
+        init_method="tcp://{}:12390".format(address),
         remote_parallel_entrance=remote_mq_entrance,
         attach_to=attach_to,
         device_maps=device_map,