/* Copyright (c) 2016 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 #include #include "glog/logging.h" // For VLOG #include "paddle/fluid/framework/threadpool.h" #include "paddle/fluid/operators/distributed/grpc_client.h" #include "paddle/fluid/operators/distributed/grpc_serde.h" #include "paddle/fluid/operators/distributed/request_handler.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { namespace distributed { void GRPCClient::InitImpl() { InitEventLoop(); } void GRPCClient::InitEventLoop() { // start the client process thread // TODO(wuyi): can make this in a threadpool client_thread_.reset(new std::thread(std::bind(&GRPCClient::Proceed, this))); } void GRPCClient::SendComplete() { std::unique_lock lk(completed_mutex_); if (!completed_) { for (auto& it : channels_) { VLOG(3) << "send complete message to " << it.first; this->AsyncSendComplete(it.first); } PADDLE_ENFORCE(this->Wait(), "internal grpc error"); completed_ = true; } } GRPCClient::~GRPCClient() { stopped_ = true; Wait(); cq_.Shutdown(); { std::lock_guard guard(chan_mutex_); for (auto& it : channels_) { it.second.reset(); } channels_.clear(); } client_thread_->join(); } VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, const platform::DeviceContext& ctx, const framework::Scope& scope, const std::string& var_name, int64_t time_out) { 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); SendProcessor* s = new SendProcessor(ch); const std::string method = "SendRPC"; VarHandlePtr h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope)); s->Prepare(h, time_out); framework::AsyncIO([var_name_val, p_scope, p_ctx, s, method, h, this] { auto* var = p_scope->FindVar(var_name_val); ::grpc::ByteBuffer req; SerializeToByteBuffer(var_name_val, var, *p_ctx, &req, "", trainer_id_); VLOG(3) << s->GetVarHandlePtr()->String() << " begin"; // stub context s->response_call_back_ = nullptr; platform::RecordRPCEvent record_event(method, p_ctx); auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); if (UNLIKELY(platform::IsProfileEnabled())) { h->Wait(); } }); req_count_++; return h; } void ProcGetResponse(const VarHandle& var_h, const ::grpc::ByteBuffer& ret_msg) { framework::Variable* outvar = nullptr; // get response's trainer_id is not used int trainer_id; DeserializeFromByteBuffer(ret_msg, *var_h.ctx(), var_h.scope(), &outvar, &trainer_id); } template void RequestToByteBuffer(const T& proto, ::grpc::ByteBuffer* result) { ::grpc::Slice slice(proto.ByteSizeLong()); proto.SerializeWithCachedSizesToArray(const_cast(slice.begin())); ::grpc::ByteBuffer tmp(&slice, 1); result->Swap(&tmp); } VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep, const platform::DeviceContext& ctx, const framework::Scope& scope, const std::string& var_name, int64_t time_out) { 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); GetProcessor* s = new GetProcessor(ch); const std::string method = "GetRPC"; VarHandlePtr h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope)); s->Prepare(h, time_out); framework::AsyncIO([var_name_val, s, method, p_ctx, h, this] { // prepare input sendrecv::VariableMessage req; req.set_varname(var_name_val); req.set_trainer_id(trainer_id_); ::grpc::ByteBuffer buf; RequestToByteBuffer(req, &buf); VLOG(3) << s->GetVarHandlePtr()->String() << " begin"; // stub context s->response_call_back_ = ProcGetResponse; platform::RecordRPCEvent record_event(method, p_ctx); auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); if (UNLIKELY(platform::IsProfileEnabled())) { h->Wait(); } }); req_count_++; return h; } VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep, const platform::DeviceContext& ctx, const framework::Scope& scope, const std::string& in_var_name, const std::string& out_var_name, int64_t time_out) { const platform::DeviceContext* p_ctx = &ctx; const std::string ep_val = ep; const std::string in_var_name_val = in_var_name; const std::string out_var_name_val = out_var_name; const framework::Scope* p_scope = &scope; const auto ch = GetChannel(ep_val); GetProcessor* s = new GetProcessor(ch); const std::string method = "PrefetchRPC"; VarHandlePtr h(new VarHandle(ep, method, out_var_name_val, p_ctx, p_scope)); s->Prepare(h, time_out); framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx, s, method, h, this] { auto* var = p_scope->FindVar(in_var_name_val); ::grpc::ByteBuffer req; SerializeToByteBuffer(in_var_name_val, var, *p_ctx, &req, out_var_name_val); VLOG(3) << s->GetVarHandlePtr()->String() << " begin"; // stub context s->response_call_back_ = ProcGetResponse; platform::RecordRPCEvent record_event(method, p_ctx); auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/PrefetchVariable", req, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, static_cast(s)); if (UNLIKELY(platform::IsProfileEnabled())) { h->Wait(); } }); req_count_++; return h; } VarHandlePtr GRPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) { const auto ch = GetChannel(ep); BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); const std::string method = "BatchBarrierRPC"; VarHandlePtr h( new VarHandle(ep, method, BATCH_BARRIER_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(BATCH_BARRIER_MESSAGE); platform::RecordRPCEvent record_event(method, nullptr); auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; if (UNLIKELY(platform::IsProfileEnabled())) { h->Wait(); } return h; } VarHandlePtr GRPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) { const auto ch = GetChannel(ep); FetchBarrierProcessor* s = new FetchBarrierProcessor(ch); const std::string method = "FetchBarrierRPC"; VarHandlePtr h( new VarHandle(ep, method, FETCH_BARRIER_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(FETCH_BARRIER_MESSAGE); platform::RecordRPCEvent record_event(method, nullptr); auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; if (UNLIKELY(platform::IsProfileEnabled())) { h->Wait(); } return h; } VarHandlePtr GRPCClient::AsyncSendComplete(const std::string& ep, int64_t time_out) { const auto ch = GetChannel(ep); BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); const std::string method = "SendCompleteRPC"; VarHandlePtr h(new VarHandle(ep, method, COMPLETE_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(COMPLETE_MESSAGE); platform::RecordRPCEvent record_event(method, nullptr); auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; if (UNLIKELY(platform::IsProfileEnabled())) { h->Wait(); } return h; } VarHandlePtr GRPCClient::AsyncCheckpointNotify(const std::string& ep, const std::string& dir, int64_t time_out) { const auto ch = GetChannel(ep); CheckpointNotifyProcessor* s = new CheckpointNotifyProcessor(ch); const std::string method = "CheckPointNotifyRPC"; VarHandlePtr h( new VarHandle(ep, method, CHECKPOINT_SAVE_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(CHECKPOINT_SAVE_MESSAGE); req.set_out_varname(dir); platform::RecordRPCEvent record_event(method, nullptr); auto rpc = s->stub_->AsyncCheckpointNotify(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; if (UNLIKELY(platform::IsProfileEnabled())) { h->Wait(); } return h; } bool GRPCClient::Wait() { std::unique_lock lk(sync_mutex_); sync_cond_.wait(lk, [this] { return (req_count_ == 0 || ok_ == false); }); return ok_; } void GRPCClient::Proceed() { void* tag = nullptr; bool ok = false; VLOG(3) << "GRPCClient Proceed begin"; while (!stopped_ && cq_.Next(&tag, &ok)) { BaseProcessor* c = static_cast(tag); GPR_ASSERT(ok); PADDLE_ENFORCE(c); if (c->status_.ok()) { VLOG(3) << c->GetVarHandlePtr()->String() << " process"; c->Process(); } else if (c->status_.error_code() == grpc::StatusCode::DEADLINE_EXCEEDED) { // FIXME(gongwb): parse error_details? LOG(ERROR) << c->GetVarHandlePtr()->String() << " meets grpc error, error_code:" << c->status_.error_code() << " error_message:" << c->status_.error_message() << " error_details:" << c->status_.error_details(); { std::lock_guard lk(sync_mutex_); ok_ = false; } c->Finish(false); } else { LOG(FATAL) << c->GetVarHandlePtr()->String() << " meets grpc error, error_code:" << c->status_.error_code() << " error_message:" << c->status_.error_message() << " error_details:" << c->status_.error_details(); c->Finish(false); } bool notify = false; { std::lock_guard lk(sync_mutex_); req_count_--; notify = (req_count_ <= 0 || !c->status_.ok()); } delete c; if (notify) { sync_cond_.notify_all(); } } VLOG(3) << "GRPCClient Proceed end"; } std::shared_ptr GRPCClient::GetChannel(const std::string& ep) { std::lock_guard guard(chan_mutex_); auto it = channels_.find(ep); if (it != channels_.end()) { return it->second; } // Channel configurations: grpc::ChannelArguments args; args.SetInt(GRPC_ARG_MAX_RECONNECT_BACKOFF_MS, 2000); args.SetCompressionAlgorithm(GRPC_COMPRESS_NONE); args.SetMaxSendMessageSize(std::numeric_limits::max()); args.SetMaxReceiveMessageSize(std::numeric_limits::max()); auto ch = grpc::CreateCustomChannel(ep, grpc::InsecureChannelCredentials(), args); channels_[ep] = ch; return ch; } } // namespace distributed } // namespace operators } // namespace paddle