fleet support elastic scale up/down (#36684)

* fleet support elastic train

* fleet support elastic train

* support elastic

* add unittest

* fix unitest bug

* fix unittest bug

* fix unittest bug

* fix unittest coverage

* fix unittest coverage

* fix unittest coverage

* fix unittest coverage

* fix unittest coverage

* fix elastic bug

* fix ci fail

* fix ci fail

* fix elastic bug

* fix elastic bug

* fix joint debugging bug

* fix joint debugging bug

* fix windows ci failed

* fix windows ci failed

python/paddle/distributed/elastic.py

@@ -50,7 +50,10 @@ if __name__ == '__main__':
     parser.add_argument(
         "--elastic_server", type=str, help="etcd server host:port")
     parser.add_argument("--job_id", type=str, help="job unique id")
-    parser.add_argument("--np", type=int, help="job pod/node number")
+    parser.add_argument(
+        "--np",
+        type=str,
+        help="job pod/node number, need to be 'MIN' or 'MIN:MAX' format")
     parser.add_argument("action", type=str, help="action to take")
 
     args = parser.parse_args()
@@ -58,7 +61,7 @@ if __name__ == '__main__':
     server = args.elastic_server or os.getenv('PADDLE_ELASTIC_SERVER')
     name = args.job_id or os.getenv('PADDLE_ELASTIC_JOB_ID')
 
-    np = args.np or int(os.getenv('PADDLE_ELASTIC_NP', 0))
+    np = int(args.np.split(":")[0]) or int(os.getenv('PADDLE_ELASTIC_NP', 0))
 
     cmd = Command(server, name)
 

python/paddle/distributed/fleet/elastic/__init__.py

@@ -33,7 +33,7 @@ def enable_elastic(args, distribute_mode):
     if not args.job_id and not os.getenv('PADDLE_ELASTIC_JOB_ID'):
         return False
 
-    if not args.np and not int(os.getenv('PADDLE_ELASTIC_NP', 0)):
+    if not args.np and not os.getenv('PADDLE_ELASTIC_NP'):
         return False
 
     return True
@@ -41,7 +41,11 @@ def enable_elastic(args, distribute_mode):
 
 def launch_elastic(args, distribute_mode):
 
-    elastic = ElasticManager(args)
+    server = args.elastic_server or os.getenv('PADDLE_ELASTIC_SERVER')
+    srv, port = server.split(':')
+    import etcd3
+    etcd_client = etcd3.client(host=srv, port=port)
+    elastic = ElasticManager(args, etcd_client)
 
     signal.signal(signal.SIGTERM, elastic.signal_handler)
     signal.signal(signal.SIGABRT, elastic.signal_handler)

python/paddle/distributed/fleet/elastic/collective.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import tempfile
 from paddle.distributed.fleet import launch_utils
 from paddle.distributed.fleet import cloud_utils
 from paddle.distributed.fleet import ascend_utils

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

+#   Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import os
+import time
+import unittest
+import argparse
+from warnings import catch_warnings
+
+from paddle.distributed.fleet.elastic import enable_elastic, launch_elastic
+from paddle.distributed.fleet.launch_utils import DistributeMode
+
+
+class TestElasticInit(unittest.TestCase):
+    def setUp(self):
+        class Argument:
+            elastic_server = "127.0.0.1:2379"
+            job_id = "test_job_id_123"
+            np = "2:4"
+
+        self.args = Argument()
+
+    def test_enable_elastic(self):
+        result = enable_elastic(self.args, DistributeMode.COLLECTIVE)
+        self.assertEqual(result, True)
+
+    def test_launch_elastic(self):
+        try:
+            launch_elastic(self.args, DistributeMode.COLLECTIVE)
+        except Exception as e:
+            pass
+
+
+if __name__ == "__main__":
+    unittest.main()

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

+#   Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import os
+import time
+import unittest
+import argparse
+
+from paddle.distributed.fleet.elastic.manager import ElasticManager
+from paddle.distributed.fleet.elastic.manager import ELASTIC_TIMEOUT
+
+
+class MockLease():
+    def refresh(self):
+        pass
+
+
+class MockEtcdClient:
+    def __init__(self, lease=None):
+        self._lease = lease
+
+    def put(self, key, value, lease=None):
+        pass
+
+    def get(self, key):
+        value = "0"
+        return value, value
+
+    def delete_prefix(self, key):
+        pass
+
+    def get_prefix(self, key_prefix):
+        hosts = ["10.10.10.1:6001", "10.10.10.2:6001"]
+        return hosts
+
+    def add_watch_callback(self, *args, **kwargs):
+        return "host_watch"
+
+    def add_watch_prefix_callback(self, key_prefix, callback, **kwargs):
+        callback(None)
+        return "host_watch"
+
+    def cancel_watch(self, watch_id):
+        pass
+
+    def delete(self, key):
+        pass
+
+    def lease(self, ttl):
+        if self._lease:
+            return self._lease
+        else:
+            return MockLease()
+
+
+class TestElasticManager(unittest.TestCase):
+    def setUp(self):
+        self.etcd_client = MockEtcdClient()
+
+    def test_elastic_manager_init(self):
+        class Argument:
+            elastic_server = "127.0.0.1:2379"
+            job_id = "test_job_id_123"
+            np = "2"
+            gpus = "0"
+            nproc_per_node = 1
+            host = None
+            host_port = None
+            scale = None
+            force = None
+            backend = 'gloo'
+
+        args = Argument()
+
+        class _MockLease():
+            def refresh(self):
+                raise ValueError("valid error, this only for unittest")
+
+        etcd_client = MockEtcdClient(lease=_MockLease())
+        elastic = ElasticManager(args, etcd_client=etcd_client)
+
+    def test_match_faulttolerance(self):
+        class Argument:
+            elastic_server = "127.0.0.1:2379"
+            job_id = "test_job_id_123"
+            np = "2"
+            gpus = "0"
+            nproc_per_node = 1
+            host = None
+            host_port = None
+            scale = None
+            force = None
+            backend = 'gloo'
+
+        args = Argument()
+        elastic = ElasticManager(args, self.etcd_client)
+        hosts = ["10.10.10.1:6001", "10.10.10.2:6001"]
+        os.environ[
+            'PADDLE_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.2:6001"
+        self.assertEqual(elastic._match(hosts), True)
+        hosts = ["10.10.10.1:6001"]
+        os.environ['PADDLE_TRAINER_ENDPOINTS'] = "10.10.10.1:6001"
+        self.assertEqual(elastic._match(hosts), False)
+
+    def test_match_elastic(self):
+        class Argument:
+            elastic_server = "127.0.0.1:2379"
+            job_id = "test_job_id_123"
+            np = "2:4"
+            gpus = "0"
+            nproc_per_node = 1
+            host = None
+            host_port = None
+            scale = None
+            force = None
+            backend = 'gloo'
+
+        os.environ['PADDLE_ELASTIC_TIMEOUT'] = "60"
+        args = Argument()
+        os.environ['FLAGS_START_PORT'] = "6001"
+        os.environ[
+            'DISTRIBUTED_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.2:6001,10.10.10.3:6001,10.10.10.4:6001"
+        os.environ[
+            'PADDLE_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.2:6001,10.10.10.3:6001,10.10.10.4:6001"
+        elastic = ElasticManager(args, self.etcd_client)
+        hosts = ["10.10.10.1:6001", "10.10.10.2:6001"]
+        self.assertEqual(elastic._match(hosts), False)
+
+        hosts = [
+            "10.10.10.1:6001", "10.10.10.2:6001", "10.10.10.3:6001",
+            "10.10.10.4:6001"
+        ]
+        self.assertEqual(elastic._match(hosts), True)
+
+        hosts = ["10.10.10.1:6001", "10.10.10.2:6001", "10.10.10.3:6001"]
+        self.assertEqual(elastic._match(hosts), False)
+
+        hosts = ["10.10.10.1:6001"]
+        self.assertEqual(elastic._match(hosts), False)
+
+        os.environ[
+            'DISTRIBUTED_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.2:6001"
+        os.environ[
+            'PADDLE_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.2:6001"
+        elastic = ElasticManager(args, self.etcd_client)
+        hosts = ["10.10.10.1:6001", "10.10.10.2:6001"]
+        self.assertEqual(elastic._match(hosts), True)
+
+        # TODO test timeout
+        #time.sleep(60)
+        #self.assertEqual(elastic._match(hosts), True)
+
+    def test_update_hosts_for_faulttolerance(self):
+        class Argument:
+            elastic_server = "127.0.0.1:2379"
+            job_id = "test_job_id_123"
+            np = "0"
+            gpus = "0"
+            nproc_per_node = 1
+            host = None
+            host_port = None
+            scale = None
+            force = None
+            backend = 'gloo'
+
+        args = Argument()
+        os.environ['FLAGS_START_PORT'] = "6001"
+        os.environ['PADDLE_ELASTIC_NP'] = "2"
+        os.environ['PADDLE_TRAINERS'] = "10.10.10.1,10.10.10.2"
+        os.environ[
+            'DISTRIBUTED_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.2:6001"
+        os.environ[
+            'PADDLE_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.2:6001"
+        elastic = ElasticManager(args, self.etcd_client)
+        # add 10.10.10.3:6001
+        os.environ['PADDLE_TRAINER_ID'] = "0"
+        elastic.host_port = "10.10.10.1:6001"
+        elastic.hosts = ["10.10.10.1:6001", "10.10.10.2:6001"]
+        elastic._update_hosts()
+        self.assertEqual(os.getenv('PADDLE_TRAINERS'), "10.10.10.1,10.10.10.2")
+
+        # add 10.10.10.3:6001
+        elastic.host_port = "10.10.10.3:6001"
+        elastic.hosts = ["10.10.10.1:6001", "10.10.10.3:6001"]
+        os.environ['PADDLE_TRAINER_ID'] = "1"
+        elastic._update_hosts()
+        self.assertEqual(os.getenv('PADDLE_TRAINERS'), "10.10.10.1,10.10.10.3")
+
+        elastic.host_port = "10.10.10.3:6001"
+        elastic.hosts = ["10.10.10.1:6001", "10.10.10.3:6001"]
+        os.environ['PADDLE_TRAINER_ID'] = "-1"
+        elastic._update_hosts()
+        self.assertEqual(os.getenv('PADDLE_TRAINERS'), "10.10.10.1,10.10.10.3")
+
+    def test_update_hosts_for_elastic(self):
+        #######################
+        #  elastic, scale up  #
+        #######################
+        class Argument:
+            elastic_server = "127.0.0.1:2379"
+            job_id = "test_job_id_123"
+            np = "2:4"
+            gpus = "0"
+            nproc_per_node = 1
+            host = None
+            host_port = None
+            scale = None
+            force = None
+            backend = 'gloo'
+
+        args = Argument()
+
+        os.environ['FLAGS_START_PORT'] = "6001"
+        os.environ['PADDLE_TRAINERS'] = "10.10.10.1,10.10.10.2"
+        os.environ[
+            'DISTRIBUTED_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.2:6001"
+        os.environ[
+            'PADDLE_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.2:6001"
+        elastic = ElasticManager(args, self.etcd_client)
+        # add 10.10.10.3:6001
+        elastic.host_port = "10.10.10.1:6001"
+        elastic.hosts = [
+            "10.10.10.1:6001", "10.10.10.2:6001", "10.10.10.3:6001"
+        ]
+        elastic._update_hosts()
+        #self.assertEqual(elastic.all_host_endpoints,
+        #                 ["10.10.10.1:6001", "10.10.10.2:6001", "10.10.10.3:6001"])
+        self.assertEqual(
+            os.getenv('PADDLE_TRAINERS'), "10.10.10.1,10.10.10.2,10.10.10.3")
+
+        #######################
+        # elastic, scale down #
+        #######################
+        os.environ[
+            'PADDLE_TRAINERS'] = "10.10.10.0,10.10.10.1,10.10.10.2,10.10.10.3"
+        os.environ[
+            'DISTRIBUTED_TRAINER_ENDPOINTS'] = "10.10.10.0:6000,10.10.10.1:6001,10.10.10.2:6001,10.10.10.3:6001"
+        os.environ[
+            'PADDLE_TRAINER_ENDPOINTS'] = "10.10.10.0:6000,10.10.10.1:6001,10.10.10.2:6001,10.10.10.3:6001"
+        elastic = ElasticManager(args, self.etcd_client)
+        # remove 10.10.10.1:6001
+        elastic.host_port = "10.10.10.1:6001"
+        elastic.hosts = [
+            "10.10.10.1:6001", "10.10.10.2:6001", "10.10.10.3:6001"
+        ]
+        elastic._update_hosts()
+        #self.assertEqual(elastic.all_host_endpoints,
+        #                 ["10.10.10.3:6001", "10.10.10.1:6001", "10.10.10.2:6001"])
+        self.assertEqual(
+            os.getenv('PADDLE_TRAINERS'), "10.10.10.3,10.10.10.1,10.10.10.2")
+        self.assertEqual(
+            os.getenv('DISTRIBUTED_TRAINER_ENDPOINTS'),
+            "10.10.10.3:6001,10.10.10.1:6001,10.10.10.2:6001")
+
+        ############
+        os.environ['PADDLE_TRAINERS'] = "10.10.10.1,10.10.10.1"
+        os.environ[
+            'DISTRIBUTED_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.1:6002,10.10.10.1:6003,10.10.10.1:6004"
+        os.environ[
+            'PADDLE_TRAINER_ENDPOINTS'] = "10.10.10.1:6001,10.10.10.1:6002,10.10.10.1:6003,10.10.10.1:6004"
+        elastic = ElasticManager(args, self.etcd_client)
+        # remove 10.10.10.1:6001
+        elastic.host_port = "10.10.10.1:6001"
+        os.environ['PADDLE_TRAINER_ID'] = "-1"
+        elastic.hosts = ["10.10.10.1:6001", "10.10.10.1:6001"]
+        elastic._update_hosts()
+        #self.assertEqual(elastic.all_host_endpoints,
+        #                 ["10.10.10.1:6001", "10.10.10.1:6001"])
+        self.assertEqual(os.getenv('PADDLE_TRAINERS'), "10.10.10.1,10.10.10.1")
+        self.assertEqual(
+            os.getenv('DISTRIBUTED_TRAINER_ENDPOINTS'),
+            "10.10.10.1:6001,10.10.10.1:6001")
+
+    def test_exit(self):
+        class Argument:
+            elastic_server = "127.0.0.1:2379"
+            job_id = "test_job_id_123"
+            np = "2"
+            gpus = "0"
+            nproc_per_node = 1
+            host = None
+            host_port = None
+            scale = None
+            force = None
+            backend = 'gloo'
+
+        args = Argument()
+        elastic = ElasticManager(args, self.etcd_client)
+        elastic.exit()
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/test_fleet_launch_elastic.sh

@@ -15,7 +15,7 @@
 echo "begin test elastic"
 
 unset GREP_OPTIONS
-rm -rf log
+rm -rf log*
 
 pids=`ps -ef | grep "python -m paddle.distributed.launch elastic_demo.[py]" | awk '{print $2}'`
 if [ -n "$pids" ]; then
@@ -28,6 +28,11 @@ fi
 
 python -m pip install --no-cache-dir etcd3 -i https://mirror.baidu.com/pypi/simple
 
+
+#############################
+#### test fault tolrance ####
+#############################
+
 # common env
 export PADDLE_ELASTIC_NP=2
 export PADDLE_ELASTIC_SERVER=127.0.0.1:2379
@@ -137,7 +142,7 @@ export PADDLE_TRAINER_ID=1
 export PADDLE_TRAINERS_NUM=2
 
 python -m paddle.distributed.launch elastic_demo.py &> log_1.log &
-p1=$!
+p1_1=$!
 
 for i in {1..10}
 do
@@ -184,7 +189,7 @@ export PADDLE_TRAINER_ID=0
 export PADDLE_TRAINERS_NUM=2
 
 python -m paddle.distributed.launch elastic_demo.py &> log_0.log &
-p0=$!
+p0_1=$!
 
 for i in {1..10}
 do
@@ -205,4 +210,102 @@ check_env
 echo "All check done"
 
 sleep 3
-kill $p0 $p1
+kill $p0 $p1 $p0_1 $p1_1
+
+#############################
+#####    test elastic   #####
+#############################
+# common env
+export PADDLE_ELASTIC_NP=2:4
+export PADDLE_ELASTIC_SERVER=127.0.0.1:2379
+export PADDLE_ELASTIC_JOB_ID=elastic-demo-2
+
+# run node 0
+export NVIDIA_VISIBLE_DEVICES=0
+export CUDA_VISIBLE_DEVICES=0
+export DISTRIBUTED_TRAINER_ENDPOINTS=10.10.10.1:8001,10.10.10.2:8001,10.10.10.3:8001
+export PADDLE_TRAINERS=10.10.10.1,10.10.10.2,10.10.10.3
+export TRAINER_PORTS_NUM=1
+export POD_IP=10.10.10.1
+export PADDLE_TRAINER_ID=0
+export PADDLE_TRAINERS_NUM=3
+
+python -m paddle.distributed.launch elastic_demo.py &> log_pe_0.log &
+pe_0=$!
+
+for i in {1..10}
+do
+    if grep -q "INFO:ELASTIC:not ready" log_pe_0.log; then
+        echo "run node 0 ok"
+        break
+    else
+        sleep 10
+    fi
+    if [ $i -eq 10 ]; then
+        echo "run node 0 error"
+        exit -1
+    fi
+done
+
+# run node 1
+export NVIDIA_VISIBLE_DEVICES=1
+export CUDA_VISIBLE_DEVICES=1
+export DISTRIBUTED_TRAINER_ENDPOINTS=10.10.10.1:8001,10.10.10.2:8001,10.10.10.3:8001
+export PADDLE_TRAINERS=10.10.10.1,10.10.10.2,10.10.10.3
+export TRAINER_PORTS_NUM=1
+export POD_IP=10.10.10.2
+export PADDLE_TRAINER_ID=1
+export PADDLE_TRAINERS_NUM=3
+
+python -m paddle.distributed.launch elastic_demo.py &> log_pe_1.log &
+pe_1=$!
+
+for i in {1..10}
+do
+    if grep -q "INFO:ELASTIC:not ready" log_pe_1.log; then
+        echo "run node 1 ok"
+        break
+    else
+        sleep 10
+    fi
+    if [ $i -eq 10 ]; then
+        echo "run node 1 error"
+        exit -1
+    fi
+done
+
+# run node 2
+export NVIDIA_VISIBLE_DEVICES=1
+export CUDA_VISIBLE_DEVICES=1
+export DISTRIBUTED_TRAINER_ENDPOINTS=10.10.10.1:8001,10.10.10.2:8001,10.10.10.3:8001
+export PADDLE_TRAINERS=10.10.10.1,10.10.10.2,10.10.10.3
+export TRAINER_PORTS_NUM=1
+export POD_IP=10.10.10.3
+export PADDLE_TRAINER_ID=2
+export PADDLE_TRAINERS_NUM=3
+
+python -m paddle.distributed.launch elastic_demo.py &> log_pe_2.log &
+pe_2=$!
+
+for i in {1..10}
+do
+    if grep -q "INFO:ELASTIC:ready with hosts" log_pe_2.log; then
+        echo "run node 2 ok"
+        break
+    else
+        sleep 10
+    fi
+    if [ $i -eq 10 ]; then
+        echo "run node 2 error"
+        exit -1
+    fi
+done
+
+lw0="log/workerlog.0"
+
+check_env
+
+echo "All check done"
+
+sleep 3
+kill $pe_0 $pe_1 $pe_2