diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc
index 20ef7e09f630140c44774147aa727780df6333fa..95e807c0afa45bc4f4feb84d450b2d0584bc3b28 100644
--- a/paddle/fluid/framework/parallel_executor.cc
+++ b/paddle/fluid/framework/parallel_executor.cc
@@ -58,7 +58,8 @@ ParallelExecutor::ParallelExecutor(
     const std::unordered_set<std::string> &bcast_vars,
     const ProgramDesc &main_program, const std::string &loss_var_name,
     Scope *scope, const std::vector<Scope *> &local_scopes, bool allow_op_delay,
-    bool use_default_grad_scale, bool balance_parameter_opt_between_cards)
+    bool use_default_grad_scale, bool balance_parameter_opt_between_cards,
+    size_t num_trainers, size_t trainer_id)
     : member_(new ParallelExecutorPrivate(places)) {
   member_->global_scope_ = scope;
 
@@ -80,7 +81,13 @@ ParallelExecutor::ParallelExecutor(
 
 // Bcast Parameters to all GPUs
 #ifdef PADDLE_WITH_CUDA
-  member_->nccl_ctxs_.reset(new platform::NCCLContextMap(member_->places_));
+  auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME);
+  ncclUniqueId *nccl_id = nullptr;
+  if (nccl_id_var != nullptr) {
+    nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
+  }
+  member_->nccl_ctxs_.reset(new platform::NCCLContextMap(
+      member_->places_, nccl_id, num_trainers, trainer_id));
 #endif
   if (platform::is_gpu_place(places[0]) && member_->local_scopes_.size() != 1 &&
       local_scopes.empty()) {  // Is CUDA
diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h
index b251fc91417a1c00e61e9c3c952460e6268d2819..9e279876cfeef20a1921f8bd1c27046a477b9f56 100644
--- a/paddle/fluid/framework/parallel_executor.h
+++ b/paddle/fluid/framework/parallel_executor.h
@@ -41,7 +41,8 @@ class ParallelExecutor {
                             const std::string& loss_var_name, Scope* scope,
                             const std::vector<Scope*>& local_scopes,
                             bool allow_op_delay, bool use_default_grad_scale,
-                            bool balance_parameter_opt_between_cards);
+                            bool balance_parameter_opt_between_cards,
+                            size_t num_trainers = 1, size_t trainer_id = 0);
 
   ~ParallelExecutor();
 
diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt
index c14a2b7786f9f7c06d59479d3bbce9c5d542e495..d38a9ce58726a1d045d6905354b0b592166c0110 100644
--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -186,6 +186,11 @@ endif()
 
 add_subdirectory(detail)
 if(WITH_DISTRIBUTE)
+    if(WITH_GPU)
+        op_library(gen_nccl_id_op DEPS nccl_common)
+    else()
+        set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op)
+    endif()
     set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf)
     set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
     op_library(send_op DEPS ${DISTRIBUTE_DEPS})
@@ -202,8 +207,9 @@ if(WITH_DISTRIBUTE)
     set_source_files_properties(send_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)
+    cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS send_op listen_and_serv_op executor)
 else()
-    set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op)
+    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)
 endif()
 
 op_library(cross_entropy_op DEPS cross_entropy)
diff --git a/paddle/fluid/operators/detail/grpc_client.cc b/paddle/fluid/operators/detail/grpc_client.cc
index 661dfa69fe1580ff3890f12defcd124225be0c06..ae60ab15325ef101feb7270a4f5d840cb2112be0 100644
--- a/paddle/fluid/operators/detail/grpc_client.cc
+++ b/paddle/fluid/operators/detail/grpc_client.cc
@@ -52,7 +52,7 @@ bool RPCClient::AsyncSendVariable(const std::string& ep,
     // stub context
     SendProcessor* s = new SendProcessor(ch);
     s->Prepare(var_h, time_out);
-    s->response_call_back_ = NULL;
+    s->response_call_back_ = nullptr;
 
     auto call = s->stub_g_.PrepareUnaryCall(
         s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_);
diff --git a/paddle/fluid/operators/detail/grpc_client.h b/paddle/fluid/operators/detail/grpc_client.h
index f6229b71bc01a6de51f50f5fe880ada6e15e74dd..dabce7414d2f0dca74193f1cd10c341793c10ec9 100644
--- a/paddle/fluid/operators/detail/grpc_client.h
+++ b/paddle/fluid/operators/detail/grpc_client.h
@@ -57,7 +57,9 @@ void ProcGetResponse(const VarHandle& var_h, const grpc::ByteBuffer& msg);
 
 class BaseProcessor {
  public:
-  explicit BaseProcessor(std::shared_ptr<grpc::Channel> ch) { context_ = NULL; }
+  explicit BaseProcessor(std::shared_ptr<grpc::Channel> ch) {
+    context_ = nullptr;
+  }
 
   virtual ~BaseProcessor() {}
 
@@ -105,7 +107,7 @@ class SendProcessor : public BaseProcessor {
 
   ::grpc::GenericStub stub_g_;
   ::grpc::ByteBuffer reply_;
-  RequestSendCallBack response_call_back_ = NULL;
+  RequestSendCallBack response_call_back_ = nullptr;
 };
 
 typedef std::function<void(const VarHandle&, const ::grpc::ByteBuffer&)>
diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h
index 7f9cae21ccca8dd51f9fbe98148d01a51ac6eb84..18f1bc53d0f561f412a5bbbe018bc3d427ac9ef9 100644
--- a/paddle/fluid/operators/detail/grpc_server.h
+++ b/paddle/fluid/operators/detail/grpc_server.h
@@ -47,6 +47,7 @@ class AsyncGRPCServer final {
   explicit AsyncGRPCServer(const std::string &address, bool sync_mode)
       : address_(address), sync_mode_(sync_mode), ready_(0) {}
 
+  ~AsyncGRPCServer() {}
   void WaitServerReady();
   void RunSyncUpdate();
 
diff --git a/paddle/fluid/operators/detail/send_recv.proto b/paddle/fluid/operators/detail/send_recv.proto
index fffa9ae7a43ea5cd7b2bda6fbbf6ef9f7d23009d..9478c5702bcbf99fc88207b8c4843dbccf8a5925 100644
--- a/paddle/fluid/operators/detail/send_recv.proto
+++ b/paddle/fluid/operators/detail/send_recv.proto
@@ -32,6 +32,7 @@ service SendRecvService {
 enum VarType {
   LOD_TENSOR = 0;
   SELECTED_ROWS = 1;
+  NCCL_ID = 2;
 }
 
 // NOTICE(gongwb):don't modify this proto if you are not
diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.cc b/paddle/fluid/operators/detail/sendrecvop_utils.cc
index 1a8a1af20fa446dbd537944409ef0ca1e3e9116f..07c43554bc6a0d71d688a5a5772d0ab3d2de319a 100644
--- a/paddle/fluid/operators/detail/sendrecvop_utils.cc
+++ b/paddle/fluid/operators/detail/sendrecvop_utils.cc
@@ -14,6 +14,9 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/detail/sendrecvop_utils.h"
 
+#ifdef PADDLE_WITH_CUDA
+#include <nccl.h>
+#endif
 #include <sys/time.h>
 #include <thread>  // NOLINT
 
@@ -129,6 +132,10 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
   } else if (var->IsType<framework::SelectedRows>()) {
     request.set_type(::sendrecv::SELECTED_ROWS);
     GetSelectedRowsPayload(var, ctx, &request, &payload, &payload_size);
+#ifdef PADDLE_WITH_CUDA
+  } else if (var->IsType<ncclUniqueId>()) {
+    request.set_type(::sendrecv::NCCL_ID);
+#endif
   } else {
     PADDLE_THROW("Serialize does not support type: %s",
                  typeid(var->Type()).name());
@@ -149,6 +156,24 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
   void* buf = buffer.get();
   ProtoEncodeHelper e(static_cast<char*>(buf), 1024);
   e.WriteRawBytes(std::string(header.data(), header.size()));
+// NCCLID is copied directly to the message, return bytebuffer
+// with only one slice if serializing NCCLID.
+#ifdef PADDLE_WITH_CUDA
+  if (var->IsType<ncclUniqueId>()) {
+    e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber,
+                              NCCL_UNIQUE_ID_BYTES);
+    const ncclUniqueId& uid = var->Get<ncclUniqueId>();
+    e.WriteRawBytes(std::string(uid.internal, NCCL_UNIQUE_ID_BYTES));
+
+    // for serialize NCCL_ID
+    ::grpc::Slice slices(e.size());
+    memcpy(const_cast<uint8_t*>(slices.begin()), e.data(), e.size());
+    ::grpc::ByteBuffer tmp(&slices, 1);
+    msg->Swap(&tmp);
+    return;
+  }
+#endif
+
   e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size);
   // steal reference of tensor data
   ::grpc::Slice slices[4];  // metadata, tensor, rows meta, rows
diff --git a/paddle/fluid/operators/detail/variable_response.cc b/paddle/fluid/operators/detail/variable_response.cc
index 99602a05d023f30c2eed8df25e7534fdc9ef2ced..462e303096e609c6797ca8cc16266ec3621623fc 100644
--- a/paddle/fluid/operators/detail/variable_response.cc
+++ b/paddle/fluid/operators/detail/variable_response.cc
@@ -17,6 +17,9 @@
 #include <string>
 #include <utility>
 #include <vector>
+#ifdef PADDLE_WITH_CUDA
+#include <nccl.h>
+#endif
 #include "paddle/fluid/platform/profiler.h"
 
 #include "paddle/fluid/operators/detail/send_recv.pb.h"
@@ -368,7 +371,8 @@ int VariableResponse::Parse(Source* source) {
       }
       case sendrecv::VariableMessage::kSerializedFieldNumber: {
         PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS ||
-                        meta_.type() == sendrecv::LOD_TENSOR) &&
+                        meta_.type() == sendrecv::LOD_TENSOR ||
+                        meta_.type() == sendrecv::NCCL_ID) &&
                            meta_.varname() != "",
                        "meta info should be got first!");
 
@@ -378,6 +382,22 @@ int VariableResponse::Parse(Source* source) {
           return tag;
         }
 
+        if (meta_.type() == sendrecv::NCCL_ID) {
+#ifdef PADDLE_WITH_CUDA
+          auto* var = scope_->FindVar(meta_.varname());
+          if (var != nullptr) {
+            ncclUniqueId* id = var->GetMutable<ncclUniqueId>();
+            if (!ReadRaw(&input, *dev_ctx_, platform::CPUPlace(), id->internal,
+                         num_bytes)) {
+              return tag;
+            }
+          }
+          break;
+#else
+          PADDLE_THROW("Not compiled with CUDA!");
+#endif
+        }
+
         framework::DDim dims = GetDims(meta_.dims());
         if (meta_.type() == sendrecv::LOD_TENSOR) {
           PADDLE_ENFORCE(meta_.lod_size() >= 0,
diff --git a/paddle/fluid/operators/gen_nccl_id_op.cc b/paddle/fluid/operators/gen_nccl_id_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..a5678f63466d368b3dd59380c18f9625cabd368b
--- /dev/null
+++ b/paddle/fluid/operators/gen_nccl_id_op.cc
@@ -0,0 +1,128 @@
+/* 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 <nccl.h>
+#include <stdint.h>
+#include <ostream>
+#include <string>
+
+#include "paddle/fluid/framework/executor.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/threadpool.h"
+#include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/operators/detail/grpc_server.h"
+#include "paddle/fluid/platform/nccl_helper.h"
+
+namespace paddle {
+namespace operators {
+
+class GenNCCLIdOp : public framework::OperatorBase {
+ public:
+  GenNCCLIdOp(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& dev_place) const override {
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    // put nccl id in CPUPlace
+    auto& dev_ctx = *pool.Get(platform::CPUPlace());
+    int trainer_id = Attr<int>("trainer_id");
+    framework::Scope& local_scope = scope.NewScope();
+
+    if (trainer_id == 0) {
+      GenerateAndSend(&local_scope, dev_ctx);
+    } else {
+      GetIdByServer(&local_scope, dev_ctx);
+    }
+  }
+
+ private:
+  void GenerateAndSend(framework::Scope* scope,
+                       const platform::DeviceContext& dev_ctx) const {
+    auto var = scope->FindVar(NCCL_ID_VARNAME);
+    PADDLE_ENFORCE_NOT_NULL(var);
+    auto id = var->GetMutable<ncclUniqueId>();
+    PADDLE_ENFORCE(platform::dynload::ncclGetUniqueId(id));
+
+    std::vector<std::string> endpoint_list =
+        Attr<std::vector<std::string>>("endpoint_list");
+    detail::RPCClient client;
+    for (auto& ep : endpoint_list) {
+      VLOG(3) << "sending nccl id to " << ep;
+      client.AsyncSendVariable(ep, dev_ctx, *scope, NCCL_ID_VARNAME);
+    }
+    client.Wait();
+    VLOG(3) << "sending completed...";
+  }
+
+  void GetIdByServer(framework::Scope* scope,
+                     const platform::DeviceContext& dev_ctx) const {
+    std::string endpoint = Attr<std::string>("endpoint");
+    // NOTE: Can not use unique_ptr here because the default
+    // deleter will call GRPC Server's base class's dtor and
+    // that will cause a wired crash.
+    detail::AsyncGRPCServer rpc_service(endpoint, true);
+    framework::ProgramDesc empty_program;
+    framework::Executor executor(dev_ctx.GetPlace());
+    rpc_service.SetScope(scope);
+    rpc_service.SetDevCtx(&dev_ctx);
+    rpc_service.SetProgram(&empty_program);
+    rpc_service.SetExecutor(&executor);
+
+    std::thread server_thread(
+        std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, &rpc_service));
+    rpc_service.SetCond(0);
+    VLOG(3) << "start getting nccl id from trainer 0...";
+    auto recv = rpc_service.Get();
+    VLOG(3) << "got nccl id and stop server...";
+    rpc_service.ShutDown();
+    VLOG(3) << "rpc server stopped";
+    server_thread.join();
+  }
+};
+
+class GenNCCLIdOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddOutput("NCCLID", "Raw variable contains a NCCL UniqueId instaces.");
+    AddComment(R"DOC(
+GenNCCLId operator
+
+For trainer 0: generate a new UniqueId and send it to all the other trainers.
+For trainer 1~n: start a gRPC server to get the UniqueId, once got, stop the server.
+)DOC");
+    AddAttr<std::string>("endpoint",
+                         "(string), e.g. 127.0.0.1:6175 "
+                         "current listen endpoint");
+    AddAttr<std::vector<std::string>>(
+        "endpoint_list",
+        "['trainer1_ip:port', 'trainer2_ip:port', ...] "
+        "list of trainer endpoints start from trainer 1")
+        .SetDefault({});
+    AddAttr<int>("trainer_id",
+                 "(int default 0) "
+                 "The index of the trainer in distributed training.")
+        .SetDefault(0);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(gen_nccl_id, ops::GenNCCLIdOp, ops::GenNCCLIdOpMaker);
diff --git a/paddle/fluid/operators/test_send_nccl_id.cc b/paddle/fluid/operators/test_send_nccl_id.cc
new file mode 100644
index 0000000000000000000000000000000000000000..bbae1d54aa3524fd45cb8ab13c86df8d54b8e643
--- /dev/null
+++ b/paddle/fluid/operators/test_send_nccl_id.cc
@@ -0,0 +1,94 @@
+/* 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 <unistd.h>
+#include <string>
+#include <thread>  // NOLINT
+
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/operators/listen_and_serv_op.h"
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/operators/math/selected_rows_functor.h"
+#include "paddle/fluid/platform/nccl_helper.h"
+#include "paddle/fluid/string/printf.h"
+
+USE_NO_KERNEL_OP(listen_and_serv);
+
+namespace f = paddle::framework;
+namespace p = paddle::platform;
+namespace m = paddle::operators::math;
+namespace detail = paddle::operators::detail;
+namespace string = paddle::string;
+
+std::unique_ptr<detail::AsyncGRPCServer> rpc_service;
+
+void StartServer(std::atomic<bool>* initialized) {
+  f::Scope scope;
+  p::CPUPlace place;
+  scope.Var(NCCL_ID_VARNAME);
+  p::DeviceContextPool& pool = p::DeviceContextPool::Instance();
+  auto& dev_ctx = *pool.Get(p::CPUPlace());
+
+  rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", true));
+
+  f::ProgramDesc empty_program;
+  f::Executor executor(dev_ctx.GetPlace());
+  rpc_service->SetScope(&scope);
+  rpc_service->SetDevCtx(&dev_ctx);
+  rpc_service->SetProgram(&empty_program);
+  rpc_service->SetExecutor(&executor);
+
+  std::thread server_thread(
+      std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, rpc_service.get()));
+  *initialized = true;
+  rpc_service->SetCond(0);
+  auto recv = rpc_service->Get();
+  LOG(INFO) << "got nccl id and stop server...";
+  rpc_service->ShutDown();
+  server_thread.join();
+}
+
+TEST(SendNcclId, Normal) {
+  std::atomic<bool> initialized{false};
+  std::thread server_thread(StartServer, &initialized);
+  while (!initialized) {
+  }
+  // wait server to start
+  // sleep(2);
+  rpc_service->WaitServerReady();
+
+  f::Scope scope;
+  p::CPUPlace place;
+  p::DeviceContextPool& pool = p::DeviceContextPool::Instance();
+  auto& dev_ctx = *pool.Get(p::CPUPlace());
+
+  auto var = scope.Var(NCCL_ID_VARNAME);
+  // var->SetType(f::proto::VarType_Type_RAW);
+  auto id = var->GetMutable<ncclUniqueId>();
+  p::dynload::ncclGetUniqueId(id);
+
+  int port = rpc_service->GetSelectedPort();
+  std::string ep = string::Sprintf("127.0.0.1:%d", port);
+  detail::RPCClient client;
+
+  client.AsyncSendVariable(ep, dev_ctx, scope, NCCL_ID_VARNAME);
+  client.Wait();
+  server_thread.join();
+  auto* ptr = rpc_service.release();
+  delete ptr;
+}
diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h
index 0013597fd516d15c7d502370eec77e1a6a5dca88..e30c1a9ebf08365a9856fb32b1ce5790869e2b33 100644
--- a/paddle/fluid/platform/nccl_helper.h
+++ b/paddle/fluid/platform/nccl_helper.h
@@ -14,12 +14,15 @@
 
 #pragma once
 
+#include <stdio.h>
 #include <thread>  // NOLINT
 #include <typeindex>
 #include <vector>
 #include "paddle/fluid/platform/dynload/nccl.h"
 #include "paddle/fluid/platform/enforce.h"
 
+#define NCCL_ID_VARNAME "NCCLID"
+
 namespace paddle {
 namespace platform {
 
@@ -73,7 +76,9 @@ struct NCCLContextMap {
   std::unordered_map<int, NCCLContext> contexts_;
   std::vector<int> order_;
 
-  explicit NCCLContextMap(const std::vector<platform::Place> &places) {
+  explicit NCCLContextMap(const std::vector<platform::Place> &places,
+                          ncclUniqueId *nccl_id = nullptr,
+                          size_t num_trainers = 1, size_t trainer_id = 0) {
     PADDLE_ENFORCE(!places.empty());
     order_.reserve(places.size());
     for (auto &p : places) {
@@ -85,18 +90,34 @@ struct NCCLContextMap {
         order_.size(), contexts_.size(),
         "NCCL Context Map does not support contain two or more same device");
 
-    if (places.size() > 1) {
-      std::unique_ptr<ncclComm_t[]> comms(new ncclComm_t[order_.size()]);
+    if (places.size() <= 1) {
+      return;
+    }
+    std::unique_ptr<ncclComm_t[]> comms(new ncclComm_t[order_.size()]);
+    // if pass nccl_id here, can assume we are doing multi node training
+    if (nccl_id == nullptr) {
+      std::lock_guard<std::mutex> guard(NCCLGroupGuard::NCCLMutex());
+      PADDLE_ENFORCE(platform::dynload::ncclCommInitAll(
+          comms.get(), static_cast<int>(order_.size()), order_.data()));
+    } else {
+      PADDLE_ENFORCE_GT(num_trainers, 1);
+      // TODO(wuyi): need to ensure each node have same number of GPUs
       {
-        std::lock_guard<std::mutex> guard(NCCLGroupGuard::NCCLMutex());
-        PADDLE_ENFORCE(platform::dynload::ncclCommInitAll(
-            comms.get(), static_cast<int>(order_.size()), order_.data()));
-      }
-      int i = 0;
-      for (auto &dev_id : order_) {
-        contexts_.at(dev_id).comm_ = comms[i++];
+        int nranks = num_trainers * order_.size();
+        NCCLGroupGuard gurad;
+        for (auto &gpu_id : order_) {
+          int rank = trainer_id * order_.size() + gpu_id;
+          VLOG(3) << "init nccl rank: " << rank << " nranks: " << nranks;
+          PADDLE_ENFORCE(cudaSetDevice(gpu_id));
+          PADDLE_ENFORCE(platform::dynload::ncclCommInitRank(
+              comms.get() + gpu_id, nranks, *nccl_id, rank));
+        }
       }
     }
+    int i = 0;
+    for (auto &dev_id : order_) {
+      contexts_.at(dev_id).comm_ = comms[i++];
+    }
   }
 
   NCCLContextMap(const NCCLContextMap &other) = delete;
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index 3e2eed31b446b83843fba943e4f2bc9e3787d7f6..b62291a99d34457dd17bf2bcafc1fc611419f086 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -503,12 +503,13 @@ All parameter, weight, gradient are variables in Paddle.
               const ProgramDesc &main_program, const std::string &loss_var_name,
               Scope *scope, std::vector<Scope *> &local_scopes,
               bool allow_op_delay, bool use_default_grad_scale,
-              bool balance_parameter_opt_between_cards) {
+              bool balance_parameter_opt_between_cards, size_t num_trainers,
+              size_t trainer_id) {
              new (&self) ParallelExecutor(
                  num_threads, use_event, places, params, bcast_vars,
                  main_program, loss_var_name, scope, local_scopes,
                  allow_op_delay, use_default_grad_scale,
-                 balance_parameter_opt_between_cards);
+                 balance_parameter_opt_between_cards, num_trainers, trainer_id);
            })
       .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs)
       // NOTE: even we return a vec<Scope*>* to Python use reference policy.
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index 28e54f5492e7b04a1406e319cecf977d4a55725e..38c765938fe9d7b2103bfdd926874c485d0ff4dc 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -489,7 +489,7 @@ class Operator(object):
             'rnn_memory_helper_grad', 'conditional_block', 'while', 'send',
             'recv', 'listen_and_serv', 'parallel_do', 'save_combine',
             'load_combine', 'ncclInit', 'channel_create', 'channel_close',
-            'channel_send', 'channel_recv', 'select'
+            'channel_send', 'channel_recv', 'select', 'gen_nccl_id'
         }
         if type not in no_kernel_op_set:
             self.desc.infer_var_type(self.block.desc)
diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py
index 5b43f860e7075745bbf6e76c2f9d0e9a87a86db0..7358c4b60e87893b9c04e3da2221dfb69d3ba0c7 100644
--- a/python/paddle/fluid/parallel_executor.py
+++ b/python/paddle/fluid/parallel_executor.py
@@ -31,7 +31,9 @@ class ParallelExecutor(object):
                  allow_op_delay=False,
                  share_vars_from=None,
                  use_default_grad_scale=True,
-                 balance_parameter_opt_between_cards=False):
+                 balance_parameter_opt_between_cards=False,
+                 num_trainers=1,
+                 trainer_id=0):
         """
         ParallelExecutor can run program in parallel.
 
@@ -55,6 +57,11 @@ class ParallelExecutor(object):
             balance_parameter_opt_between_cards(bool, default True): Whether
                 updating different gradients on different cards. Currently, it
                 is not recommended.
+            num_trainers(int, default 1): If greater than 1, NCCL will be
+                initialized with multpile rank of nodes, each node should have
+                same number of GPUs. Distributed training will be enabled then.
+            trainer_id(int, default 0): Must use together with num_trainers.
+                trainer_id is the "rank" of current node starts from 0.
 
         Returns:
             A ParallelExecutor object.
@@ -134,8 +141,9 @@ class ParallelExecutor(object):
             local_scopes,
             allow_op_delay,
             use_default_grad_scale,
-            balance_parameter_opt_between_cards)
-
+            balance_parameter_opt_between_cards,
+            num_trainers,
+            trainer_id)
         self.scope = scope
 
     def run(self, fetch_list, feed=None, feed_dict=None):