pull_dense_worker.cc 4.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
/* 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 <time.h>
#include "paddle/fluid/framework/device_worker.h"

namespace paddle {
namespace framework {

std::shared_ptr<PullDenseWorker> PullDenseWorker::s_instance_ = NULL;

void PullDenseWorker::Initialize(const TrainerDesc& param) {
23
  LOG(WARNING) << "going to initialize pull dense worker";
24 25 26 27 28
  running_ = false;
  param_ = param.pull_dense_param();
  threshold_ = param_.threshold();
  thread_num_ = param_.device_num();
  sleep_time_ms_ = param_.sleep_time_ms();
29 30
  LOG(WARNING) << "dense table size: " << param_.dense_table_size();
  LOG(WARNING) << "thread num: " << thread_num_;
31 32 33
  for (size_t i = 0; i < param_.dense_table_size(); ++i) {
    // setup dense variables for each table
    int var_num = param_.dense_table(i).dense_value_name_size();
34
    LOG(WARNING) << "var num: " << var_num;
35 36 37 38
    uint64_t tid = static_cast<uint64_t>(param_.dense_table(i).table_id());
    dense_value_names_[tid].resize(var_num);
    for (int j = 0; j < var_num; ++j) {
      dense_value_names_[tid][j] = param_.dense_table(i).dense_value_name(j);
39 40
      LOG(WARNING) << "dense value names " << j << " "
                   << dense_value_names_[tid][j];
41 42 43 44 45 46
    }
    // setup training version for each table
    training_versions_[tid].resize(thread_num_, 0);
    last_versions_[tid] = 0;
    current_version_[tid] = 0;
  }
47
  LOG(WARNING) << "initialize pull dense worker done.";
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
}

void PullDenseWorker::Wait(std::vector<::std::future<int32_t>>* status_vec) {
  for (auto& t : *status_vec) {
    t.wait();
    auto status = t.get();
    if (status != 0) {
      LOG(WARNING) << "Current Pull Dense Thread Failed Times"
                   << ++pull_dense_fail_times_;
    }
  }

  int MAX_FAIL_NUM = 20;
  if (pull_dense_fail_times_ > MAX_FAIL_NUM) {
    LOG(FATAL) << "Pull Dense Failed Times More Than " << MAX_FAIL_NUM
               << " Times";
    exit(-1);
  }
66
  status_vec->resize(0);
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
}

void PullDenseWorker::Stop() {
  if (running_) {
    running_ = false;
    t_.join();
  }
}

int PullDenseWorker::Start() {
  running_ = true;
  t_ = std::thread(&PullDenseWorker::Run, this);
  return 0;
}

void PullDenseWorker::Run() {
  while (running_) {
    pull_dense_status_.resize(0);
    for (size_t i = 0; i < param_.dense_table_size(); ++i) {
      uint64_t tid = static_cast<uint64_t>(param_.dense_table(i).table_id());
      if (CheckUpdateParam(tid)) {
        fleet_ptr_->PullDenseVarsAsync(
            *root_scope_, tid, dense_value_names_[tid], &pull_dense_status_);
        ResetThreadVersion(tid);
      }
    }
    if (pull_dense_status_.size() != 0) {
      Wait(&pull_dense_status_);
    }
    usleep(sleep_time_ms_ * 1000);
  }
}

void PullDenseWorker::IncreaseThreadVersion(int thread_id, uint64_t table_id) {
101 102
  LOG(WARNING) << "increase thread version input: " << thread_id << " table id "
               << table_id;
103
  std::lock_guard<std::mutex> lock(mutex_for_version_);
104
  LOG(WARNING) << "going to increase";
105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
  training_versions_[table_id][thread_id]++;
}

bool PullDenseWorker::CheckUpdateParam(uint64_t table_id) {
  std::lock_guard<std::mutex> lock(mutex_for_version_);
  auto& version = training_versions_[table_id];
  current_version_[table_id] =
      *(std::min_element(version.begin(), version.end()));
  if (current_version_[table_id] - last_versions_[table_id] < threshold_) {
    return false;
  }
  return true;
}

void PullDenseWorker::ResetThreadVersion(uint64_t table_id) {
  std::lock_guard<std::mutex> lock(mutex_for_version_);
  last_versions_[table_id] = current_version_[table_id];
}

}  // namespace framework
}  // namespace paddle