request_handler_impl.cc 12.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// 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.

W
Wang Guibao 已提交
15
#include "paddle/fluid/operators/distributed/request_handler_impl.h"
16 17 18 19 20 21 22 23
#include <iostream>
#include <string>
#include <vector>

#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows.h"
W
Wang Guibao 已提交
24
#include "paddle/fluid/framework/variable_helper.h"
25
#include "paddle/fluid/operators/distributed/rpc_server.h"
26
#include "paddle/fluid/string/piece.h"
T
tangwei12 已提交
27
#include "paddle/fluid/string/printf.h"
28
#include "paddle/fluid/string/split.h"
29

30 31 32
#include "paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h"
#include "paddle/fluid/operators/distributed/heart_beat_monitor.h"

33 34
namespace paddle {
namespace operators {
35
namespace distributed {
36

T
tangwei12 已提交
37 38
// define LOOKUP_TABLE_PATH for checkpoint notify to save lookup table variables
// to directory specified.
T
tangwei12 已提交
39
constexpr char LOOKUP_TABLE_PATH[] = "kLookupTablePath";
T
tangwei12 已提交
40

41 42 43
bool RequestSendHandler::Handle(const std::string& varname,
                                framework::Scope* scope,
                                framework::Variable* invar,
Q
qiaolongfei 已提交
44
                                framework::Variable** outvar,
W
Wu Yi 已提交
45
                                const int trainer_id,
Q
Qiao Longfei 已提交
46 47
                                const std::string& out_var_name,
                                const std::string& table_name) {
M
minqiyang 已提交
48
  VLOG(4) << "RequestSendHandler:" << varname;
49 50 51

  // Sync
  if (varname == BATCH_BARRIER_MESSAGE) {
M
minqiyang 已提交
52
    VLOG(3) << "sync: recv BATCH_BARRIER_MESSAGE";
53
    rpc_server_->IncreaseBatchBarrier(kRequestSend);
Y
Yancey1989 已提交
54
  } else if (varname == COMPLETE_MESSAGE) {
M
minqiyang 已提交
55
    VLOG(3) << "sync: recv complete message";
56 57 58 59 60

    if (HeartBeatMonitor::GetInstance() != nullptr) {
      HeartBeatMonitor::GetInstance()->Update(trainer_id, "", COMPLETED);
    }

Y
Yancey1989 已提交
61
    rpc_server_->Complete();
62
  } else {
63
    // Async
1
123malin 已提交
64
    if (distributed_mode_ != DistributedMode::kSync) {
M
minqiyang 已提交
65
      VLOG(3) << "async process var: " << varname;
Q
Qiao Longfei 已提交
66
      if (varname == BATCH_BARRIER_MESSAGE) {
Q
Qiao Longfei 已提交
67 68 69 70
        PADDLE_THROW(
            "async mode should not recv BATCH_BARRIER_MESSAGE or "
            "COMPLETE_MESSAGE");
      }
71
      HeartBeatMonitor::GetInstance()->Update(trainer_id, varname, RUNNING);
72 73 74 75 76 77 78 79 80 81 82 83 84

      std::string run_varname = varname;

      string::Piece part_piece("@PIECE");
      string::Piece var_name_piece = string::Piece(varname);

      if (string::Contains(var_name_piece, part_piece)) {
        auto varname_splits = paddle::string::Split(varname, '@');
        PADDLE_ENFORCE_EQ(varname_splits.size(), 3);
        run_varname = varname_splits[0];
        scope->Rename(varname, run_varname);
      }

1
123malin 已提交
85 86
      if (distributed_mode_ == DistributedMode::kGeo &&
          AsyncSparseParamUpdateRecorder::GetInstance()->HasGrad(run_varname)) {
Q
Qiao Longfei 已提交
87
        auto& grad_slr =
88 89
            scope->FindVar(run_varname)->Get<framework::SelectedRows>();
        AsyncSparseParamUpdateRecorder::GetInstance()->Update(run_varname,
Q
Qiao Longfei 已提交
90 91
                                                              grad_slr.rows());
      }
92
      executor_->RunPreparedContext((*grad_to_prepared_ctx_)[run_varname].get(),
Q
Qiao Longfei 已提交
93
                                    scope);
94

95 96 97
      return true;
    } else {  // sync
      rpc_server_->WaitCond(kRequestSend);
M
minqiyang 已提交
98
      VLOG(3) << "sync: processing received var: " << varname;
99 100 101
      PADDLE_ENFORCE_NOT_NULL(
          invar, platform::errors::NotFound(
                     "sync: Can not find server side var %s.", varname));
Y
Yancey1989 已提交
102
    }
103 104 105 106 107 108 109
  }
  return true;
}

bool RequestGetHandler::Handle(const std::string& varname,
                               framework::Scope* scope,
                               framework::Variable* invar,
Q
qiaolongfei 已提交
110
                               framework::Variable** outvar,
W
Wu Yi 已提交
111
                               const int trainer_id,
Q
Qiao Longfei 已提交
112 113
                               const std::string& out_var_name,
                               const std::string& table_name) {
Q
Qiao Longfei 已提交
114 115 116
  VLOG(3) << "RequestGetHandler:" << varname
          << " out_var_name: " << out_var_name << " trainer_id: " << trainer_id
          << " table_name: " << table_name;
117

1
123malin 已提交
118
  if (distributed_mode_ == DistributedMode::kSync) {
Y
Yancey1989 已提交
119
    if (varname == FETCH_BARRIER_MESSAGE) {
M
minqiyang 已提交
120
      VLOG(3) << "sync: recv fetch barrier message";
Y
Yancey1989 已提交
121 122
      rpc_server_->IncreaseBatchBarrier(kRequestGet);
    } else {
123
      rpc_server_->WaitCond(kRequestGet);
Y
Yancey1989 已提交
124 125 126
      *outvar = scope_->FindVar(varname);
    }
  } else {
Y
Yancey1989 已提交
127
    if (varname != FETCH_BARRIER_MESSAGE && varname != COMPLETE_MESSAGE) {
W
Wu Yi 已提交
128
      if (enable_dc_asgd_) {
T
tangwei12 已提交
129
        // NOTE: the format is determined by distribute_transpiler.py
W
Wu Yi 已提交
130 131
        std::string param_bak_name =
            string::Sprintf("%s.trainer_%d_bak", varname, trainer_id);
M
minqiyang 已提交
132
        VLOG(3) << "getting " << param_bak_name << " trainer_id " << trainer_id;
W
Wu Yi 已提交
133 134 135 136 137
        auto var = scope_->FindVar(varname);
        auto t_orig = var->Get<framework::LoDTensor>();
        auto param_bak = scope_->Var(param_bak_name);
        auto t = param_bak->GetMutable<framework::LoDTensor>();
        t->mutable_data(dev_ctx_->GetPlace(), t_orig.type());
M
minqiyang 已提交
138
        VLOG(3) << "copying " << varname << " to " << param_bak_name;
W
Wu Yi 已提交
139 140
        framework::TensorCopy(t_orig, dev_ctx_->GetPlace(), t);
      }
141
      VLOG(1) << "Table name empty? " << table_name.empty();
1
123malin 已提交
142 143 144 145 146 147 148
      if (distributed_mode_ == DistributedMode::kGeo) {
        VLOG(1) << "AsyncSparseParamUpdateRecorder " << varname << " exist "
                << AsyncSparseParamUpdateRecorder::GetInstance()->HasParam(
                       varname);
      }
      if (distributed_mode_ == DistributedMode::kGeo &&
          AsyncSparseParamUpdateRecorder::GetInstance()->HasParam(varname) &&
149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
          !table_name.empty()) {
        std::vector<int64_t> updated_rows;
        AsyncSparseParamUpdateRecorder::GetInstance()->GetAndClear(
            varname, trainer_id, &updated_rows);
        if (VLOG_IS_ON(3)) {
          std::ostringstream sstream;
          sstream << "[";
          for (auto& row_id : updated_rows) {
            sstream << row_id << ", ";
          }
          sstream << "]";
          VLOG(3) << "updated_rows size: " << updated_rows.size() << " "
                  << sstream.str();
        }
        auto& origin_tensor =
            scope_->FindVar(varname)->Get<framework::LoDTensor>();
        auto* origin_tensor_data = origin_tensor.data<float>();
        auto& dims = origin_tensor.dims();
        *outvar = scope->Var();
        auto* out_slr = (*outvar)->GetMutable<framework::SelectedRows>();
        out_slr->set_rows(updated_rows);
        out_slr->set_height(dims[0]);
        auto out_dims = framework::make_ddim(
            {static_cast<int64_t>(updated_rows.size()), dims[1]});
        auto* data = out_slr->mutable_value()->mutable_data<float>(
            out_dims, origin_tensor.place());
        auto width = dims[1];
176
        for (size_t i = 0; i < updated_rows.size(); ++i) {
177 178 179 180 181 182 183
          PADDLE_ENFORCE_LT(updated_rows[i], dims[0]);
          memcpy(data + i * width, origin_tensor_data + updated_rows[i] * width,
                 sizeof(float) * width);
        }
      } else {
        *outvar = scope_->FindVar(varname);
      }
184 185 186 187 188
    }
  }
  return true;
}

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
bool RequestGetNoBarrierHandler::Handle(const std::string& varname,
                                        framework::Scope* scope,
                                        framework::Variable* invar,
                                        framework::Variable** outvar,
                                        const int trainer_id,
                                        const std::string& out_var_name,
                                        const std::string& table_name) {
  VLOG(4) << "RequestGetNoBarrierHandler:" << varname
          << " out_var_name: " << out_var_name;

  // get var from pserver immediately without barriers
  string::Piece without_barrier_piece(WITHOUT_BARRIER_MESSAGE);
  string::Piece var_name_piece = string::Piece(varname);

  if (string::Contains(var_name_piece, without_barrier_piece)) {
    var_name_piece = string::TrimSuffix(var_name_piece, without_barrier_piece);
    VLOG(4) << "Get var " << var_name_piece << " with "
            << WITHOUT_BARRIER_MESSAGE;
    *outvar = scope_->FindVar(var_name_piece.ToString());
    return true;
  } else {
    PADDLE_THROW("GetNoBarrier must contain %s", WITHOUT_BARRIER_MESSAGE);
  }
  return true;
}

215 216 217
bool RequestPrefetchHandler::Handle(const std::string& varname,
                                    framework::Scope* scope,
                                    framework::Variable* invar,
Q
qiaolongfei 已提交
218
                                    framework::Variable** outvar,
W
Wu Yi 已提交
219
                                    const int trainer_id,
Q
Qiao Longfei 已提交
220 221
                                    const std::string& out_var_name,
                                    const std::string& table_name) {
M
minqiyang 已提交
222
  VLOG(4) << "RequestPrefetchHandler " << varname;
223

Q
Qiao Longfei 已提交
224
  if (table_name.empty()) {
Q
Qiao Longfei 已提交
225 226
    auto var_desc = program_->Block(0).FindVar(out_var_name);
    InitializeVariable(*outvar, var_desc->GetType());
Q
Qiao Longfei 已提交
227 228 229
    executor_->RunPreparedContext(
        (*prefetch_var_name_to_prepared_ctx_)[varname].get(), scope);
  } else {
Q
Qiao Longfei 已提交
230
    (*outvar)->GetMutable<framework::LoDTensor>();
Q
Qiao Longfei 已提交
231 232 233 234 235
    auto lookup_table_op =
        BuildLookupTableOp(table_name, varname, out_var_name);
    paddle::platform::CPUPlace cpu_place;
    lookup_table_op->Run(*scope, cpu_place);
  }
236 237 238
  return true;
}

T
tangwei12 已提交
239 240 241 242
bool RequestCheckpointHandler::Handle(const std::string& varname,
                                      framework::Scope* scope,
                                      framework::Variable* invar,
                                      framework::Variable** outvar,
W
Wu Yi 已提交
243
                                      const int trainer_id,
Q
Qiao Longfei 已提交
244 245
                                      const std::string& out_var_name,
                                      const std::string& table_name) {
246 247 248
  PADDLE_ENFORCE(
      checkpoint_notify_id != -1,
      "when checkpoint_notify_id = -1, there should be no RPC invoke.");
T
tangwei12 已提交
249

T
tangwei12 已提交
250
  // TODO(tangwei12): find out why scope will be error.
T
bug fix  
tangwei12 已提交
251
  auto* lt_var = scope_->FindVar(LOOKUP_TABLE_PATH)->GetMutable<std::string>();
T
tangwei12 已提交
252 253
  lt_var->clear();
  lt_var->append(out_var_name);
M
minqiyang 已提交
254 255
  VLOG(4) << "RequestCheckpointHandler update var kLookupTablePath to: "
          << out_var_name;
T
bug fix  
tangwei12 已提交
256
  executor_->RunPreparedContext(checkpoint_prepared_ctx_.get(), scope_);
T
bug fix  
tangwei12 已提交
257 258
  return true;
}
T
tangwei12 已提交
259

260 261 262 263 264 265 266
bool RequestNotifyHandler::Handle(const std::string& varname,
                                  framework::Scope* scope,
                                  framework::Variable* invar,
                                  framework::Variable** outvar,
                                  const int trainer_id,
                                  const std::string& out_var_name,
                                  const std::string& table_name) {
1
123malin 已提交
267 268 269 270 271 272 273
  VLOG(4) << "RequestNotifyHandler: " << varname;
  VLOG(3) << "async process var: " << varname << ", trainer_id: " << trainer_id;

  string::Piece decay_piece(LEARNING_RATE_DECAY_COUNTER);
  string::Piece var_name_piece = string::Piece(varname);
  if (string::Contains(var_name_piece, decay_piece)) {
    VLOG(3) << "LearningRate Decay Counter Update";
274 275 276
    PADDLE_ENFORCE_NE(
        lr_decay_block_id, -1,
        "when lr_decay_block_id = -1, there should be no RPC invoke.");
1
123malin 已提交
277 278 279 280 281 282 283 284 285
    auto* origin_var = scope_->FindVar(varname);
    auto origin_var_tensor = origin_var->Get<framework::LoDTensor>();
    auto* send_var = scope->FindVar(varname);
    auto send_var_tensor = send_var->Get<framework::LoDTensor>();
    int64_t* origin_value =
        origin_var_tensor.mutable_data<int64_t>(origin_var_tensor.place());
    int64_t* send_value =
        send_var_tensor.mutable_data<int64_t>(send_var_tensor.place());
    origin_value[0] += send_value[0];
286 287 288 289 290
    executor_->RunPreparedContext(lr_decay_prepared_ctx_.get(), scope_);
  }
  return true;
}

291
}  // namespace distributed
292 293
}  // namespace operators
}  // namespace paddle