fleet_wrapper.cc 12.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// Copyright (c) 2019 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 17 18 19 20 21 22 23 24 25 26 27 28 29
/* 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. */

#include "paddle/fluid/framework/fleet/fleet_wrapper.h"
X
xujiaqi01 已提交
30
#include <utility>
31
#include "paddle/fluid/framework/data_feed.h"
32
#include "paddle/fluid/framework/scope.h"
33 34 35 36 37 38

namespace paddle {
namespace framework {

const uint32_t MAX_FEASIGN_NUM = 1024 * 100 * 100;
std::shared_ptr<FleetWrapper> FleetWrapper::s_instance_ = NULL;
39 40
bool FleetWrapper::is_initialized_ = false;

41
#ifdef PADDLE_WITH_PSLIB
D
dongdaxiang 已提交
42 43 44
template <class AR>
paddle::ps::Archive<AR>& operator<<(paddle::ps::Archive<AR>& ar,
                                    const MultiSlotType& ins) {
45 46 47 48
  ar << ins.GetType();
  ar << ins.GetOffset();
  ar << ins.GetFloatData();
  ar << ins.GetUint64Data();
X
xujiaqi01 已提交
49
  return ar;
50 51
}

D
dongdaxiang 已提交
52 53 54
template <class AR>
paddle::ps::Archive<AR>& operator>>(paddle::ps::Archive<AR>& ar,
                                    MultiSlotType& ins) {
55 56 57 58
  ar >> ins.MutableType();
  ar >> ins.MutableOffset();
  ar >> ins.MutableFloatData();
  ar >> ins.MutableUint64Data();
X
xujiaqi01 已提交
59
  return ar;
60 61 62
}
#endif

63 64 65
#ifdef PADDLE_WITH_PSLIB
std::shared_ptr<paddle::distributed::PSlib> FleetWrapper::pslib_ptr_ = NULL;
#endif
66 67 68 69

void FleetWrapper::InitServer(const std::string& dist_desc, int index) {
#ifdef PADDLE_WITH_PSLIB
  if (!is_initialized_) {
D
dongdaxiang 已提交
70
    VLOG(3) << "Going to init server";
71 72 73 74 75
    pslib_ptr_ = std::shared_ptr<paddle::distributed::PSlib>(
        new paddle::distributed::PSlib());
    pslib_ptr_->init_server(dist_desc, index);
    is_initialized_ = true;
  } else {
D
dongdaxiang 已提交
76
    VLOG(3) << "Server can be initialized only once";
77 78 79 80 81 82 83 84 85
  }
#endif
}

void FleetWrapper::InitWorker(const std::string& dist_desc,
                              const std::vector<uint64_t>& host_sign_list,
                              int node_num, int index) {
#ifdef PADDLE_WITH_PSLIB
  if (!is_initialized_) {
D
dongdaxiang 已提交
86
    VLOG(3) << "Going to init worker";
87 88 89 90 91 92 93
    pslib_ptr_ = std::shared_ptr<paddle::distributed::PSlib>(
        new paddle::distributed::PSlib());
    pslib_ptr_->init_worker(dist_desc,
                            const_cast<uint64_t*>(host_sign_list.data()),
                            node_num, index);
    is_initialized_ = true;
  } else {
D
dongdaxiang 已提交
94
    VLOG(3) << "Worker can be initialized only once";
95 96 97 98 99 100
  }
#endif
}

void FleetWrapper::StopServer() {
#ifdef PADDLE_WITH_PSLIB
D
dongdaxiang 已提交
101
  VLOG(3) << "Going to stop server";
102 103 104 105 106 107
  pslib_ptr_->stop_server();
#endif
}

uint64_t FleetWrapper::RunServer() {
#ifdef PADDLE_WITH_PSLIB
D
dongdaxiang 已提交
108
  VLOG(3) << "Going to run server";
109 110 111 112 113 114 115 116 117
  return pslib_ptr_->run_server();
#else
  return 0;
#endif
}

void FleetWrapper::GatherServers(const std::vector<uint64_t>& host_sign_list,
                                 int node_num) {
#ifdef PADDLE_WITH_PSLIB
D
dongdaxiang 已提交
118
  VLOG(3) << "Going to gather server ips";
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172
  pslib_ptr_->gather_servers(const_cast<uint64_t*>(host_sign_list.data()),
                             node_num);
#endif
}

void FleetWrapper::PullSparseVarsSync(
    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) {
#ifdef PADDLE_WITH_PSLIB
  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 (auto name : var_names) {
    Variable* var = scope.FindVar(name);
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    int64_t* ids = tensor->data<int64_t>();
    int 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());
    }
    auto status = pslib_ptr_->_worker_ptr->pull_sparse(
        pull_result_ptr.data(), table_id, fea_keys->data(), fea_keys->size());
    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 << "]";
      exit(-1);
    }
  }
#endif
}

void FleetWrapper::PullDenseVarsAsync(
    const Scope& scope, const uint64_t tid,
    const std::vector<std::string>& var_names,
    std::vector<::std::future<int32_t>>* pull_dense_status) {
#ifdef PADDLE_WITH_PSLIB
X
xujiaqi01 已提交
173 174
  auto& regions = _regions[tid];
  regions.clear();
175 176 177
  regions.resize(var_names.size());
  for (auto i = 0u; i < var_names.size(); ++i) {
    Variable* var = scope.FindVar(var_names[i]);
178 179 180
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    float* w = tensor->data<float>();
    paddle::ps::Region reg(w, tensor->numel());
181
    regions[i] = std::move(reg);
182 183 184 185 186 187 188 189 190 191 192
  }
  auto status =
      pslib_ptr_->_worker_ptr->pull_dense(regions.data(), regions.size(), tid);
  pull_dense_status->push_back(std::move(status));
#endif
}

void FleetWrapper::PullDenseVarsSync(
    const Scope& scope, const uint64_t tid,
    const std::vector<std::string>& var_names) {
#ifdef PADDLE_WITH_PSLIB
X
xujiaqi01 已提交
193 194
  auto& regions = _regions[tid];
  regions.clear();
195 196 197 198 199 200 201 202 203 204 205 206 207 208
  regions.reserve(var_names.size());
  for (auto& t : var_names) {
    Variable* var = scope.FindVar(t);
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    float* w = tensor->data<float>();
    paddle::ps::Region reg(w, tensor->numel());
    regions.emplace_back(std::move(reg));
  }
  auto status =
      pslib_ptr_->_worker_ptr->pull_dense(regions.data(), regions.size(), tid);
  status.wait();
#endif
}

209 210 211 212 213 214 215 216 217
void FleetWrapper::PushDenseParamSync(
    const ProgramDesc& program, const uint64_t table_id,
    const std::vector<std::string>& var_names) {
#ifdef PADDLE_WITH_PSLIB
  auto place = platform::CPUPlace();
  std::vector<paddle::ps::Region> regions;
  for (auto& t : var_names) {
    Variable* var = scope.FindVar(t);
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
218
    float* g = tensor->mutable_data<float>(place);
219 220 221
    paddle::ps::Region reg(g, tensor->numel());
    regions.emplace_back(std::move(reg));
  }
222 223 224 225 226
  auto push_status = pslib_ptr_->_worker_ptr->push_dense_param(
      regions.data(), regions.size(), table_id);
  push_status.wait();
  auto status = push_status.get();
  CHECK(status == 0) << "push dense param failed, status[" << status << "]";
227 228 229
#endif
}

D
dongdaxiang 已提交
230 231 232 233
void FleetWrapper::PushDenseVarsSync(
    Scope* scope, const uint64_t table_id,
    const std::vector<std::string>& var_names) {}

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
void FleetWrapper::PushDenseVarsAsync(
    const Scope& scope, const uint64_t table_id,
    const std::vector<std::string>& var_names,
    std::vector<::std::future<int32_t>>* push_sparse_status) {
#ifdef PADDLE_WITH_PSLIB
  std::vector<paddle::ps::Region> regions;
  for (auto& t : var_names) {
    Variable* var = scope.FindVar(t);
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    int count = tensor->numel();
    float* g = tensor->data<float>();
    paddle::ps::Region reg(g, count);
    regions.emplace_back(std::move(reg));
  }
  auto status = pslib_ptr_->_worker_ptr->push_dense(regions.data(),
                                                    regions.size(), table_id);
  push_sparse_status->push_back(std::move(status));
#endif
}

void FleetWrapper::PushSparseVarsWithLabelAsync(
    const Scope& scope, const uint64_t table_id,
    const std::vector<uint64_t>& fea_keys, const std::vector<float>& fea_labels,
    const std::vector<std::string>& sparse_key_names,
    const std::vector<std::string>& sparse_grad_names, const int emb_dim,
    std::vector<std::vector<float>>* push_values,
    std::vector<::std::future<int32_t>>* push_sparse_status) {
#ifdef PADDLE_WITH_PSLIB
  int offset = 2;
  uint64_t fea_idx = 0u;
  for (size_t i = 0; i < sparse_key_names.size(); ++i) {
265 266
    Variable* g_var = scope.FindVar(sparse_grad_names[i]);
    CHECK(g_var != nullptr) << "var[" << sparse_grad_names[i] << "] not found";
267 268 269 270 271 272 273 274 275 276 277 278 279 280 281
    LoDTensor* g_tensor = g_var->GetMutable<LoDTensor>();
    if (g_tensor == NULL) {
      LOG(ERROR) << "var[" << sparse_key_names[i] << "] not found";
      exit(-1);
    }
    float* g = g_tensor->data<float>();
    Variable* var = scope.FindVar(sparse_key_names[i]);
    CHECK(var != nullptr) << "var[" << sparse_key_names[i] << "] not found";
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    if (tensor == NULL) {
      LOG(ERROR) << "var[" << sparse_key_names[i] << "] not found";
      exit(-1);
    }
    int len = tensor->numel();
    int64_t* ids = tensor->data<int64_t>();
282 283 284 285 286
    push_values->resize(fea_keys.size() + 1);
    for (auto& t : *push_values) {
      t.resize(emb_dim + offset);
    }

287 288 289 290 291
    for (auto id_idx = 0u; id_idx < len; ++id_idx) {
      if (ids[id_idx] == 0) {
        g += emb_dim;
        continue;
      }
292 293
      CHECK(fea_idx < (*push_values).size());
      CHECK(fea_idx < fea_labels.size());
294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315
      memcpy((*push_values)[fea_idx].data() + offset, g,
             sizeof(float) * emb_dim);
      (*push_values)[fea_idx][0] = 1.0f;
      (*push_values)[fea_idx][1] = static_cast<float>(fea_labels[fea_idx]);
      g += emb_dim;
      fea_idx++;
    }
  }
  CHECK(fea_idx == fea_keys.size()) << "fea_idx: " << fea_idx
                                    << "features size: " << fea_keys.size();
  std::vector<float*> push_g_vec;
  for (auto i = 0u; i < fea_keys.size(); ++i) {
    push_g_vec.push_back((*push_values)[i].data());
  }
  auto status = pslib_ptr_->_worker_ptr->push_sparse(
      table_id, fea_keys.data(), (const float**)push_g_vec.data(),
      fea_keys.size());
  push_sparse_status->push_back(std::move(status));

#endif
}

316 317
int FleetWrapper::RegisterClientToClientMsgHandler(int msg_type,
                                                   MsgHandlerFunc handler) {
318
#ifdef PADDLE_WITH_PSLIB
X
xujiaqi01 已提交
319 320 321
  VLOG(3) << "calling FleetWrapper::RegisterClientToClientMsgHandler";
  VLOG(3) << "pslib_ptr_=" << pslib_ptr_;
  VLOG(3) << "_worker_ptr=" << pslib_ptr_->_worker_ptr;
322 323
  return pslib_ptr_->_worker_ptr->registe_client2client_msg_handler(msg_type,
                                                                    handler);
324 325 326 327
#else
  VLOG(0) << "FleetWrapper::RegisterClientToClientMsgHandler"
          << " does nothing when no pslib";
#endif
X
xujiaqi01 已提交
328
  return 0;
329 330
}

331 332
std::future<int32_t> FleetWrapper::SendClientToClientMsg(
    int msg_type, int to_client_id, const std::string& msg) {
333
#ifdef PADDLE_WITH_PSLIB
334 335
  return pslib_ptr_->_worker_ptr->send_client2client_msg(msg_type, to_client_id,
                                                         msg);
336 337 338 339
#else
  VLOG(0) << "FleetWrapper::SendClientToClientMsg"
          << " does nothing when no pslib";
#endif
340
  return std::future<int32_t>();
X
xujiaqi01 已提交
341 342
}

D
dongdaxiang 已提交
343
template <typename T>
344
void FleetWrapper::Serialize(const std::vector<T*>& t, std::string* str) {
345 346
#ifdef PADDLE_WITH_PSLIB
  paddle::ps::BinaryArchive ar;
347 348 349
  for (size_t i = 0; i < t.size(); ++i) {
    ar << *(t[i]);
  }
X
xujiaqi01 已提交
350
  *str = std::string(ar.buffer(), ar.length());
351
#else
352
  VLOG(0) << "FleetWrapper::Serialize does nothing when no pslib";
353 354 355
#endif
}

D
dongdaxiang 已提交
356
template <typename T>
357
void FleetWrapper::Deserialize(std::vector<T>* t, const std::string& str) {
358
#ifdef PADDLE_WITH_PSLIB
359 360 361
  if (str.length() == 0) {
    return;
  }
362 363
  paddle::ps::BinaryArchive ar;
  ar.set_read_buffer(const_cast<char*>(str.c_str()), str.length(), nullptr);
364 365 366 367 368 369 370 371
  if (ar.cursor() == ar.finish()) {
    return;
  }
  while (ar.cursor() < ar.finish()) {
    t->push_back(ar.get<T>());
  }
  CHECK(ar.cursor() == ar.finish());
  VLOG(3) << "Deserialize size " << t->size();
372
#else
373
  VLOG(0) << "FleetWrapper::Deserialize does nothing when no pslib";
374 375 376 377
#endif
}

template void FleetWrapper::Serialize<std::vector<MultiSlotType>>(
378 379 380
    const std::vector<std::vector<MultiSlotType>*>&, std::string*);
template void FleetWrapper::Deserialize<std::vector<MultiSlotType>>(
    std::vector<std::vector<MultiSlotType>>*, const std::string&);
381

382 383
}  // end namespace framework
}  // end namespace paddle