ps_gpu_worker.cc 9.9 KB
Newer Older
T
Thunderbrook 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
/* 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. */

#include "paddle/fluid/framework/device_worker.h"
#include "paddle/fluid/framework/device_worker_factory.h"
#include "paddle/fluid/platform/cpu_helper.h"
18
#include "paddle/fluid/platform/lodtensor_printer.h"
T
Thunderbrook 已提交
19 20
#include "paddle/fluid/string/string_helper.h"

21 22
#if (defined PADDLE_WITH_NCCL || defined PADDLE_WITH_RCCL) && \
    (defined PADDLE_WITH_PSLIB)
F
Fan Zhang 已提交
23
#ifdef PADDLE_WITH_CUDA
T
Thunderbrook 已提交
24
#include "paddle/fluid/platform/cuda_device_guard.h"
F
Fan Zhang 已提交
25
#endif
T
Thunderbrook 已提交
26 27 28 29 30 31 32 33 34 35 36

#if defined _WIN32 || defined __APPLE__
#else
#define _LINUX
#endif

namespace paddle {
namespace framework {

void PSGPUWorker::Initialize(const TrainerDesc& desc) {
  param_ = desc.downpour_param();
37
  dev_ctx_ = platform::DeviceContextPool::Instance().Get(place_);
T
Thunderbrook 已提交
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123
  mpi_rank_ = desc.mpi_rank();
  trainer_desc_ = desc;
  for (int i = 0; i < param_.sparse_table_size(); ++i) {
    uint64_t table_id =
        static_cast<uint64_t>(param_.sparse_table(i).table_id());
    TableParameter table = param_.sparse_table(i);
    sparse_key_names_[table_id].resize(table.sparse_key_name_size());
    for (int j = 0; j < table.sparse_key_name_size(); ++j) {
      sparse_key_names_[table_id][j] = table.sparse_key_name(j);
    }
    sparse_value_names_[table_id].resize(table.sparse_value_name_size());
    for (int j = 0; j < table.sparse_value_name_size(); ++j) {
      sparse_value_names_[table_id][j] = table.sparse_value_name(j);
    }
    sparse_grad_names_[table_id].resize(table.sparse_grad_name_size());
    for (int j = 0; j < table.sparse_grad_name_size(); ++j) {
      sparse_grad_names_[table_id][j] = table.sparse_grad_name(j);
    }
    label_var_name_[table_id] = table.label_var_name();
    sparse_push_keys_[table_id] = std::vector<uint64_t>();
  }

  for (int i = 0; i < param_.dense_table_size(); ++i) {
    uint64_t table_id = static_cast<uint64_t>(param_.dense_table(i).table_id());
    auto table = param_.dense_table(i);
    dense_value_names_[table_id].resize(table.dense_value_name_size());
    for (int j = 0; j < table.dense_value_name_size(); ++j) {
      dense_value_names_[table_id][j] = table.dense_value_name(j);
    }
    dense_grad_names_[table_id].resize(table.dense_grad_name_size());
    for (int j = 0; j < table.dense_grad_name_size(); ++j) {
      dense_grad_names_[table_id][j] = table.dense_grad_name(j);
    }
  }

  skip_ops_.resize(param_.skip_ops_size());
  for (int i = 0; i < param_.skip_ops_size(); ++i) {
    skip_ops_[i] = param_.skip_ops(i);
  }
  for (int i = 0; i < param_.stat_var_names_size(); ++i) {
    stat_var_name_map_[param_.stat_var_names(i)] = 1;
  }

  need_to_push_sparse_ = param_.push_sparse();
  need_to_push_dense_ = param_.push_dense();

  fetch_config_ = desc.fetch_config();
  use_cvm_ = desc.use_cvm();
  // for sparse value accessor, embedding only
  no_cvm_ = desc.no_cvm();
  scale_datanorm_ = desc.scale_datanorm();
  dump_slot_ = desc.dump_slot();
  adjust_ins_weight_config_ = desc.adjust_ins_weight_config();
  for (int i = 0; i < desc.check_nan_var_names_size(); ++i) {
    check_nan_var_names_.push_back(desc.check_nan_var_names(i));
  }
  copy_table_config_ = desc.copy_table_config();
  for (int i = 0; i < copy_table_config_.src_sparse_tables_size(); ++i) {
    uint64_t src_table = copy_table_config_.src_sparse_tables(i);
    uint64_t dest_table = copy_table_config_.dest_sparse_tables(i);
    VLOG(3) << "copy_sparse_tables_ push back " << src_table << "->"
            << dest_table;
    copy_sparse_tables_.push_back(std::make_pair(src_table, dest_table));
  }
  for (int i = 0; i < copy_table_config_.src_dense_tables_size(); ++i) {
    uint64_t src_table = copy_table_config_.src_dense_tables(i);
    uint64_t dest_table = copy_table_config_.dest_dense_tables(i);
    VLOG(3) << "copy_dense_tables_ push back " << src_table << "->"
            << dest_table;
    copy_dense_tables_.push_back(std::make_pair(src_table, dest_table));
  }
  for (auto& m : copy_table_config_.table_denpendency_map()) {
    if (sparse_key_names_.find(m.key()) != sparse_key_names_.end()) {
      // currently only support one dependency
      for (auto& value : m.values()) {
        table_dependency_[m.key()] = value;
      }
    }
  }
}

void PSGPUWorker::SetChannelWriter(ChannelObject<std::string>* queue) {
  writer_.Reset(queue);
}

void PSGPUWorker::TrainFiles() {
124
  VLOG(0) << "Begin to train files";
T
Thunderbrook 已提交
125
  platform::SetNumThreads(1);
126 127 128 129
  platform::Timer timeline;
  timeline.Start();

  int total_ins_num = 0;
T
Thunderbrook 已提交
130 131 132 133

  // how to accumulate fetched values here
  device_reader_->Start();
  int cur_batch;
T
Thunderbrook 已提交
134
  int batch_cnt = 0;
135 136

  platform::SetDeviceId(thread_id_);
T
Thunderbrook 已提交
137
  while ((cur_batch = device_reader_->Next()) > 0) {
138
    total_ins_num += cur_batch;
T
Thunderbrook 已提交
139 140 141 142 143 144 145 146 147 148 149 150
    for (auto& op : ops_) {
      bool need_skip = false;
      for (auto t = 0u; t < skip_ops_.size(); ++t) {
        if (op->Type().find(skip_ops_[t]) != std::string::npos) {
          need_skip = true;
          break;
        }
      }
      if (!need_skip) {
        op->Run(*thread_scope_, place_);
      }
    }
T
Thunderbrook 已提交
151 152 153 154 155 156
    if (need_dump_field_) {
      DumpField(*thread_scope_, dump_mode_, dump_interval_);
    }
    if (need_dump_param_ && thread_id_ == 0) {
      DumpParam(*thread_scope_, batch_cnt);
    }
T
Thunderbrook 已提交
157

158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189
    for (std::string& var_name : check_nan_var_names_) {
      Variable* var = thread_scope_->FindVar(var_name);
      if (var == nullptr) {
        continue;
      }
      LoDTensor* tensor = var->GetMutable<LoDTensor>();
      if (tensor == nullptr || !tensor->IsInitialized()) {
        continue;
      }
      if (framework::TensorContainsInf(*tensor) ||
          framework::TensorContainsNAN(*tensor)) {
        static std::mutex mutex;
        {
          std::lock_guard<std::mutex> lock(mutex);
          VLOG(0) << "worker " << thread_id_ << ": " << var_name
                  << " cantains inf or nan";
          auto all_vars = thread_scope_->LocalVarNames();
          std::stringstream ss;
          ss << "====== worker " << thread_id_ << "======\n";
          for (auto& local_var : all_vars) {
            platform::PrintVar(thread_scope_, local_var, local_var, &ss);
            ss << "\n";
          }
          std::cout << ss.str() << std::endl;
          VLOG(0) << "worker " << thread_id_ << "print nan var done....";
        }
        sleep(600);
        exit(-1);
      }
    }

    dev_ctx_->Wait();
T
Thunderbrook 已提交
190 191
    PrintFetchVars();
    thread_scope_->DropKids();
T
Thunderbrook 已提交
192 193 194 195
    ++batch_cnt;
  }
  if (need_dump_field_ || need_dump_param_) {
    writer_.Flush();
T
Thunderbrook 已提交
196
  }
197
  timeline.Pause();
198
  VLOG(0) << "GpuPs worker " << thread_id_ << " train cost "
199
          << timeline.ElapsedSec() << " seconds, ins_num: " << total_ins_num;
T
Thunderbrook 已提交
200 201 202
  return;
}

203 204
void PSGPUWorker::TrainFilesWithProfiler() {
  platform::SetNumThreads(1);
205
  VLOG(0) << "Begin to train files with profiler";
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
  device_reader_->Start();
  std::vector<double> op_total_time;
  std::vector<std::string> op_name;
  for (auto& op : ops_) {
    bool need_skip = false;
    for (auto t = 0u; t < skip_ops_.size(); ++t) {
      if (op->Type().find(skip_ops_[t]) != std::string::npos) {
        need_skip = true;
        break;
      }
    }
    if (!need_skip) {
      op_name.push_back(op->Type());
    }
  }

  VLOG(3) << "op name size: " << op_name.size();
  op_total_time.resize(op_name.size());
  for (size_t i = 0; i < op_total_time.size(); ++i) {
    op_total_time[i] = 0.0;
  }
  platform::Timer timeline;
  double total_time = 0.0;
  double read_time = 0.0;
  int total_ins_num = 0;
  int cur_batch;
  timeline.Start();
233
  platform::SetDeviceId(thread_id_);
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
  while ((cur_batch = device_reader_->Next()) > 0) {
    total_ins_num += cur_batch;
    timeline.Pause();
    read_time += timeline.ElapsedSec();
    total_time += timeline.ElapsedSec();

    int run_op_idx = 0;
    dev_ctx_->Wait();
    for (auto& op : ops_) {
      bool need_skip = false;
      for (auto t = 0u; t < skip_ops_.size(); ++t) {
        if (op->Type().find(skip_ops_[t]) != std::string::npos) {
          need_skip = true;
          break;
        }
      }
      if (!need_skip) {
        timeline.Start();
        VLOG(3) << "Going to run op " << op_name[run_op_idx];
        op->Run(*thread_scope_, place_);
        dev_ctx_->Wait();
        VLOG(3) << "Op " << op_name[run_op_idx] << " Finished";
        timeline.Pause();
        op_total_time[run_op_idx++] += timeline.ElapsedSec();
        total_time += timeline.ElapsedSec();
      }
    }
    timeline.Start();
    PrintFetchVars();
    thread_scope_->DropKids();
    dev_ctx_->Wait();
    timeline.Pause();
    total_time += timeline.ElapsedSec();
    timeline.Start();
  }
269
  VLOG(0) << "GpuPs worker " << thread_id_ << " train cost " << total_time
270 271
          << " seconds, ins_num: " << total_ins_num;
  for (size_t i = 0; i < op_name.size(); ++i) {
272
    VLOG(0) << "card:" << thread_id_ << ", op: " << op_name[i]
273 274 275
            << ", mean time: " << op_total_time[i] / total_ins_num
            << "s, totol time:" << op_total_time[i] << "sec";
  }
276 277
  VLOG(0) << "card: " << thread_id_ << " read time: " << read_time
          << ", percent: " << read_time / total_time * 100;
278 279 280
  return;
}

T
Thunderbrook 已提交
281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302
void PSGPUWorker::ResetStat() {
  total_time_ = 0;
  read_time_ = 0;
  pack_time_ = 0;
  pull_sparse_local_time_ = 0;
  op_all_time_ = 0;
  xpu_op_time_ = 0;
  xpu_wait_time_ = 0;
  cpu_op_time_ = 0;
  collect_label_time_ = 0;
  fill_sparse_time_ = 0;
  push_sparse_time_ = 0;
  gpu_2_cpu_time_ = 0;
  cpu_2_gpu_time_ = 0;
  total_inst_ = 0;
}

void PSGPUWorker::ProduceTasks() { return; }

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