fleet_wrapper.cc 11.5 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 33 34 35 36 37

namespace paddle {
namespace framework {

const uint32_t MAX_FEASIGN_NUM = 1024 * 100 * 100;
std::shared_ptr<FleetWrapper> FleetWrapper::s_instance_ = NULL;
38 39
bool FleetWrapper::is_initialized_ = false;

40 41 42 43 44 45 46 47 48
#ifdef PADDLE_WITH_PSLIB
template<class AR>
paddle::ps::Archive<AR>& operator << (
    paddle::ps::Archive<AR>& ar,
    const MultiSlotType& ins) {
  ar << ins.GetType();
  ar << ins.GetOffset();
  ar << ins.GetFloatData();
  ar << ins.GetUint64Data();
X
xujiaqi01 已提交
49
  return ar;
50 51 52 53 54 55 56 57 58 59
}

template<class AR>
paddle::ps::Archive<AR>& operator >> (
    paddle::ps::Archive<AR>& ar,
    MultiSlotType& ins) {
  ar >> ins.MutableType();
  ar >> ins.MutableOffset();
  ar >> ins.MutableFloatData();
  ar >> ins.MutableUint64Data();
X
xujiaqi01 已提交
60
  return ar;
61 62 63
}
#endif

64 65 66
#ifdef PADDLE_WITH_PSLIB
std::shared_ptr<paddle::distributed::PSlib> FleetWrapper::pslib_ptr_ = NULL;
#endif
67 68 69 70

void FleetWrapper::InitServer(const std::string& dist_desc, int index) {
#ifdef PADDLE_WITH_PSLIB
  if (!is_initialized_) {
D
dongdaxiang 已提交
71
    VLOG(3) << "Going to init server";
72 73 74 75 76
    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
77
    VLOG(3) << "Server can be initialized only once";
78 79 80 81 82 83 84 85 86
  }
#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
87
    VLOG(3) << "Going to init worker";
88 89 90 91 92 93 94
    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
95
    VLOG(3) << "Worker can be initialized only once";
96 97 98 99 100 101
  }
#endif
}

void FleetWrapper::StopServer() {
#ifdef PADDLE_WITH_PSLIB
D
dongdaxiang 已提交
102
  VLOG(3) << "Going to stop server";
103 104 105 106 107 108
  pslib_ptr_->stop_server();
#endif
}

uint64_t FleetWrapper::RunServer() {
#ifdef PADDLE_WITH_PSLIB
D
dongdaxiang 已提交
109
  VLOG(3) << "Going to run server";
110 111 112 113 114 115 116 117 118
  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
119
  VLOG(3) << "Going to gather server ips";
120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174
  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
  std::vector<paddle::ps::Region> regions;
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);
182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238
  }
  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
  std::vector<paddle::ps::Region> regions;
  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
}

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) {
239 240 241 242
    LOG(WARNING) << "sparse key names[" << i << "]: " << sparse_key_names[i];
    LOG(WARNING) << "sparse grad names[" << i << "]: " << sparse_grad_names[i];
    Variable* g_var = scope.FindVar(sparse_grad_names[i]);
    CHECK(g_var != nullptr) << "var[" << sparse_grad_names[i] << "] not found";
243 244 245 246 247 248 249 250 251 252 253 254 255 256
    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();
257
    LOG(WARNING) << " tensor len: " << len;
258
    int64_t* ids = tensor->data<int64_t>();
259 260 261 262 263
    push_values->resize(fea_keys.size() + 1);
    for (auto& t : *push_values) {
      t.resize(emb_dim + offset);
    }

264 265 266 267 268
    for (auto id_idx = 0u; id_idx < len; ++id_idx) {
      if (ids[id_idx] == 0) {
        g += emb_dim;
        continue;
      }
269
      LOG(WARNING) << "going to memcpy";
270 271
      memcpy((*push_values)[fea_idx].data() + offset, g,
             sizeof(float) * emb_dim);
272
      LOG(WARNING) << "show";
273
      (*push_values)[fea_idx][0] = 1.0f;
274
      LOG(WARNING) << "click";
275
      (*push_values)[fea_idx][1] = static_cast<float>(fea_labels[fea_idx]);
276
      LOG(WARNING) << "offset";
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294
      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
}

X
xujiaqi01 已提交
295 296 297 298 299
int FleetWrapper::registe_client2client_msg_handler(
    int msg_type, MsgHandlerFunc handler) {
  pslib_ptr_->_worker_ptr->registe_client2client_msg_handler(
      msg_type, handler);
  return 0;
300 301
}

X
xujiaqi01 已提交
302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323
int FleetWrapper::send_client2client_msg(
    int msg_type, int to_client_id, const std::string& msg) {
  pslib_ptr_->_worker_ptr->send_client2client_msg(
      msg_type, to_client_id, msg);
  return 0;
}

std::default_random_engine& FleetWrapper::local_random_engine() {
  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;
324 325 326
}

template<typename T>
X
xujiaqi01 已提交
327
void FleetWrapper::Serialize(const T& t, std::string* str) {
328 329 330
#ifdef PADDLE_WITH_PSLIB
  paddle::ps::BinaryArchive ar;
  ar << t;
X
xujiaqi01 已提交
331
  *str = std::string(ar.buffer(), ar.length());
332 333 334 335 336 337
#else
  VLOG(0) << "FleetWrapper::Serialize do nothing when no pslib";
#endif
}

template<typename T>
X
xujiaqi01 已提交
338
void FleetWrapper::Deserialize(T* t, const std::string& str) {
339 340 341
#ifdef PADDLE_WITH_PSLIB
  paddle::ps::BinaryArchive ar;
  ar.set_read_buffer(const_cast<char*>(str.c_str()), str.length(), nullptr);
X
xujiaqi01 已提交
342
  *t = ar.get<T>();
343 344 345 346 347 348
#else
  VLOG(0) << "FleetWrapper::Deserialize do nothing when no pslib";
#endif
}

template void FleetWrapper::Serialize<std::vector<MultiSlotType>>(
X
xujiaqi01 已提交
349
    const std::vector<MultiSlotType>&, std::string*);
350
template void FleetWrapper::Deserialize(
X
xujiaqi01 已提交
351
    std::vector<MultiSlotType>*, const std::string&);
352

353 354
}  // end namespace framework
}  // end namespace paddle