grpc_client.cc 4.8 KB
Newer Older
G
gongweibao 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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 "grpc_client.h"
16
#include "paddle/framework/threadpool.h"
G
gongweibao 已提交
17 18 19 20 21 22 23 24 25
namespace paddle {
namespace operators {
namespace detail {

bool RPCClient::AsyncSendVariable(const std::string& ep,
                                  const platform::DeviceContext& ctx,
                                  const framework::Scope& scope,
                                  const std::string& var_name,
                                  int64_t time_out) {
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
  const platform::DeviceContext* p_ctx = &ctx;
  const std::string ep_val = ep;
  const std::string var_name_val = var_name;
  const framework::Scope* p_scope = &scope;
  const auto ch = GetChannel(ep_val);

  framework::Async([var_name_val, p_ctx, ep_val, p_scope, time_out, ch, this] {
    auto* var = p_scope->FindVar(var_name_val);
    sendrecv::VariableMessage req;
    SerializeToMessage(var_name_val, var, *p_ctx, &req);

    // varhandle
    VarHandle var_h;
    var_h.ep = ep_val;
    var_h.scope = p_scope;
    var_h.name = var_name_val;
    var_h.ctx = p_ctx;

    // stub context
    SendProcessor* s = new SendProcessor(ch);
    s->Prepare(var_h, time_out);
    s->response_call_back_ = NULL;

    auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
    rpc->Finish(&s->reply_, &s->status_, (void*)s);
  });
G
gongweibao 已提交
52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68

  req_count_++;

  return true;
}

void ProcGetResponse(const VarHandle& var_h,
                     const sendrecv::VariableMessage& ret_msg) {
  auto* outvar = var_h.scope->FindVar(var_h.name);
  DeserializeFromMessage(ret_msg, *var_h.ctx, outvar);
}

bool RPCClient::AsyncGetVariable(const std::string& ep,
                                 const platform::DeviceContext& ctx,
                                 const framework::Scope& scope,
                                 const std::string& var_name,
                                 int64_t time_out) {
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
  const platform::DeviceContext* p_ctx = &ctx;
  const std::string ep_val = ep;
  const std::string var_name_val = var_name;
  const framework::Scope* p_scope = &scope;
  const auto ch = GetChannel(ep_val);

  framework::Async([var_name_val, ep_val, p_scope, p_ctx, time_out, ch, this] {
    sendrecv::VariableMessage req;
    req.set_varname(var_name_val);

    // varhandle
    VarHandle var_h;
    var_h.ep = ep_val;
    var_h.scope = p_scope;
    var_h.name = var_name_val;
    var_h.ctx = p_ctx;

    // stub context
    GetProcessor* s = new GetProcessor(ch);
    s->Prepare(var_h, time_out);
    s->response_call_back_ = ProcGetResponse;

    auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_);
    rpc->Finish(&s->reply_, &s->status_, (void*)s);
  });
G
gongweibao 已提交
94 95 96 97 98 99

  req_count_++;

  return true;
}

T
typhoonzero 已提交
100
bool RPCClient::Wait() {
101 102 103
  if (req_count_ <= 0) {
    return true;
  }
G
gongweibao 已提交
104

105 106
  std::vector<bool> a(req_count_);
  std::vector<std::future<void>> waits(req_count_);
G
gongweibao 已提交
107

108 109 110 111 112 113 114 115 116 117 118 119 120
  for (int i = 0; i < req_count_; i++) {
    waits[i] = framework::Async([i, &a, this] { a[i] = Proceed(); });
  }

  for (int i = 0; i < req_count_; i++) {
    waits[i].wait();
  }

  int last_req_count = req_count_;
  req_count_ = 0;

  for (int i = 0; i < last_req_count; i++) {
    if (!a[i]) {
G
gongweibao 已提交
121 122 123 124
      return false;
    }
  }

125
  return true;
G
gongweibao 已提交
126 127 128 129 130 131 132 133
}

bool RPCClient::Proceed() {
  void* tag = NULL;
  bool ok = false;

  // request counts.
  if (!cq_.Next(&tag, &ok)) {
G
gongweibao 已提交
134
    LOG(ERROR) << "Get meets CompletionQueue error";
G
gongweibao 已提交
135 136 137 138 139 140 141 142 143
    return false;
  }

  GPR_ASSERT(ok);
  PADDLE_ENFORCE(tag);

  // TODO(gongwb): add more retries.
  ClientBase* c = static_cast<ClientBase*>(tag);
  if (!c->status_.ok()) {
G
gongweibao 已提交
144 145
    LOG(ERROR) << "proc param error:" << c->var_h_.String()
               << " grpc error:" << c->status_.error_message();
G
gongweibao 已提交
146
    delete c;
G
gongweibao 已提交
147
    return false;
G
gongweibao 已提交
148 149 150 151 152 153 154 155 156 157 158 159 160
  }

  c->Process();
  delete c;
  return true;
}

std::shared_ptr<grpc::Channel> RPCClient::GetChannel(const std::string& ep) {
  auto it = channels_.find(ep);
  if (it != channels_.end()) {
    return it->second;
  }

G
gongweibao 已提交
161 162 163 164
  grpc::ChannelArguments args;
  args.SetMaxSendMessageSize(std::numeric_limits<int>::max());
  args.SetMaxReceiveMessageSize(std::numeric_limits<int>::max());

G
gongweibao 已提交
165
  auto ch = std::shared_ptr<grpc::Channel>(
G
gongweibao 已提交
166
      grpc::CreateCustomChannel(ep, grpc::InsecureChannelCredentials(), args));
G
gongweibao 已提交
167 168 169 170 171 172 173 174

  channels_[ep] = ch;
  return ch;
}

}  // namespace detail
}  // namespace operators
}  // namespace paddle