[cherry-pick]Fix communicator slow bug & fix communicator stop bug (#20366) (#20646)

* Fix communicator slow bug & fix communicator stop bug (#20366)

* test=develop,Fix communicator slow bug

* test=develop, delete if() in stop_worker()

* test=develop

* fix UT, test=develop

* fix bug in fetch handler, test=develop

* fix bug in fetch handler, test=develop

* test=develop, fix fetch barrier bug

* test=develop, bug fix

* test=develop, bug fix

* test=develop, fix bug

* test=develop,test=release/1.6

Author: MrChengmo

paddle/fluid/framework/dist_multi_trainer.cc

@@ -144,6 +144,10 @@ void DistMultiTrainer::Run() {
   }
 }
 
+Scope *DistMultiTrainer::GetWorkerScope(int thread_id) {
+  return workers_[thread_id]->GetThreadScope();
+}
+
 void DistMultiTrainer::Finalize() {
   for (auto &th : threads_) {
     th.join();
@@ -199,5 +203,5 @@ void DistMultiTrainer::MergeToRootScope(LoDTensor *root_tensor,
     root_data[i] += data[i];
   }
 }
-}  // end namespace framework
-}  // end namespace paddle
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/trainer.h

@@ -93,8 +93,8 @@ class DistMultiTrainer : public MultiTrainer {
   void MergeToRootScope(LoDTensor* root_tensor, LoDTensor* thread_tensor);
   virtual void FinalizeDumpEnv();
   virtual void InitDumpEnv();
+  virtual Scope* GetWorkerScope(int thread_id);
   virtual void DumpWork(int tid);
-  virtual Scope* GetWorkerScope(int thread_id) { return root_scope_; }
 
  protected:
   std::shared_ptr<paddle::framework::PullDenseWorker> pull_dense_worker_;

paddle/fluid/operators/distributed/communicator.cc

@@ -923,6 +923,7 @@ void GeoSgdCommunicator::RpcSend(const std::string &origin_var_name,
   auto &cpu_ctx_send = *pool.Get(platform::CPUPlace());
   distributed::RPCClient *rpc_client =
       distributed::RPCClient::GetInstance<RPCCLIENT_T>(trainer_id);
+
   rpc_client->AsyncSendVar(endpoint, cpu_ctx_send, *delta_scope_.get(),
                            splited_var_name);
 }

paddle/fluid/operators/distributed/rpc_client.cc

@@ -17,7 +17,7 @@
 
 // default to 3min to avoid temprary network failures.
 DEFINE_int32(rpc_deadline, 180000, "deadline timeouts for rpc");
-DEFINE_int32(rpc_retry_times, 3, "retry times for rpc");
+DEFINE_int32(rpc_retry_times, 0, "retry times for rpc");
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/distributed_ops/fetch_barrier_op.cc

@@ -55,6 +55,9 @@ class FetchBarrierOp : public framework::OperatorBase {
 class FetchBarrierOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() {
+    AddInput("X", "(Any) Dummy inputs, used for control dependency")
+        .AsDispensable()
+        .AsDuplicable();
     AddOutput("Out", "(Any) Dummy outputs, used for control dependency")
         .AsDuplicable();
     AddComment(R"DOC(

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

@@ -129,8 +129,7 @@ def stop_worker(self):
         Returns:
             None
         """
-        if not self._transpile_config.sync_mode and self._communicator.is_running(
-        ):
+        if not self._transpile_config.sync_mode:
             self._communicator.stop()
         self._executor.close()
         if isinstance(self._role_maker, MPISymetricRoleMaker):

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

@@ -67,7 +67,7 @@ def get_rand(low=0.0, high=1.0):
             return random.random()
 
         def iter():
-            if get_rand() < 0.1:
+            if get_rand() < 0.05:
                 fs = line.strip().split('\t')
                 dnn_input = load_dnn_input_record(fs[0])
                 lr_input = load_lr_input_record(fs[1])

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

@@ -139,7 +139,7 @@ def do_training(self, fleet):
         dataset.set_filelist(filelist)
         dataset.set_thread(thread_num)
 
-        for epoch_id in range(2):
+        for epoch_id in range(1):
             pass_start = time.time()
             dataset.set_filelist(filelist)
             exe.train_from_dataset(
@@ -157,7 +157,7 @@ def handler(self, fetch_target_vars):
                     print("{}: \n {}\n".format(self.fetch_target_names[0],
                                                fetch_target_vars[0]))
 
-        for epoch_id in range(2):
+        for epoch_id in range(1):
             pass_start = time.time()
             dataset.set_filelist(filelist)
             exe.train_from_dataset(

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

@@ -30,7 +30,6 @@ def __func__(*args, **kwargs):
     return __func__
 
 
-@skip_ci
 class TestDistMnist2x2(TestFleetBase):
     def _setup_config(self):
         self._sync_mode = False

python/paddle/fluid/trainer_factory.py

@@ -84,6 +84,9 @@ def handler_decorator(self, fetch_scope, fetch_handler):
                     for varname in fetch_target_names
                 ]
 
+                if None in fetch_vars:
+                    continue
+
                 fetch_tensors = [var.get_tensor() for var in fetch_vars]
 
                 if self.fetch_instance.return_np: