communicator.h 4.9 KB
Newer Older
Q
Qiao Longfei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
/* 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. */

#pragma once

#include <deque>
#include <memory>
#include <string>
#include <vector>

Q
Qiao Longfei 已提交
22 23
#include <ThreadPool.h>

Q
Qiao Longfei 已提交
24 25 26 27 28 29 30 31 32 33 34 35 36 37 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
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/variable.h"
#include "paddle/fluid/operators/distributed/rpc_common.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/place.h"

namespace paddle {
namespace operators {
namespace distributed {

using Scope = framework::Scope;
using Variable = framework::Variable;

template <typename T>
class BlockingQueue {
 public:
  explicit BlockingQueue(size_t capacity) : capacity_(capacity) {
    PADDLE_ENFORCE_GT(capacity_, 0, "The capacity must be greater than 0.");
  }

  bool Push(const T& elem) {
    std::unique_lock<std::mutex> lock(mutex_);
    send_cv_.wait(lock, [&] { return queue_.size() < capacity_; });
    PADDLE_ENFORCE_LT(queue_.size(), capacity_);
    queue_.push_back(elem);
    recv_cv_.notify_one();
    return true;
  }

  bool Push(T&& elem) {
    std::unique_lock<std::mutex> lock(mutex_);
    send_cv_.wait(lock, [&] { return queue_.size() < capacity_; });
    PADDLE_ENFORCE_LT(queue_.size(), capacity_);
    queue_.emplace_back(std::move(elem));
    recv_cv_.notify_one();
    return true;
  }

  T Pop() {
    std::unique_lock<std::mutex> lock(mutex_);
    recv_cv_.wait(lock, [=] { return !queue_.empty(); });
    T rc(std::move(queue_.front()));
    queue_.pop_front();
    return rc;
  }

  size_t Cap() const {
    std::lock_guard<std::mutex> lock(mutex_);
    return capacity_;
  }

  size_t Size() const {
    std::lock_guard<std::mutex> lock(mutex_);
    return queue_.size();
  }

 private:
  const size_t capacity_;
  std::deque<T> queue_;

  mutable std::mutex mutex_;
  std::condition_variable recv_cv_;
  std::condition_variable send_cv_;
};

Q
Qiao Longfei 已提交
90 91
using RpcCtxMap = std::unordered_map<std::string, RpcContext>;

Q
Qiao Longfei 已提交
92 93
class Communicator {
 public:
Q
Qiao Longfei 已提交
94 95
  Communicator(const RpcCtxMap& send_varname_to_ctx,
               const RpcCtxMap& recv_varname_to_ctx, Scope* recv_scope)
Q
Qiao Longfei 已提交
96 97 98 99 100 101 102 103 104
      : send_varname_to_ctx_(send_varname_to_ctx),
        recv_varname_to_ctx_(recv_varname_to_ctx),
        recv_scope_(recv_scope) {
    // get all send information from graph, build vars_to_send
    send_scope_.reset(new Scope());
    for (auto& iter : send_varname_to_ctx_) {
      send_varname_to_queue_[iter.first] =
          std::make_shared<BlockingQueue<std::shared_ptr<Variable>>>(10);
    }
Q
Qiao Longfei 已提交
105 106 107
    // TODO(qiao): default 5, need to config
    send_threadpool_.reset(new ::ThreadPool(5));
    recv_threadpool_.reset(new ::ThreadPool(5));
Q
Qiao Longfei 已提交
108 109
  }

Q
Qiao Longfei 已提交
110 111 112 113 114 115 116
  ~Communicator() {
    VLOG(3) << "~Communicator";
    running_ = false;
    send_thread_->join();
    recv_thread_->join();
    VLOG(3) << "~Communicator done";
  }
Q
Qiao Longfei 已提交
117 118 119 120 121 122 123 124 125 126

  void Start();

  // send grad
  void Send(const std::string& var_name, const framework::Scope& scope);

 private:
  void SendThread();
  void RecvThread();

Q
Qiao Longfei 已提交
127
  bool running_ = false;
Q
Qiao Longfei 已提交
128 129 130
  std::unordered_map<std::string,
                     std::shared_ptr<BlockingQueue<std::shared_ptr<Variable>>>>
      send_varname_to_queue_;
Q
Qiao Longfei 已提交
131 132
  RpcCtxMap send_varname_to_ctx_;
  RpcCtxMap recv_varname_to_ctx_;
Q
Qiao Longfei 已提交
133 134 135 136
  std::unique_ptr<std::thread> send_thread_;
  std::unique_ptr<std::thread> recv_thread_;
  Scope* recv_scope_;                  // should be global scope
  std::unique_ptr<Scope> send_scope_;  // an independent scope
Q
Qiao Longfei 已提交
137 138
  std::unique_ptr<::ThreadPool> send_threadpool_{nullptr};
  std::unique_ptr<::ThreadPool> recv_threadpool_{nullptr};
Q
Qiao Longfei 已提交
139 140 141 142 143 144 145 146

  // the following code is for initialize the commnunicator
 public:
  static void Init(const RpcCtxMap& send_varname_to_ctx,
                   const RpcCtxMap& recv_varname_to_ctx, Scope* recv_scope) {
    InitImpl(send_varname_to_ctx, recv_varname_to_ctx, recv_scope);
  }

Q
can run  
Qiao Longfei 已提交
147
  static Communicator* GetInstance();
Q
Qiao Longfei 已提交
148 149 150 151 152 153 154 155 156 157 158 159 160 161 162

 private:
  // Init is called by GetInstance.
  static void InitImpl(const RpcCtxMap& send_varname_to_ctx,
                       const RpcCtxMap& recv_varname_to_ctx,
                       Scope* recv_scope) {
    if (communicator_ == nullptr) {
      communicator_.reset(new Communicator(send_varname_to_ctx,
                                           recv_varname_to_ctx, recv_scope));
    }
  }

 private:
  static std::once_flag init_flag_;
  static std::unique_ptr<Communicator> communicator_;
Q
Qiao Longfei 已提交
163 164 165 166 167
};

}  // namespace distributed
}  // namespace operators
}  // namespace paddle