SYNC with communicaotor (#22344)

* add sync communicator and implement

paddle/fluid/framework/details/async_ssa_graph_executor.cc

@@ -180,8 +180,6 @@ FeedFetchList AsyncSSAGraphExecutor::Run(
 
   if (places_.size() == 1) {
     exception_holder_.Clear();
-  } else {
-    HandleException();
   }
 
   FeedFetchList fetch_data;

paddle/fluid/operators/distributed/communicator.cc

@@ -27,6 +27,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/distributed/distributed.h"
 #include "paddle/fluid/operators/distributed/parameter_recv.h"
 #include "paddle/fluid/operators/distributed/parameter_send.h"
+#include "paddle/fluid/string/printf.h"
 #include "paddle/fluid/string/split.h"
 
 namespace paddle {
@@ -64,7 +65,6 @@ std::shared_ptr<Communicator> Communicator::communicator_(nullptr);
 void AsyncCommunicator::InitImpl(const RpcCtxMap &send_varname_to_ctx,
                                  const RpcCtxMap &recv_varname_to_ctx,
                                  Scope *recv_scope) {
-  VLOG(0) << "AsyncCommunicator Initializing";
   send_varname_to_ctx_ = std::move(send_varname_to_ctx);
   recv_varname_to_ctx_ = std::move(recv_varname_to_ctx);
   recv_scope_ = std::move(recv_scope);
@@ -90,7 +90,6 @@ void AsyncCommunicator::InitImpl(const RpcCtxMap &send_varname_to_ctx,
 
 void AsyncCommunicator::InitImpl(const paddle::framework::ProgramDesc &program,
                                  Scope *param_scope) {
-  VLOG(0) << "AsyncCommunicator Initializing";
   RpcCtxMap send_varname_to_ctx;
   RpcCtxMap recv_varname_to_ctx;
   for (auto *op : program.Block(0).AllOps()) {
@@ -332,8 +331,6 @@ GeoSgdCommunicator::~GeoSgdCommunicator() {
 
 void GeoSgdCommunicator::InitImpl(const paddle::framework::ProgramDesc &program,
                                   Scope *recv_scope) {
-  VLOG(0) << "GeoCommunicator Initializing";
-
   training_scope_ = std::move(recv_scope);
 
   auto geo_send_varnames = envs["geo_send_varnames"];
@@ -954,7 +951,6 @@ void GeoSgdCommunicator::Recv() {}
 void HalfAsyncCommunicator::InitImpl(const RpcCtxMap &send_varname_to_ctx,
                                      const RpcCtxMap &recv_varname_to_ctx,
                                      Scope *recv_scope) {
-  VLOG(0) << "HalfAsyncCommunicator Initializing";
   send_varname_to_ctx_ = std::move(send_varname_to_ctx);
   recv_varname_to_ctx_ = std::move(recv_varname_to_ctx);
   recv_scope_ = std::move(recv_scope);
@@ -1011,6 +1007,8 @@ void HalfAsyncCommunicator::InitImpl(
       auto trainer_id = boost::get<int>(op->GetNullableAttr("trainer_id"));
       recv_varname_to_ctx[recv_var_name] = operators::distributed::RpcContext(
           recv_var_name, recv_varnames, epmap, {}, trainer_id);
+      VLOG(3) << "find and init an recv op: "
+              << recv_varname_to_ctx[recv_var_name];
     }
   }
 
@@ -1032,7 +1030,8 @@ void HalfAsyncCommunicator::ConsumeThread() {
   VLOG(3) << "ConsumeThread start!";
   while (running_) {
     while (running_) {
-      if (barrier_counter_.load() >= barrier_trigger_.load()) {
+      if (barrier_counter_.load() >= barrier_trigger_.load() &&
+          barrier_trigger_.load() != 0) {
         break;
       } else {
         std::this_thread::sleep_for(std::chrono::milliseconds(10));
@@ -1096,8 +1095,10 @@ void HalfAsyncCommunicator::ConsumeThread() {
 
     VLOG(3) << "run send graph use time "
             << after_run_send_graph - before_run_send_graph;
-    Recv();
 
+    BarrierSend();
+    Recv();
+    BarrierRecv();
     BarrierWeakUp();
   }
   VLOG(0) << "communicator stopped, send thread exit";
@@ -1200,6 +1201,49 @@ void HalfAsyncCommunicator::Stop() {
   VLOG(0) << "Communicator stop done";
 }
 
+void SyncCommunicator::BarrierSend() {
+  if (!running_) return;
+
+  distributed::RPCClient *rpc_client =
+      distributed::RPCClient::GetInstance<RPCCLIENT_T>(trainer_id_);
+
+  std::vector<distributed::VarHandlePtr> rets;
+
+  for (auto &ep : pserver_endpoints_) {
+    rets.push_back(rpc_client->AsyncSendBatchBarrier(ep));
+  }
+
+  for (size_t i = 0; i < rets.size(); i++) {
+    PADDLE_ENFORCE_NE(rets[i]->Wait(), 0U, platform::errors::External(
+                                               "internal error in RPCClient"));
+  }
+
+  VLOG(4) << "BarrierSend with SyncCommunicator";
+}
+
+void SyncCommunicator::BarrierRecv() {
+  if (!running_) return;
+
+  distributed::RPCClient *rpc_client =
+      distributed::RPCClient::GetInstance<RPCCLIENT_T>(trainer_id_);
+
+  std::vector<distributed::VarHandlePtr> rets;
+  for (auto &ep : pserver_endpoints_) {
+    rets.push_back(rpc_client->AsyncSendFetchBarrier(ep));
+  }
+
+  for (size_t i = 0; i < rets.size(); i++) {
+    PADDLE_ENFORCE_NE(rets[i]->Wait(), 0U, platform::errors::External(
+                                               "internal error in RPCClient"));
+  }
+
+  VLOG(4) << "BarrierRecv with SyncCommunicator";
+}
+
+SyncCommunicator::~SyncCommunicator() {
+  running_ = false;
+  if (consume_thread_) consume_thread_->join();
+}
 }  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/distributed/communicator.h

@@ -37,6 +37,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/place.h"
+#include "paddle/fluid/string/split.h"
 
 DECLARE_bool(communicator_is_sgd_optimizer);
 
@@ -246,6 +247,7 @@ class AsyncCommunicator : public Communicator {
     send_queue_size_ = std::stoi(envs.at("communicator_send_queue_size"));
     is_sgd_optimizer_ =
         static_cast<bool>(std::stoi(envs.at("communicator_is_sgd_optimizer")));
+    VLOG(0) << "AsyncCommunicator Initialized";
   }
   ~AsyncCommunicator();
   void Start() override;
@@ -301,6 +303,7 @@ class HalfAsyncCommunicator : public Communicator {
     send_wait_times_ = std::stoi(envs.at("communicator_send_wait_times"));
     thread_pool_size_ = std::stoi(envs.at("communicator_thread_pool_size"));
     send_queue_size_ = std::stoi(envs.at("communicator_send_queue_size"));
+    VLOG(0) << "HalfAsyncCommunicator Initialized";
   }
   ~HalfAsyncCommunicator();
   void Start() override;
@@ -326,14 +329,17 @@ class HalfAsyncCommunicator : public Communicator {
                 Scope* recv_scope) override;
 
   void ConsumeThread();
+  virtual void BarrierSend() {}
+  virtual void BarrierRecv() {}
 
- private:
+ protected:
   int max_merge_var_num_;
   int send_wait_times_;
   int thread_pool_size_;
   int send_queue_size_;
+  int trainer_id_ = 0;
 
- private:
+ protected:
   std::unordered_map<std::string,
                      std::shared_ptr<BlockingQueue<std::shared_ptr<Variable>>>>
       send_varname_to_queue_;
@@ -352,6 +358,24 @@ class HalfAsyncCommunicator : public Communicator {
   std::atomic<int64_t> barrier_counter_{0};
 };
 
+class SyncCommunicator : public HalfAsyncCommunicator {
+ public:
+  SyncCommunicator() : HalfAsyncCommunicator() {}
+  explicit SyncCommunicator(const std::map<std::string, std::string>& envs)
+      : HalfAsyncCommunicator(envs) {
+    trainer_id_ = std::stoi(envs.at("trainer_id"));
+    auto pserver_strings = envs.at("pserver_endpoints");
+    pserver_endpoints_ = paddle::string::Split(pserver_strings, ',');
+    VLOG(0) << "SyncCommunicator Initialized";
+  }
+  ~SyncCommunicator();
+  void BarrierSend();
+  void BarrierRecv();
+
+ private:
+  std::vector<std::string> pserver_endpoints_{};
+};
+
 class GeoSgdCommunicator : public Communicator {
  public:
   GeoSgdCommunicator() : Communicator() {}
@@ -361,6 +385,7 @@ class GeoSgdCommunicator : public Communicator {
     trainer_nums_ = std::stoi(envs.at("geo_trainer_nums"));
     thread_pool_size_ = std::stoi(envs.at("communicator_thread_pool_size"));
     send_wait_times_ = std::stoi(envs.at("communicator_send_wait_times"));
+    VLOG(0) << "GeoSgdCommunicator Initialized";
   }
 
   ~GeoSgdCommunicator();

paddle/fluid/operators/distributed/parameter_send.cc

@@ -115,6 +115,11 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
         *out = send_tensor.Slice(row_offset, row_offset + outs_dims[i][0]);
         row_offset += outs_dims[i][0];
       }
+    } else {
+      auto &send_tensor = send_var->Get<framework::LoDTensor>();
+      framework::Tensor *out = local_scope->Var(rpc_ctx.splited_var_names[0])
+                                   ->GetMutable<framework::LoDTensor>();
+      out->ShareDataWith(send_tensor);
     }
     if (rpc_ctx.use_send_handler) {
       for (size_t i = 0; i < rpc_ctx.splited_var_names.size(); i++) {

paddle/fluid/operators/distributed_ops/fetch_barrier_op.cc

@@ -36,6 +36,7 @@ class FetchBarrierOp : public framework::OperatorBase {
   void RunImpl(const framework::Scope& scope,
                const platform::Place& place) const override {
     std::vector<std::string> eps = Attr<std::vector<std::string>>("endpoints");
+
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>(
             Attr<int>("trainer_id"));

paddle/fluid/pybind/communicator_py.cc

@@ -32,6 +32,7 @@ using paddle::operators::distributed::AsyncCommunicator;
 using paddle::operators::distributed::Communicator;
 using paddle::operators::distributed::GeoSgdCommunicator;
 using paddle::operators::distributed::HalfAsyncCommunicator;
+using paddle::operators::distributed::SyncCommunicator;
 
 namespace paddle {
 namespace pybind {
@@ -52,6 +53,9 @@ void BindCommunicator(py::module* m) {
         } else if (mode == "GEO") {
           Communicator::InitInstance<GeoSgdCommunicator>(program, param_scope,
                                                          envs);
+        } else if (mode == "SYNC") {
+          Communicator::InitInstance<SyncCommunicator>(program, param_scope,
+                                                       envs);
         } else {
           PADDLE_THROW(platform::errors::InvalidArgument(
               "unsuported communicator MODE"));

python/paddle/fluid/communicator.py

@@ -70,6 +70,10 @@ class Communicator(object):
             envs["geo_need_push_nums"] = str(kwargs["push_nums"])
             envs["geo_send_varnames"] = '#'.join(push_var_names)
 
+        if mode == DistributedMode.SYNC:
+            envs["pserver_endpoints"] = ','.join(kwargs["pserver_endpoints"])
+            envs["trainer_id"] = str(kwargs["trainer_id"])
+
         mode_str = None
 
         if mode == DistributedMode.SYNC:

python/paddle/fluid/incubate/fleet/parameter_server/distribute_transpiler/__init__.py

@@ -73,9 +73,6 @@ class DistributedTranspiler(Fleet):
         trainer_communicator_config = self._transpile_config.get_trainer_runtime_config(
         )
 
-        if isinstance(self._transpile_config, SyncStrategy):
-            return
-
         print(trainer_communicator_config)
 
         if isinstance(self._transpile_config, GeoStrategy):
@@ -98,6 +95,17 @@ class DistributedTranspiler(Fleet):
             self._communicator = Communicator(
                 self.main_program, DistributedMode.HALF_ASYNC, None,
                 trainer_communicator_config.get_communicator_flags())
+
+        elif isinstance(self._transpile_config, SyncStrategy):
+            kwargs = {}
+            kwargs[
+                "pserver_endpoints"] = self._role_maker.get_pserver_endpoints()
+            kwargs["trainer_id"] = self._role_maker.worker_index()
+
+            self._communicator = Communicator(
+                self.main_program, DistributedMode.SYNC, kwargs,
+                trainer_communicator_config.get_communicator_flags())
+
         else:
             raise TypeError("Training MODE do not supported")
 
@@ -156,8 +164,7 @@ class DistributedTranspiler(Fleet):
             None
         """
 
-        if not isinstance(self._transpile_config, SyncStrategy):
-            self._communicator.stop()
+        self._communicator.stop()
         if isinstance(self._role_maker, MPISymetricRoleMaker):
             self._role_maker._finalize()
         self._executor.close()

python/paddle/fluid/incubate/fleet/parameter_server/distribute_transpiler/distributed_strategy.py

@@ -181,9 +181,12 @@ class DistributedStrategy(object):
 class SyncStrategy(DistributedStrategy):
     def __init__(self):
         super(SyncStrategy, self).__init__()
-        self._program_config.sync_mode = True
-        self._program_config.runtime_split_send_recv = False
-        self._build_strategy.async_mode = False
+        self._program_config.sync_mode = False
+        self._program_config.runtime_split_send_recv = True
+        self._build_strategy.async_mode = True
+        self._program_config.half_async = True
+        self._program_config.completely_not_async = True
+        self._execute_strategy.use_thread_barrier = True
 
         num_threads = os.getenv("CPU_NUM", "1")
 

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -26,6 +26,7 @@ list(APPEND MIXED_DIST_TEST_OPS test_launch_ps)
 list(APPEND MIXED_DIST_TEST_OPS test_communicator_async)
 list(APPEND MIXED_DIST_TEST_OPS test_communicator_geo)
 list(APPEND MIXED_DIST_TEST_OPS test_communicator_half_async)
+list(APPEND MIXED_DIST_TEST_OPS test_communicator_sync)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_api_input)
 foreach(TEST_OP ${MIXED_DIST_TEST_OPS})
   list(REMOVE_ITEM TEST_OPS ${TEST_OP})
@@ -284,6 +285,8 @@ if(WITH_DISTRIBUTE)
     py_test_modules(test_communicator_async MODULES test_communicator_async ENVS ${dist_ENVS})
     py_test_modules(test_communicator_geo MODULES test_communicator_geo ENVS ${dist_ENVS})
     py_test_modules(test_communicator_half_async MODULES test_communicator_half_async ENVS ${dist_ENVS} FLAGS_communicator_send_queue_size=1 FLAGS_communicator_max_merge_var_num=1)
+    py_test_modules(test_communicator_sync MODULES test_communicator_sync ENVS ${dist_ENVS} FLAGS_communicator_send_queue_size=1 FLAGS_communicator_max_merge_var_num=1)
+
     if(WITH_DGC)
         # if with dgc, test all dgc tests.
         # NOTE. dist dgc tests is already in DIST_TEST_OPS

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

@@ -0,0 +1,64 @@
+#   Copyright (c) 2019 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 unittest
+import time
+import threading
+import numpy
+
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.communicator import Communicator
+
+import paddle.fluid.incubate.fleet.base.role_maker as role_maker
+from paddle.fluid.incubate.fleet.parameter_server.distribute_transpiler import fleet
+from paddle.fluid.incubate.fleet.parameter_server.distribute_transpiler.distributed_strategy import StrategyFactory
+
+
+class TestCommunicator(unittest.TestCase):
+    def net(self):
+        x = fluid.layers.data(name='x', shape=[13], dtype='float32')
+        y_predict = fluid.layers.fc(input=x, size=1, act=None)
+        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+
+        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+        avg_cost = fluid.layers.mean(cost)
+        return avg_cost
+
+    def test_communicator_sync(self):
+        role = role_maker.UserDefinedRoleMaker(
+            current_id=0,
+            role=role_maker.Role.WORKER,
+            worker_num=2,
+            server_endpoints=["127.0.0.1:6001", "127.0.0.1:6002"])
+
+        fleet.init(role)
+        avg_cost = self.net()
+
+        optimizer = fluid.optimizer.SGD(0.01)
+
+        strategy = StrategyFactory.create_sync_strategy()
+        strategy._program_config.wait_port = False
+        optimizer = fleet.distributed_optimizer(optimizer, strategy)
+        optimizer.minimize(avg_cost)
+
+        fleet.init_worker()
+        time.sleep(10)
+        fleet.stop_worker()
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -25,10 +25,9 @@ class TestStrategyFactor(unittest.TestCase):
     def test_sync_strategy(self):
         os.environ['CPU_NUM'] = "2"
         strategy = StrategyFactory.create_sync_strategy()
-        self.assertEqual(strategy._program_config.sync_mode, True)
-        self.assertEqual(strategy._program_config.runtime_split_send_recv,
-                         False)
-        self.assertEqual(strategy._build_strategy.async_mode, False)
+        self.assertEqual(strategy._program_config.sync_mode, False)
+        self.assertEqual(strategy._program_config.runtime_split_send_recv, True)
+        self.assertEqual(strategy._build_strategy.async_mode, True)
         self.assertEqual(strategy._execute_strategy.num_threads, 2)
 
         # test set_program_config using DistributeTranspilerConfig()

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -192,6 +192,7 @@ class DistributeTranspilerConfig(object):
 
     # half_async
     half_async = False
+    completely_not_async = False
 
     # Geo-sgd algorithm
     geo_sgd_mode = False
@@ -323,7 +324,7 @@ class DistributeTranspiler(object):
         if self.config.split_method is None:
             self.config.split_method = RoundRobin
 
-        if self.config.sync_mode:
+        if self.config.sync_mode or self.config.completely_not_async:
             self.distributed_mode = DistributedMode.SYNC
         elif self.config.runtime_split_send_recv:
             self.distributed_mode = DistributedMode.ASYNC
@@ -728,7 +729,14 @@ WIKI: https://github.com/PaddlePaddle/Fleet/blob/develop/markdown_doc/transpiler
                     program.global_block().vars[splited_grad_varname]
                 ]
                 sections = self._get_splited_var_sections(splited_vars)
-                send_varnames = [var.name for var in splited_vars]
+
+                if self.config.completely_not_async:
+                    send_varnames = [
+                        "{}.trainer_{}".format(var.name, self.trainer_id)
+                        for var in splited_vars
+                    ]
+                else:
+                    send_varnames = [var.name for var in splited_vars]
             else:
                 send_input_vars = splited_vars
                 sections = []
@@ -1199,7 +1207,7 @@ WIKI: https://github.com/PaddlePaddle/Fleet/blob/develop/markdown_doc/transpiler
                     type=v.type,
                     dtype=v.dtype,
                     shape=v.shape)
-            if self.sync_mode and self.trainer_num > 1:
+            if self.sync_mode or self.config.completely_not_async and self.trainer_num > 1:
                 for trainer_id in range(self.trainer_num):
                     var = pserver_program.global_block().create_var(
                         name="%s.trainer_%d" % (orig_var_name, trainer_id),
@@ -2204,7 +2212,7 @@ WIKI: https://github.com/PaddlePaddle/Fleet/blob/develop/markdown_doc/transpiler
 
         merged_var = pserver_block.vars[merged_var_name]
         grad_to_block_id.append(merged_var.name + ":" + str(optimize_block.idx))
-        if self.sync_mode and self.trainer_num > 1:
+        if self.sync_mode or self.config.completely_not_async and self.trainer_num > 1:
             vars2merge = []
             for i in range(self.trainer_num):
                 per_trainer_name = "%s.trainer_%d" % \