diff --git a/paddle/framework/block_desc.cc b/paddle/framework/block_desc.cc
index 2d7db382a60b23eba0924519a0cea9ad3eb90142..0668b08ff7ab3c8ca4f1e989fc7af45a8ec5f63c 100644
--- a/paddle/framework/block_desc.cc
+++ b/paddle/framework/block_desc.cc
@@ -90,6 +90,21 @@ OpDesc *BlockDesc::PrependOp() {
   return ops_.front().get();
 }
 
+void BlockDesc::RemoveOp(size_t s, size_t e) {
+  if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) {
+    return;
+  }
+  need_update_ = true;
+  for (auto it = ops_.begin() + s; it != ops_.begin() + e; it++) {
+    auto names = (*it)->InputArgumentNames();
+    for (auto n : names) {
+      // TODO(typhoonzero): delete vars if no other op use it.
+      VLOG(3) << "deleting var " << n;
+    }
+  }
+  ops_.erase(ops_.begin() + s, ops_.begin() + e);
+}
+
 std::vector<OpDesc *> BlockDesc::AllOps() const {
   std::vector<OpDesc *> res;
   for (const auto &op : ops_) {
diff --git a/paddle/framework/block_desc.h b/paddle/framework/block_desc.h
index 513fc54f24c4760b27936e2ac205b2410e9861a4..6c8c81b332d99e52db41018e117aa837be6745bc 100644
--- a/paddle/framework/block_desc.h
+++ b/paddle/framework/block_desc.h
@@ -79,6 +79,8 @@ class BlockDesc {
 
   OpDesc *PrependOp();
 
+  void RemoveOp(size_t s, size_t e);
+
   std::vector<OpDesc *> AllOps() const;
 
   size_t OpSize() const { return ops_.size(); }
diff --git a/paddle/framework/executor.cc b/paddle/framework/executor.cc
index c4b76911a67fe02d74dbc1917b0832a39604a366..14ae37ec49c12203381e74b3f9174a460e41c18e 100644
--- a/paddle/framework/executor.cc
+++ b/paddle/framework/executor.cc
@@ -65,7 +65,7 @@ static void CreateTensor(Variable* var, proto::VarDesc::VarType var_type) {
 }
 
 void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
-                   bool create_local_scope) {
+                   bool create_local_scope, bool create_vars) {
   // TODO(tonyyang-svail):
   //    - only runs on the first device (i.e. no interdevice communication)
   //    - will change to use multiple blocks for RNN op and Cond Op
@@ -74,33 +74,35 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
   auto& device = device_contexts_[0];
 
   Scope* local_scope = scope;
-  if (create_local_scope) {
-    local_scope = &scope->NewScope();
-    for (auto& var : block.AllVars()) {
-      if (var->Name() == framework::kEmptyVarName) {
-        continue;
+  if (create_vars) {
+    if (create_local_scope) {
+      local_scope = &scope->NewScope();
+      for (auto& var : block.AllVars()) {
+        if (var->Name() == framework::kEmptyVarName) {
+          continue;
+        }
+
+        if (var->Persistable()) {
+          auto* ptr = scope->Var(var->Name());
+          CreateTensor(ptr, var->GetType());
+          VLOG(3) << "Create Variable " << var->Name()
+                  << " global, which pointer is " << ptr;
+        } else {
+          auto* ptr = local_scope->Var(var->Name());
+          CreateTensor(ptr, var->GetType());
+          VLOG(3) << "Create Variable " << var->Name()
+                  << " locally, which pointer is " << ptr;
+        }
       }
-
-      if (var->Persistable()) {
-        auto* ptr = scope->Var(var->Name());
-        CreateTensor(ptr, var->GetType());
-        VLOG(3) << "Create Variable " << var->Name()
-                << " global, which pointer is " << ptr;
-      } else {
+    } else {
+      for (auto& var : block.AllVars()) {
         auto* ptr = local_scope->Var(var->Name());
         CreateTensor(ptr, var->GetType());
-        VLOG(3) << "Create Variable " << var->Name()
-                << " locally, which pointer is " << ptr;
+        VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
+                << ptr;
       }
-    }
-  } else {
-    for (auto& var : block.AllVars()) {
-      auto* ptr = local_scope->Var(var->Name());
-      CreateTensor(ptr, var->GetType());
-      VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
-              << ptr;
-    }
-  }
+    }  // if (create_local_scope)
+  }    // if (create_vars)
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
diff --git a/paddle/framework/executor.h b/paddle/framework/executor.h
index fb861d47126c4397792c63d411b973ffd979244d..a3d1609293a0d687c33447ca7a0df95c6aac3bc5 100644
--- a/paddle/framework/executor.h
+++ b/paddle/framework/executor.h
@@ -124,7 +124,8 @@ class Executor {
    *  ProgramDesc
    *  Scope
    */
-  void Run(const ProgramDesc&, Scope*, int, bool create_local_scope = true);
+  void Run(const ProgramDesc&, Scope*, int, bool create_local_scope = true,
+           bool create_vars = true);
 
  private:
   std::vector<const platform::DeviceContext*> device_contexts_;
diff --git a/paddle/operators/detail/recv_impl.cc b/paddle/operators/detail/recv_impl.cc
index 89dc5045221156eed7aa9411bc96ad86f91136d2..517a1946a0c25081b20c320f4104c81503a4249e 100644
--- a/paddle/operators/detail/recv_impl.cc
+++ b/paddle/operators/detail/recv_impl.cc
@@ -20,25 +20,57 @@ namespace detail {
 
 Status SendRecvServerImpl::SendVariable(ServerContext *context,
                                         const VariableMessage *in_var,
-                                        VariableMessage *out_var) {
-  framework::LoDTensor t;
-  // TODO(typhoonzero): desirealize in_tensor and run pserver network.
+                                        VoidMessage *out_var) {
+  // TODO(typhoonzero): support different variable types.
   std::istringstream iss(in_var->serialized());
+  framework::LoDTensor t;
   framework::DeserializeFromStream(iss, &t);
-  lodtensor_queue_.Push(std::move(t));
-  // Block util the sub graph is done.
-  t = lodtensor_return_queue_.Pop();
+  TensorWithName tensor_with_name =
+      std::make_pair(in_var->varname(), std::move(t));
+
+  var_recv_queue_.Push(std::move(tensor_with_name));
+  return Status::OK;
+}
+
+Status SendRecvServerImpl::GetVariable(ServerContext *context,
+                                       const VariableMessage *in_var,
+                                       VariableMessage *out_var) {
+  std::string get_var_name = in_var->varname();
+  auto *var = scope_->FindVar(get_var_name);
+  auto tensor = var->Get<framework::LoDTensor>();
   std::ostringstream oss;
-  // FIXME(typhoonzero): get context from op.
-  framework::SerializeToStream(oss, t, platform::CPUDeviceContext());
+  framework::SerializeToStream(oss, tensor, platform::CPUDeviceContext());
+
   std::string *varname = out_var->mutable_varname();
-  *varname = in_var->varname();
+  *varname = get_var_name;
   std::string *serialized = out_var->mutable_serialized();
   *serialized = oss.str();
+  return Status::OK;
+}
 
+Status SendRecvServerImpl::Wait(ServerContext *context,
+                                const VoidMessage *in_var,
+                                VoidMessage *out_var) {
+  {
+    std::unique_lock<std::mutex> lock(this->mutex_);
+    condition_.wait(lock, [=] { return this->done_ == true; });
+  }
   return Status::OK;
 }
 
+void SendRecvServerImpl::Reset() {
+  std::lock_guard<std::mutex> lock(this->mutex_);
+  done_ = false;
+}
+
+void SendRecvServerImpl::Done() {
+  {
+    std::lock_guard<std::mutex> lock(this->mutex_);
+    done_ = true;
+  }
+  condition_.notify_all();
+}
+
 }  // namespace detail
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/detail/send_impl.cc b/paddle/operators/detail/send_impl.cc
index da1ddf75d2afb85670c5ea0c9884376415f28208..d7165e13db961f7719139b1a40211d6a2ca67a9f 100644
--- a/paddle/operators/detail/send_impl.cc
+++ b/paddle/operators/detail/send_impl.cc
@@ -19,10 +19,10 @@ namespace operators {
 namespace detail {
 
 bool RPCClient::SendVariable(const framework::Scope& scope,
-                             const std::string& inname,
-                             const std::string& outname) {
+                             const std::string& inname) {
   ClientContext context;
-  VariableMessage msg, out_msg;
+  VariableMessage msg;
+  VoidMessage out_msg;
   // FIXME(typhoonzero): pass device context to here.
   auto ctx = platform::CPUDeviceContext();
   auto* var = scope.FindVar(inname);
@@ -37,9 +37,26 @@ bool RPCClient::SendVariable(const framework::Scope& scope,
   msg.set_serialized(oss.str());
   Status status = stub_->SendVariable(&context, msg, &out_msg);
   if (!status.ok()) {
+    LOG(ERROR) << "gRPC error: " << status.error_message();
     return false;
   }
-  std::istringstream iss(out_msg.serialized());
+  return true;
+}
+
+bool RPCClient::GetVariable(const framework::Scope& scope,
+                            const std::string& outname) {
+  ClientContext context;
+  VariableMessage call_msg, ret_msg;
+  call_msg.set_varname(outname);
+  auto ctx = platform::CPUDeviceContext();
+  Status status = stub_->GetVariable(&context, call_msg, &ret_msg);
+  if (!status.ok()) {
+    LOG(ERROR) << "gRPC error: " << status.error_message();
+    return false;
+  }
+
+  std::istringstream iss(ret_msg.serialized());
+
   framework::LoDTensor ret_tensor;
   framework::DeserializeFromStream(iss, &ret_tensor);
   auto* outvar = scope.FindVar(outname);
@@ -49,6 +66,12 @@ bool RPCClient::SendVariable(const framework::Scope& scope,
   return true;
 }
 
+void RPCClient::Wait() {
+  ClientContext context;
+  VoidMessage call_msg, ret_msg;
+  stub_->Wait(&context, call_msg, &ret_msg);
+}
+
 }  // namespace detail
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/detail/send_recv.proto b/paddle/operators/detail/send_recv.proto
index 07ff9d2c621a2dfb51792821a0d3fc398c315835..ce729908062ad442e66cc00001e14ceb6f268560 100644
--- a/paddle/operators/detail/send_recv.proto
+++ b/paddle/operators/detail/send_recv.proto
@@ -19,7 +19,12 @@ package sendrecv;
 service SendRecvService {
   // For parameter server round-robin like hashing, do not split tensors.
   // Send and recv only one tensor
-  rpc SendVariable(VariableMessage) returns (VariableMessage) {}
+  // TODO(typhoonzero): add streaming API
+  rpc SendVariable(VariableMessage) returns (VoidMessage) {}
+  // Argument VariableMessage for GetVariable should only contain varname.
+  rpc GetVariable(VariableMessage) returns (VariableMessage) {}
+  // wait for one execution of the program
+  rpc Wait(VoidMessage) returns (VoidMessage) {}
 }
 
 // VariableMessage is serialized paddle variable message.
diff --git a/paddle/operators/detail/send_recv_impl.h b/paddle/operators/detail/send_recv_impl.h
index b9a5340a8636db7b5d6ec7b21368632d3916b4aa..eec9dd38d188247cba4da2a377038a28c847e40e 100644
--- a/paddle/operators/detail/send_recv_impl.h
+++ b/paddle/operators/detail/send_recv_impl.h
@@ -20,10 +20,6 @@
 #include "paddle/framework/selected_rows.h"
 #include "paddle/operators/detail/simple_block_queue.h"
 
-// #include <grpc++/channel.h>
-// #include <grpc++/client_context.h>
-// #include <grpc++/create_channel.h>
-// #include <grpc++/security/credentials.h>
 #include "paddle/operators/detail/send_recv.grpc.pb.h"
 #include "paddle/operators/detail/send_recv.pb.h"
 
@@ -48,24 +44,32 @@ namespace paddle {
 namespace operators {
 namespace detail {
 
+typedef std::pair<std::string, framework::LoDTensor> TensorWithName;
+
 class SendRecvServerImpl final : public SendRecvService::Service {
  public:
   explicit SendRecvServerImpl() {}
 
   Status SendVariable(ServerContext *context, const VariableMessage *in_var,
-                      VariableMessage *out_var) override;
-
-  const framework::LoDTensor Get() { return this->lodtensor_queue_.Pop(); }
+                      VoidMessage *out_var) override;
+  Status GetVariable(ServerContext *context, const VariableMessage *in_var,
+                     VariableMessage *out_var) override;
+  Status Wait(ServerContext *context, const VoidMessage *in_var,
+              VoidMessage *out_var) override;
+  void Reset();
+  void Done();
+  void SetScope(framework::Scope *scope) { scope_ = scope; };
 
-  void Push(const framework::LoDTensor &tensor) {
-    this->lodtensor_return_queue_.Push(tensor);
-  }
+  const TensorWithName Get() { return this->var_recv_queue_.Pop(); }
 
  private:
-  SimpleBlockQueue<framework::LoDTensor> lodtensor_queue_;
-  SimpleBlockQueue<framework::LoDTensor> lodtensor_return_queue_;
-  SimpleBlockQueue<framework::SelectedRows> selected_rows_queue_;
-  SimpleBlockQueue<framework::SelectedRows> selected_rows_return_queue_;
+  // received variable from RPC, operators fetch variable from this queue.
+  SimpleBlockQueue<TensorWithName> var_recv_queue_;
+  framework::Scope *scope_;
+  // condition of the sub program
+  std::mutex mutex_;
+  bool done_;
+  std::condition_variable condition_;
 };
 
 // RPCClient is a class to send tensors to pserver sub-network
@@ -75,8 +79,9 @@ class RPCClient {
   RPCClient(std::shared_ptr<Channel> channel)
       : stub_(SendRecvService::NewStub(channel)) {}
 
-  bool SendVariable(const framework::Scope &scope, const std::string &inname,
-                    const std::string &outname);
+  bool SendVariable(const framework::Scope &scope, const std::string &inname);
+  bool GetVariable(const framework::Scope &scope, const std::string &outname);
+  void Wait();
 
  private:
   std::unique_ptr<SendRecvService::Stub> stub_;
diff --git a/paddle/operators/recv_op.cc b/paddle/operators/recv_op.cc
index 2cc6cf6947b601e326ed563082800946b3ff221d..4e91d1151ebf7a0cd520f2f1fa58a0cc4a0d1bef 100644
--- a/paddle/operators/recv_op.cc
+++ b/paddle/operators/recv_op.cc
@@ -24,6 +24,7 @@
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/framework/proto_desc.h"
 #include "paddle/operators/detail/send_recv_impl.h"
 #include "paddle/operators/detail/simple_block_queue.h"
 
@@ -61,29 +62,76 @@ class RecvOp : public framework::OperatorBase {
     server_thread_->join();
   }
 
+  std::string GetGradVarNameForTrainer(const std::string &varname) const {
+    if (grads_counter_.find(varname) == grads_counter_.end()) {
+      grads_counter_[varname] = 0;
+    }
+    char ret[256];
+    snprintf(ret, sizeof(ret), "%s.trainer_%d", varname.c_str(),
+             grads_counter_[varname]++);
+    return std::string(ret);
+  }
+
   void Run(const framework::Scope &scope,
            const platform::DeviceContext &dev_ctx) const override {
-    // blocking get one var from client.
-    const framework::LoDTensor &t = rpc_service_->Get();
+    // FIXME(typhoonzero): no new scopes for every run.
     framework::Scope &recv_scope = scope.NewScope();
-    // set graph input var
-    auto *var = recv_scope.Var(Input("RX"));
-    auto *tensor = var->GetMutable<framework::LoDTensor>();
-    // FIXME(typhoonzero): do not copy
-    framework::CopyFrom(t, dev_ctx.GetPlace(), dev_ctx, tensor);
-
-    std::string program_str = Attr<std::string>("OptimizeProgram");
-    framework::ProgramDesc program_desc;
-    program_desc.ParseFromString(program_str);
-    framework::ProgramDescBind program(program_desc);
-    framework::Executor executor(dev_ctx);
-    // Run sub graph to get optimized tensor
-    executor.Run(program, &recv_scope, 0, /*global_block*/
-                 false /*create_local_scope*/);
-
-    auto *out_var = recv_scope.FindVar("Out");
-    // push back
-    rpc_service_->Push(out_var->Get<framework::LoDTensor>());
+    rpc_service_->SetScope(&recv_scope);
+    auto param_list = Attr<std::vector<std::string>>("ParamList");
+    auto grad_list = Attr<std::vector<std::string>>("GradList");
+    auto trainer_count = Attr<int>("Trainers");
+    size_t param_count = param_list.size();
+    rpc_service_->Reset();
+    // TODO(typhoonzero): change this to a while_op for every cluster-batch.
+    while (true) {
+      // Get from multiple trainers, we don't care about order in which
+      // the gradient arrives, just add suffix 0~n then average the gradient.
+      for (size_t i = 0; i < param_count * trainer_count; ++i) {
+        // blocking get one var from client.
+        const detail::TensorWithName &v = rpc_service_->Get();
+        auto grad_var_name = v.first;
+        auto it = std::find(grad_list.begin(), grad_list.end(), grad_var_name);
+        std::string param_var_name;
+        if (it != grad_list.end()) {
+          param_var_name = param_list[it - grad_list.begin()];
+        } else {
+          LOG(ERROR) << "grad have no paired param found!";
+        }
+        VLOG(3) << "recved grad: " << grad_var_name
+                << " updating param: " << param_var_name;
+        auto *merged_grad = recv_scope.FindVar(grad_var_name);
+        if (merged_grad == nullptr) {
+          // create output of merged var.
+          auto merged_var = recv_scope.Var(grad_var_name);
+          merged_var->GetMutable<framework::LoDTensor>();
+        }
+
+        if (trainer_count > 1) {
+          grad_var_name = this->GetGradVarNameForTrainer(grad_var_name);
+        }
+
+        auto *var = recv_scope.Var(grad_var_name);
+        auto *tensor = var->GetMutable<framework::LoDTensor>();
+        // FIXME(typhoonzero): do not copy
+        framework::CopyFrom(v.second, dev_ctx.GetPlace(), dev_ctx, tensor);
+      }
+      rpc_service_->Reset();
+
+      std::string program_str = Attr<std::string>("OptimizeProgram");
+      framework::proto::ProgramDesc program_desc;
+      program_desc.ParseFromString(program_str);
+      framework::ProgramDesc program(program_desc);
+      framework::Executor executor(dev_ctx);
+      // Run sub graph to get optimized tensor
+      try {
+        executor.Run(program, &recv_scope, 0, /*global_block*/
+                     false /*create_local_scope*/, false /*create_vars*/);
+      } catch (std::exception &e) {
+        LOG(ERROR) << "run sub program error " << e.what();
+      }
+      rpc_service_->Done();
+      grads_counter_.clear();
+    }  // while(true)
   }
 
  protected:
@@ -93,13 +141,14 @@ class RecvOp : public framework::OperatorBase {
   // grpc send/recv service implement to register.
   std::shared_ptr<detail::SendRecvServerImpl> rpc_service_;
   std::shared_ptr<std::thread> server_thread_;
+  mutable std::unordered_map<std::string, int> grads_counter_;
 };
 
 class RecvOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   RecvOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("RX", "(Tensor) Input tensor to be saved");
+    AddInput("RX", "(Tensor) Input tensor to be optimized").AsDuplicable();
     AddComment(R"DOC(
 Recv operator
 
@@ -112,6 +161,17 @@ This operator will recv tensor from send_op
         .AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
     AddAttr<std::string>("OptimizeProgram", "type string",
                          "Serialized ProgramDesc string for recv to run.");
+    AddAttr<std::vector<std::string>>(
+        "ParamList", "type list of string",
+        "grad->param name mapping to find which param to optimize.")
+        .SetDefault({});
+    AddAttr<std::vector<std::string>>(
+        "GradList", "type list of string",
+        "grad->param name mapping to find which param to optimize.")
+        .SetDefault({});
+    AddAttr<int>("Trainers", "type int",
+                 "Number of trainers in the current cluster job")
+        .SetDefault(1);
   }
 };
 
diff --git a/paddle/operators/send_op.cc b/paddle/operators/send_op.cc
index 0d121fb48dc2dc8bd0312aa2091f4e058caa4515..a5681910708bae584bdf4217b61ca63a3988e1ff 100644
--- a/paddle/operators/send_op.cc
+++ b/paddle/operators/send_op.cc
@@ -34,45 +34,56 @@ class SendOp : public framework::OperatorBase {
          const framework::AttributeMap &attrs)
       : OperatorBase(type, inputs, outputs, attrs) {
     // init client when the operator is created at runtime.
-    if (!client_) {
-      std::string endpoint = Attr<std::string>("endpoint");
-      client_.reset(new detail::RPCClient(
-          grpc::CreateChannel(endpoint, grpc::InsecureChannelCredentials())));
-      // TODO(typhoonzero): how to call InitVariables
+    std::vector<std::string> endpoints =
+        Attr<std::vector<std::string>>("endpoints");
+    for (auto ep : endpoints) {
+      client_map_[ep].reset(new detail::RPCClient(
+          grpc::CreateChannel(ep, grpc::InsecureChannelCredentials())));
     }
   }
   void Run(const framework::Scope &scope,
            const platform::DeviceContext &dev_ctx) const override {
-    auto iname = Input("X");
-    auto oname = Output("Out");
-    // TODO(typhoonzero): currently it's non-blocking,
-    // should block until server responds.
-    bool ret = client_->SendVariable(scope, iname, oname);
-    if (!ret) {
-      LOG(ERROR) << "send variable error";
+    auto ins = Inputs("X");
+    std::vector<std::string> epmap = Attr<std::vector<std::string>>("epmap");
+    // TODO(typhoonzero): use async calls to send multiple variable asyncly.
+    for (size_t i = 0; i < ins.size(); ++i) {
+      bool ret = client_map_[epmap[i]]->SendVariable(scope, ins[i]);
+      if (!ret) {
+        LOG(ERROR) << "send variable error: " << ins[i];
+      }
+    }
+    // TODO(typhoonzero): support async optimization
+    client_map_[epmap[0]]->Wait();
+    for (size_t i = 0; i < ins.size(); ++i) {
+      bool ret = client_map_[epmap[i]]->GetVariable(scope, ins[i]);
+      if (!ret) {
+        LOG(ERROR) << "GetVariable error: " << ins[i];
+      }
     }
   }
 
  protected:
-  std::shared_ptr<detail::RPCClient> client_{nullptr};
+  mutable std::unordered_map<std::string, std::shared_ptr<detail::RPCClient>>
+      client_map_;
 };
 
 class SendOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   SendOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "(Tensor) Input tensor to be saved");
-    AddOutput("Out", "(Tensor) Output fetched from server");
+    AddInput("X", "(Tensor) Input tensor to be send").AsDuplicable();
     AddComment(R"DOC(
 Recv operator
 
 This operator will recv tensor from send_op
 )DOC");
-    AddAttr<std::string>("endpoint",
-                         "(string, default 127.0.0.1:6164)"
-                         "IP address to listen on.")
-        .SetDefault("127.0.0.1:6164")
-        .AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
+    AddAttr<std::vector<std::string>>("endpoints",
+                                      "(string vector, default 127.0.0.1:6164)"
+                                      "Server endpoints to send variables to.");
+    AddAttr<std::vector<std::string>>("epmap",
+                                      "(string vector, default 127.0.0.1:6164)"
+                                      "Server endpoints in the order of input "
+                                      "variables for mapping");
   }
 };
 
diff --git a/paddle/operators/send_recv_op_test.cc b/paddle/operators/send_recv_op_test.cc
index 3e2e2051afacb748877e3b0c3dec8d6662ac4e72..d899d8154cce57a88de0e1d19e3393528ce367e2 100644
--- a/paddle/operators/send_recv_op_test.cc
+++ b/paddle/operators/send_recv_op_test.cc
@@ -16,12 +16,14 @@
 // a RemoteOptimizer.
 
 #include <unistd.h>
+#include <string>
 #include <thread>
 
 #include "gtest/gtest.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/program_desc.h"
+#include "paddle/string/printf.h"
 
 USE_NO_KERNEL_OP(send);
 USE_NO_KERNEL_OP(recv);
@@ -33,30 +35,33 @@ std::unique_ptr<paddle::framework::OperatorBase> recv_op;
 void InitTensorsInScope(paddle::framework::Scope &scope,
                         paddle::platform::CPUPlace &place) {
   paddle::platform::CPUDeviceContext ctx(place);
-  auto var = scope.Var("X");
-  auto tensor = var->GetMutable<paddle::framework::LoDTensor>();
-  tensor->Resize({10, 10});
-  float *expect = tensor->mutable_data<float>(place);
-  for (int64_t i = 0; i < tensor->numel(); ++i) {
-    expect[i] = static_cast<float>(i);
+  for (int i = 0; i < 2; ++i) {
+    auto var_name = paddle::string::Sprintf("x%d", i);
+    auto var = scope.Var(var_name);
+    auto tensor = var->GetMutable<paddle::framework::LoDTensor>();
+    tensor->Resize({10, 10});
+    float *expect = tensor->mutable_data<float>(place);
+    for (int64_t i = 0; i < tensor->numel(); ++i) {
+      expect[i] = static_cast<float>(i);
+    }
   }
 
   auto out_var = scope.Var("Out");
   auto out_tensor = out_var->GetMutable<paddle::framework::LoDTensor>();
   out_tensor->Resize({10, 10});
-  tensor->mutable_data<float>(place);  // allocate
+  out_tensor->mutable_data<float>(place);  // allocate
 }
 
 void AddOp(const std::string &type,
            const paddle::framework::VariableNameMap &inputs,
            const paddle::framework::VariableNameMap &outputs,
            paddle::framework::AttributeMap attrs,
-           paddle::framework::BlockDescBind *block) {
+           paddle::framework::BlockDesc *block) {
   // insert output
   for (auto kv : outputs) {
     for (auto v : kv.second) {
       auto var = block->Var(v);
-      var->SetDataType(paddle::framework::DataType::FP32);
+      var->SetDataType(paddle::framework::proto::DataType::FP32);
     }
   }
 
@@ -78,10 +83,10 @@ void StartServerNet() {
   InitTensorsInScope(scope, place);
 
   // sub program run in recv_op, for simple test we use sum
-  paddle::framework::ProgramDescBind program;
-  paddle::framework::BlockDescBind *block = program.MutableBlock(0);
+  paddle::framework::ProgramDesc program;
+  paddle::framework::BlockDesc *block = program.MutableBlock(0);
   // X for server side tensors, RX for received tensers, must be of same shape.
-  AddOp("sum", {{"X", {"X", "RX"}}}, {{"Out", {"Out"}}}, {}, block);
+  AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, block);
 
   paddle::framework::AttributeMap attrs;
   attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
@@ -89,8 +94,8 @@ void StartServerNet() {
   PADDLE_ENFORCE(program.Proto()->SerializeToString(&program_proto));
 
   attrs.insert({"OptimizeProgram", program_proto});
-  recv_op = paddle::framework::OpRegistry::CreateOp("recv", {{"RX", {"RX"}}},
-                                                    {{"Out", {"Out"}}}, attrs);
+  recv_op = paddle::framework::OpRegistry::CreateOp(
+      "recv", {{"RX", {"x0", "x1"}}}, {{"Out", {"Out"}}}, attrs);
   paddle::platform::CPUDeviceContext ctx(place);
   recv_op->Run(scope, ctx);
 }
@@ -107,11 +112,11 @@ TEST(SendRecvOp, CPU) {
   attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
 
   auto send_op = paddle::framework::OpRegistry::CreateOp(
-      "send", {{"X", {"X"}}}, {{"Out", {"Out"}}}, attrs);
+      "send", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, attrs);
   paddle::platform::CPUDeviceContext ctx(place);
   send_op->Run(scope, ctx);
 
-  auto in_var = scope.Var("X");
+  auto in_var = scope.Var("x0");
   auto tensor = in_var->GetMutable<paddle::framework::LoDTensor>();
   float *expected = tensor->data<float>();
 
diff --git a/paddle/pybind/protobuf.cc b/paddle/pybind/protobuf.cc
index 88e9cdadd8650504706baf07e8d29fa0663f2905..f105370f226e2cceaac685f280d55134d4291028 100644
--- a/paddle/pybind/protobuf.cc
+++ b/paddle/pybind/protobuf.cc
@@ -159,6 +159,7 @@ void BindBlockDesc(py::module &m) {
            py::return_value_policy::reference)
       .def("prepend_op", &BlockDesc::PrependOp,
            py::return_value_policy::reference)
+      .def("remove_op", &BlockDesc::RemoveOp)
       .def("var",
            [](BlockDesc &self, py::bytes byte_name) {
              std::string name = byte_name;
@@ -249,6 +250,12 @@ void BindOpDesc(py::module &m) {
       .def("set_attr", &OpDesc::SetAttr)
       .def("attr", &OpDesc::GetAttr)
       .def("set_block_attr", &OpDesc::SetBlockAttr)
+      .def("set_serialized_attr",
+           [](OpDesc &self, const std::string &name,
+              const py::bytes &seriralized) {
+             std::string ser(seriralized);
+             self.SetAttr(name, ser);
+           })
       .def("block_attr", &OpDesc::GetBlockAttr)
       .def("check_attrs", &OpDesc::CheckAttrs)
       .def("infer_shape", &OpDesc::InferShape)
diff --git a/python/paddle/v2/fluid/__init__.py b/python/paddle/v2/fluid/__init__.py
index 9b3792ee9e3e4c6f319b3e2b13c4aa3a05cce8be..471255ef50a3be4739a89efbd978cdb4304d992d 100644
--- a/python/paddle/v2/fluid/__init__.py
+++ b/python/paddle/v2/fluid/__init__.py
@@ -16,13 +16,14 @@ import regularizer
 from param_attr import ParamAttr
 from data_feeder import DataFeeder
 from core import LoDTensor, CPUPlace, GPUPlace
+from distribute_transpiler import DistributeTranspiler
 import clip
 
 Tensor = LoDTensor
 __all__ = framework.__all__ + executor.__all__ + [
     'io', 'initializer', 'layers', 'nets', 'optimizer', 'backward',
     'regularizer', 'LoDTensor', 'CPUPlace', 'GPUPlace', 'Tensor', 'ParamAttr'
-    'DataFeeder', 'clip'
+    'DataFeeder', 'clip', 'DistributeTranspiler'
 ]
 
 
diff --git a/python/paddle/v2/fluid/distribute_transpiler.py b/python/paddle/v2/fluid/distribute_transpiler.py
new file mode 100644
index 0000000000000000000000000000000000000000..50364c64bec46aabc7be4f4b4370a3ad5b0eb07c
--- /dev/null
+++ b/python/paddle/v2/fluid/distribute_transpiler.py
@@ -0,0 +1,238 @@
+import framework
+from framework import Program, default_main_program, Parameter, Variable
+import optimizer
+from layer_helper import LayerHelper
+
+
+def hash_name_to_server(params_grads, pserver_endpoints):
+    """
+    :param param_grads:
+    :return: a map of pserver endpoint -> 
+                    params -> [param list]
+                    grads  -> [grad list]
+    """
+
+    def _hash_param(param_name, total):
+        return hash(param_name) % total
+
+    param_grad_map = dict()
+    for param, grad in params_grads:
+        if param.trainable is True and grad is not None:
+            server_id = _hash_param(param.name, len(pserver_endpoints))
+            server_for_param = pserver_endpoints[server_id]
+            if not param_grad_map.has_key(server_for_param):
+                param_grad_map[server_for_param] = {"params": [], "grads": []}
+            param_grad_map[server_for_param]["params"].append(param)
+            param_grad_map[server_for_param]["grads"].append(grad)
+
+    return param_grad_map
+
+
+def round_robin(params_grads, pserver_endpoints):
+    assert (len(params_grads) > len(pserver_endpoints))
+
+    param_grad_map = dict()
+    pserver_idx = 0
+    for param, grad in params_grads:
+        if param.trainable is True:
+            server_for_param = pserver_endpoints[pserver_idx]
+            if not param_grad_map.has_key(server_for_param):
+                param_grad_map[server_for_param] = {"params": [], "grads": []}
+
+            param_grad_map[server_for_param]["params"].append(param)
+            param_grad_map[server_for_param]["grads"].append(grad)
+
+            pserver_idx += 1
+            if pserver_idx >= len(pserver_endpoints):
+                pserver_idx = 0
+    return param_grad_map
+
+
+class DistributeTranspiler:
+    def transpile(self,
+                  optimize_ops,
+                  params_grads,
+                  program=None,
+                  pservers="127.0.0.1:6174",
+                  trainers=1,
+                  split_method=round_robin):
+        """
+            Transpile the program to a distributed data-parallelism programs.
+
+            The main_program will be transform to use a remote parameter server
+            to do parameter optimization. And the optimization graph will be put
+            in to a parameter server program.
+
+            Use different methods to split trainable varialbles to different
+            parameter servers.
+
+            Example to run:
+
+            exe = fluid.Executor(place)
+            t = fluid.DistributeTranspiler()
+            t.transpile(optimize_ops, params_grads, pservers="127.0.0.1:6174", trainers=1)
+
+            pserver_endpoint = os.getenv("PSERVER")
+            if pserver_endpoint:
+                pserver_prog = t.get_pserver_program(pserver_endpoint, optimize_ops)
+                exe.run(fluid.default_startup_program())
+                exe.run(pserver_prog)
+            else:
+                feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
+                exe.run(fluid.default_startup_program())
+
+                for pass_id in range(PASS_NUM):
+                    ...
+
+            :param optimize_ops: op list of optimization, should be the
+                                 return value of Optimizer.minimize
+            :type optimize_ops: list
+            :param program: program to optimize, default default_main_program
+            :param pservers: parameter server endpoints like "m1:6174,m2:6174"
+            :type pservers: string
+
+            :return: return a list of programs
+        """
+        if program is None:
+            program = default_main_program()
+        self.trainers = trainers
+        self._optimize_distributed(
+            optimize_ops,
+            program,
+            params_grads,
+            pservers=pservers,
+            trainers=trainers,
+            split_method=split_method)
+
+    def _clone_param(self, block, v):
+        assert isinstance(v, Parameter)
+        new_p = Parameter(
+            block=block,
+            shape=v.shape,
+            dtype=v.dtype,
+            type=v.type,
+            lod_level=v.lod_level,
+            stop_gradient=v.stop_gradient,
+            trainable=v.trainable,
+            optimize_attr=v.optimize_attr,
+            regularizer=v.regularizer,
+            name=v.name)
+        block.vars[new_p.name] = new_p
+
+    def _clone_var(self, block, var):
+        assert isinstance(var, Variable)
+        return block.create_var(
+            name=var.name,
+            shape=var.shape,
+            dtype=var.dtype,
+            type=var.type,
+            lod_level=var.lod_level,
+            persistable=var.persistable)
+
+    def _optimize_distributed(self, optimize_ops, program, params_and_grads,
+                              **kwargs):
+        if kwargs.has_key("split_method"):
+            split_method = kwargs["split_method"]
+        else:
+            split_method = round_robin
+
+        assert (callable(split_method))
+        pserver_endpoints = kwargs["pservers"].split(",")
+        self.param_grad_map = split_method(params_and_grads, pserver_endpoints)
+
+        send_op_ordered_inputs = []
+        epmap = []
+        for ep, v in self.param_grad_map.iteritems():
+            send_op_ordered_inputs.extend(v["grads"])
+            for i in v["grads"]:
+                epmap.append(ep)
+        send_op = program.global_block().append_op(
+            type="send",
+            inputs={"X": send_op_ordered_inputs
+                    },  # inputs is a list of tensors to be send
+            outputs={},
+            attrs={"endpoints": pserver_endpoints,
+                   "epmap": epmap})
+
+    def get_trainer_program(optimize_ops, program):
+        # remove optimize ops and add a send op to main_program
+        program.global_block().delete_ops(optimize_ops)
+
+    def _create_var_for_trainers(self, block, var, trainers):
+        var_list = []
+        for i in xrange(trainers):
+            var_each = block.create_var(
+                name="%s.trainer_%d" % (var.name, i),
+                psersistable=var.persistable,
+                dtype=var.dtype,
+                shape=var.shape)
+            var_list.append(var_each)
+        return var_list
+
+    def get_pserver_program(self, endpoint, optimize_ops):
+        pserver_program = Program()
+        for v in self.param_grad_map[endpoint]["params"]:
+            self._clone_param(pserver_program.global_block(), v)
+
+        optimize_sub_program = Program()
+        grad_var_names = [
+            var.name for var in self.param_grad_map[endpoint]["grads"]
+        ]
+        for opt_op in optimize_ops:
+            for _, var in opt_op.inputs.iteritems():
+                # NOTE: append operators to merge gradients from multiple
+                # trainers. If trainers == 1, this is not needed.
+                if self.trainers > 1 and var.name in grad_var_names:
+                    vars2merge = self._create_var_for_trainers(
+                        optimize_sub_program.global_block(), var, self.trainers)
+                    merged_var = optimize_sub_program.global_block().create_var(
+                        name=var.name,
+                        persistable=var.persistable,
+                        dtype=var.dtype,
+                        shape=var.shape)
+                    optimize_sub_program.global_block().append_op(
+                        type="sum",
+                        inputs={"X": vars2merge},
+                        outputs={"Out": merged_var})
+                    optimize_sub_program.global_block().append_op(
+                        type="scale",
+                        inputs={"X": merged_var},
+                        outputs={"Out": merged_var},
+                        attrs={"scale": 1.0 / float(self.trainers)})
+                else:
+                    optimize_sub_program.global_block().create_var(
+                        name=var.name,
+                        persistable=var.persistable,
+                        dtype=var.dtype,
+                        shape=var.shape)
+
+            if opt_op.inputs.has_key("Grad"):
+                if opt_op.inputs["Grad"].name in grad_var_names:
+                    print "appending ", opt_op.type, opt_op.inputs
+                    optimize_sub_program.global_block().append_op(
+                        type=opt_op.type,
+                        inputs=opt_op.inputs,
+                        outputs=opt_op.outputs,
+                        attrs=opt_op.attrs)
+            else:
+                optimize_sub_program.global_block().append_op(
+                    type=opt_op.type,
+                    inputs=opt_op.inputs,
+                    outputs=opt_op.outputs,
+                    attrs=opt_op.attrs)
+        pserver_program.global_block().append_op(
+            type="recv",
+            inputs={"RX":
+                    self.param_grad_map[endpoint]["grads"]},  # grads to recv
+            outputs={},
+            attrs={
+                "OptimizeProgram": optimize_sub_program.desc,
+                "endpoint": endpoint,
+                "ParamList":
+                [p.name for p in self.param_grad_map[endpoint]["params"]],
+                "GradList":
+                [p.name for p in self.param_grad_map[endpoint]["grads"]],
+                "Trainers": self.trainers
+            })
+        pserver_program.sync_with_cpp()
+        return pserver_program
diff --git a/python/paddle/v2/fluid/executor.py b/python/paddle/v2/fluid/executor.py
index 9a99b045dc70a9e4662a6f4da141183ffc8f1846..4b4a0820abb9a85f3e9936190c835c2f186107b3 100644
--- a/python/paddle/v2/fluid/executor.py
+++ b/python/paddle/v2/fluid/executor.py
@@ -1,6 +1,6 @@
 import numpy as np
 from . import core
-from framework import Program, default_main_program
+from framework import Program, default_main_program, Parameter, Variable
 
 __all__ = ['Executor', 'g_scope']
 
@@ -148,7 +148,7 @@ class Executor(object):
                 outputs={'Out': [fetch_var]},
                 attrs={'col': i})
 
-        self.executor.run(program.desc, scope, 0, True)
+        self.executor.run(program.desc, scope, 0, True, True)
         outs = [
             core.get_fetch_variable(scope, fetch_var_name, i)
             for i in xrange(len(fetch_list))
diff --git a/python/paddle/v2/fluid/framework.py b/python/paddle/v2/fluid/framework.py
index d1b12a8f097ed674b6b6384fe8ba5db950a94da5..920a49b69c77ccb98214ab76b4ee2aa125e01f02 100644
--- a/python/paddle/v2/fluid/framework.py
+++ b/python/paddle/v2/fluid/framework.py
@@ -359,6 +359,10 @@ class Operator(object):
         """
         self.block = block
         self.desc = desc
+        # for clone a new operator
+        self.inputs = inputs
+        self.outputs = outputs
+        self.attrs = attrs
         if len(self.desc.type()) != 0:
             return
         if type is None:
@@ -430,13 +434,18 @@ class Operator(object):
                     continue
                 if isinstance(attrs[attr_name], Block):
                     self.desc.set_block_attr(attr_name, attrs[attr_name].desc)
+                elif isinstance(attrs[attr_name], core.BlockDesc) or \
+                   isinstance(attrs[attr_name], core.ProgramDesc):
+                    self.desc.set_serialized_attr(
+                        attr_name, attrs[attr_name].serialize_to_string())
                 else:
                     self.desc.set_attr(attr_name, attrs[attr_name])
 
         self.desc.check_attrs()
         no_kernel_op_set = {
             'feed', 'fetch', 'save', 'load', 'recurrent',
-            'rnn_memory_helper_grad', 'conditional_block', 'while'
+            'rnn_memory_helper_grad', 'conditional_block', 'while', 'send',
+            'recv'
         }
         if type not in no_kernel_op_set:
             self.desc.infer_var_type(self.block.desc)
@@ -582,6 +591,7 @@ class Block(object):
         self.vars = dict()  # var_name --> var
         self.ops = collections.deque()  # operator list
         self.program = program
+        self.removed_vars = dict()
 
     def __str__(self):
         return self.to_string(True)
@@ -638,6 +648,16 @@ class Block(object):
         self.ops.append(op)
         return op
 
+    def delete_ops(self, ops):
+        # remove from cpp
+        # FIXME(typhoonzero): remove only the first occuracy.
+        try:
+            start = list(self.ops).index(ops[0])
+            end = list(self.ops).index(ops[-1])
+        except Exception, e:
+            raise e
+        self.desc.remove_op(start, end)
+
     def prepend_op(self, *args, **kwargs):
         op_desc = self.desc.prepend_op()
         op = Operator(self, op_desc, *args, **kwargs)
diff --git a/python/paddle/v2/fluid/optimizer.py b/python/paddle/v2/fluid/optimizer.py
index 84fcbcdc2f2868a84bad5e145a934e33485b1fef..c56a531ed531cf0219e94854ba66c7399e003292 100644
--- a/python/paddle/v2/fluid/optimizer.py
+++ b/python/paddle/v2/fluid/optimizer.py
@@ -207,7 +207,7 @@ class Optimizer(object):
 
         optimize_ops = self.create_optimization_pass(params_grads, loss,
                                                      startup_program)
-        return optimize_ops
+        return optimize_ops, params_grads
 
 
 class SGDOptimizer(Optimizer):
diff --git a/python/paddle/v2/fluid/tests/book/notest_recognize_digits_conv_dist.py b/python/paddle/v2/fluid/tests/book/notest_recognize_digits_conv_dist.py
new file mode 100644
index 0000000000000000000000000000000000000000..2680502efb91061be37a77fbe5b451960fdd15f7
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/book/notest_recognize_digits_conv_dist.py
@@ -0,0 +1,72 @@
+from __future__ import print_function
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+import os
+
+images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype='float32')
+label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+conv_pool_1 = fluid.nets.simple_img_conv_pool(
+    input=images,
+    filter_size=5,
+    num_filters=20,
+    pool_size=2,
+    pool_stride=2,
+    act="relu")
+conv_pool_2 = fluid.nets.simple_img_conv_pool(
+    input=conv_pool_1,
+    filter_size=5,
+    num_filters=50,
+    pool_size=2,
+    pool_stride=2,
+    act="relu")
+
+predict = fluid.layers.fc(input=conv_pool_2, size=10, act="softmax")
+cost = fluid.layers.cross_entropy(input=predict, label=label)
+avg_cost = fluid.layers.mean(x=cost)
+optimizer = fluid.optimizer.Adam(learning_rate=0.01)
+optimize_ops, params_grads = optimizer.minimize(avg_cost)
+
+accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
+
+BATCH_SIZE = 50
+PASS_NUM = 3
+train_reader = paddle.batch(
+    paddle.reader.shuffle(
+        paddle.dataset.mnist.train(), buf_size=500),
+    batch_size=BATCH_SIZE)
+
+place = fluid.CPUPlace()
+exe = fluid.Executor(place)
+t = fluid.DistributeTranspiler()
+pserver_endpoints = os.getenv("PSERVERS")
+training_role = os.getenv("TRAINING_ROLE",
+                          "TRAINER")  # get the training role: trainer/pserver
+t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=1)
+
+if training_role == "PSERVER":
+    pserver_prog = t.get_pserver_program(pserver_endpoints, optimize_ops)
+    exe.run(fluid.default_startup_program())
+    exe.run(pserver_prog)
+elif training_role == "TRAINER":
+    feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
+    exe.run(fluid.default_startup_program())
+
+    for pass_id in range(PASS_NUM):
+        accuracy.reset(exe)
+        for data in train_reader():
+            loss, acc = exe.run(fluid.default_main_program(),
+                                feed=feeder.feed(data),
+                                fetch_list=[avg_cost] + accuracy.metrics)
+            pass_acc = accuracy.eval(exe)
+            # print loss, acc
+            if loss < 10.0 and pass_acc > 0.9:
+                # if avg cost less than 10.0 and accuracy is larger than 0.9, we think our code is good.
+                exit(0)
+
+        pass_acc = accuracy.eval(exe)
+        print("pass_id=" + str(pass_id) + " pass_acc=" + str(pass_acc))
+else:
+    print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
+
+exit(1)
diff --git a/python/paddle/v2/fluid/tests/test_optimizer.py b/python/paddle/v2/fluid/tests/test_optimizer.py
index 2459dfd664300d405edb36c4ca906c1769b5e7d2..29694be58bce0eb41b05439da35ef07a542ef12a 100644
--- a/python/paddle/v2/fluid/tests/test_optimizer.py
+++ b/python/paddle/v2/fluid/tests/test_optimizer.py
@@ -27,7 +27,7 @@ class TestOptimizer(unittest.TestCase):
         block.append_op(
             type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out})
         sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.01)
-        opts = sgd_optimizer.minimize(mean_out, init_program)
+        opts, _ = sgd_optimizer.minimize(mean_out, init_program)
         self.assertEqual(len(opts), 1)
         sgd_op = opts[0]
         self.assertEqual(sgd_op.type, "sgd")
@@ -57,7 +57,7 @@ class TestOptimizer(unittest.TestCase):
         learning_rate = 0.01
         sgd_optimizer = optimizer.SGDOptimizer(
             learning_rate=learning_rate, global_step=global_step)
-        opts = sgd_optimizer.minimize(mean_out, init_program)
+        opts, _ = sgd_optimizer.minimize(mean_out, init_program)
         self.assertEqual(len(opts), 2)
         sgd_op = opts[0]
         self.assertEqual(sgd_op.type, "sgd")