Merge pull request #10550 from Yancey1989/overlap_send_op

overlap send ops and backward ops

paddle/fluid/framework/details/CMakeLists.txt

@@ -3,7 +3,7 @@ cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context
 cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
 cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
 cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
-cc_library(send_op_handle SRCS send_op_handle.cc DEPS framework_proto scope place operator op_registry)
+cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry)
 
 cc_library(ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base)
 cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph)
@@ -26,7 +26,7 @@ endif()
 cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
 
 cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle
-        scale_loss_grad_op_handle send_op_handle ${multi_devices_graph_builder_deps} reduce_op_handle broadcast_op_handle)
+        scale_loss_grad_op_handle rpc_op_handle ${multi_devices_graph_builder_deps} reduce_op_handle broadcast_op_handle)
 
 cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph framework_proto)
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -12,12 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 #include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
+#include <fstream>
 #include <utility>
 #include "paddle/fluid/framework/details/broadcast_op_handle.h"
 #include "paddle/fluid/framework/details/computation_op_handle.h"
 #include "paddle/fluid/framework/details/reduce_op_handle.h"
+#include "paddle/fluid/framework/details/rpc_op_handle.h"
 #include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
-#include "paddle/fluid/framework/details/send_op_handle.h"
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/scope.h"
 
@@ -28,6 +29,10 @@
 #include <string>
 #include <vector>
 
+DEFINE_string(ssa_graph_path, "/tmp/ssa_graph.dot",
+              "the ssa graph path only print with GLOG_v=10,"
+              "default /tmp/graph.dot");
+
 namespace paddle {
 namespace framework {
 namespace details {
@@ -79,9 +84,44 @@ void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result,
   }
 }
 
-bool MultiDevSSAGraphBuilder::IsDistTrainOp(const OpDesc &op,
-                                            OpDesc *send_op) const {
-  if (send_op == nullptr) {
+std::vector<std::string> MultiDevSSAGraphBuilder::FindDistTrainSendVars(
+    const ProgramDesc &program) const {
+  std::vector<std::string> send_vars;
+  // since parameters are all in block 0,
+  // it's enough to only scan send ops in block 0
+  for (auto *op : program.Block(0).AllOps()) {
+    // TODO(Yancey1989): use a graceful method to find send op,
+    // instead of the the hard code string
+    if (op->Type() == "send_vars") {
+      auto op_vars = op->InputArgumentNames();
+      send_vars.reserve(send_vars.size() +
+                        std::distance(op_vars.begin(), op_vars.end()));
+      send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end());
+    }
+  }
+  return send_vars;
+}
+
+std::vector<std::string> MultiDevSSAGraphBuilder::FindDistTrainRecvVars(
+    const ProgramDesc &program) const {
+  std::vector<std::string> recv_vars;
+  for (auto *op : program.Block(0).AllOps()) {
+    // TODO(Yancey1989): use a graceful method to find recv op,
+    // instead of the hard code string
+    if (op->Type() == "recv") {
+      auto op_vars = op->OutputArgumentNames();
+      recv_vars.reserve(recv_vars.size() +
+                        std::distance(op_vars.begin(), op_vars.end()));
+      recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end());
+    }
+  }
+  return recv_vars;
+}
+
+bool MultiDevSSAGraphBuilder::IsDistTrainOp(
+    const OpDesc &op, const std::vector<std::string> &send_vars,
+    const std::vector<std::string> &recv_vars) const {
+  if (send_vars.size() == 0 || recv_vars.size() == 0) {
     return false;
   }
 
@@ -89,22 +129,21 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp(const OpDesc &op,
    * Check any of opvars contains `.block` and in sendvars
    */
   auto checker = [](const std::vector<std::string> &opvars,
-                    const std::vector<std::string> &sendvars) -> bool {
+                    const std::vector<std::string> &rpc_vars) -> bool {
     for (auto &var : opvars) {
+      // a variable name with the suffix `.block` means it's a splited
+      // variable by (DistributeTranspiler)
+      // [python/paddle/fluid/transpiler/distribute_transpiler.py]
       if (var.find(".block") != std::string::npos &&
-          std::find(sendvars.begin(), sendvars.end(), var) != sendvars.end()) {
+          std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) {
         return true;
       }
     }
     return false;
   };
 
-  if (op.Type() == "split" || op.Type() == "split_byref") {
-    return checker(op.OutputArgumentNames(), send_op->InputArgumentNames());
-  } else if (op.Type() == "concat") {
-    return checker(op.InputArgumentNames(), send_op->OutputArgumentNames());
-  }
-  return false;
+  return checker(op.OutputArgumentNames(), send_vars) ||
+         checker(op.InputArgumentNames(), recv_vars);
 }
 
 std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
@@ -123,8 +162,10 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
       std::unordered_map<std::string, std::vector<std::unique_ptr<VarHandle>>>>(
       places_.size());
 
-  // Find "send" op first for split is in front of send.
-  OpDesc *send_op = GetSendOpDesc(program);
+  // find send/recv vars so that we can place the distributed training
+  // realted op in the place 0
+  auto send_vars = FindDistTrainSendVars(program);
+  auto recv_vars = FindDistTrainRecvVars(program);
 
   size_t cur_device_id = 0;
   std::vector<std::unordered_set<std::string>> var_name_on_devices;
@@ -134,12 +175,14 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
 
   bool is_forwarding = true;
   for (auto *op : program.Block(0).AllOps()) {
-    if (op->Type() == "send") {
-      // append send op if program is distributed trainer main program.
+    if (boost::get<int>(
+            op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) ==
+        static_cast<int>(OpRole::kRPC)) {
+      // append rpc op if program is distributed trainer main program.
       // always use the first device
-      CreateSendOp(&result, *op);
-    } else if (IsDistTrainOp(*op, send_op)) {
-      CreateComputationalOps(&result, *op, 1);
+      CreateRPCOp(&result, *op);
+    } else if (IsDistTrainOp(*op, send_vars, recv_vars)) {
+      CreateDistTrainOp(&result, *op);
     } else if (IsScaleLossOp(*op)) {
       // user can customize loss@grad if not use_default_grad_scale_
       if (strategy_.gradient_scale_ !=
@@ -218,9 +261,8 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   AddOutputToLeafOps(&result);
 
   if (VLOG_IS_ON(10)) {
-    std::ostringstream sout;
-    PrintGraphviz(*graph, sout);
-    VLOG(10) << sout.str();
+    std::ofstream fout(FLAGS_ssa_graph_path);
+    PrintGraphviz(*graph, fout);
   }
 
   return std::unique_ptr<SSAGraph>(graph);
@@ -270,15 +312,6 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(SSAGraph *result,
   CreateOpHandleIOs(result, op, dev_id);
 }
 
-OpDesc *MultiDevSSAGraphBuilder::GetSendOpDesc(
-    const ProgramDesc &program) const {
-  for (auto *op : program.Block(0).AllOps()) {
-    if (op->Type() == "send") {
-      return op;
-    }
-  }
-  return nullptr;
-}
 void MultiDevSSAGraphBuilder::InsertNCCLAllReduceOp(
     SSAGraph *result, const std::string &og) const {
 #ifdef PADDLE_WITH_CUDA
@@ -401,14 +434,48 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result,
   return var;
 }
 
-void MultiDevSSAGraphBuilder::CreateSendOp(SSAGraph *result,
-                                           const OpDesc &op) const {
+void MultiDevSSAGraphBuilder::ConnectOp(SSAGraph *result, OpHandleBase *op,
+                                        const std::string &prev_op_name) const {
+  for (auto &prev_op : result->ops_) {
+    if (prev_op->Name() == prev_op_name) {
+      auto *dep_var = new DummyVarHandle();
+      prev_op->AddOutput(dep_var);
+      result->dep_vars_.emplace(dep_var);
+      op->AddInput(dep_var);
+    }
+  }
+}
+
+void MultiDevSSAGraphBuilder::CreateDistTrainOp(SSAGraph *result,
+                                                const OpDesc &op) const {
+  CreateComputationalOp(result, op, 0);
+  if (op.Type() == "concat") {
+    ConnectOp(result, result->ops_.back().get(), "fetch_barrier");
+  }
+}
+
+void MultiDevSSAGraphBuilder::CreateRPCOp(SSAGraph *result,
+                                          const OpDesc &op) const {
   auto &p = places_[0];
   auto *s = local_scopes_[0];
-  // FIXME(wuyi): send op always copy from GPU 0
-  result->ops_.emplace_back(new SendOpHandle(op, s, p));
-  // Create inputs for output on original place and no ssa output
-  // is created for send op.
+  result->ops_.emplace_back(new RPCOpHandle(op, s, p, op.Type()));
+
+  if (op.Type() == "send_barrier") {
+    ConnectOp(result, result->ops_.back().get(), "send_vars");
+  } else if (op.Type() == "recv") {
+    ConnectOp(result, result->ops_.back().get(), "send_barrier");
+  } else if (op.Type() == "fetch_barrier") {
+    ConnectOp(result, result->ops_.back().get(), "recv");
+  } else if (op.Type() == "send_vars") {
+    // do nothing
+  } else {
+    PADDLE_THROW(
+        "rpc op should be in ["
+        "send_vars, send_barrier. recv, fetch_barrier]");
+  }
+
+  // TODO(Yancey1989): schedule rpc op on different place may
+  // increate throughput
   CreateOpHandleIOs(result, op, 0);
 }
 

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -64,12 +64,24 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 
   bool IsScaleLossOp(const OpDesc &op) const;
 
-  void CreateSendOp(SSAGraph *result, const OpDesc &op) const;
+  void CreateRPCOp(SSAGraph *result, const OpDesc &op) const;
+  void CreateDistTrainOp(SSAGraph *result, const OpDesc &op) const;
 
   /**
    * Is this operator as the end-point operator before/after send operator.
    */
-  bool IsDistTrainOp(const OpDesc &op, OpDesc *send_op) const;
+  bool IsDistTrainOp(const OpDesc &op,
+                     const std::vector<std::string> &send_vars,
+                     const std::vector<std::string> &recv_vars) const;
+
+  std::vector<std::string> FindDistTrainSendVars(
+      const ProgramDesc &program) const;
+
+  std::vector<std::string> FindDistTrainRecvVars(
+      const ProgramDesc &program) const;
+
+  void ConnectOp(SSAGraph *result, OpHandleBase *op,
+                 const std::string &prev_op_name) const;
 
   void CreateComputationalOps(SSAGraph *result, const OpDesc &op,
                               size_t num_places) const;
@@ -93,12 +105,6 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
   void CreateBroadcastOp(SSAGraph *result, const std::string &p_name,
                          size_t src_dev_id) const;
 
-  /**
-   * Get send op in the global block of program.
-   * nullptr if not found.
-   */
-  OpDesc *GetSendOpDesc(const ProgramDesc &program) const;
-
   bool IsSparseGradient(
       const std::unordered_map<std::string, proto::VarType::Type> &var_types,
       const std::string &og) const;

paddle/fluid/framework/details/send_op_handle.cc → paddle/fluid/framework/details/rpc_op_handle.cc

@@ -12,24 +12,26 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/details/send_op_handle.h"
+#include "paddle/fluid/framework/details/rpc_op_handle.h"
 
 namespace paddle {
 namespace framework {
 namespace details {
 
-SendOpHandle::SendOpHandle(const framework::OpDesc &op_desc,
-                           const Scope *local_scope,
-                           const platform::Place &place)
+RPCOpHandle::RPCOpHandle(const framework::OpDesc &op_desc,
+                         const Scope *local_scope, const platform::Place &place,
+                         const std::string &name)
     : op_(framework::OpRegistry::CreateOp(op_desc)),
       local_scope_(local_scope),
-      place_(place) {}
+      place_(place),
+      name_(name) {}
 
-void SendOpHandle::RunImpl() {
+void RPCOpHandle::RunImpl() {
   // TODO(wuyi): need further analysis whether wait VarDummyHandle.
   // Wait input done
   for (auto *in : inputs_) {
     auto &p = static_cast<VarHandle *>(in)->place_;
+    // FIXME(Yancey1989): need a better solution instead of use DebugString()
     if (in->DebugString() == "dummy") {  // HACK
       continue;
     }
@@ -43,7 +45,7 @@ void SendOpHandle::RunImpl() {
   op_->Run(*tmp_scope, place_);
 }
 
-std::string SendOpHandle::Name() const { return "send"; }
+std::string RPCOpHandle::Name() const { return name_; }
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/send_op_handle.h → paddle/fluid/framework/details/rpc_op_handle.h

@@ -27,9 +27,9 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-struct SendOpHandle : public OpHandleBase {
-  SendOpHandle(const framework::OpDesc& op_desc, const Scope* local_scope,
-               const platform::Place& place);
+struct RPCOpHandle : public OpHandleBase {
+  RPCOpHandle(const framework::OpDesc& op_desc, const Scope* local_scope,
+              const platform::Place& place, const std::string& name);
 
   std::string Name() const override;
 
@@ -44,6 +44,7 @@ struct SendOpHandle : public OpHandleBase {
   std::unique_ptr<OperatorBase> op_;
   const Scope* local_scope_;
   const platform::Place& place_;
+  const std::string name_;
 };
 
 }  // namespace details

paddle/fluid/framework/op_proto_maker.cc

@@ -66,7 +66,7 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
       .InEnum(
           {static_cast<int>(OpRole::kForward),
            static_cast<int>(OpRole::kBackward),
-           static_cast<int>(OpRole::kOptimize),
+           static_cast<int>(OpRole::kOptimize), static_cast<int>(OpRole::kRPC),
            static_cast<int>(OpRole::kLoss) | static_cast<int>(OpRole::kForward),
            static_cast<int>(OpRole::kLoss) |
                static_cast<int>(OpRole::kBackward),

paddle/fluid/framework/op_proto_maker.h

@@ -24,6 +24,7 @@ enum class OpRole {
   kForward = 0x0000,
   kBackward = 0x0001,
   kOptimize = 0x0002,
+  kRPC = 0x0003,
 
   kLoss = 0x0100,
   // The default value of op's role. This should be only used for unittests and

paddle/fluid/inference/analysis/data_flow_graph_tester.cc

@@ -35,7 +35,7 @@ TEST(DataFlowGraph, BFS) {
 
   GraphTraits<DataFlowGraph> trait(&dfg);
   auto nodes = trait.nodes();
-  int count = 0;
+  size_t count = 0;
   for (auto it = nodes.begin(); it != nodes.end(); ++it) {
     LOG(INFO) << "visiting " << it->name();
     ++count;
@@ -49,7 +49,7 @@ TEST(DataFlowGraph, DFS) {
   dfg.Build();
   GraphTraits<DataFlowGraph> trait(&dfg);
   auto nodes = trait.nodes_in_DFS();
-  int count = 0;
+  size_t count = 0;
   for (auto it = nodes.begin(); it != nodes.end(); ++it) {
     LOG(INFO) << "visiting " << it->name();
     ++count;

paddle/fluid/operators/CMakeLists.txt

@@ -200,7 +200,9 @@ if(WITH_DISTRIBUTE)
     op_library(send_vars_op DEPS ${DISTRIBUTE_DEPS})
     set_source_files_properties(send_vars_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     op_library(send_barrier_op DEPS ${DISTRIBUTE_DEPS})
+    op_library(fetch_barrier_op DEPS ${DISTRIBUTE_DEPS})
     set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    set_source_files_properties(fetch_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op
     #        listen_and_serv_op sum_op executor SERIAL)
@@ -214,7 +216,7 @@ if(WITH_DISTRIBUTE)
         set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op)
     endif()
 else()
-    set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op gen_nccl_id_op)
+    set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op fetch_barrier_op gen_nccl_id_op)
 endif()
 
 op_library(cross_entropy_op DEPS cross_entropy)

paddle/fluid/operators/detail/grpc_client.cc

@@ -25,6 +25,21 @@ namespace paddle {
 namespace operators {
 namespace detail {
 
+std::once_flag RPCClient::init_flag_;
+
+std::unique_ptr<RPCClient> RPCClient::rpc_client_(nullptr);
+
+RPCClient* RPCClient::GetInstance() {
+  std::call_once(init_flag_, &RPCClient::Init);
+  return rpc_client_.get();
+}
+
+void RPCClient::Init() {
+  if (rpc_client_.get() == nullptr) {
+    rpc_client_.reset(new RPCClient());
+  }
+}
+
 bool RPCClient::AsyncSendVariable(const std::string& ep,
                                   const platform::DeviceContext& ctx,
                                   const framework::Scope& scope,
@@ -60,7 +75,6 @@ bool RPCClient::AsyncSendVariable(const std::string& ep,
     call->StartCall();
     call->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
   });
-
   req_count_++;
 
   return true;
@@ -249,8 +263,9 @@ bool RPCClient::Proceed() {
   delete c;
   return true;
 }
-
 std::shared_ptr<grpc::Channel> RPCClient::GetChannel(const std::string& ep) {
+  // TODO(Yancey1989): make grpc client completely thread-safe
+  std::unique_lock<std::mutex> lock(mutex_);
   auto it = channels_.find(ep);
   if (it != channels_.end()) {
     return it->second;
@@ -263,7 +278,6 @@ std::shared_ptr<grpc::Channel> RPCClient::GetChannel(const std::string& ep) {
 
   auto ch =
       grpc::CreateCustomChannel(ep, grpc::InsecureChannelCredentials(), args);
-
   channels_[ep] = ch;
   return ch;
 }

paddle/fluid/operators/detail/grpc_client.h

@@ -21,6 +21,7 @@ limitations under the License. */
 #include <functional>
 #include <iostream>
 #include <map>
+#include <mutex>  // NOLINT
 #include <string>
 #include <vector>
 
@@ -35,6 +36,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/operators/detail/sendrecvop_utils.h"
+#include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
 
 namespace paddle {
 namespace operators {
@@ -161,6 +163,10 @@ class FetchBarrierProcessor : public BaseProcessor {
 
 class RPCClient {
  public:
+  RPCClient() {}
+
+  static RPCClient* GetInstance();
+
   bool AsyncSendVariable(const std::string& ep,
                          const platform::DeviceContext& ctx,
                          const framework::Scope& scope,
@@ -191,11 +197,17 @@ class RPCClient {
  private:
   bool Proceed();
   std::shared_ptr<grpc::Channel> GetChannel(const std::string& ep);
+  // Init is called by GetInstance.
+  static void Init();
 
  private:
   grpc::CompletionQueue cq_;
   std::map<std::string, std::shared_ptr<grpc::Channel>> channels_;
-  int64_t req_count_ = 0;
+  std::atomic<int64_t> req_count_{0};
+  std::mutex mutex_;
+  static std::unique_ptr<RPCClient> rpc_client_;
+  static std::once_flag init_flag_;
+  DISABLE_COPY_AND_ASSIGN(RPCClient);
 };
 
 }  // namespace detail

paddle/fluid/operators/detail/grpc_server_test.cc

@@ -121,10 +121,10 @@ TEST(PREFETCH, DISABLED_CPU) {
   std::string in_var_name("ids");
   std::string out_var_name("out");
 
-  detail::RPCClient client;
-  client.AsyncPrefetchVariable("127.0.0.1:8889", ctx, scope, in_var_name,
-                               out_var_name);
-  client.Wait();
+  auto client = detail::RPCClient::GetInstance();
+  client->AsyncPrefetchVariable("127.0.0.1:8889", ctx, scope, in_var_name,
+                                out_var_name);
+  client->Wait();
 
   auto var = scope.Var(out_var_name);
   auto value = var->GetMutable<framework::SelectedRows>()->value();

paddle/fluid/operators/fetch_barrier_op.cc

+/* Copyright (c) 2016 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. */
+
+#include <future>  // NOLINT
+#include <ostream>
+
+#include "paddle/fluid/framework/data_type.h"
+#include "paddle/fluid/framework/framework.pb.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+#include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/platform/profiler.h"
+
+namespace paddle {
+namespace operators {
+
+class FetchBarrierOp : public framework::OperatorBase {
+ public:
+  FetchBarrierOp(const std::string& type,
+                 const framework::VariableNameMap& inputs,
+                 const framework::VariableNameMap& outputs,
+                 const framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+
+  void RunImpl(const framework::Scope& scope,
+               const platform::Place& place) const override {
+    std::vector<std::string> eps = Attr<std::vector<std::string>>("endpoints");
+
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    auto& ctx = *pool.Get(place);
+    // For profiling
+    platform::RecordEvent record_event(Type(), &ctx);
+
+    auto rpc_client = detail::RPCClient::GetInstance();
+
+    PADDLE_ENFORCE(rpc_client->Wait());
+
+    for (auto& ep : eps) {
+      VLOG(3) << "fetch barrier, ep: " << ep;
+      rpc_client->AsyncSendFetchBarrier(ep);
+    }
+    PADDLE_ENFORCE(rpc_client->Wait());
+  }
+};
+
+class FetchBarrierOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddComment(R"DOC(
+SendBarrier operator
+
+This operator will send a send barrier signal to list_and_serv op, so that
+the Parameter Server would knew all variables have been sent.
+)DOC");
+
+    AddAttr<std::vector<std::string>>("endpoints",
+                                      "(string vector, default 127.0.0.1:6164)"
+                                      "Server endpoints to send variables to.")
+        .SetDefault({"127.0.0.1:6164"});
+  }
+};
+
+class FetchBarrierOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {}
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(fetch_barrier, ops::FetchBarrierOp,
+                  paddle::framework::EmptyGradOpMaker, ops::FetchBarrierOpMaker,
+                  ops::FetchBarrierOpShapeInference);

paddle/fluid/operators/prefetch_op.cc

@@ -41,12 +41,7 @@ class PrefetchOp : public framework::OperatorBase {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
 
-    auto client_var_name = Output("RPCClient");
-    PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name),
-                            "Can not find variable '%s' in the scope.",
-                            client_var_name);
-    auto* client_var = scope.FindVar(client_var_name);
-    detail::RPCClient* rpc_client = client_var->GetMutable<detail::RPCClient>();
+    auto rpc_client = detail::RPCClient::GetInstance();
 
     for (size_t i = 0; i < ins.size(); i++) {
       if (NeedSend(scope, ins[i])) {
@@ -66,9 +61,6 @@ class PrefetchOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() {
     AddInput("X", "(LoDTensor) Input Id variables to be sent").AsDuplicable();
-    AddOutput("RPCClient",
-              "(RPCClient) The RPC client object which will be"
-              "initialized at most once.");
     AddOutput("Out",
               "(LoDTensor) result "
               "to be fetched from parameter server")
@@ -87,17 +79,6 @@ the parameter server and fetch result back.
   }
 };
 
-class PrefetchOpVarTypeInference : public framework::VarTypeInference {
- public:
-  void operator()(const framework::OpDesc& op_desc,
-                  framework::BlockDesc* block) const override {
-    auto out_var_name = op_desc.Output("RPCClient").front();
-    auto& out_var = block->FindRecursiveOrCreateVar(out_var_name);
-    auto var_type = framework::proto::VarType::RAW;
-    out_var.SetType(var_type);
-  }
-};
-
 class PrefetchOpShapeInference : public framework::InferShapeBase {
  public:
   void operator()(framework::InferShapeContext* ctx) const override {}
@@ -110,5 +91,4 @@ namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(prefetch, ops::PrefetchOp,
                   paddle::framework::EmptyGradOpMaker, ops::PrefetchOpMaker,
-                  ops::PrefetchOpVarTypeInference,
                   ops::PrefetchOpShapeInference);

paddle/fluid/operators/recv_op.cc

@@ -21,6 +21,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 
 #include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
@@ -36,19 +37,23 @@ class RecvOp : public framework::OperatorBase {
                const platform::Place& place) const override {
     auto outs = Outputs("Out");
     std::vector<std::string> epmap = Attr<std::vector<std::string>>("epmap");
+    int sync_mode = Attr<int>("sync_mode");
 
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
+    // For profiling
+    platform::RecordEvent record_event(Type(), &ctx);
+
+    auto rpc_client = detail::RPCClient::GetInstance();
 
     for (size_t i = 0; i < outs.size(); i++) {
-      VLOG(3) << "getting " << outs[i];
-      client_.AsyncGetVariable(epmap[i], ctx, scope, outs[i]);
+      VLOG(3) << "getting " << outs[i] << " from " << epmap[i];
+      rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]);
+    }
+    if (sync_mode) {
+      PADDLE_ENFORCE(rpc_client->Wait());
     }
-    PADDLE_ENFORCE(client_.Wait());
   }
-
- private:
-  mutable detail::RPCClient client_;
 };
 
 class RecvOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -65,6 +70,10 @@ This operator can get variables from server side.
                                       "Server endpoints in the order of input "
                                       "variables for mapping")
         .SetDefault({});
+    AddAttr<int>("sync_mode",
+                 "(int, default 0)"
+                 "sync recv or async recv.")
+        .SetDefault(0);
   }
 };
 

paddle/fluid/operators/send_barrier_op.cc

@@ -21,6 +21,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 
 #include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
@@ -36,31 +37,30 @@ class SendBarrierOp : 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");
+    bool sync_mode = Attr<bool>("sync_mode");
 
-    auto client_var_name = Output("RPCClient");
-    PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name),
-                            "Can not find variable '%s' in the scope.",
-                            client_var_name);
-    auto* client_var = scope.FindVar(client_var_name);
-    detail::RPCClient* rpc_client = client_var->GetMutable<detail::RPCClient>();
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    auto& ctx = *pool.Get(place);
+    // For profiling
+    platform::RecordEvent record_event(Type(), &ctx);
+
+    auto rpc_client = detail::RPCClient::GetInstance();
 
     // need to wait before sending send_barrier message
     PADDLE_ENFORCE(rpc_client->Wait());
-
-    for (auto& ep : eps) {
-      VLOG(3) << "send barrier, ep: " << ep;
-      rpc_client->AsyncSendBatchBarrier(ep);
+    if (sync_mode) {
+      for (auto& ep : eps) {
+        VLOG(3) << "send barrier, ep: " << ep;
+        rpc_client->AsyncSendBatchBarrier(ep);
+      }
+      PADDLE_ENFORCE(rpc_client->Wait());
     }
-    PADDLE_ENFORCE(rpc_client->Wait());
   }
 };
 
 class SendBarrierOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() {
-    AddOutput("RPCClient",
-              "(RPCClient) The RPC client object which is"
-              "initialized at most once.");
     AddComment(R"DOC(
 SendBarrier operator
 
@@ -72,17 +72,7 @@ the Parameter Server would knew all variables have been sent.
                                       "(string vector, default 127.0.0.1:6164)"
                                       "Server endpoints to send variables to.")
         .SetDefault({"127.0.0.1:6164"});
-  }
-};
-
-class SendBarrierOpVarTypeInference : public framework::VarTypeInference {
- public:
-  void operator()(const framework::OpDesc& op_desc,
-                  framework::BlockDesc* block) const override {
-    auto out_var_name = op_desc.Output("RPCClient").front();
-    auto& out_var = block->FindRecursiveOrCreateVar(out_var_name);
-    auto var_type = framework::proto::VarType::RAW;
-    out_var.SetType(var_type);
+    AddAttr<bool>("sync_mode", "work in sync_mode or not").SetDefault(true);
   }
 };
 
@@ -98,5 +88,4 @@ namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(send_barrier, ops::SendBarrierOp,
                   paddle::framework::EmptyGradOpMaker, ops::SendBarrierOpMaker,
-                  ops::SendBarrierOpVarTypeInference,
                   ops::SendBarrierOpShapeInference);

paddle/fluid/operators/send_op.cc

@@ -49,12 +49,7 @@ class SendOp : public framework::OperatorBase {
     // For profiling
     platform::RecordEvent record_event(Type(), &ctx);
 
-    auto client_var_name = Output("RPCClient");
-    PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name),
-                            "Can not find variable '%s' in the scope.",
-                            client_var_name);
-    auto* client_var = scope.FindVar(client_var_name);
-    detail::RPCClient* rpc_client = client_var->GetMutable<detail::RPCClient>();
+    auto rpc_client = detail::RPCClient::GetInstance();
 
     for (size_t i = 0; i < ins.size(); i++) {
       if (NeedSend(scope, ins[i])) {
@@ -96,9 +91,6 @@ class SendOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "(Tensor) Input tensor to be sent").AsDuplicable();
     AddOutput("Out", "(Tensor) Output tensor to be received from server")
         .AsDuplicable();
-    AddOutput("RPCClient",
-              "(RPCClient) The RPC client object which is"
-              "initialized at most once.");
     AddComment(R"DOC(
 Send operator
 
@@ -119,17 +111,6 @@ This operator will send tensor to recv_op at the parameter server.
   }
 };
 
-class SendOpVarTypeInference : public framework::VarTypeInference {
- public:
-  void operator()(const framework::OpDesc& op_desc,
-                  framework::BlockDesc* block) const override {
-    auto out_var_name = op_desc.Output("RPCClient").front();
-    auto& out_var = block->FindRecursiveOrCreateVar(out_var_name);
-    auto var_type = framework::proto::VarType::RAW;
-    out_var.SetType(var_type);
-  }
-};
-
 class SendOpShapeInference : public framework::InferShapeBase {
  public:
   void operator()(framework::InferShapeContext* ctx) const override {}
@@ -141,5 +122,4 @@ class SendOpShapeInference : public framework::InferShapeBase {
 namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(send, ops::SendOp, paddle::framework::EmptyGradOpMaker,
-                  ops::SendOpMaker, ops::SendOpVarTypeInference,
-                  ops::SendOpShapeInference);
+                  ops::SendOpMaker, ops::SendOpShapeInference);

paddle/fluid/operators/send_recv_op_test.cc

@@ -156,6 +156,7 @@ TEST(SendRecvOp, CPUDense) {
   std::thread server_thread(StartServerNet, false, &initialized);
   while (!initialized) {
   }
+
   static_cast<paddle::operators::ListenAndServOp *>(listen_and_serv_op.get())
       ->WaitServerReady();
 
@@ -175,9 +176,10 @@ TEST(SendRecvOp, CPUDense) {
   std::string endpoint = paddle::string::Sprintf("127.0.0.1:%d", selected_port);
   attrs.insert({"endpoints", std::vector<std::string>({endpoint})});
   attrs.insert({"epmap", std::vector<std::string>({endpoint})});
-  auto send_op = f::OpRegistry::CreateOp(
-      "send", {{"X", {"x1"}}},
-      {{"Out", {"Out"}}, {"RPCClient", {"RPC_CLIENT_VAR"}}}, attrs);
+  const f::VariableNameMap &inputs = {{"X", {"x1"}}};
+  const f::VariableNameMap &outputs = {{"Out", {"Out"}}};
+
+  auto send_op = f::OpRegistry::CreateOp("send", inputs, outputs, attrs);
   send_op->Run(scope, place);
 
   auto in_var = scope.Var("x1");
@@ -220,9 +222,8 @@ TEST(SendRecvOp, CPUSparse) {
   std::string endpoint = paddle::string::Sprintf("127.0.0.1:%d", selected_port);
   attrs.insert({"endpoints", std::vector<std::string>({endpoint})});
   attrs.insert({"epmap", std::vector<std::string>({endpoint})});
-  auto send_op = f::OpRegistry::CreateOp(
-      "send", {{"X", {"x1"}}},
-      {{"Out", {"Out"}}, {"RPCClient", {"RPC_CLIENT_VAR"}}}, attrs);
+  auto send_op = f::OpRegistry::CreateOp("send", {{"X", {"x1"}}},
+                                         {{"Out", {"Out"}}}, attrs);
   send_op->Run(scope, place);
 
   auto x0 = scope.Var("x0")->GetMutable<f::SelectedRows>();

paddle/fluid/operators/send_recv_util.h

@@ -20,6 +20,9 @@ namespace operators {
 
 inline bool NeedSend(const framework::Scope& scope,
                      const std::string& varname) {
+  // dummy variable is only used in parallel executor to represent
+  // some dependency relationship, we don't need to send/recv it.
+  if (varname == "dummy") return false;
   auto* var = scope.FindVar(varname);
   PADDLE_ENFORCE_NOT_NULL(var, "Can not find variable '%s' in the send side.",
                           varname);

paddle/fluid/operators/send_vars_op.cc

@@ -20,6 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/grpc_client.h"
 #include "paddle/fluid/operators/send_recv_util.h"
+#include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
@@ -41,12 +42,10 @@ class SendVarsOp : public framework::OperatorBase {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
 
-    auto client_var_name = Output("RPCClient");
-    PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name),
-                            "Can not find variable '%s' in the scope.",
-                            client_var_name);
-    auto* client_var = scope.FindVar(client_var_name);
-    detail::RPCClient* rpc_client = client_var->GetMutable<detail::RPCClient>();
+    // For profiling
+    platform::RecordEvent record_event(Type(), &ctx);
+
+    auto rpc_client = detail::RPCClient::GetInstance();
 
     for (size_t i = 0; i < ins.size(); i++) {
       if (NeedSend(scope, ins[i])) {
@@ -69,9 +68,6 @@ class SendVarsOpMaker : public framework::OpProtoAndCheckerMaker {
   void Make() {
     AddInput("X", "(Tensor, SelectedRows) Input variables to be sent")
         .AsDuplicable();
-    AddOutput("RPCClient",
-              "(RPCClient) The RPC client object which will be"
-              "initialized at most once.");
     AddComment(R"DOC(
 Send operator
 
@@ -89,17 +85,6 @@ This operator will send variables to listen_and_serve op at the parameter server
   }
 };
 
-class SendVarsOpVarTypeInference : public framework::VarTypeInference {
- public:
-  void operator()(const framework::OpDesc& op_desc,
-                  framework::BlockDesc* block) const override {
-    auto out_var_name = op_desc.Output("RPCClient").front();
-    auto& out_var = block->FindRecursiveOrCreateVar(out_var_name);
-    auto var_type = framework::proto::VarType::RAW;
-    out_var.SetType(var_type);
-  }
-};
-
 class SendVarsOpShapeInference : public framework::InferShapeBase {
  public:
   void operator()(framework::InferShapeContext* ctx) const override {}
@@ -112,5 +97,4 @@ namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(send_vars, ops::SendVarsOp,
                   paddle::framework::EmptyGradOpMaker, ops::SendVarsOpMaker,
-                  ops::SendVarsOpVarTypeInference,
                   ops::SendVarsOpShapeInference);

paddle/fluid/pybind/const_value.cc

@@ -32,7 +32,8 @@ void BindConstValue(pybind11::module* m) {
       .value("Forward", framework::OpRole::kForward)
       .value("Backward", framework::OpRole::kBackward)
       .value("Optimize", framework::OpRole::kOptimize)
-      .value("Loss", framework::OpRole::kLoss);
+      .value("Loss", framework::OpRole::kLoss)
+      .value("RPC", framework::OpRole::kRPC);
 
   op_proto_and_checker_maker.def(
       "kOpRoleAttrName", framework::OpProtoAndCheckerMaker::OpRoleAttrName);

python/paddle/fluid/layers/io.py

@@ -195,21 +195,23 @@ def Send(endpoints, send_vars, get_vars=None):
     endpoints = list(set(epmap))
 
     helper = LayerHelper("Send", **locals())
-    rpc_client_var = default_main_program().global_block().create_var(
-        name="RPC_CLIENT_VAR", persistable=True, type=core.VarDesc.VarType.RAW)
     if not get_vars:
         get_vars = []
         for s in send_vars:
             v = helper.create_tmp_variable(dtype=s.dtype, stop_gradient=True)
             get_vars.append(v)
+    rpc_op_role_name = core.op_proto_and_checker_maker.kOpRoleAttrName()
 
     helper.append_op(
         type="send",
         inputs={"X": send_vars},
-        outputs={"Out": get_vars,
-                 "RPCClient": rpc_client_var},
-        attrs={"endpoints": endpoints,
-               "epmap": epmap})
+        outputs={"Out": get_vars},
+        attrs={
+            "endpoints": endpoints,
+            "epmap": epmap,
+            rpc_op_role_name: core.op_proto_and_checker_maker.OpRole.RPC
+        })
+
     return get_vars
 
 

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

@@ -49,7 +49,6 @@ class TestDistTranspiler(unittest.TestCase):
     def test_transpiler(self):
         trainer = self.get_trainer()
         pserver, startup = self.get_pserver(self.current_pserver_ep)
-
         self.assertEqual([op.type for op in trainer.global_block().ops],
                          self.get_expect_trainer_ops())
 
@@ -67,7 +66,7 @@ class TestDistTranspiler(unittest.TestCase):
             "fill_constant", "fill_constant", "uniform_random", "uniform_random"
         ])
 
-        # the variable #fc_w will be split into two blocks 
+        # the variable #fc_w will be split into two blocks
         fc_w_var = startup.global_block().var("fc_w.block1")
         self.assertEqual(fc_w_var.shape, (500, 1000))
 
@@ -86,8 +85,12 @@ class TestDistTranspiler(unittest.TestCase):
             optimize_ops, params_grads = self.net_conf()
 
         delete_ops(trainer.global_block(), optimize_ops)
-        return [op.type for op in trainer.global_block().ops
-                ] + ["split_byref", "send", "concat"]
+        ops = [op.type for op in trainer.global_block().ops] + [
+            "split_byref", "send_vars", "send_barrier", "recv", "recv",
+            "fetch_barrier", "concat"
+        ]
+        ops.insert(ops.index("elementwise_add_grad") + 1, "send_vars")
+        return ops
 
     def get_trainer(self):
         return self._transpiler_instance().get_trainer_program()

python/paddle/fluid/transpiler/__init__.py

@@ -16,8 +16,9 @@ from distribute_transpiler import DistributeTranspiler
 from inference_transpiler import InferenceTranspiler
 from memory_optimization_transpiler import memory_optimize, release_memory
 from distribute_transpiler_simple import SimpleDistributeTranspiler
+from ps_dispatcher import HashName, RoundRobin
 
 __all__ = [
     "DistributeTranspiler", "InferenceTranspiler", "SimpleDistributeTranspiler",
-    "memory_optimize", "release_memory"
+    "memory_optimize", "release_memory", "HashName", "RoundRobin"
 ]

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import math
 
-import distributed_splitter as splitter
+from ps_dispatcher import RoundRobin, HashName, PSDispatcher
 from .. import core, framework
 from ..framework import Program, default_main_program, \
                         default_startup_program, \
@@ -24,7 +24,9 @@ from ..framework import Program, default_main_program, \
 
 LOOKUP_TABLE_TYPE = "lookup_table"
 LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad"
-RPC_CLIENT_VAR_NAME = "RPC_CLIENT_VAR"
+RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName(
+)
+RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC
 
 
 class VarBlock:
@@ -149,13 +151,27 @@ def delete_ops(block, ops):
     block.program.sync_with_cpp()
 
 
+def find_op_by_input_arg(block, arg_name):
+    for index, op in enumerate(block.ops):
+        if arg_name in op.input_arg_names:
+            return index
+    return -1
+
+
+def find_op_by_output_arg(block, arg_name):
+    for index, op in enumerate(block.ops):
+        if arg_name in op.output_arg_names:
+            return index
+    return -1
+
+
 class DistributeTranspiler:
     def transpile(self,
                   trainer_id,
                   program=None,
                   pservers="127.0.0.1:6174",
                   trainers=1,
-                  split_method=splitter.round_robin,
+                  split_method=RoundRobin,
                   sync_mode=True):
         """
         Transpile the program to distributed data-parallelism programs.
@@ -196,7 +212,7 @@ class DistributeTranspiler:
         :param sync_mode: if sync_mode is set True, it means that dist transpiler
         will transpile the program into sync_mode pserver and trainer program.
         """
-        assert (callable(split_method))
+        assert (split_method.__bases__[0] == PSDispatcher)
         if program is None:
             program = default_main_program()
         self.origin_program = program
@@ -209,6 +225,7 @@ class DistributeTranspiler:
         pserver_endpoints = pservers.split(",")
         self.pserver_endpoints = pserver_endpoints
         self.optimize_ops, params_grads = self._get_optimize_pass()
+        ps_dispatcher = split_method(pserver_endpoints)
 
         # process lookup_table_op
         # 1. check all lookup_table_op is distributed
@@ -268,54 +285,110 @@ class DistributeTranspiler:
 
         grad_blocks = split_dense_variable(grad_list, len(pserver_endpoints))
         param_blocks = split_dense_variable(param_list, len(pserver_endpoints))
+        assert (len(grad_blocks) == len(param_blocks))
         # step2: Create new vars for the parameters and gradients blocks and
         # add ops to do the split.
-        grad_var_mapping = self._append_split_op(program, grad_blocks)
         param_var_mapping = self._create_vars_from_blocklist(program,
                                                              param_blocks)
+        grad_var_mapping = self._create_vars_from_blocklist(
+            program, grad_blocks, add_trainer_suffix=self.trainer_num > 1)
+        grad_param_mapping = dict()
+        for g, p in zip(grad_blocks, param_blocks):
+            g_name, g_bid, _ = g.split(":")
+            p_name, p_bid, _ = p.split(":")
+            grad_param_mapping[grad_var_mapping[g_name][int(g_bid)]] =  \
+                    param_var_mapping[p_name][int(p_bid)]
+
+        # step 3: transpile trainer side program, insert recv op and send op.
 
-        # step3: Add gradients as send op inputs and parameters as send
-        # op outputs.
-        send_inputs = []
-        send_outputs = []
-        for b in grad_blocks:  # append by order
-            varname, block_id, _ = b.split(":")
-            send_inputs.append(grad_var_mapping[varname][int(block_id)])
-
-        for b in param_blocks:
-            varname, block_id, _ = b.split(":")
-            send_outputs.append(param_var_mapping[varname][int(block_id)])
-
-        # let send_op know which endpoint to send which var to, eplist has the same
-        # order as send_inputs.
-        eplist = split_method(send_inputs, pserver_endpoints)
         # create mapping of endpoint -> split var to create pserver side program
         self.param_grad_ep_mapping = dict()
+        [
+            self.param_grad_ep_mapping.update({
+                ep: {
+                    "params": [],
+                    "grads": []
+                }
+            }) for ep in self.pserver_endpoints
+        ]
+
+        # step 3.1: insert send op to send gradient vars to parameter servers
+        ps_dispatcher.reset()
+        send_vars = []
+        for orig_varname, splited_vars in grad_var_mapping.items():
+            eplist = ps_dispatcher.dispatch(splited_vars)
+            if len(splited_vars) == 1:
+                orig_varname = splited_vars[0].name
+                index = find_op_by_output_arg(program.global_block(),
+                                              orig_varname)
+            elif len(splited_vars) > 1:
+                orig_var = program.global_block().vars[orig_varname]
+                index = find_op_by_output_arg(program.global_block(),
+                                              orig_varname)
+                self._insert_split_op(program, orig_var, index, splited_vars)
+                index += 1
+            else:
+                AssertionError("Can not insert the send op by original "
+                               "variable name :", orig_varname)
+
+            program.global_block().insert_op(
+                index=index + 1,
+                type="send_vars",
+                inputs={"X": splited_vars},
+                outputs={},
+                attrs={
+                    "epmap": eplist,
+                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
+                })
+            for _, var in enumerate(splited_vars):
+                send_vars.append(var)
+
+        if self.sync_mode:
+            program.global_block().append_op(
+                type="send_barrier",
+                inputs={},
+                outputs={},
+                attrs={
+                    "endpoints": pserver_endpoints,
+                    "sync_mode": self.sync_mode,
+                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
+                })
+
+        # step 3.2: insert recv op to receive parameters from parameter server
+        recv_vars = []
+        for _, var in enumerate(send_vars):
+            recv_vars.append(grad_param_mapping[var])
+        ps_dispatcher.reset()
+        eplist = ps_dispatcher.dispatch(recv_vars)
+
         for i, ep in enumerate(eplist):
-            param = send_outputs[i]
-            grad = send_inputs[i]
-            if not self.param_grad_ep_mapping.has_key(ep):
-                self.param_grad_ep_mapping[ep] = {"params": [], "grads": []}
-            self.param_grad_ep_mapping[ep]["params"].append(param)
-            self.param_grad_ep_mapping[ep]["grads"].append(grad)
-
-        rpc_client_var = program.global_block().create_var(
-            name=RPC_CLIENT_VAR_NAME,
-            persistable=True,
-            type=core.VarDesc.VarType.RAW)
-
-        # create send_op
+            self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i])
+            self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i])
+        # step4: Concat the parameters splits together after recv.
+        for varname, splited_var in param_var_mapping.iteritems():
+            eps = []
+            for var in splited_var:
+                index = [v.name for v in recv_vars].index(var.name)
+                eps.append(eplist[index])
+
+            program.global_block().append_op(
+                type="recv",
+                inputs={},
+                outputs={"Out": splited_var},
+                attrs={
+                    "epmap": eps,
+                    RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
+                })
+
         program.global_block().append_op(
-            type="send",
-            inputs={"X": send_inputs},
-            outputs={"Out": send_outputs,
-                     "RPCClient": rpc_client_var},
+            type="fetch_barrier",
+            inputs={},
+            outputs={},
             attrs={
                 "endpoints": pserver_endpoints,
-                "epmap": eplist,
-                "sync_mode": self.sync_mode
+                RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
             })
-        # step4: Concat the parameters splits together after recv.
+
         for varname, splited_var in param_var_mapping.iteritems():
             if len(splited_var) <= 1:
                 continue
@@ -327,10 +400,8 @@ class DistributeTranspiler:
                 attrs={"axis": 0})
 
         if self.has_distributed_lookup_table:
-            self._replace_lookup_table_op_with_prefetch(program, rpc_client_var,
-                                                        eplist)
-            self._split_table_grad_and_add_send_vars(program, rpc_client_var,
-                                                     pserver_endpoints)
+            self._replace_lookup_table_op_with_prefetch(program, eplist)
+            self._split_table_grad_and_add_send_vars(program, pserver_endpoints)
 
     def get_trainer_program(self):
         # remove optimize ops and add a send op to main_program
@@ -550,8 +621,7 @@ class DistributeTranspiler:
         return s_prog
 
     # transpiler function for dis lookup_table
-    def _replace_lookup_table_op_with_prefetch(self, program, rpc_client_var,
-                                               eplist):
+    def _replace_lookup_table_op_with_prefetch(self, program, eplist):
         # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op
         self.prefetch_input_vars = None
         self.prefetch_output_vars = None
@@ -598,11 +668,11 @@ class DistributeTranspiler:
                         index=op_index + 1,
                         type="prefetch",
                         inputs={'X': self.prefetch_input_vars},
-                        outputs={
-                            "Out": self.prefetch_output_vars,
-                            "RPCClient": rpc_client_var
-                        },
-                        attrs={"epmap": eplist})
+                        outputs={"Out": self.prefetch_output_vars},
+                        attrs={
+                            "epmap": eplist,
+                            RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
+                        })
 
                     # insert concat_op
                     program.global_block().insert_op(
@@ -622,8 +692,7 @@ class DistributeTranspiler:
                     # break for loop
                     break
 
-    def _split_table_grad_and_add_send_vars(self, program, rpc_client_var,
-                                            pserver_endpoints):
+    def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints):
         # 2. add split_ids_op and send_vars_op to send gradient to pservers
         # there should only be one table_name
         all_ops = program.global_block().ops
@@ -643,9 +712,12 @@ class DistributeTranspiler:
                     index=op_index + 2,
                     type="send_vars",
                     inputs={'X': self.table_grad_list},
-                    outputs={"RPCClient": rpc_client_var},
-                    attrs={"sync_send": True,
-                           "epmap": pserver_endpoints})
+                    outputs={},
+                    attrs={
+                        "sync_send": True,
+                        "epmap": pserver_endpoints,
+                        RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
+                    })
                 break
 
     def _create_prefetch_block(self, pserver_index, pserver_program,
@@ -838,50 +910,31 @@ class DistributeTranspiler:
             lod_level=var.lod_level,
             persistable=persistable)
 
-    def _append_split_op(self, program, gradblocks):
-        """
-        Split variables that need to be split and append respective ops
-        Args:
-            program (ProgramDesc): ProgramDesc that gradients blong.
-            gradblocks (list[(varname, block_id, block_size)]): List of gradient blocks.
-        Returns:
-            var_mapping (dict(varname->[new_splitted_variable])):A dict mapping 
-                from original var name to each var split.
-        """
-
-        add_suffix = False
-        if self.trainer_num > 1:
-            add_suffix = True
-        var_mapping = self._create_vars_from_blocklist(
-            program, gradblocks, add_trainer_suffix=add_suffix)
-        for varname, splited_vars in var_mapping.iteritems():
-            # variable that don't need to split have empty splited_vars
-            if len(splited_vars) <= 1:
-                continue
-            orig_var = program.global_block().vars[varname]
-            if orig_var.type == core.VarDesc.VarType.SELECTED_ROWS:
-                height_sections = []
-                for v in splited_vars:
-                    height_sections.append(v.shape[0])
-                program.global_block().append_op(
-                    type="split_selected_rows",
-                    inputs={"X": orig_var},
-                    outputs={"Out": splited_vars},
-                    attrs={"height_sections": height_sections})
-            elif orig_var.type == core.VarDesc.VarType.LOD_TENSOR:
-                sections = []
-                for v in splited_vars:
-                    sections.append(v.shape[0])
-                program.global_block().append_op(
-                    type="split_byref",
-                    inputs={"X": orig_var},
-                    outputs={"Out": splited_vars},
-                    attrs={"sections": sections}  # assume split evenly
-                )
-            else:
-                AssertionError("Variable type should be in set "
-                               "[LOD_TENSOR, SELECTED_ROWS]")
-        return var_mapping
+    def _insert_split_op(self, program, orig_var, index, splited_vars):
+        if orig_var.type == core.VarDesc.VarType.SELECTED_ROWS:
+            height_sections = []
+            for v in splited_vars:
+                height_sections.append(v.shape[0])
+            program.global_block().insert_op(
+                index=index + 1,
+                type="split_selected_rows",
+                inputs={"X": orig_var},
+                outputs={"Out": splited_vars},
+                attrs={"height_sections": height_sections})
+        elif orig_var.type == core.VarDesc.VarType.LOD_TENSOR:
+            sections = []
+            for v in splited_vars:
+                sections.append(v.shape[0])
+            program.global_block().insert_op(
+                index=index + 1,
+                type="split_byref",
+                inputs={"X": orig_var},
+                outputs={"Out": splited_vars},
+                attrs={"sections": sections}  # assume split evenly
+            )
+        else:
+            AssertionError("Variable type should be in set "
+                           "[LOD_TENSOR, SELECTED_ROWS]")
 
     def _get_optimizer_input_shape(self, op_type, varkey, orig_shape,
                                    param_shape):

python/paddle/fluid/transpiler/distributed_splitter.py → python/paddle/fluid/transpiler/ps_dispatcher.py

@@ -13,45 +13,66 @@
 # limitations under the License.
 
 
-def hash_name(varlist, pserver_endpoints):
+class PSDispatcher(object):
     """
-    hash variable names to several endpoints.
+    PSDispatcher is the base class for dispatching vars
+    into different pserver instance.
+    You need to implement the `dispatch` inferface.
+    """
+
+    def __init__(self, pserver_endpoints):
+        self._eps = pserver_endpoints
+        self._step = 0
+
+    @property
+    def eps(self):
+        return self._eps
+
+    def reset(self):
+        self._step = 0
+
+    def dispatch(self, varlist):
+        """
+        :param varlist: a list of Variables
+        :return: a map of pserver endpoint -> varname 
+        """
+        AssertionError("Interface has not been implemented.")
 
-    Args:
-        varlist(list): a list of Variables
 
-    Returns(dict): a map of pserver endpoint -> varname
+class HashName(PSDispatcher):
     """
+      Hash variable names to several endpoints
+    """
+
+    def __init__(self, pserver_endpoints):
+        super(self.__class__, self).__init__(pserver_endpoints)
 
-    def _hash_block(block_str, total):
+    def _hash_block(self, block_str, total):
         return hash(block_str) % total
 
-    eplist = []
-    for var in varlist:
-        server_id = _hash_block(var.name(), len(pserver_endpoints))
-        server_for_param = pserver_endpoints[server_id]
-        eplist.append(server_for_param)
-    return eplist
+    def dispatch(self, varlist):
+        eplist = []
+        for var in varlist:
+            server_id = self._hash_block(var.name(), len(self._eps))
+            server_for_param = self._eps[server_id]
+            eplist.append(server_for_param)
+        return eplist
 
 
-def round_robin(varlist, pserver_endpoints):
+class RoundRobin(PSDispatcher):
     """
-    Distribute variables to several endpoints.
-    Args:
-        varlist(list): a list of variables
-        pserver_endpoints(list): a list of pserver endpoints
-
-    Returns(list[int]): the endpoint for each variable
+    Distribute variables to serveral endpoints.
     """
-    assert (len(varlist) >= len(pserver_endpoints))
-
-    eplist = []
-    pserver_idx = 0
-    for var in varlist:
-        server_for_param = pserver_endpoints[pserver_idx]
-        eplist.append(server_for_param)
-
-        pserver_idx += 1
-        if pserver_idx >= len(pserver_endpoints):
-            pserver_idx = 0
-    return eplist
+
+    def __init__(self, pserver_endpoints):
+        super(self.__class__, self).__init__(pserver_endpoints)
+
+    def dispatch(self, varlist):
+        eplist = []
+        for var in varlist:
+            server_for_param = self._eps[self._step]
+            eplist.append(server_for_param)
+            self._step += 1
+            if self._step >= len(self._eps):
+                self._step = 0
+        return eplist