【paddle.distributed.fleet】Optimize ParameterServer's Async Mode (#28442)

* test=develop, optimize global_step

paddle/fluid/operators/distributed/communicator.h

@@ -302,16 +302,13 @@ class AsyncCommunicator : public Communicator {
             const std::vector<std::string> &var_tables,
             const framework::Scope &scope) override;
 
-  virtual void SendByCommunicator(int batches);
-
+  virtual void SendByCommunicator();
   virtual void SendGlobalStep(int batches);
 
   virtual void RecvByCommunicator();
 
   virtual void RecvNoBarrier();
 
-  virtual int BatchesCounter();
-
   virtual void BarrierSend() {}
 
   virtual void BarrierRecv() {}
@@ -359,6 +356,10 @@ class HalfAsyncCommunicator : public AsyncCommunicator {
     VLOG(0) << "HalfAsyncCommunicator Initialized";
   }
 
+  void MainThread() override;
+
+  void SendByCommunicator() override;
+
   void Clean() override;
 
   void Barrier() override;
@@ -438,7 +439,7 @@ class GeoCommunicator : public AsyncCommunicator {
             const std::vector<std::string> &var_tables,
             const framework::Scope &scope) override;
 
-  void SendByCommunicator(int batches) { return; }
+  void SendByCommunicator() { return; }
 
   std::vector<int64_t> MergeSparseIds(const std::string &send_varname);
 
@@ -475,6 +476,7 @@ class GeoCommunicator : public AsyncCommunicator {
   std::shared_ptr<Scope> pserver_scope_;
 
   int send_var_nums_ = 0;
+
   std::unordered_map<std::string, std::shared_ptr<SparseValue>> old_sparses_;
 
   std::unordered_map<

python/paddle/distributed/fleet/runtime/parameter_server_runtime.py

@@ -207,6 +207,7 @@ class ParameterServerRuntime(RuntimeBase):
             SyncStrategy, GeoStrategy
 
         trainer_config = self.async_strategy.get_trainer_runtime_config()
+        print(trainer_config)
 
         dist_strategy = self.context["valid_strategy"]
         launch_barrier = dist_strategy.a_sync_configs["launch_barrier"]