You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Paddle/paddle/fluid/operators/distributed/request_handler_impl.cc

199 lines
8.0 KiB

// 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.
AsyncExecutor (#14627) * AsyncExecutor: C++ side * Google naming conventions * Rename MultiExecutor to AsyncExecutor * pybind with async_executor * Naming convention * remove some flags and unused code * add refactored file of async_executor and data_feed * clear async executor interface and add data feed factory * split async executor into executor_thread_worker and async_executor, refactor pybind, add datafeed and corresponding proto * Fix async_executor interfaces: 1) Remove all protobufs; 2) Stop after each epoch * refine async_executor_refactor.cc * add some files about datafeed * Revert "add some files about datafeed" This reverts commit 8ee8133ab841196925a2812b76f18d2812a6701d. * Interface rework * add MultiSlotDataFeed * Creating DataFeedDesc from .proto file, then manipulate it (add/del fields etc) from python side * update data_feed for add MultiSlotDataFeed * update datafeed and async_executor to run bow_net demo * fix bug that finish_set_filelist failed in multithread * delete finish_binding_memory_(flag), because it can not be marked under the current interface * Fix bug * update async_executor.py for support set_use_slots * update async_executor.py for support set_use_slots and set set_dense_slots * fix bug that when the number of files is less than the number of threads, it will fetch nan * remove redundant code, and make executor exit when set a illegal queue size * add batch_size check * add MultiSlotDesc * Revert "add MultiSlotDesc" This reverts commit 2e72ebfad364ed6b5dcc75f38ffb2a1fdec83d8e. * add some checkpoint in DataFeedDesc * add CheckFile function in MultiSlotDataFeed * update something error info * fix deaded lock bug * Fix fetch variable * Merge error * fix code style in async_executor * using one lock blocking queue replace two lock blocking queue because of some bugs * update code style * add utest for data_feed * Fix fetch var * update utest for data_feed for multithread * update SetFileList info * fix bug in utest of data_feed * Add comments for python * Add comments for python code * Fix pybind.cc with new pybind11 version * add note for DataFeedDesc's set_use_slots function * Add save_model * update data_feed_test for multi-type * add comment for executor_thread_worker * Remove unused code * update data_feed_test for generate test data file * removed unnecessary interfaces and add comments * c++ style check * update data_feed.cc * AsyncExecutor: C++ side Google naming conventions Rename MultiExecutor to AsyncExecutor pybind with async_executor Naming convention remove some flags and unused code add refactored file of async_executor and data_feed clear async executor interface and add data feed factory split async executor into executor_thread_worker and async_executor, refactor pybind, add datafeed and corresponding proto Fix async_executor interfaces: 1) Remove all protobufs; 2) Stop after each epoch refine async_executor_refactor.cc add some files about datafeed Revert "add some files about datafeed" This reverts commit 8ee8133ab841196925a2812b76f18d2812a6701d. add MultiSlotDataFeed Interface rework Creating DataFeedDesc from .proto file, then manipulate it (add/del fields etc) from python side update datafeed and async_executor to run bow_net demo update async_executor.py for support set_use_slots Fix bug update async_executor.py for support set_use_slots and set set_dense_slots fix bug that when the number of files is less than the number of threads, it will fetch nan remove redundant code, and make executor exit when set a illegal queue size add MultiSlotDesc Revert "add MultiSlotDesc" This reverts commit 2e72ebfad364ed6b5dcc75f38ffb2a1fdec83d8e. add some checkpoint in DataFeedDesc Fix fetch variable fix code style in async_executor Fix fetch var add utest for data_feed Add comments for python update utest for data_feed for multithread fix bug in utest of data_feed Add comments for python code Fix pybind.cc with new pybind11 version add note for DataFeedDesc's set_use_slots function update data_feed_test for multi-type Add save_model update data_feed_test for generate test data file removed unnecessary interfaces and add comments add comment for executor_thread_worker Remove unused code update data_feed.cc c++ style check * commit for code style * commit for code style * commit for code style * commit for code style * Comment away __init__ in async_executor.py * clang-format fix test=develop * use PADDLE_THROW instead of exit(-1); use unique_ptr to manage scope var in data_feed_test.cc * commit for update code style * commit for update code style * Add async_executor demo; Remove some methods test=develop * commit for update code style * commit for update code style * commit for update code style * update API.spec * AsyncExecutor test=develop * AsyncExecutor test=develop * AsyncExecutor test=develop * AsyncExecutor test=develop * Fix API.spec test=develop * Fix API.spec test=develop * Fix windows build error test=develop * FIx windows build error test=develop * FIx windows build error test=develop * FIx windows build error test=develop * Fix Windows Build test=develop * Fix Windows Build test=develop * Fix Windows Build test=develop * Fix code style test=develop * Fix code style test=develop * update datafeed * Fix code style test=develop * update data_feed_test for test Tensor test=develop * Fix code style test=develop * Fix windows build failure test=develop * Fix code style and windows build failure test=develop * Fix PYTHON3.5 build failure test=develop * AsyncExecutor API test=develop
6 years ago
#include "paddle/fluid/operators/distributed/request_handler_impl.h"
#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"
AsyncExecutor (#14627) * AsyncExecutor: C++ side * Google naming conventions * Rename MultiExecutor to AsyncExecutor * pybind with async_executor * Naming convention * remove some flags and unused code * add refactored file of async_executor and data_feed * clear async executor interface and add data feed factory * split async executor into executor_thread_worker and async_executor, refactor pybind, add datafeed and corresponding proto * Fix async_executor interfaces: 1) Remove all protobufs; 2) Stop after each epoch * refine async_executor_refactor.cc * add some files about datafeed * Revert "add some files about datafeed" This reverts commit 8ee8133ab841196925a2812b76f18d2812a6701d. * Interface rework * add MultiSlotDataFeed * Creating DataFeedDesc from .proto file, then manipulate it (add/del fields etc) from python side * update data_feed for add MultiSlotDataFeed * update datafeed and async_executor to run bow_net demo * fix bug that finish_set_filelist failed in multithread * delete finish_binding_memory_(flag), because it can not be marked under the current interface * Fix bug * update async_executor.py for support set_use_slots * update async_executor.py for support set_use_slots and set set_dense_slots * fix bug that when the number of files is less than the number of threads, it will fetch nan * remove redundant code, and make executor exit when set a illegal queue size * add batch_size check * add MultiSlotDesc * Revert "add MultiSlotDesc" This reverts commit 2e72ebfad364ed6b5dcc75f38ffb2a1fdec83d8e. * add some checkpoint in DataFeedDesc * add CheckFile function in MultiSlotDataFeed * update something error info * fix deaded lock bug * Fix fetch variable * Merge error * fix code style in async_executor * using one lock blocking queue replace two lock blocking queue because of some bugs * update code style * add utest for data_feed * Fix fetch var * update utest for data_feed for multithread * update SetFileList info * fix bug in utest of data_feed * Add comments for python * Add comments for python code * Fix pybind.cc with new pybind11 version * add note for DataFeedDesc's set_use_slots function * Add save_model * update data_feed_test for multi-type * add comment for executor_thread_worker * Remove unused code * update data_feed_test for generate test data file * removed unnecessary interfaces and add comments * c++ style check * update data_feed.cc * AsyncExecutor: C++ side Google naming conventions Rename MultiExecutor to AsyncExecutor pybind with async_executor Naming convention remove some flags and unused code add refactored file of async_executor and data_feed clear async executor interface and add data feed factory split async executor into executor_thread_worker and async_executor, refactor pybind, add datafeed and corresponding proto Fix async_executor interfaces: 1) Remove all protobufs; 2) Stop after each epoch refine async_executor_refactor.cc add some files about datafeed Revert "add some files about datafeed" This reverts commit 8ee8133ab841196925a2812b76f18d2812a6701d. add MultiSlotDataFeed Interface rework Creating DataFeedDesc from .proto file, then manipulate it (add/del fields etc) from python side update datafeed and async_executor to run bow_net demo update async_executor.py for support set_use_slots Fix bug update async_executor.py for support set_use_slots and set set_dense_slots fix bug that when the number of files is less than the number of threads, it will fetch nan remove redundant code, and make executor exit when set a illegal queue size add MultiSlotDesc Revert "add MultiSlotDesc" This reverts commit 2e72ebfad364ed6b5dcc75f38ffb2a1fdec83d8e. add some checkpoint in DataFeedDesc Fix fetch variable fix code style in async_executor Fix fetch var add utest for data_feed Add comments for python update utest for data_feed for multithread fix bug in utest of data_feed Add comments for python code Fix pybind.cc with new pybind11 version add note for DataFeedDesc's set_use_slots function update data_feed_test for multi-type Add save_model update data_feed_test for generate test data file removed unnecessary interfaces and add comments add comment for executor_thread_worker Remove unused code update data_feed.cc c++ style check * commit for code style * commit for code style * commit for code style * commit for code style * Comment away __init__ in async_executor.py * clang-format fix test=develop * use PADDLE_THROW instead of exit(-1); use unique_ptr to manage scope var in data_feed_test.cc * commit for update code style * commit for update code style * Add async_executor demo; Remove some methods test=develop * commit for update code style * commit for update code style * commit for update code style * update API.spec * AsyncExecutor test=develop * AsyncExecutor test=develop * AsyncExecutor test=develop * AsyncExecutor test=develop * Fix API.spec test=develop * Fix API.spec test=develop * Fix windows build error test=develop * FIx windows build error test=develop * FIx windows build error test=develop * FIx windows build error test=develop * Fix Windows Build test=develop * Fix Windows Build test=develop * Fix Windows Build test=develop * Fix code style test=develop * Fix code style test=develop * update datafeed * Fix code style test=develop * update data_feed_test for test Tensor test=develop * Fix code style test=develop * Fix windows build failure test=develop * Fix code style and windows build failure test=develop * Fix PYTHON3.5 build failure test=develop * AsyncExecutor API test=develop
6 years ago
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h"
#include "paddle/fluid/operators/distributed/rpc_server.h"
#include "paddle/fluid/string/piece.h"
#include "paddle/fluid/string/printf.h"
namespace paddle {
namespace operators {
namespace distributed {
// define LOOKUP_TABLE_PATH for checkpoint notify to save lookup table variables
// to directory specified.
constexpr char LOOKUP_TABLE_PATH[] = "kLookupTablePath";
bool RequestSendHandler::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) << "RequestSendHandler:" << varname;
// Sync
if (varname == BATCH_BARRIER_MESSAGE) {
VLOG(3) << "sync: recv BATCH_BARRIER_MESSAGE";
rpc_server_->IncreaseBatchBarrier(kRequestSend);
} else if (varname == COMPLETE_MESSAGE) {
VLOG(3) << "sync: recv complete message";
rpc_server_->Complete();
} else {
// Async
if (!sync_mode_) {
VLOG(3) << "async process var: " << varname;
if (varname == BATCH_BARRIER_MESSAGE) {
PADDLE_THROW(
"async mode should not recv BATCH_BARRIER_MESSAGE or "
"COMPLETE_MESSAGE");
}
if (AsyncSparseParamUpdateRecorder::GetInstance()->HasGrad(varname)) {
auto& grad_slr =
scope->FindVar(varname)->Get<framework::SelectedRows>();
AsyncSparseParamUpdateRecorder::GetInstance()->Update(varname,
grad_slr.rows());
}
6 years ago
executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(),
scope);
return true;
} else { // sync
rpc_server_->WaitCond(kRequestSend);
VLOG(3) << "sync: processing received var: " << varname;
if (invar == nullptr) {
LOG(FATAL) << "sync: Can not find server side var: " << varname;
return false;
}
}
}
return true;
}
bool RequestGetHandler::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(3) << "RequestGetHandler:" << varname
<< " out_var_name: " << out_var_name << " trainer_id: " << trainer_id
<< " table_name: " << table_name;
if (sync_mode_) {
if (varname == FETCH_BARRIER_MESSAGE) {
VLOG(3) << "sync: recv fetch barrier message";
rpc_server_->IncreaseBatchBarrier(kRequestGet);
} else {
rpc_server_->WaitCond(kRequestGet);
*outvar = scope_->FindVar(varname);
}
} else {
if (varname != FETCH_BARRIER_MESSAGE && varname != COMPLETE_MESSAGE) {
if (enable_dc_asgd_) {
// NOTE: the format is determined by distribute_transpiler.py
std::string param_bak_name =
string::Sprintf("%s.trainer_%d_bak", varname, trainer_id);
VLOG(3) << "getting " << param_bak_name << " trainer_id " << trainer_id;
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());
VLOG(3) << "copying " << varname << " to " << param_bak_name;
framework::TensorCopy(t_orig, dev_ctx_->GetPlace(), t);
}
*outvar = scope_->FindVar(varname);
}
}
return true;
}
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;
}
bool RequestPrefetchHandler::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) << "RequestPrefetchHandler " << varname;
if (table_name.empty()) {
auto var_desc = program_->Block(0).FindVar(out_var_name);
InitializeVariable(*outvar, var_desc->GetType());
executor_->RunPreparedContext(
(*prefetch_var_name_to_prepared_ctx_)[varname].get(), scope);
} else {
(*outvar)->GetMutable<framework::LoDTensor>();
auto lookup_table_op =
BuildLookupTableOp(table_name, varname, out_var_name);
paddle::platform::CPUPlace cpu_place;
lookup_table_op->Run(*scope, cpu_place);
}
return true;
}
bool RequestCheckpointHandler::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) {
PADDLE_ENFORCE(
checkpoint_notify_id != -1,
"when checkpoint_notify_id = -1, there should be no RPC invoke.");
7 years ago
// TODO(tangwei12): find out why scope will be error.
7 years ago
auto* lt_var = scope_->FindVar(LOOKUP_TABLE_PATH)->GetMutable<std::string>();
lt_var->clear();
lt_var->append(out_var_name);
VLOG(4) << "RequestCheckpointHandler update var kLookupTablePath to: "
<< out_var_name;
7 years ago
executor_->RunPreparedContext(checkpoint_prepared_ctx_.get(), scope_);
7 years ago
return true;
}
} // namespace distributed
} // namespace operators
} // namespace paddle