fleet.cc 31.0 KB
Newer Older
T
tangwei12 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2020 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. */

15 16
#include "paddle/fluid/distributed/ps/wrapper/fleet.h"

17 18
#include <google/protobuf/text_format.h>

19 20
#include "paddle/fluid/distributed/ps/service/communicator/communicator.h"
#include "paddle/fluid/distributed/ps/table/table.h"
T
tangwei12 已提交
21 22 23 24 25 26 27 28 29 30 31 32 33 34

namespace paddle {
namespace distributed {

using framework::LoDTensor;
using framework::ProgramDesc;
using framework::VarDesc;
using framework::Variable;

const uint32_t MAX_FEASIGN_NUM = 1024 * 100 * 100;
std::shared_ptr<FleetWrapper> FleetWrapper::s_instance_ = NULL;
bool FleetWrapper::is_initialized_ = false;

std::shared_ptr<paddle::distributed::PSCore> FleetWrapper::pserver_ptr_ = NULL;
35 36 37 38 39 40 41 42 43 44 45 46 47 48
std::shared_ptr<paddle::distributed::PSClient> FleetWrapper::worker_ptr_ = NULL;

int FleetWrapper::RegisterHeterCallback(HeterCallBackFunc handler) {
  VLOG(0) << "RegisterHeterCallback support later";
  return 0;
}

int32_t FleetWrapper::CopyTable(const uint64_t src_table_id,
                                const uint64_t dest_table_id) {
  VLOG(0) << "CopyTable support later";
  return 0;
}

int32_t FleetWrapper::CopyTableByFeasign(
49 50
    const uint64_t src_table_id,
    const uint64_t dest_table_id,
51 52 53 54
    const std::vector<uint64_t>& feasign_list) {
  VLOG(0) << "CopyTableByFeasign support later";
  return 0;
}
T
tangwei12 已提交
55 56 57 58 59 60 61 62 63 64 65 66 67 68

void FleetWrapper::SetClient2ClientConfig(int request_timeout_ms,
                                          int connect_timeout_ms,
                                          int max_retry) {
  client2client_request_timeout_ms_ = request_timeout_ms;
  client2client_connect_timeout_ms_ = connect_timeout_ms;
  client2client_max_retry_ = max_retry;
}

void FleetWrapper::LoadSparseOnServer(const std::string& path,
                                      const std::string& meta,
                                      uint32_t table_id) {
  VLOG(3) << "load sparse table " << table_id << " with " << path << " meta "
          << meta;
Z
zhaocaibei123 已提交
69
  pserver_ptr_->_server_ptr->GetTable(table_id)->Load(path, meta);
T
tangwei12 已提交
70 71
}

72 73
void FleetWrapper::InitServer(
    const std::string& dist_desc,
74 75 76
    const std::vector<std::string>& host_sign_list,
    int index,
    int trainers,
77
    const std::vector<framework::ProgramDesc>& server_sub_program) {
T
tangwei12 已提交
78 79 80 81
  if (!is_initialized_) {
    VLOG(3) << "Going to init server";
    pserver_ptr_ = std::shared_ptr<paddle::distributed::PSCore>(
        new paddle::distributed::PSCore());
82 83 84 85 86 87
    pserver_ptr_->InitServer(dist_desc,
                             &host_sign_list,
                             host_sign_list.size(),
                             index,
                             trainers,
                             server_sub_program);
T
tangwei12 已提交
88 89 90 91 92 93
    is_initialized_ = true;
  } else {
    VLOG(3) << "Server can be initialized only once";
  }
}

94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112
void FleetWrapper::InitGFlag(const std::string& gflags) {
  VLOG(3) << "Init With Gflags:" << gflags;
  std::vector<std::string> flags = paddle::string::split_string(gflags);
  if (flags.size() < 1) {
    flags.push_back("-max_body_size=314217728");
    flags.push_back("-bthread_concurrency=40");
    flags.push_back("-socket_max_unwritten_bytes=2048000000");
    flags.push_back("-max_connection_pool_size=1950");
  }
  auto it = flags.begin();
  flags.insert(it, "exe default");
  char* flags_ptr[flags.size()];
  for (size_t i = 0; i < flags.size(); ++i) {
    flags_ptr[i] = (char*)(flags[i].c_str());  // NOLINT
  }
  int params_cnt = flags.size();
  char** params_ptr = &(flags_ptr[0]);
  ::GFLAGS_NAMESPACE::ParseCommandLineFlags(&params_cnt, &params_ptr, true);
}
T
tangwei12 已提交
113

114 115 116 117 118 119 120 121 122 123 124 125 126 127
void FleetWrapper::InitWorker(const std::string& dist_desc,
                              const std::vector<std::string>& host_sign_list,
                              int index) {
  if (!is_initialized_) {
    // not used, just for psclient's init
    // TODO(zhaocaibei123): remove this later
    std::map<uint64_t, std::vector<paddle::distributed::Region>>
        dense_pull_regions;

    if (worker_ptr_.get() == nullptr) {
      paddle::distributed::PSParameter ps_param;
      google::protobuf::TextFormat::ParseFromString(dist_desc, &ps_param);
      InitGFlag(ps_param.init_gflags());
      int servers = host_sign_list.size();
Z
zhaocaibei123 已提交
128
      ps_env_.SetPsServers(&host_sign_list, servers);
129
      worker_ptr_ = std::shared_ptr<paddle::distributed::PSClient>(
Z
zhaocaibei123 已提交
130 131
          paddle::distributed::PSClientFactory::Create(ps_param));
      worker_ptr_->Configure(ps_param, dense_pull_regions, ps_env_, index);
132
    }
T
tangwei12 已提交
133
  } else {
134
    VLOG(3) << "Client can be initialized only once";
T
tangwei12 已提交
135 136 137 138 139
  }
}

void FleetWrapper::StopServer() {
  VLOG(3) << "Going to stop server";
Z
zhaocaibei123 已提交
140
  auto status = worker_ptr_->StopServer();
T
tangwei12 已提交
141 142 143 144 145
  status.wait();
}

void FleetWrapper::FinalizeWorker() {
  VLOG(3) << "Going to finalize worker";
Z
zhaocaibei123 已提交
146
  worker_ptr_->FinalizeWorker();
T
tangwei12 已提交
147 148 149 150 151 152 153 154 155 156
}

void FleetWrapper::BarrierWithTable(uint32_t barrier_type) {
  VLOG(3) << "Going to Barrier worker";
  auto* communicator = Communicator::GetInstance();
  communicator->BarrierWithTable(barrier_type);
}

uint64_t FleetWrapper::RunServer(const std::string& ip, uint32_t port) {
  VLOG(3) << "Going to run server with ip " << ip << " port " << port;
Z
zhaocaibei123 已提交
157
  auto ret = pserver_ptr_->RunServer(ip, port);
T
tangwei12 已提交
158 159 160 161 162
  return ret;
}

std::vector<uint64_t> FleetWrapper::GetClientsInfo() {
  VLOG(3) << "Going to get client info";
Z
zhaocaibei123 已提交
163
  std::vector<uint64_t> res = ps_env_.GetClientInfo();
164 165 166
  for (auto rr : res) {
    VLOG(2) << "FleetWrapper::GetClientInfo " << rr;
  }
Z
zhaocaibei123 已提交
167
  return res;
T
tangwei12 已提交
168 169
}

170 171
int FleetWrapper::SetClients(std::vector<uint64_t>& host_sign_list) {
  int node = host_sign_list.size();
Z
zhaocaibei123 已提交
172
  return ps_env_.SetPsClients(host_sign_list.data(), node);
173 174
}

T
tangwei12 已提交
175
void FleetWrapper::CreateClient2ClientConnection() {
Z
zhaocaibei123 已提交
176
  VLOG(1) << "Going to create client2client connection";
Z
zhaocaibei123 已提交
177 178 179
  worker_ptr_->CreateClient2ClientConnection(client2client_request_timeout_ms_,
                                             client2client_connect_timeout_ms_,
                                             client2client_max_retry_);
T
tangwei12 已提交
180 181
}

182
std::future<int32_t> FleetWrapper::PullSparseVarsAsync(
183 184 185 186 187 188
    const Scope& scope,
    const uint64_t table_id,
    const std::vector<std::string>& var_names,
    std::vector<uint64_t>* fea_keys,
    std::vector<std::vector<float>>* fea_values,
    int fea_value_dim) {
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217
  fea_keys->clear();
  fea_keys->resize(0);
  fea_keys->reserve(MAX_FEASIGN_NUM);
  for (auto name : var_names) {
    Variable* var = scope.FindVar(name);
    if (var == nullptr) {
      continue;
    }
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    CHECK(tensor != nullptr) << "tensor of var " << name << " is null";
    int64_t* ids = tensor->data<int64_t>();
    size_t len = tensor->numel();
    for (auto i = 0u; i < len; ++i) {
      if (ids[i] == 0u) {
        continue;
      }
      fea_keys->push_back(static_cast<uint64_t>(ids[i]));
    }
  }
  fea_values->resize(fea_keys->size() + 1);
  for (auto& t : *fea_values) {
    t.resize(fea_value_dim);
  }
  std::vector<float*> pull_result_ptr;
  for (auto& t : *fea_values) {
    pull_result_ptr.push_back(t.data());
  }

  bool training = true;
218 219
  return pserver_ptr_->_worker_ptr->PullSparse(pull_result_ptr.data(),
                                               table_id,
Z
zhaocaibei123 已提交
220
                                               fea_keys->data(),
221 222
                                               fea_keys->size(),
                                               training);
223 224
}

T
tangwei12 已提交
225
void FleetWrapper::PullSparseVarsSync(
226 227 228 229 230 231
    const Scope& scope,
    const uint64_t table_id,
    const std::vector<std::string>& var_names,
    std::vector<uint64_t>* fea_keys,
    std::vector<std::vector<float>>* fea_values,
    int fea_value_dim,
T
tangwei12 已提交
232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270
    const std::vector<std::string>& var_emb_names) {
  std::vector<std::future<int32_t>> pull_sparse_status;
  pull_sparse_status.resize(0);
  fea_keys->clear();
  fea_keys->resize(0);
  fea_keys->reserve(MAX_FEASIGN_NUM);
  for (size_t var_index = 0; var_index < var_names.size(); ++var_index) {
    const std::string& name = var_names[var_index];
    Variable* var = scope.FindVar(name);
    if (var == nullptr) {
      continue;
    }
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    CHECK(tensor != nullptr) << "tensor of var " << name << " is null";
    int64_t* ids = tensor->data<int64_t>();
    size_t len = tensor->numel();

    // skip slots which do not have embedding
    const std::string& emb_name = var_emb_names[var_index];
    Variable* emb_var = scope.FindVar(emb_name);
    if (emb_var == nullptr) {
      continue;
    }

    for (auto i = 0u; i < len; ++i) {
      if (ids[i] == 0u) {
        continue;
      }
      fea_keys->push_back(static_cast<uint64_t>(ids[i]));
    }
  }
  fea_values->resize(fea_keys->size() + 1);
  for (auto& t : *fea_values) {
    t.resize(fea_value_dim);
  }
  std::vector<float*> pull_result_ptr;
  for (auto& t : *fea_values) {
    pull_result_ptr.push_back(t.data());
  }
271
  bool training = true;
272 273 274 275 276
  auto status = pserver_ptr_->_worker_ptr->PullSparse(pull_result_ptr.data(),
                                                      table_id,
                                                      fea_keys->data(),
                                                      fea_keys->size(),
                                                      training);
T
tangwei12 已提交
277 278 279 280 281 282 283 284 285 286 287 288
  pull_sparse_status.push_back(std::move(status));
  for (auto& t : pull_sparse_status) {
    t.wait();
    auto status = t.get();
    if (status != 0) {
      LOG(ERROR) << "fleet pull sparse failed, status[" << status << "]";
      sleep(sleep_seconds_before_fail_exit_);
      exit(-1);
    }
  }
}

289 290 291
// is_training is true means training, false means inference, the behavior is
// different on pserver

292 293
void FleetWrapper::PullSparseToTensorSync(const uint64_t table_id,
                                          int fea_dim,
T
tangwei12 已提交
294 295
                                          uint64_t padding_id,
                                          platform::Place place,
296
                                          bool is_training,
T
tangwei12 已提交
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
                                          std::vector<const LoDTensor*>* inputs,
                                          std::vector<LoDTensor*>* outputs) {
  std::vector<uint64_t> fea_keys;
  std::vector<float*> pull_result_ptr;
  fea_keys.reserve(MAX_FEASIGN_NUM / 100);
  pull_result_ptr.reserve(MAX_FEASIGN_NUM / 100);
  std::vector<float> init_value(fea_dim, 0);
  framework::LoDTensor* output = nullptr;
  float* output_data = nullptr;
  size_t output_index = -1;
  size_t output_len = 0;
  for (size_t index = 0; index < inputs->size(); ++index) {
    const framework::LoDTensor* tensor = inputs->at(index);
    const int64_t* ids = tensor->data<int64_t>();
    size_t len = tensor->numel();
    for (size_t i = 0; i < len; ++i, output_len += fea_dim) {
      if (!output || output_len == size_t(output->numel())) {
        ++output_index;
        CHECK(output_index < outputs->size());  // NOLINT
        output = outputs->at(output_index);
        output->set_lod(tensor->lod());
        output_data = output->mutable_data<float>(place);
        output_len = 0;
        CHECK(output->numel() % fea_dim == 0);  // NOLINT
        CHECK(output_data != nullptr);          // NOLINT
      }
      uint64_t real_id = static_cast<uint64_t>(ids[i]);
      if (real_id == padding_id) {
325 326
        memcpy(output_data + output_len,
               init_value.data(),
T
tangwei12 已提交
327 328 329 330 331 332 333
               sizeof(float) * fea_dim);
        continue;
      }
      fea_keys.push_back(real_id);
      pull_result_ptr.push_back(output_data + output_len);
    }
  }
Z
zhaocaibei123 已提交
334

335 336 337 338 339
  auto status = worker_ptr_->PullSparse(pull_result_ptr.data(),
                                        table_id,
                                        fea_keys.data(),
                                        fea_keys.size(),
                                        is_training);
T
tangwei12 已提交
340 341 342 343 344 345 346 347 348
  status.wait();
  auto ret = status.get();
  if (ret != 0) {
    LOG(ERROR) << "fleet pull sparse failed, status[" << ret << "]";
    sleep(sleep_seconds_before_fail_exit_);
  }
}

void FleetWrapper::PullDenseVarsAsync(
349 350
    const Scope& scope,
    const uint64_t tid,
T
tangwei12 已提交
351
    const std::vector<std::string>& var_names,
352 353
    std::vector<std::future<int32_t>>* pull_dense_status,
    bool in_cpu) {
Z
zhaocaibei123 已提交
354
  auto& regions = regions_[tid];
T
tangwei12 已提交
355 356 357 358 359 360 361 362 363 364 365 366 367
  regions.clear();
  regions.resize(var_names.size());
  for (auto i = 0u; i < var_names.size(); ++i) {
    std::string varname = var_names[i];
    if (!in_cpu) {
      varname = var_names[i] + "pin";
    }
    Variable* var = scope.FindVar(varname);
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    float* w = tensor->data<float>();
    paddle::distributed::Region reg(w, tensor->numel());
    regions[i] = std::move(reg);
  }
Z
zhaocaibei123 已提交
368 369

  auto status = worker_ptr_->PullDense(regions.data(), regions.size(), tid);
T
tangwei12 已提交
370 371 372 373
  pull_dense_status->push_back(std::move(status));
}

void FleetWrapper::PullDenseVarsSync(
374 375
    const Scope& scope,
    const uint64_t tid,
T
tangwei12 已提交
376
    const std::vector<std::string>& var_names) {
Z
zhaocaibei123 已提交
377
  auto& regions = regions_[tid];
T
tangwei12 已提交
378 379 380 381 382
  regions.clear();
  regions.reserve(var_names.size());
  for (auto& t : var_names) {
    Variable* var = scope.FindVar(t);
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
383 384 385 386 387
    if (!platform::is_gpu_place(tensor->place())) {
      float* w = tensor->data<float>();
      paddle::distributed::Region reg(w, tensor->numel());
      regions.emplace_back(std::move(reg));
    }
T
tangwei12 已提交
388
  }
Z
zhaocaibei123 已提交
389
  auto status = worker_ptr_->PullDense(regions.data(), regions.size(), tid);
T
tangwei12 已提交
390 391 392 393
  status.wait();
}

void FleetWrapper::PushDenseParamSync(
394 395
    const Scope& scope,
    const uint64_t table_id,
T
tangwei12 已提交
396 397 398 399 400 401 402
    const std::vector<std::string>& var_names) {
  auto place = platform::CPUPlace();
  std::vector<paddle::distributed::Region> regions;
  for (auto& t : var_names) {
    Variable* var = scope.FindVar(t);
    CHECK(var != nullptr) << "var[" << t << "] not found";
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
403 404 405 406 407
    if (!platform::is_gpu_place(tensor->place())) {
      float* g = tensor->mutable_data<float>(place);
      paddle::distributed::Region reg(g, tensor->numel());
      regions.emplace_back(std::move(reg));
    }
T
tangwei12 已提交
408
  }
409
  auto push_status =
Z
zhaocaibei123 已提交
410
      worker_ptr_->PushDenseParam(regions.data(), regions.size(), table_id);
T
tangwei12 已提交
411 412 413 414 415 416
  push_status.wait();
  auto status = push_status.get();
  CHECK(status == 0) << "push dense param failed, status[" << status << "]";
}

void FleetWrapper::PushDenseVarsSync(
417 418
    Scope* scope,
    const uint64_t table_id,
T
tangwei12 已提交
419 420 421
    const std::vector<std::string>& var_names) {}

void FleetWrapper::PushDenseVarsAsync(
422 423
    const Scope& scope,
    const uint64_t table_id,
T
tangwei12 已提交
424
    const std::vector<std::string>& var_names,
425 426
    std::vector<std::future<int32_t>>* push_sparse_status,
    float scale_datanorm,
T
tangwei12 已提交
427
    int batch_size) {
Z
zhaocaibei123 已提交
428 429 430 431 432 433
  auto place = platform::CPUPlace();
  std::vector<paddle::distributed::Region> regions;
  for (auto& t : var_names) {
    Variable* var = scope.FindVar(t);
    CHECK(var != nullptr) << "var[" << t << "] not found";
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
434
    int count = tensor->numel();
Z
zhaocaibei123 已提交
435
    float* g = tensor->mutable_data<float>(place);
436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451
    // TODO(zhaocaibei123): how to get batch_size in op?
    if (scale_datanorm >= 0) {
      if (t.find(".batch_size@GRAD") != std::string::npos ||
          t.find(".batch_sum@GRAD") != std::string::npos) {
        Eigen::Map<Eigen::MatrixXf> mat(g, 1, count);
        float scale = 1.0 / batch_size;
        mat *= scale;
      } else if (t.find(".batch_square_sum@GRAD") != std::string::npos) {
        VLOG(3) << "epsilon: " << scale_datanorm;
        for (int i = 0; i < count; ++i) {
          g[i] = (g[i] - batch_size * scale_datanorm) / batch_size +
                 batch_size * scale_datanorm;
        }
      }
    }

Z
zhaocaibei123 已提交
452 453 454 455 456 457 458
    paddle::distributed::Region reg(g, tensor->numel());
    regions.emplace_back(std::move(reg));
    VLOG(3) << "FleetWrapper::PushDenseVarsAsync Var " << t << " talbe_id "
            << table_id << " Temp_data[0] " << g[0] << " Temp_data[-1] "
            << g[tensor->numel() - 1];
  }

Z
zhaocaibei123 已提交
459 460
  auto push_status =
      worker_ptr_->PushDense(regions.data(), regions.size(), table_id);
T
tangwei12 已提交
461 462 463
}

void FleetWrapper::PushSparseVarsAsync(
464 465
    const Scope& scope,
    const uint64_t table_id,
T
tangwei12 已提交
466 467 468 469 470 471 472
    const std::string& grad_varname,
    std::vector<std::future<int32_t>>* push_sparse_status) {
  std::vector<std::string> varnames;
  varnames.push_back(grad_varname);

  auto* communicator = Communicator::GetInstance();
  PADDLE_ENFORCE_EQ(
473 474
      communicator->Check(table_id),
      true,
T
tangwei12 已提交
475 476 477 478 479 480
      platform::errors::InvalidArgument(
          "can not find table: %s, please check your config", table_id));
  communicator->Send(varnames, scope);
}

void FleetWrapper::PushSparseVarsWithLabelAsync(
481 482 483 484
    const Scope& scope,
    const uint64_t table_id,
    const std::vector<uint64_t>& fea_keys,
    const std::vector<float>& fea_labels,
T
tangwei12 已提交
485
    const std::vector<std::string>& sparse_key_names,
486 487
    const std::vector<std::string>& sparse_grad_names,
    const int emb_dim,
T
tangwei12 已提交
488
    std::vector<std::vector<float>>* push_values,
489 490 491 492 493 494
    std::vector<std::future<int32_t>>* push_sparse_status,
    const int batch_size,
    const bool use_cvm,
    const bool dump_slot,
    std::vector<uint64_t>* sparse_push_keys,
    const bool no_cvm) {
T
tangwei12 已提交
495 496 497 498 499
  // not support
  return;
}

void FleetWrapper::PushSparseFromTensorWithLabelAsync(
500 501 502 503 504 505 506 507
    const Scope& scope,
    const uint64_t table_id,
    int fea_dim,
    uint64_t padding_id,
    bool scale_sparse,
    const std::string& accesor,
    const std::string& click_name,
    platform::Place place,
T
tangwei12 已提交
508 509 510 511 512 513 514
    const std::vector<std::string>& input_names,
    std::vector<const LoDTensor*>* inputs,
    std::vector<const LoDTensor*>* outputs) {
  // not support
  return;
}

Z
zhaocaibei123 已提交
515
void FleetWrapper::PushSparseFromTensorAsync(
516 517 518 519 520 521 522 523 524
    const uint64_t table_id,
    int fea_dim,
    uint64_t padding_id,
    platform::Place place,
    std::vector<const LoDTensor*>* inputs,
    const LoDTensor* shows,
    const LoDTensor* clks,
    std::vector<LoDTensor*>* outputs,
    bool use_cvm_op) {
Z
zhaocaibei123 已提交
525
  int batch_size = -1;
Z
zhaocaibei123 已提交
526
  bool batch_size_consist = true;
Z
zhaocaibei123 已提交
527 528 529 530 531
  for (auto* input : *inputs) {
    int cur_batch_size =
        input->lod().size() ? input->lod()[0].size() - 1 : input->dims()[0];
    if (batch_size == -1) {
      batch_size = cur_batch_size;
532
    } else if (batch_size != cur_batch_size) {
Z
zhaocaibei123 已提交
533 534 535
      // CHECK(batch_size == cur_batch_size);  // NOLINT
      batch_size_consist = false;
      break;
Z
zhaocaibei123 已提交
536 537 538 539 540 541 542 543 544 545 546
    }
  }
  CHECK(batch_size > 0);  // NOLINT

  int show_size =
      shows->lod().size() ? shows->lod()[0].size() - 1 : shows->dims()[0];
  CHECK(show_size == batch_size || show_size == 1);
  int clk_size =
      clks->lod().size() ? clks->lod()[0].size() - 1 : clks->dims()[0];
  CHECK(clk_size == batch_size || clk_size == 1);

547
  CHECK(outputs->size() == inputs->size());
Z
zhaocaibei123 已提交
548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563
  std::vector<uint64_t> push_keys;
  push_keys.reserve(MAX_FEASIGN_NUM / 100);
  std::vector<std::vector<float>> push_values;
  push_values.reserve(MAX_FEASIGN_NUM / 100);
  size_t output_len = 0;
  size_t input_idx = 0;

  VLOG(2) << "fleet.cc::emb_dim: " << fea_dim;

  // TODO(zhaocaibei123): check type of show/clk is int? float? uint64?
  // const long int* show_tensor = shows->data<int64_t>();
  // const long int* clk_tensor = clks->data<int64_t>();
  const int64_t* show_tensor = shows->data<int64_t>();
  const int64_t* clk_tensor = clks->data<int64_t>();

  for (size_t index = 0; index < inputs->size(); ++index) {
564 565 566 567 568 569 570 571
    framework::LoDTensor* g_tensor = outputs->at(index);
    float* g = g_tensor->data<float>();
    // no cvm
    if (batch_size_consist) {  // TODO(zhaocaibei123): add config
                               // scale_sparse_gradient_with_batch_size_
      Eigen::Map<
          Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>>
          g_mat(g, g_tensor->numel() / fea_dim, fea_dim);
572 573 574 575 576
      if (use_cvm_op) {
        g_mat.rightCols(fea_dim - 2) *= batch_size;
      } else {
        g_mat.rightCols(fea_dim) *= batch_size;
      }
577 578
    }

Z
zhaocaibei123 已提交
579 580 581
    const framework::LoDTensor* tensor = inputs->at(index);
    const int64_t* ids = tensor->data<int64_t>();
    size_t len = tensor->numel();
582
    output_len = 0;
Z
zhaocaibei123 已提交
583 584

    if (tensor->lod().size() > 0) {
Z
zhangchunle 已提交
585
      for (size_t i = 0; i < tensor->lod()[0].size() - 1; ++i) {
586
        for (size_t j = tensor->lod()[0][i]; j < tensor->lod()[0][i + 1];
Z
zhaocaibei123 已提交
587 588 589 590 591 592
             ++j, output_len += fea_dim) {
          uint64_t real_id = static_cast<uint64_t>(ids[j]);
          if (real_id == padding_id) {
            continue;
          }
          push_keys.emplace_back(real_id);
593 594 595 596 597 598 599 600 601 602 603
          if (use_cvm_op) {
            push_values.emplace_back(fea_dim + 1);
            push_values.back()[0] = 2;  // TODO(zhaocaibei123): slot
            float* data = push_values.back().data() + 1;
            memcpy(data, g + output_len, sizeof(float) * fea_dim);
          } else {
            push_values.emplace_back(fea_dim + 3);
            // slot show clk grad... consistent with CtrCommonPushValue defined
            // in
            // ctr_accessor.h
            push_values.back()[0] = 2;  // TODO(zhaocaibei123): slot
Z
zhangchunle 已提交
604 605 606 607 608 609
            push_values.back()[1] = (static_cast<int>(i) >= show_size
                                         ? 1
                                         : static_cast<float>(show_tensor[i]));
            push_values.back()[2] = (static_cast<int>(i) >= clk_size
                                         ? 0
                                         : static_cast<float>(clk_tensor[i]));
610 611 612 613 614 615 616
            float* data = push_values.back().data() + 3;
            memcpy(data, g + output_len, sizeof(float) * fea_dim);
          }
          ++input_idx;
        }
      }
    } else {
Z
zhangchunle 已提交
617
      for (size_t i = 0; i < len; ++i, output_len += fea_dim) {
618 619 620 621 622 623 624 625 626 627 628
        uint64_t real_id = static_cast<uint64_t>(ids[i]);
        if (real_id == padding_id) {
          continue;
        }
        push_keys.emplace_back(real_id);
        if (use_cvm_op) {
          push_values.emplace_back(fea_dim + 1);
          push_values.back()[0] = 2;  // TODO(zhaocaibei123): slot
          float* data = push_values.back().data() + 1;
          memcpy(data, g + output_len, sizeof(float) * fea_dim);
        } else {
Z
zhaocaibei123 已提交
629 630 631 632
          push_values.emplace_back(fea_dim + 3);
          // slot show clk grad... consistent with CtrCommonPushValue defined in
          // ctr_accessor.h
          push_values.back()[0] = 2;  // TODO(zhaocaibei123): slot
Z
zhangchunle 已提交
633 634 635 636 637 638
          push_values.back()[1] = (static_cast<int>(i) >= show_size
                                       ? 1
                                       : static_cast<float>(show_tensor[i]));
          push_values.back()[2] = (static_cast<int>(i) >= clk_size
                                       ? 0
                                       : static_cast<float>(clk_tensor[i]));
Z
zhaocaibei123 已提交
639
          float* data = push_values.back().data() + 3;
640
          memcpy(data, g + output_len, sizeof(float) * fea_dim);
Z
zhaocaibei123 已提交
641 642 643 644
        }
        ++input_idx;
      }
    }
Z
zhangchunle 已提交
645
    CHECK(static_cast<int64_t>(output_len) == g_tensor->numel());
Z
zhaocaibei123 已提交
646 647 648 649 650 651 652 653
  }

  std::vector<float*> push_g_vec(input_idx, nullptr);

  for (auto i = 0u; i < push_keys.size(); ++i) {
    push_g_vec[i] = push_values.at(i).data();
  }

654 655
  auto status = worker_ptr_->PushSparse(table_id,
                                        push_keys.data(),
Z
zhaocaibei123 已提交
656 657
                                        (const float**)push_g_vec.data(),
                                        push_keys.size());
Z
zhaocaibei123 已提交
658 659 660
}

void FleetWrapper::LoadModel(const std::string& path, const int mode) {
Z
zhaocaibei123 已提交
661
  auto ret = worker_ptr_->Load(path, std::to_string(mode));
T
tangwei12 已提交
662 663 664 665 666 667 668
  ret.wait();
  if (ret.get() != 0) {
    LOG(ERROR) << "load model from path:" << path << " failed";
  }
}

void FleetWrapper::LoadModelOneTable(const uint64_t table_id,
669 670
                                     const std::string& path,
                                     const int mode) {
Z
zhaocaibei123 已提交
671
  auto ret = worker_ptr_->Load(table_id, path, std::to_string(mode));
T
tangwei12 已提交
672 673 674 675 676 677 678 679
  ret.wait();
  if (ret.get() != 0) {
    LOG(ERROR) << "load model of table id: " << table_id
               << ", from path: " << path << " failed";
  }
}

void FleetWrapper::SaveModel(const std::string& path, const int mode) {
Z
zhaocaibei123 已提交
680
  auto ret = worker_ptr_->Save(path, std::to_string(mode));
T
tangwei12 已提交
681 682 683 684 685 686 687 688
  ret.wait();
  int32_t feasign_cnt = ret.get();
  if (feasign_cnt == -1) {
    LOG(ERROR) << "save model failed";
  }
}

void FleetWrapper::SaveModelOneTable(const uint64_t table_id,
689 690
                                     const std::string& path,
                                     const int mode) {
Z
zhaocaibei123 已提交
691
  auto ret = worker_ptr_->Save(table_id, path, std::to_string(mode));
T
tangwei12 已提交
692 693 694 695 696 697 698
  ret.wait();
  if (ret.get() != 0) {
    LOG(ERROR) << "save model of table id: " << table_id
               << ", to path: " << path << " failed";
  }
}

699 700
void FleetWrapper::RecvAndSaveTable(const uint64_t table_id,
                                    const std::string& path) {
Z
zhaocaibei123 已提交
701
  auto ret = worker_ptr_->RecvAndSaveTable(table_id, path);
702 703 704 705 706 707
  if (ret != 0) {
    LOG(ERROR) << "save model of table id: " << table_id
               << ", to path: " << path << " failed";
  }
}

T
tangwei12 已提交
708
void FleetWrapper::PrintTableStat(const uint64_t table_id) {
Z
zhaocaibei123 已提交
709
  auto ret = worker_ptr_->PrintTableStat(table_id);
T
tangwei12 已提交
710 711 712 713 714 715 716
  ret.wait();
  int32_t err_code = ret.get();
  if (err_code == -1) {
    LOG(ERROR) << "print table stat failed";
  }
}

717
void FleetWrapper::ShrinkSparseTable(int table_id, int threshold) {
Z
zhaocaibei123 已提交
718
  auto ret = worker_ptr_->Shrink(table_id, std::to_string(threshold));
T
tangwei12 已提交
719
  ret.wait();
720 721 722 723
  int32_t err_code = ret.get();
  if (err_code == -1) {
    LOG(ERROR) << "shrink sparse table stat failed";
  }
T
tangwei12 已提交
724 725 726
}

void FleetWrapper::ClearModel() {
Z
zhaocaibei123 已提交
727
  auto ret = pserver_ptr_->_worker_ptr->Clear();
T
tangwei12 已提交
728 729 730 731
  ret.wait();
}

void FleetWrapper::ClearOneTable(const uint64_t table_id) {
Z
zhaocaibei123 已提交
732
  auto ret = pserver_ptr_->_worker_ptr->Clear(table_id);
T
tangwei12 已提交
733 734 735
  ret.wait();
}

736 737
void FleetWrapper::ShrinkDenseTable(int table_id,
                                    Scope* scope,
T
tangwei12 已提交
738
                                    std::vector<std::string> var_list,
739 740
                                    float decay,
                                    int emb_dim) {
T
tangwei12 已提交
741 742 743 744 745
  std::vector<paddle::distributed::Region> regions;
  for (std::string& name : var_list) {
    if (name.find("batch_sum") != std::string::npos) {
      Variable* var = scope->FindVar(name);
      CHECK(var != nullptr) << "var[" << name << "] not found";
746
      VLOG(3) << "prepare shrink dense batch_sum";
T
tangwei12 已提交
747 748 749 750 751
      LoDTensor* tensor = var->GetMutable<LoDTensor>();
      float* g = tensor->data<float>();

      // show_batch_sum += N * log(decay)
      std::string size_name = name;
752 753
      size_name.replace(
          size_name.find("batch_sum"), size_name.length(), "batch_size");
T
tangwei12 已提交
754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772
      Variable* var_size = scope->FindVar(size_name);
      CHECK(var_size != nullptr) << "var[" << size_name << "] not found";
      VLOG(3) << "shrink dense batch_sum: " << name << ", " << size_name;
      float* g_size = var_size->GetMutable<LoDTensor>()->data<float>();

      for (int k = 0; k < tensor->numel(); k += emb_dim) {
        g[k] = g[k] + g_size[k] * log(decay);
      }
      paddle::distributed::Region reg(g, tensor->numel());
      regions.emplace_back(std::move(reg));
    } else {
      Variable* var = scope->FindVar(name);
      CHECK(var != nullptr) << "var[" << name << "] not found";
      LoDTensor* tensor = var->GetMutable<LoDTensor>();
      float* g = tensor->data<float>();
      paddle::distributed::Region reg(g, tensor->numel());
      regions.emplace_back(std::move(reg));
    }
  }
Z
zhaocaibei123 已提交
773
  auto push_status = pserver_ptr_->_worker_ptr->PushDenseParam(
T
tangwei12 已提交
774 775 776 777 778 779 780 781 782 783 784 785
      regions.data(), regions.size(), table_id);
  push_status.wait();
  auto status = push_status.get();
  if (status != 0) {
    // PADDLE_THORW(platform::errors::Fatal(
    //    "push shrink dense param failed, status is [%d].", status));
    sleep(sleep_seconds_before_fail_exit_);
    exit(-1);
  }
}

void FleetWrapper::ClientFlush() {
786 787 788 789
  if (worker_ptr_.get() == nullptr) {
    VLOG(0) << "worker_ptr null, do nothing";
    return;
  }
Z
zhaocaibei123 已提交
790
  auto ret = worker_ptr_->Flush();
T
tangwei12 已提交
791
  ret.wait();
792 793 794 795
  int32_t err_code = ret.get();
  if (err_code == -1) {
    LOG(ERROR) << "Client Flush failed";
  }
T
tangwei12 已提交
796 797 798 799
}

int FleetWrapper::RegisterClientToClientMsgHandler(int msg_type,
                                                   MsgHandlerFunc handler) {
800 801
  if (worker_ptr_.get() == nullptr) {
    VLOG(0) << "FleetWrapper::Client is null";
Z
zhaocaibei123 已提交
802 803
    return -1;
  } else {
Z
zhaocaibei123 已提交
804
    return worker_ptr_->RegisteClient2ClientMsgHandler(msg_type, handler);
Z
zhaocaibei123 已提交
805
  }
T
tangwei12 已提交
806 807 808 809
}

std::future<int32_t> FleetWrapper::SendClientToClientMsg(
    int msg_type, int to_client_id, const std::string& msg) {
Z
zhaocaibei123 已提交
810
  return worker_ptr_->SendClient2ClientMsg(msg_type, to_client_id, msg);
T
tangwei12 已提交
811 812
}

Z
zhaocaibei123 已提交
813 814 815 816 817 818 819 820 821 822 823 824 825 826
double FleetWrapper::GetCacheThreshold(int table_id) {
  double cache_threshold = 0.0;
  auto ret = worker_ptr_->Flush();
  ret.wait();
  ret = worker_ptr_->GetCacheThreshold(table_id, cache_threshold);
  ret.wait();
  if (cache_threshold < 0) {
    LOG(ERROR) << "get cache threshold failed";
    sleep(sleep_seconds_before_fail_exit_);
    exit(-1);
  }
  return cache_threshold;
}

827 828 829 830 831 832
void FleetWrapper::CacheShuffle(int table_id,
                                const std::string& path,
                                const int mode,
                                const double cache_threshold) {
  auto ret = worker_ptr_->CacheShuffle(
      table_id, path, std::to_string(mode), std::to_string(cache_threshold));
Z
zhaocaibei123 已提交
833 834 835 836 837 838 839 840 841
  ret.wait();
  int32_t feasign_cnt = ret.get();
  if (feasign_cnt == -1) {
    LOG(ERROR) << "cache shuffle failed";
    sleep(sleep_seconds_before_fail_exit_);
    exit(-1);
  }
}

842 843
int32_t FleetWrapper::SaveCache(int table_id,
                                const std::string& path,
Z
zhaocaibei123 已提交
844 845 846 847 848 849 850 851 852 853 854 855
                                const int mode) {
  auto ret = worker_ptr_->SaveCache(table_id, path, std::to_string(mode));
  ret.wait();
  int32_t feasign_cnt = ret.get();
  if (feasign_cnt == -1) {
    LOG(ERROR) << "table save cache failed";
    sleep(sleep_seconds_before_fail_exit_);
    exit(-1);
  }
  return feasign_cnt;
}

Z
zhaocaibei123 已提交
856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873
void FleetWrapper::Revert() {
  auto ret = worker_ptr_->Revert();
  ret.wait();
  if (ret.get() == -1) {
    LOG(ERROR) << "table revert failed";
    exit(-1);
  }
}

void FleetWrapper::CheckSavePrePatchDone() {
  auto ret = worker_ptr_->CheckSavePrePatchDone();
  ret.wait();
  if (ret.get() == -1) {
    LOG(ERROR) << "table revert failed";
    exit(-1);
  }
}

T
tangwei12 已提交
874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890
std::default_random_engine& FleetWrapper::LocalRandomEngine() {
  struct engine_wrapper_t {
    std::default_random_engine engine;

    engine_wrapper_t() {
      struct timespec tp;
      clock_gettime(CLOCK_REALTIME, &tp);
      double cur_time = tp.tv_sec + tp.tv_nsec * 1e-9;
      static std::atomic<uint64_t> x(0);
      std::seed_seq sseq = {x++, x++, x++, (uint64_t)(cur_time * 1000)};
      engine.seed(sseq);
    }
  };
  thread_local engine_wrapper_t r;
  return r.engine;
}

891 892 893
size_t FleetWrapper::GetAbsoluteSum(size_t start,
                                    size_t end,
                                    size_t level,
T
tangwei12 已提交
894 895 896 897 898 899 900 901 902 903 904 905 906 907 908
                                    const framework::LoD& lod) {
  if (level >= lod.size() - 1) {
    return end - start;
  }
  size_t ret = 0;
  for (size_t i = start; i < end - 1; ++i) {
    size_t pos1 = lod[level][i];
    size_t pos2 = lod[level][i + 1];
    ret += GetAbsoluteSum(pos1, pos2, level + 1, lod);
  }
  return ret;
}

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