ps_gpu_wrapper.h 15.5 KB
Newer Older
T
Thunderbrook 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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. */

#pragma once

T
Thunderbrook 已提交
17
#ifdef PADDLE_WITH_HETERPS
T
Thunderbrook 已提交
18 19 20 21 22 23 24 25

#include <atomic>
#include <ctime>
#include <map>
#include <memory>
#include <random>
#include <string>
#include <unordered_map>
Y
yaoxuefeng 已提交
26
#include <unordered_set>
T
Thunderbrook 已提交
27
#include <vector>
28 29 30 31
#ifdef PADDLE_WITH_GLOO
#include <gloo/broadcast.h>
#include "paddle/fluid/framework/fleet/gloo_wrapper.h"
#endif
32
#include "paddle/fluid/distributed/ps/thirdparty/round_robin.h"
Y
yaoxuefeng 已提交
33
#include "paddle/fluid/framework/data_set.h"
T
Thunderbrook 已提交
34 35 36
#include "paddle/fluid/framework/fleet/heter_context.h"
#include "paddle/fluid/framework/fleet/heter_ps/heter_ps_base.h"
#include "paddle/fluid/framework/fleet/heter_ps/heter_resource.h"
Y
yaoxuefeng 已提交
37
#include "paddle/fluid/framework/fleet/heter_ps/mem_pool.h"
T
Thunderbrook 已提交
38 39 40
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/framework/variable_helper.h"
41
#include "paddle/fluid/platform/device/gpu/gpu_info.h"
42
#include "paddle/fluid/platform/dynload/nccl.h"
T
Thunderbrook 已提交
43 44
#include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
#include "paddle/fluid/platform/place.h"
T
Thunderbrook 已提交
45
#ifdef PADDLE_WITH_PSCORE
46
#include "paddle/fluid/distributed/ps/wrapper/fleet.h"
T
Thunderbrook 已提交
47
#endif
T
Thunderbrook 已提交
48 49 50
#ifdef PADDLE_WITH_PSLIB
#include "afs_api.h"
#endif
T
Thunderbrook 已提交
51 52 53 54

namespace paddle {
namespace framework {

Y
yaoxuefeng 已提交
55 56 57
#define TYPEALIGN(ALIGNVAL, LEN) \
  (((uint64_t)(LEN) + ((ALIGNVAL)-1)) & ~((uint64_t)((ALIGNVAL)-1)))

T
Thunderbrook 已提交
58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73
#ifdef PADDLE_WITH_PSLIB
class AfsWrapper {
 public:
  AfsWrapper() {}
  virtual ~AfsWrapper() {}
  void init(const std::string& fs_name, const std::string& fs_user,
            const std::string& pass_wd, const std::string& conf);
  int remove(const std::string& path);
  int mkdir(const std::string& path);
  std::vector<std::string> list(const std::string& path);

  int exist(const std::string& path);
  int upload(const std::string& local_file, const std::string& afs_file);

  int download(const std::string& local_file, const std::string& afs_file);

74 75 76 77
  int touchz(const std::string& path);
  std::string cat(const std::string& path);
  int mv(const std::string& old_path, const std::string& dest_path);

T
Thunderbrook 已提交
78 79 80 81 82
 private:
  paddle::ps::AfsApiWrapper afs_handler_;
};
#endif

T
Thunderbrook 已提交
83 84 85 86 87 88 89
class PSGPUWrapper {
 public:
  virtual ~PSGPUWrapper() { delete HeterPs_; }

  PSGPUWrapper() {
    HeterPs_ = NULL;
    sleep_seconds_before_fail_exit_ = 300;
T
Thunderbrook 已提交
90 91 92 93
    hbm_thread_pool_.resize(thread_keys_shard_num_);
    for (size_t i = 0; i < hbm_thread_pool_.size(); i++) {
      hbm_thread_pool_[i].reset(new ::ThreadPool(1));
    }
T
Thunderbrook 已提交
94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121
  }

  void PullSparse(const paddle::platform::Place& place, const int table_id,
                  const std::vector<const uint64_t*>& keys,
                  const std::vector<float*>& values,
                  const std::vector<int64_t>& slot_lengths,
                  const int hidden_size);
  void PushSparseGrad(const paddle::platform::Place& place, const int table_id,
                      const std::vector<const uint64_t*>& keys,
                      const std::vector<const float*>& grad_values,
                      const std::vector<int64_t>& slot_lengths,
                      const int hidden_size, const int batch_size);
  void CopyKeys(const paddle::platform::Place& place, uint64_t** origin_keys,
                uint64_t* total_keys, const int64_t* gpu_len, int slot_num,
                int total_len);
  void CopyForPull(const paddle::platform::Place& place, uint64_t** gpu_keys,
                   const std::vector<float*>& values,
                   const FeatureValue* total_values_gpu, const int64_t* gpu_len,
                   const int slot_num, const int hidden_size,
                   const int64_t total_length);

  void CopyForPush(const paddle::platform::Place& place,
                   const std::vector<const float*>& grad_values,
                   FeaturePushValue* total_grad_values_gpu,
                   const std::vector<int64_t>& slot_lengths,
                   const int hidden_size, const int64_t total_length,
                   const int batch_size);

122
  void BuildGPUTask(std::shared_ptr<HeterContext> gpu_task);
123 124
  void PreBuildTask(std::shared_ptr<HeterContext> gpu_task);
  void BuildPull(std::shared_ptr<HeterContext> gpu_task);
125 126 127 128
  void LoadIntoMemory(bool is_shuffle);
  void BeginPass();
  void EndPass();
  void start_build_thread();
129
  void pre_build_thread();
130
  void build_task();
131 132 133 134 135 136 137 138 139 140

  void Finalize() {
    VLOG(3) << "PSGPUWrapper Begin Finalize.";
    if (s_instance_ == nullptr) {
      return;
    }
    data_ready_channel_->Close();
    buildcpu_ready_channel_->Close();
    gpu_free_channel_->Close();
    running_ = false;
141 142
    VLOG(3) << "begin stop pre_build_threads_";
    pre_build_threads_.join();
143 144 145 146
    s_instance_ = nullptr;
    VLOG(3) << "PSGPUWrapper Finalize Finished.";
  }

T
Thunderbrook 已提交
147
  void InitializeGPU(const std::vector<int>& dev_ids) {
148
    if (s_instance_ != NULL && is_initialized_ == false) {
T
Thunderbrook 已提交
149
      VLOG(3) << "PSGPUWrapper Begin InitializeGPU";
150
      is_initialized_ = true;
T
Thunderbrook 已提交
151 152
      resource_ = std::make_shared<HeterPsResource>(dev_ids);
      resource_->enable_p2p();
153
      keys_tensor.resize(resource_->total_device());
Y
yaoxuefeng 已提交
154 155 156 157 158 159 160 161 162
#ifdef PADDLE_WITH_GLOO
      auto gloo = paddle::framework::GlooWrapper::GetInstance();
      if (gloo->Size() > 1) {
        multi_node_ = 1;
      }
#else
      PADDLE_THROW(
          platform::errors::Unavailable("heter ps need compile with GLOO"));
#endif
163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197
      if (multi_node_) {
        int dev_size = dev_ids.size();
        // init inner comm
        inner_comms_.resize(dev_size);
        inter_ncclids_.resize(dev_size);
        platform::dynload::ncclCommInitAll(&(inner_comms_[0]), dev_size,
                                           &dev_ids[0]);
// init inter comm
#ifdef PADDLE_WITH_GLOO
        inter_comms_.resize(dev_size);
        if (gloo->Rank() == 0) {
          for (int i = 0; i < dev_size; ++i) {
            platform::dynload::ncclGetUniqueId(&inter_ncclids_[i]);
          }
        }

        PADDLE_ENFORCE_EQ(
            gloo->IsInitialized(), true,
            platform::errors::PreconditionNotMet(
                "You must initialize the gloo environment first to use it."));
        gloo::BroadcastOptions opts(gloo->GetContext());
        opts.setOutput(&inter_ncclids_[0], dev_size);
        opts.setRoot(0);
        gloo::broadcast(opts);

        for (int i = 0; i < dev_size; ++i) {
          platform::dynload::ncclCommInitRank(&inter_comms_[i], gloo->Size(),
                                              inter_ncclids_[i], gloo->Rank());
        }
        node_size_ = gloo->Size();
#else
        PADDLE_THROW(
            platform::errors::Unavailable("heter ps need compile with GLOO"));
#endif
      }
Y
yaoxuefeng 已提交
198
      heter_devices_ = dev_ids;
199 200 201 202 203 204 205 206 207 208 209
      data_ready_channel_->Open();
      data_ready_channel_->SetCapacity(3);
      buildcpu_ready_channel_->Open();
      buildcpu_ready_channel_->SetCapacity(3);
      gpu_free_channel_->Open();
      gpu_free_channel_->SetCapacity(1);

      current_task_ = nullptr;
      gpu_free_channel_->Put(current_task_);

      table_id_ = 0;
210

211 212
      // start build cpu&gpu ps thread
      start_build_thread();
T
Thunderbrook 已提交
213 214
    }
  }
Y
yaoxuefeng 已提交
215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264

  void SetSparseSGD(float nonclk_coeff, float clk_coeff, float min_bound,
                    float max_bound, float learning_rate, float initial_g2sum,
                    float initial_range);
  void SetEmbedxSGD(float mf_create_thresholds, float mf_learning_rate,
                    float mf_initial_g2sum, float mf_initial_range,
                    float mf_min_bound, float mf_max_bound);
  void InitializeGPUServer(std::unordered_map<std::string, float> config) {
    float nonclk_coeff = (config.find("nonclk_coeff") == config.end())
                             ? 1.0
                             : config["nonclk_coeff"];
    float clk_coeff =
        (config.find("clk_coeff") == config.end()) ? 1.0 : config["clk_coeff"];
    float min_bound = (config.find("min_bound") == config.end())
                          ? -10000.0
                          : config["min_bound"];
    float max_bound = (config.find("max_bound") == config.end())
                          ? 10000.0
                          : config["max_bound"];
    float learning_rate = (config.find("learning_rate") == config.end())
                              ? 1.0
                              : config["learning_rate"];
    float initial_g2sum = (config.find("initial_g2sum") == config.end())
                              ? 1.0
                              : config["initial_g2sum"];
    float initial_range = (config.find("initial_range") == config.end())
                              ? 1.0
                              : config["initial_range"];

    // mf config settings
    float mf_create_thresholds =
        (config.find("mf_create_thresholds") == config.end())
            ? static_cast<float>(1.0)
            : config["mf_create_thresholds"];
    float mf_learning_rate = (config.find("mf_learning_rate") == config.end())
                                 ? 1.0
                                 : config["mf_learning_rate"];
    float mf_initial_g2sum = (config.find("mf_initial_g2sum") == config.end())
                                 ? 1.0
                                 : config["mf_initial_g2sum"];
    float mf_initial_range = (config.find("mf_initial_range") == config.end())
                                 ? 1.0
                                 : config["mf_initial_range"];
    float mf_min_bound = (config.find("mf_min_bound") == config.end())
                             ? 1.0
                             : config["mf_min_bound"];
    float mf_max_bound = (config.find("mf_max_bound") == config.end())
                             ? 1.0
                             : config["mf_max_bound"];
    for (size_t i = 0; i < heter_devices_.size(); i++) {
265
      PADDLE_ENFORCE_GPU_SUCCESS(cudaSetDevice(heter_devices_[i]));
Y
yaoxuefeng 已提交
266 267 268 269 270 271 272
      this->SetSparseSGD(nonclk_coeff, clk_coeff, min_bound, max_bound,
                         learning_rate, initial_g2sum, initial_range);
      this->SetEmbedxSGD(mf_create_thresholds, mf_learning_rate,
                         mf_initial_g2sum, mf_initial_range, mf_min_bound,
                         mf_max_bound);
    }
  }
273 274 275 276 277 278
  void SetDate(int year, int month, int day) {
    year_ = year;
    month_ = month;
    day_ = day;
  }

Y
yaoxuefeng 已提交
279 280
  void SetDataset(Dataset* dataset) { dataset_ = dataset; }

T
Thunderbrook 已提交
281 282 283 284 285 286 287 288 289 290 291 292 293 294 295
  // PSGPUWrapper singleton
  static std::shared_ptr<PSGPUWrapper> GetInstance() {
    if (NULL == s_instance_) {
      s_instance_.reset(new paddle::framework::PSGPUWrapper());
    }
    return s_instance_;
  }
  std::vector<std::unordered_map<uint64_t, std::vector<float>>>& GetLocalTable(
      int table_id) {
    return local_tables_[table_id];
  }
  void SetSlotVector(const std::vector<int>& slot_vector) {
    slot_vector_ = slot_vector;
  }

Y
yaoxuefeng 已提交
296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318
  void SetSlotOffsetVector(const std::vector<int>& slot_offset_vector) {
    slot_offset_vector_ = slot_offset_vector;
  }

  void SetSlotDimVector(const std::vector<int>& slot_mf_dim_vector) {
    slot_mf_dim_vector_ = slot_mf_dim_vector;
    assert(slot_mf_dim_vector_.size() == slot_vector_.size());
    for (size_t i = 0; i < slot_mf_dim_vector.size(); i++) {
      slot_dim_map_[slot_vector_[i]] = slot_mf_dim_vector_[i];
    }

    std::unordered_set<int> dims_set;
    for (auto& it : slot_dim_map_) {
      dims_set.insert(it.second);
    }
    size_t num_of_dim = dims_set.size();
    index_dim_vec_.resize(num_of_dim);
    index_dim_vec_.assign(dims_set.begin(), dims_set.end());
    std::sort(index_dim_vec_.begin(), index_dim_vec_.end());
    std::unordered_map<int, int> dim_index_map;
    for (size_t i = 0; i < num_of_dim; i++) {
      dim_index_map[index_dim_vec_[i]] = i;
    }
319 320
    hbm_pools_.resize(resource_->total_device() * num_of_dim);
    mem_pools_.resize(resource_->total_device() * num_of_dim);
Y
yaoxuefeng 已提交
321 322 323 324 325 326 327 328 329 330 331 332 333
    max_mf_dim_ = index_dim_vec_.back();
    multi_mf_dim_ = (dim_index_map.size() >= 1) ? dim_index_map.size() : 0;
    resource_->set_multi_mf(multi_mf_dim_, max_mf_dim_);
    slot_index_vec_.resize(slot_mf_dim_vector_.size());
    for (size_t i = 0; i < slot_index_vec_.size(); i++) {
      slot_index_vec_[i] = dim_index_map[slot_mf_dim_vector_[i]];
    }
    val_type_size_ =
        TYPEALIGN(8, sizeof(FeatureValue) + sizeof(float) * (max_mf_dim_ + 1));
    grad_type_size_ =
        TYPEALIGN(8, sizeof(FeaturePushValue) + (max_mf_dim_ * sizeof(float)));
  }

T
Thunderbrook 已提交
334 335
  void ShowOneTable(int index) { HeterPs_->show_one_table(index); }

T
Thunderbrook 已提交
336 337 338 339 340 341 342 343 344 345 346 347
  int UseAfsApi() { return use_afs_api_; }

#ifdef PADDLE_WITH_PSLIB
  std::shared_ptr<paddle::ps::AfsReader> OpenReader(
      const std::string& filename) {
    return afs_handler_.open_reader(filename);
  }

  void InitAfsApi(const std::string& fs_name, const std::string& fs_user,
                  const std::string& pass_wd, const std::string& conf);
#endif

T
Thunderbrook 已提交
348 349
 private:
  static std::shared_ptr<PSGPUWrapper> s_instance_;
Y
yaoxuefeng 已提交
350
  Dataset* dataset_;
T
Thunderbrook 已提交
351 352 353
#ifdef PADDLE_WITH_PSLIB
  paddle::ps::AfsApiWrapper afs_handler_;
#endif
T
Thunderbrook 已提交
354 355 356 357 358 359 360 361
  std::unordered_map<
      uint64_t, std::vector<std::unordered_map<uint64_t, std::vector<float>>>>
      local_tables_;
  HeterPsBase* HeterPs_;
  std::vector<LoDTensor> keys_tensor;  // Cache for pull_sparse
  std::shared_ptr<HeterPsResource> resource_;
  int32_t sleep_seconds_before_fail_exit_;
  std::vector<int> slot_vector_;
Y
yaoxuefeng 已提交
362 363 364 365 366 367 368 369 370
  std::vector<int> slot_offset_vector_;
  std::vector<int> slot_mf_dim_vector_;
  std::unordered_map<int, int> slot_dim_map_;
  std::vector<int> slot_index_vec_;
  std::vector<int> index_dim_vec_;
  int multi_mf_dim_{0};
  int max_mf_dim_{0};
  size_t val_type_size_{0};
  size_t grad_type_size_{0};
T
Thunderbrook 已提交
371
  int multi_node_{0};
372
  int node_size_;
373
  uint64_t table_id_;
374 375 376
  std::vector<ncclComm_t> inner_comms_;
  std::vector<ncclComm_t> inter_comms_;
  std::vector<ncclUniqueId> inter_ncclids_;
Y
yaoxuefeng 已提交
377 378 379
  std::vector<int> heter_devices_;
  std::unordered_set<std::string> gpu_ps_config_keys_;
  HeterObjectPool<HeterContext> gpu_task_pool_;
380
  std::vector<std::vector<robin_hood::unordered_set<uint64_t>>> thread_keys_;
381 382
  std::vector<std::vector<std::vector<robin_hood::unordered_set<uint64_t>>>>
      thread_dim_keys_;
Y
yaoxuefeng 已提交
383 384 385
  int thread_keys_thread_num_ = 37;
  int thread_keys_shard_num_ = 37;
  uint64_t max_fea_num_per_pass_ = 5000000000;
386 387 388
  int year_;
  int month_;
  int day_;
T
Thunderbrook 已提交
389
  int use_afs_api_ = 0;
T
Thunderbrook 已提交
390

Y
yaoxuefeng 已提交
391 392 393 394
  std::vector<MemoryPool*> mem_pools_;
  std::vector<HBMMemoryPool*> hbm_pools_;  // in multi mfdim, one table need hbm
                                           // pools of totol dims number

395 396 397 398 399 400 401 402 403 404 405 406 407
  std::shared_ptr<
      paddle::framework::ChannelObject<std::shared_ptr<HeterContext>>>
      data_ready_channel_ =
          paddle::framework::MakeChannel<std::shared_ptr<HeterContext>>();
  std::shared_ptr<
      paddle::framework::ChannelObject<std::shared_ptr<HeterContext>>>
      buildcpu_ready_channel_ =
          paddle::framework::MakeChannel<std::shared_ptr<HeterContext>>();
  std::shared_ptr<
      paddle::framework::ChannelObject<std::shared_ptr<HeterContext>>>
      gpu_free_channel_ =
          paddle::framework::MakeChannel<std::shared_ptr<HeterContext>>();
  std::shared_ptr<HeterContext> current_task_ = nullptr;
408
  std::thread pre_build_threads_;
409
  bool running_ = false;
T
Thunderbrook 已提交
410
  std::vector<std::shared_ptr<ThreadPool>> hbm_thread_pool_;
411

T
Thunderbrook 已提交
412 413 414 415 416 417 418
 protected:
  static bool is_initialized_;
};

}  // end namespace framework
}  // end namespace paddle
#endif