heter_server.h 21.2 KB
Newer Older
T
tangwei12 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
/* Copyright (c) 2018 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. */

#pragma once
#include <atomic>
#include <ctime>
#include <map>
#include <memory>
#include <random>
#include <string>
#include <unordered_map>
T
tangwei12 已提交
23
#include <unordered_set>
T
tangwei12 已提交
24
#include <vector>
25

T
tangwei12 已提交
26 27 28
#include "brpc/channel.h"
#include "brpc/controller.h"
#include "brpc/server.h"
29
#include "paddle/fluid/distributed/ps/service/brpc_utils.h"
30
#include "paddle/fluid/distributed/ps/service/heter_client.h"
31
#include "paddle/fluid/distributed/ps/service/sendrecv.pb.h"
32
#include "paddle/fluid/distributed/ps/table/depends/feature_value.h"
33
#include "paddle/fluid/framework/blocking_queue.h"
T
tangwei12 已提交
34 35 36 37 38 39 40 41
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
#include "paddle/fluid/platform/profiler.h"

42 43 44 45 46 47 48 49 50 51
namespace google {
namespace protobuf {
class Closure;
class RpcController;
}  // namespace protobuf
}  // namespace google
namespace paddle {
namespace framework {
class Executor;
class ProgramDesc;
52
class Scope;
53 54 55
}  // namespace framework
}  // namespace paddle

T
tangwei12 已提交
56
DECLARE_double(eager_delete_tensor_gb);
57 58
DECLARE_int32(pserver_timeout_ms);
DECLARE_int32(heter_world_size);
T
tangwei12 已提交
59 60 61
namespace paddle {
namespace distributed {

62 63
using MultiVarMsg = MultiVariableMessage;
using VarMsg = VariableMessage;
64

65
using serviceHandler = std::function<int32_t(
66
    const PsRequestMessage& request, PsResponseMessage& response,  // NOLINT
67 68 69
    brpc::Controller* cntl)>;
using HeterServiceHandler =
    std::function<int32_t(const MultiVarMsg*, MultiVarMsg*, brpc::Controller*)>;
T
tangwei12 已提交
70

71
using HeterRpcCallbackFunc = std::function<void(void*)>;
T
tangwei12 已提交
72

73
class ServiceHandlerBase {
T
tangwei12 已提交
74
 public:
75
  ServiceHandlerBase() : dev_ctx_(nullptr), scope_(nullptr) {}
T
tangwei12 已提交
76

77
  virtual ~ServiceHandlerBase() {}
T
tangwei12 已提交
78

79 80
  void SetScope(const framework::Scope* scope) { scope_ = scope; }
  void SetDevCtx(const platform::DeviceContext* dev_ctx) { dev_ctx_ = dev_ctx; }
T
tangwei12 已提交
81

82 83
  virtual int Handle(const MultiVarMsg* request, MultiVarMsg* response,
                     brpc::Controller* cntl) = 0;
T
tangwei12 已提交
84

85 86 87
 protected:
  const platform::DeviceContext* dev_ctx_;
  const framework::Scope* scope_;
T
tangwei12 已提交
88 89
};

90 91 92 93 94 95 96 97
using SharedMiniScope =
    std::shared_ptr<std::unordered_map<int, ::paddle::framework::Scope*>>;
using SharedMicroScope = std::shared_ptr<std::unordered_map<
    int, std::shared_ptr<std::vector<::paddle::framework::Scope*>>>>;
using SharedTaskQueue = std::shared_ptr<
    std::unordered_map<int, std::shared_ptr<::paddle::framework::BlockingQueue<
                                std::pair<std::string, int>>>>>;

98
class SendAndRecvVariableHandler final : public ServiceHandlerBase {
99
 public:
100
  SendAndRecvVariableHandler() {
101 102
    this->num_microbatch_ = 0;
    this->num_minibatch_ = 0;
103
    _local_shards.reset(new shard_type[FLAGS_heter_world_size]);
104 105
  }

106
  virtual ~SendAndRecvVariableHandler() {}
107

108 109 110 111 112
  void SetMiniScopes(SharedMiniScope mini_scopes) {
    mini_scopes_ = mini_scopes;
    num_minibatch_ = mini_scopes_->size();
  }

113 114
  void SetMicroScopes(SharedMicroScope micro_scopes) {
    micro_scopes_ = micro_scopes;
115 116 117 118 119 120 121 122 123 124 125
    for (auto& scope_pair : (*micro_scopes_)) {
      // auto mini_idx = scope_pair.first;
      auto& micro_scopes = scope_pair.second;
      num_microbatch_ = micro_scopes->size();
      break;
    }
  }

  int GetThreadNum() {
    std::unique_lock<std::mutex> lk(scope_mutex_);
    return (*task_queue_).size();
126 127
  }

128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162
  int SaveInSwitchWithScope(const MultiVarMsg* request,
                            PsResponseMessage* response,
                            brpc::Controller* cntl);

  void WaitForVarsConsumed(int32_t group_id, const std::string& var_name) {
    auto& local_shard = _local_shards[group_id];
    while (local_shard.find(var_name) != local_shard.end()) {
      if (local_shard[var_name].size() == 0) {
        break;
      }
      VLOG(4) << "waiting consume result......";
      sleep(1);
    }
    return;
  }

  void WaitForVarsProduced(int32_t group_id, const std::string& var_name) {
    auto& local_shard = _local_shards[group_id];
    while (local_shard.find(var_name) == local_shard.end()) {
      VLOG(4) << "waiting produce result......";
      sleep(1);
    }
    return;
  }

  int SaveInSwitchWithShard(const MultiVarMsg* request,
                            PsResponseMessage* response,
                            brpc::Controller* cntl);

  int QueryInSwitchWithShard(const MultiVarMsg* request, MultiVarMsg* response,
                             brpc::Controller* cntl);

  int QueryInSwitchWithScope(const MultiVarMsg* request, MultiVarMsg* response,
                             brpc::Controller* cntl);

163 164 165 166
  void SetTaskQueue(SharedTaskQueue task_queue) { task_queue_ = task_queue; }

  int Handle(const MultiVarMsg* request, MultiVarMsg* response,
             brpc::Controller* cntl) override {
167 168
    LOG(INFO) << "entered Handle";
    platform::RecordEvent record_event("SendAndRecvVariableHandler->Handle",
169 170
                                       platform::TracerEventType::Communication,
                                       1);
171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193
    FLAGS_eager_delete_tensor_gb = -1;

    // get microID from request
    // deserialize variable to micro scope
    // Push to heter worker's task_queue
    std::unique_ptr<paddle::framework::Scope> local_scope_ptr(
        new paddle::framework::Scope());
    auto& local_scope = *(local_scope_ptr.get());
    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
    platform::CPUPlace cpu_place;
    auto& cpu_dev_ctx = *pool.Get(cpu_place);

    auto message_name = request->message_name();
    auto& request_io_buffer = cntl->request_attachment();

    distributed::DeserializeFromMultiVarMsgAndIOBuf(
        *request, &request_io_buffer, cpu_dev_ctx, &local_scope);

    auto* var = local_scope.FindVar("microbatch_id");
    PADDLE_ENFORCE_NE(var, nullptr,
                      platform::errors::InvalidArgument(
                          "Not find variable microbatch_id in scope."));
    auto* tensor = var->GetMutable<framework::LoDTensor>();
194
    auto data = reinterpret_cast<const float*>(tensor->data());
195 196 197 198
    auto micro_id = static_cast<int>(data[0]);
    int minibatch_index = micro_id / 10;
    int microbatch_index = micro_id % 10;

199 200 201 202
    // check minibatch_index is in mini_scopes_
    std::unique_lock<std::mutex> lk(scope_mutex_);
    if ((*mini_scopes_).find(minibatch_index) != (*mini_scopes_).end()) {
      lk.unlock();
203

204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
      PADDLE_ENFORCE_EQ(
          (*micro_scopes_).find(minibatch_index) != (*micro_scopes_).end(), 1,
          platform::errors::InvalidArgument(
              "minibatch index should in current trainer"));

    } else {
      // create mini scope & micro scopes
      auto* minibatch_scope = &(scope_->NewScope());
      (*mini_scopes_)[minibatch_index] = minibatch_scope;
      (*micro_scopes_)[minibatch_index].reset(
          new std::vector<paddle::framework::Scope*>{});
      for (int i = 0; i < num_microbatch_; i++) {
        auto* micro_scope = &(minibatch_scope->NewScope());
        (*((*micro_scopes_)[minibatch_index])).push_back(micro_scope);
      }
      (*task_queue_)[minibatch_index].reset(
          new ::paddle::framework::BlockingQueue<
              std::pair<std::string, int>>());
      lk.unlock();
    }
224 225 226 227 228 229 230 231 232

    auto* micro_scope =
        (*((*micro_scopes_)[minibatch_index]))[microbatch_index];

    distributed::DeserializeFromMultiVarMsgAndIOBuf(
        *request, &request_io_buffer, *dev_ctx_, micro_scope);
    // blocking queue handles multi thread
    (*task_queue_)[minibatch_index]->Push(
        std::make_pair(message_name, microbatch_index));
233

234 235 236 237 238 239 240 241 242 243 244 245 246
    auto response_var_nums = request->recv_var_names_size();
    std::vector<std::string> response_var_names(response_var_nums),
        empty_var_names{};
    for (int var_idx = 0; var_idx < response_var_nums; ++var_idx) {
      response_var_names[var_idx] = request->recv_var_names(var_idx);
    }
    auto& response_io_buffer = cntl->response_attachment();
    distributed::SerializeToMultiVarMsgAndIOBuf(
        message_name, response_var_names, empty_var_names, *dev_ctx_,
        &local_scope, response, &response_io_buffer);
    return 0;
  }

247 248 249 250 251 252
 public:
  using shard_type = SparseTableShard<std::string, FixedFeatureValue>;
  std::shared_ptr<paddle::framework::Scope> local_scope_ptr;  // for switch
  std::unordered_map<std::string, uint32_t> vars_table;
  std::unique_ptr<shard_type[]> _local_shards;

253 254
 private:
  // share with HeterPipelineTrainer
255
  SharedMiniScope mini_scopes_{nullptr};
256 257 258 259
  SharedMicroScope micro_scopes_{nullptr};

  int num_microbatch_;
  int num_minibatch_;
260
  std::mutex scope_mutex_;
261 262 263 264 265 266 267

  bool is_first_stage_ = false;
  bool is_last_stage_ = false;

  SharedTaskQueue task_queue_;
};

268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505
class HeterService : public PsService {
 public:
  HeterService() {
    _service_handler_map[PS_STOP_SERVER] =
        std::bind(&HeterService::stop_heter_worker, this, std::placeholders::_1,
                  std::placeholders::_2, std::placeholders::_3);
    _service_handler_map[PS_START_PROFILER] =
        std::bind(&HeterService::start_profiler, this, std::placeholders::_1,
                  std::placeholders::_2, std::placeholders::_3);
    _service_handler_map[PS_STOP_PROFILER] =
        std::bind(&HeterService::stop_profiler, this, std::placeholders::_1,
                  std::placeholders::_2, std::placeholders::_3);

    service_handler_.local_scope_ptr =
        std::make_shared<paddle::framework::Scope>();
  }

  virtual ~HeterService() {}

  virtual void service(::google::protobuf::RpcController* controller,
                       const PsRequestMessage* request,
                       PsResponseMessage* response,
                       ::google::protobuf::Closure* done) {
    brpc::ClosureGuard done_guard(done);

    response->set_err_code(0);
    response->set_err_msg("");
    brpc::Controller* cntl = static_cast<brpc::Controller*>(controller);
    auto itr = _service_handler_map.find(request->cmd_id());
    if (itr == _service_handler_map.end()) {
      std::string err_msg(
          "undefined cmd_id, should match PsCmdID in ps.proto, cmd_id:");
      err_msg.append(std::to_string(request->cmd_id()));
      return;
    }
    serviceHandler handler = itr->second;
    int service_ret = handler(*request, *response, cntl);
    VLOG(4) << "handler in service ret: " << service_ret;
    if (service_ret != 0) {
      response->set_err_code(service_ret);
      response->set_err_msg("server internal error");
    }
  }

  virtual void SendAndRecvVariable(
      ::google::protobuf::RpcController* controller, const MultiVarMsg* request,
      MultiVarMsg* response, ::google::protobuf::Closure* done) {
    // This object helps you to call done->Run() in RAII style. If you need
    // to process the request asynchronously, pass done_guard.release().
    brpc::ClosureGuard done_guard(done);
    std::string message_name = request->message_name();
    VLOG(0) << "SendAndRecvVariable message_name: " << message_name;
    auto itr = handler_map_.find(message_name);
    brpc::Controller* cntl = static_cast<brpc::Controller*>(controller);
    LOG(INFO) << "SendAndRecvVariable(client addr) =" << cntl->remote_side();
    PADDLE_ENFORCE_NE(
        itr, handler_map_.end(),
        platform::errors::InvalidArgument(
            "HeterService::SendAndRecvVariable Get illegal message_name: %s "
            "which is not in HeterService::handler_map_",
            message_name));
    itr->second(request, response, cntl);
    // We don't want to call done->Run() here, release the guard.
    // done_guard.release();
  }

  virtual void RecvFromSwitch(::google::protobuf::RpcController* controller,
                              const MultiVarMsg* request, MultiVarMsg* response,
                              ::google::protobuf::Closure* done) {
    brpc::ClosureGuard done_guard(done);
    brpc::Controller* cntl = static_cast<brpc::Controller*>(controller);
    // int ret = service_handler_.QueryInSwitchWithScope(request, response,
    // cntl);
    int ret = service_handler_.QueryInSwitchWithShard(request, response, cntl);
    // std::string message_name = request->message_name();
    // auto itr = handler_map_.find(message_name);
    // int ret = itr->second(request, response, cntl);
    if (ret != 0) {
      LOG(ERROR) << "QueryInSwitchWithScope failed!";
    }
    // response->set_message_name(message_name);
  }

  virtual void SendToSwitch(::google::protobuf::RpcController* controller,
                            const MultiVarMsg* request,
                            PsResponseMessage* response,
                            ::google::protobuf::Closure* done) {
    VLOG(4) << "entering SendToSwitch";
    brpc::ClosureGuard done_guard(done);
    auto& switch_client_ptr_ =
        HeterClient::GetSwitchInstance(peer_endpoints_, PEER_ROLE_IS_SWITCH);
    if (switch_client_ptr_.peer_switch_channels_.empty()) {
      LOG(ERROR) << "switch_client_ptr_.peer_switch_channels_ null";
    }
    brpc::Channel* channel = switch_client_ptr_.peer_switch_channels_[0].get();
    brpc::Controller* cntl = static_cast<brpc::Controller*>(controller);
    // proxy: 定义新的 OnHeterRpcDone 对象(或者在类 OnHeterRpcDone 中 reset)
    OnHeterRpcDone* closure2 = new OnHeterRpcDone([](void* done) {
      auto* closure = reinterpret_cast<OnHeterRpcDone*>(done);
      int ret = closure->CheckResponse();
      closure->set_promise_value(ret);
      if (closure->cntl.Failed()) {
        PADDLE_ENFORCE_NE(
            closure->cntl.Failed(), true,
            platform::errors::Unimplemented(
                "HeterClient::SendS2S meets brpc error, error message is %s",
                closure->cntl.ErrorText()));
      }
    });
    auto& std_cntl = closure2->cntl;
    std_cntl.set_timeout_ms(FLAGS_pserver_timeout_ms);
    std_cntl.request_attachment().append(cntl->request_attachment().movable());

    auto promise = std::make_shared<std::promise<int32_t>>();
    closure2->add_promise(promise);
    std::future<int> fut = promise->get_future();
    // brpc::Controller std_cntl;
    // std_cntl.request_attachment().append(cntl->request_attachment().movable());
    PsService_Stub stub(channel);
    stub.SendS2S(&std_cntl, request, response, closure2);
    cntl->response_attachment().append(
        std_cntl.response_attachment().movable());
    fut.wait();
    VLOG(4) << "SendToSwitch done";
  }

  void SendS2S(::google::protobuf::RpcController* controller,
               const MultiVarMsg* request, PsResponseMessage* response,
               ::google::protobuf::Closure* done) {
    VLOG(4) << "entering SendS2S";
    brpc::ClosureGuard done_guard(done);
    brpc::Controller* cntl = static_cast<brpc::Controller*>(controller);
    // int ret = service_handler_.SaveInSwitchWithScope(request, response,
    // cntl);
    int ret = service_handler_.SaveInSwitchWithShard(request, response, cntl);
    // std::string message_name = request->message_name();
    // auto itr = handler_map_.find(message_name);
    // if (itr == handler_map_.end()) {
    //    LOG(ERROR) << "can not find func handler";
    //}
    // int ret = itr->second(request, response, cntl);
    if (ret != 0) {
      LOG(ERROR) << "SaveInSwitchWithScope failed";
    }
    std::string err_msg = "ok";
    response->set_err_msg(err_msg.c_str());
    response->set_err_code(ret);
    VLOG(4) << "heter server SendS2S done";
  }

  void SendToWorker(::google::protobuf::RpcController* controller,
                    const MultiVarMsg* request, PsResponseMessage* response,
                    ::google::protobuf::Closure* done) {
    brpc::ClosureGuard done_guard(done);
    brpc::Controller* cntl = static_cast<brpc::Controller*>(controller);
    VLOG(4) << "SendToWorker(client addr) =" << cntl->remote_side();
    auto& switch_client_ptr_ =
        HeterClient::GetSwitchInstance(peer_endpoints_, PEER_ROLE_IS_WORKER);
    VLOG(4) << "in switch client, peer worker 0: "
            << switch_client_ptr_.peer_worker_list_[0];
    brpc::Channel* channel = switch_client_ptr_.peer_worker_channels_[0].get();

    auto* closure = reinterpret_cast<OnHeterRpcDone*>(done);
    PsService_Stub stub(channel);
    stub.SendAndRecvVariable(controller, request, &closure->response, done);
    // fill response content
    std::string err_msg("pass to worker");
    response->set_err_msg(err_msg.c_str());
    response->set_err_code(0);
  }

  void RegisterServiceHandler(std::string message_name,
                              HeterServiceHandler func) {
    handler_map_[message_name] = func;
  }

  void SetEndpoint(const std::string& end_point) { endpoint_ = end_point; }

  void SetInterEndpoint(const std::string& end_point) {
    endpoint_inter_ = end_point;
  }

  void SetPeerEndPoints(const std::vector<std::string>& peer_endpoints) {
    peer_endpoints_ = peer_endpoints;
  }

  void SetFanin(const int& fan_in) { fan_in_ = fan_in; }

  void ForceExit() {
    VLOG(3) << "heter service force exit";
    is_exit_ = true;
    return;
  }

  bool IsExit() { return is_exit_; }

 private:
  int32_t stop_profiler(const PsRequestMessage& request,
                        PsResponseMessage& response,  // NOLINT
                        brpc::Controller* cntl) {
    platform::DisableProfiler(
        platform::EventSortingKey::kDefault,
        string::Sprintf("heter_worker_%s_profile", endpoint_));
    return 0;
  }

  int32_t start_profiler(const PsRequestMessage& request,
                         PsResponseMessage& response,  // NOLINT
                         brpc::Controller* cntl) {
    platform::EnableProfiler(platform::ProfilerState::kAll);
    return 0;
  }

  int32_t stop_heter_worker(const PsRequestMessage& request,
                            PsResponseMessage& response,  // NOLINT
                            brpc::Controller* cntl) {
    auto client_id = request.client_id();
    stop_cpu_worker_set_.insert(client_id);
    if (stop_cpu_worker_set_.size() == fan_in_) {
      is_exit_ = true;
    }
    return 0;
  }

 private:
  SendAndRecvVariableHandler service_handler_;
  std::string endpoint_;
  std::string endpoint_inter_;
  // for switch
  std::vector<std::string> peer_endpoints_;

  std::unordered_map<int32_t, serviceHandler> _service_handler_map;
  std::unordered_map<std::string, HeterServiceHandler> handler_map_;
  std::unordered_set<int> stop_cpu_worker_set_;
  uint32_t fan_in_;
  bool is_exit_ = false;
};

T
tangwei12 已提交
506 507
class HeterServer {
 public:
508
  HeterServer() : ready_(0) {}
T
tangwei12 已提交
509 510
  virtual ~HeterServer() {}
  void Stop() {
T
tangwei12 已提交
511
    std::unique_lock<std::mutex> lock(mutex_);
512
    if (stoped_ == true) return;
513 514 515
    if (!IsExit()) {
      service_.ForceExit();
    }
T
tangwei12 已提交
516 517
    stoped_ = true;
    cv_.notify_all();
T
tangwei12 已提交
518 519 520 521
    server_.Stop(1000);
    server_.Join();
  }

522 523
  bool IsStop() {
    std::unique_lock<std::mutex> lock(mutex_);
524
    return stoped_;
525 526
  }

T
tangwei12 已提交
527 528 529 530 531
  bool IsExit() { return service_.IsExit(); }

  void RegisterServiceHandler(std::string message_name,
                              HeterServiceHandler func);

532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554
  void StartHeterService(bool need_encrypt = false);

  void StartHeterInterService(bool need_encrypt = false);

  void SetEndPoint(const std::string& endpoint) {
    this->endpoint_ = endpoint;
    service_.SetEndpoint(endpoint);
  }

  void SetLocalScope() {
    request_handler_->local_scope_ptr =
        std::make_shared<paddle::framework::Scope>();
  }

  void SetInterEndpoint(const std::string& endpoint) {
    this->endpoint_inter_ = endpoint;
    service_.SetInterEndpoint(endpoint);
  }

  void SetPeerEndPoints(const std::vector<std::string>& peer_endpoints) {
    this->peer_endpoints_ = peer_endpoints;
    service_.SetPeerEndPoints(peer_endpoints);
  }
T
tangwei12 已提交
555

556 557
  void SetFanin(const int& fan_in);

558 559
  void SetServiceHandler(
      std::shared_ptr<SendAndRecvVariableHandler> request_handler) {
560 561 562
    request_handler_ = request_handler;
  }

563 564 565
  void SetMiniBatchScopes(SharedMiniScope mini_scopes) {
    request_handler_->SetMiniScopes(mini_scopes);
  }
566 567 568 569 570

  void SetMicroBatchScopes(SharedMicroScope micro_scopes) {
    request_handler_->SetMicroScopes(micro_scopes);
  }

571 572
  int GetThreadNum() { return request_handler_->GetThreadNum(); }

573 574 575
  void SetTaskQueue(SharedTaskQueue task_queue) {
    request_handler_->SetTaskQueue(task_queue);
  }
T
tangwei12 已提交
576 577 578 579 580 581 582 583 584 585 586 587 588

  // HeterWrapper singleton
  static std::shared_ptr<HeterServer> GetInstance() {
    if (NULL == s_instance_) {
      s_instance_.reset(new HeterServer());
    }
    return s_instance_;
  }

  void WaitServerReady();

 private:
  static std::shared_ptr<HeterServer> s_instance_;
T
tangwei12 已提交
589 590 591
  mutable std::mutex mutex_;
  std::condition_variable cv_;
  std::condition_variable condition_ready_;
592
  bool stoped_ = true;
T
tangwei12 已提交
593
  std::string endpoint_;
594 595 596
  std::string endpoint_inter_;
  // for switch
  std::vector<std::string> peer_endpoints_;
T
tangwei12 已提交
597 598 599

 protected:
  brpc::Server server_;
600
  brpc::Server server_inter_;
T
tangwei12 已提交
601
  HeterService service_;
602
  std::shared_ptr<SendAndRecvVariableHandler> request_handler_;
603

T
tangwei12 已提交
604 605
  DISABLE_COPY_AND_ASSIGN(HeterServer);
  std::mutex mutex_ready_;
T
tangwei12 已提交
606

Z
zmxdream 已提交
607
  int ready_;
T
tangwei12 已提交
608 609 610 611
};

}  // end namespace distributed
}  // end namespace paddle