Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into feature/exec_strategy

doc/fluid/design/dist_train/async_update.md

@@ -4,34 +4,37 @@
 
 For the typical synchronous distributed training, some significant steps are as follows:
 
-1. A Trainer will compute the gradients and SEND them to the Parameter Server(PServer) nodes.
+1. A trainer process will compute the gradients and **send** them to the parameter server (PS) nodes.
-1. After the PServer node received gradients came from all the Trainers, It will aggregate the
+1. After the PS node received gradients came from all the Trainers, It will aggregate the
 gradient variables for the same parameter into one gradient variable and then apply the aggregated
 gradient to the respective parameter, finally using an optimize algorithms(SGD, Monument...)
 to update the parameters.
-1. The Trainer would wait for the PServers finished the optimize stage, and GET the parameters from PServer,
+1. The Trainer would wait for the PS finished the optimize stage, and GET the parameters from PS,
 so all the Trainers would get the same parameters.
 
-In the synchronously distributed training, there should be a `Barrier` to synchronise the
+In Synchronous Distributed Training, there is a **barrier** on each PS to wait until all trainers processes
-parameters after the optimizing stage. The performance of a distributed training job would
+have completed running current mini-batch. After that, all trainers can continue to run the next
-depend on the slowest node if there were hundreds or thousands of training nodes in a
+mini-batch. So, we can find that the overall performance of Synchronous Distributed Training depends 
-Job, the performance of synchronously distributed training might be very poor because of
+on the slowest node.
-the slow node. So this design doc would introduce an approach to implement
+
-*asynchronously* distributed training in PaddlePaddle Fluid.
+In Asynchronous Distributed Training, we don't need to wait for a global mini-bach, the optimizer on
+the PS will run immediately when the gradient is uploaded to the PS from one trainer. This mode would
+train such models that achieve scaling, better throughput. In this design doc, we will introduce how to 
+implement the Asynchronous Distributed Training base on PaddlePaddle Fluid.
 
 ## Design
 
 <img src="./src/async_update.png" width="600"/>
 
-As the figure above, we describe a global view of asynchronously update process and use
+As the figure above, we describe a global view of the asynchronous update process and use
 the parameter `w1` as an example to introduce the steps:
 1. For each gradient variables, they may distribute on different GPU card and aggregate
 them while they are all calculated.
-1. Split the gradient variable into multiple blocks according to the number of PServer
+1. Split the gradient variable into multiple blocks according to the number of PS
 instances and then send them.
-1. PServer would run an `Optimize Block` using a specified optimize algorithm to update
+1. PS would run an `Optimize Block` using a specified optimize algorithm to update
 the specified parameter.
-1. The trainer will fetch latest parameter from PServer before running forward Op which depends
+1. The trainer will fetch the latest parameter from PS before running forward Op which depends
 on the specified parameter.
 1. Broadcast the received variable into multiple GPU cards and continue to run the next
 mini-batch.
@@ -40,8 +43,8 @@ mini-batch.
 
 - For the multiple devices distributed training, we need to aggregate the gradient
 variables which placed on different devices firstly and then schedule a `SendVars` Operator to
-send the gradient variables to the multiple PServer instances.
+send the gradient variables to the multiple PS instances.
-- Schedule `FetchVars` operator to fetch the latest parameter from PServer before running
+- Schedule `FetchVars` operator to fetch the latest parameter from PS before running
 the forward ops.
 - There could be a large number of gradient variables to be sent, so we need to use another
 thread pool(IO Threadpool) whose a number of the schedulable threads is larger than the

paddle/fluid/framework/CMakeLists.txt

@@ -5,11 +5,11 @@ proto_library(framework_proto SRCS framework.proto)
 cc_library(ddim SRCS ddim.cc DEPS eigen3 boost)
 cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
-
+cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context)
 if(WITH_GPU)
-  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place memory device_context framework_proto)
+  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type)
 else()
-  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place memory device_context framework_proto)
+  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type)
 endif()
 
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)

paddle/fluid/framework/data_type.cc

@@ -0,0 +1,101 @@
+//   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 "paddle/fluid/framework/data_type.h"
+#include <stdint.h>
+#include <string>
+#include <unordered_map>
+
+namespace paddle {
+namespace framework {
+
+struct DataTypeMap {
+  std::unordered_map<std::type_index, proto::VarType::Type> cpp_to_proto_;
+  std::unordered_map<int, std::type_index> proto_to_cpp_;
+  std::unordered_map<int, std::string> proto_to_str_;
+  std::unordered_map<std::type_index, size_t> cpp_to_size_;
+};
+
+static DataTypeMap* InitDataTypeMap();
+static DataTypeMap& gDataTypeMap() {
+  static DataTypeMap* g_data_type_map_ = InitDataTypeMap();
+  return *g_data_type_map_;
+}
+
+template <typename T>
+static inline void RegisterType(DataTypeMap* map,
+                                proto::VarType::Type proto_type,
+                                const std::string& name) {
+  map->proto_to_cpp_.emplace(static_cast<int>(proto_type), typeid(T));
+  map->cpp_to_proto_.emplace(typeid(T), proto_type);
+  map->proto_to_str_.emplace(static_cast<int>(proto_type), name);
+  map->cpp_to_size_.emplace(typeid(T), sizeof(T));
+}
+
+static DataTypeMap* InitDataTypeMap() {
+  auto retv = new DataTypeMap();
+
+#define RegType(cc_type, proto_type) \
+  RegisterType<cc_type>(retv, proto_type, #cc_type)
+
+  // NOTE: Add your customize type here.
+  RegType(platform::float16, proto::VarType::FP16);
+  RegType(float, proto::VarType::FP32);
+  RegType(double, proto::VarType::FP64);
+  RegType(int, proto::VarType::INT32);
+  RegType(int64_t, proto::VarType::INT64);
+  RegType(bool, proto::VarType::BOOL);
+  RegType(size_t, proto::VarType::SIZE_T);
+  RegType(int16_t, proto::VarType::INT16);
+
+#undef RegType
+  return retv;
+}
+
+proto::VarType::Type ToDataType(std::type_index type) {
+  auto it = gDataTypeMap().cpp_to_proto_.find(type);
+  if (it != gDataTypeMap().cpp_to_proto_.end()) {
+    return it->second;
+  }
+  PADDLE_THROW("Not support %s as tensor type", type.name());
+}
+
+std::type_index ToTypeIndex(proto::VarType::Type type) {
+  auto it = gDataTypeMap().proto_to_cpp_.find(static_cast<int>(type));
+  if (it != gDataTypeMap().proto_to_cpp_.end()) {
+    return it->second;
+  }
+  PADDLE_THROW("Not support proto::VarType::Type(%d) as tensor type",
+               static_cast<int>(type));
+}
+
+std::string DataTypeToString(const proto::VarType::Type type) {
+  auto it = gDataTypeMap().proto_to_str_.find(static_cast<int>(type));
+  if (it != gDataTypeMap().proto_to_str_.end()) {
+    return it->second;
+  }
+  PADDLE_THROW("Not support proto::VarType::Type(%d) as tensor type",
+               static_cast<int>(type));
+}
+
+size_t SizeOfType(std::type_index type) {
+  auto it = gDataTypeMap().cpp_to_size_.find(type);
+  if (it != gDataTypeMap().cpp_to_size_.end()) {
+    return it->second;
+  }
+  PADDLE_THROW("Not support %s as tensor type", type.name());
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/data_type.h

@@ -17,51 +17,14 @@ limitations under the License. */
 #include <typeindex>
 #include "paddle/fluid/framework/framework.pb.h"
 #include "paddle/fluid/platform/enforce.h"
+
 #include "paddle/fluid/platform/float16.h"
 
 namespace paddle {
 namespace framework {
 
-inline proto::VarType::Type ToDataType(std::type_index type) {
+extern proto::VarType::Type ToDataType(std::type_index type);
-  if (typeid(platform::float16).hash_code() == type.hash_code()) {
+extern std::type_index ToTypeIndex(proto::VarType::Type type);
-    return proto::VarType::FP16;
-  } else if (typeid(const float).hash_code() == type.hash_code()) {
-    // CPPLint complains Using C-style cast.  Use static_cast<float>() instead
-    // One fix to this is to replace float with const float because
-    // typeid(T) == typeid(const T)
-    // http://en.cppreference.com/w/cpp/language/typeid
-    return proto::VarType::FP32;
-  } else if (typeid(const double).hash_code() == type.hash_code()) {
-    return proto::VarType::FP64;
-  } else if (typeid(const int).hash_code() == type.hash_code()) {
-    return proto::VarType::INT32;
-  } else if (typeid(const int64_t).hash_code() == type.hash_code()) {
-    return proto::VarType::INT64;
-  } else if (typeid(const bool).hash_code() == type.hash_code()) {
-    return proto::VarType::BOOL;
-  } else {
-    PADDLE_THROW("Not supported");
-  }
-}
-
-inline std::type_index ToTypeIndex(proto::VarType::Type type) {
-  switch (type) {
-    case proto::VarType::FP16:
-      return typeid(platform::float16);
-    case proto::VarType::FP32:
-      return typeid(float);
-    case proto::VarType::FP64:
-      return typeid(double);
-    case proto::VarType::INT32:
-      return typeid(int);
-    case proto::VarType::INT64:
-      return typeid(int64_t);
-    case proto::VarType::BOOL:
-      return typeid(bool);
-    default:
-      PADDLE_THROW("Not support type %d", type);
-  }
-}
 
 template <typename Visitor>
 inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
@@ -89,32 +52,12 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
   }
 }
 
-inline std::string DataTypeToString(const proto::VarType::Type type) {
+extern std::string DataTypeToString(const proto::VarType::Type type);
-  switch (type) {
+extern size_t SizeOfType(std::type_index type);
-    case proto::VarType::FP16:
-      return "float16";
-    case proto::VarType::FP32:
-      return "float32";
-    case proto::VarType::FP64:
-      return "float64";
-    case proto::VarType::INT16:
-      return "int16";
-    case proto::VarType::INT32:
-      return "int32";
-    case proto::VarType::INT64:
-      return "int64";
-    case proto::VarType::BOOL:
-      return "bool";
-    default:
-      PADDLE_THROW("Not support type %d", type);
-  }
-}
-
 inline std::ostream& operator<<(std::ostream& out,
                                 const proto::VarType::Type& type) {
   out << DataTypeToString(type);
   return out;
 }
-
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/framework.proto

@@ -101,6 +101,8 @@ message VarType {
     FP16 = 4;
     FP32 = 5;
     FP64 = 6;
+    // Tensor<size_t> is used in C++.
+    SIZE_T = 19;
 
     // Other types that may need additional descriptions
     LOD_TENSOR = 7;

paddle/fluid/framework/op_kernel_type_test.cc

@@ -27,7 +27,7 @@ TEST(OpKernelType, ToString) {
                               LibraryType::kCUDNN);
 
   ASSERT_EQ(paddle::framework::KernelTypeToString(op_kernel_type),
-            "data_type[float32]:data_layout[NCHW]:place[CPUPlace]:library_type["
+            "data_type[float]:data_layout[NCHW]:place[CPUPlace]:library_type["
             "CUDNN]");
 }
 

paddle/fluid/framework/tensor_impl.h

@@ -13,54 +13,14 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/float16.h"
 
 namespace paddle {
 namespace framework {
-
+extern size_t SizeOfType(std::type_index type);
-template <typename... T>
-struct SizeOfTypeFunctor;
-
-template <typename T>
-struct SizeOfTypeFunctor<T> {
-  size_t operator()(std::type_index type) const {
-    if (typeid(T).hash_code() == type.hash_code()) {
-      return sizeof(T);
-    } else {
-      return 0UL;
-    }
-  }
-};
-
-template <>
-struct SizeOfTypeFunctor<> {
-  size_t operator()(std::type_index type) const { return 0UL; }
-};
-
-template <typename HEAD, typename... TAIL>
-struct SizeOfTypeFunctor<HEAD, TAIL...> {
-  size_t operator()(std::type_index type) const {
-    SizeOfTypeFunctor<HEAD> head;
-    size_t head_size = head(type);
-    if (head_size != 0) {
-      return head_size;
-    }
-    SizeOfTypeFunctor<TAIL...> tail;
-    return tail(type);
-  }
-};
-
-static inline size_t SizeOfType(std::type_index type) {
-  SizeOfTypeFunctor<int, float, double, int16_t, int64_t, bool, size_t,
-                    platform::float16>
-      functor;
-  size_t size = functor(type);
-  PADDLE_ENFORCE(size != 0UL, "Cannot get size of type %s", type.name());
-  return size;
-}
-
 inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
       holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");

paddle/fluid/operators/detail/grpc_server.cc

@@ -306,7 +306,7 @@ void AsyncGRPCServer::TryToRegisterNewPrefetchOne() {
   }
   RequestPrefetch* prefetch =
       new RequestPrefetch(&service_, cq_prefetch_.get(), sync_mode_, scope_,
-                          dev_ctx_, executor_, program_, prefetch_ctx_);
+                          dev_ctx_, executor_, program_, prefetch_ctx_.get());
 
   VLOG(4) << "Create RequestPrefetch status:" << prefetch->Status();
 }

paddle/fluid/operators/detail/grpc_server.h

@@ -64,8 +64,9 @@ class AsyncGRPCServer final {
 
   void SetExecutor(framework::Executor *executor) { executor_ = executor; }
 
-  void SetPrefetchPreparedCtx(framework::ExecutorPrepareContext *prepared) {
+  void SetPrefetchPreparedCtx(
-    prefetch_ctx_ = prepared;
+      std::unique_ptr<framework::ExecutorPrepareContext> prepared) {
+    prefetch_ctx_.reset(prepared.release());
   }
 
   int GetSelectedPort() const { return selected_port_; }
@@ -116,7 +117,7 @@ class AsyncGRPCServer final {
   std::unique_ptr<std::thread> t_get_;
   std::unique_ptr<std::thread> t_prefetch_;
 
-  framework::ExecutorPrepareContext *prefetch_ctx_;
+  std::unique_ptr<framework::ExecutorPrepareContext> prefetch_ctx_;
   framework::ProgramDesc *program_;
   framework::Executor *executor_;
   int selected_port_;

paddle/fluid/operators/detail/grpc_server_test.cc

@@ -100,7 +100,7 @@ void StartServer(const std::string& endpoint) {
   InitTensorsOnServer(&scope, &place, 10);
 
   rpc_service_->SetProgram(&program);
-  rpc_service_->SetPrefetchPreparedCtx(prepared.get());
+  rpc_service_->SetPrefetchPreparedCtx(std::move(prepared));
   rpc_service_->SetDevCtx(&ctx);
   rpc_service_->SetScope(&scope);
   rpc_service_->SetExecutor(&exe);

paddle/fluid/operators/listen_and_serv_op.cc

@@ -322,8 +322,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
   // prepare for prefetch
   VLOG(3) << "prefetch block id is " << prefetch_block->ID();
   auto prefetch_prepared = executor.Prepare(*program, prefetch_block->ID());
-  rpc_service_->SetPrefetchPreparedCtx(prefetch_prepared.get());
+  rpc_service_->SetPrefetchPreparedCtx(std::move(prefetch_prepared));
-  prefetch_prepared.release();
 
   // start the server listening after all member initialized.
   server_thread_.reset(new std::thread(RunServer, rpc_service_));

python/paddle/fluid/backward.py

@@ -480,6 +480,8 @@ def append_backward(loss, parameter_list=None, no_grad_set=None,
 
     program.current_block_idx = current_block_idx
     program.sync_with_cpp()
+    # FIXME(zcd): prevent loss.grad optimized by mem_opt.
+    loss.block.var(_append_grad_suffix_(loss.name)).persistable = True
 
     if parameter_list is not None:
         parameters = parameter_list

python/paddle/fluid/tests/unittests/test_split_var.py

@@ -21,15 +21,7 @@ import random
 
 
 class TestSplitVar(unittest.TestCase):
-    def test_check_output(self):
+    def check_split_output(self, shapes, expected_sizes, min_size):
-        # split below shapes to 10 servers
-        shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10]]
-        expected_sizes = [
-            [15], [1024],
-            [2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 784],
-            [2040, 2040, 2040, 2040],
-            [1150, 1150, 1150, 1150, 1150, 1150, 1100]
-        ]
         var_list = []
         program = fluid.Program()
         for shape in shapes:
@@ -39,7 +31,7 @@ class TestSplitVar(unittest.TestCase):
                 # dtype=core.VarDesc.VarType.LOD_TENSOR,
                 shape=shape)
             var_list.append(var)
-        blocks = split_dense_variable(var_list, 10)
+        blocks = split_dense_variable(var_list, 10, min_size)
         all_sizes = []
         for s in expected_sizes:
             for s2 in s:
@@ -48,6 +40,25 @@ class TestSplitVar(unittest.TestCase):
             varname, block_id, size = block_str.split(":")
             self.assertEqual(int(size), all_sizes[i])
 
+    def test_1k(self):
+        shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10]]
+        expected_sizes = [
+            [15], [1024],
+            [2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 784],
+            [2040, 2040, 2040, 2040],
+            [1150, 1150, 1150, 1150, 1150, 1150, 1100]
+        ]
+
+        self.check_split_output(shapes, expected_sizes, 1024)
+
+    def test_check_output_8k(self):
+        shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10],
+                  [6, 33, 33, 33]]
+        expected_sizes = [[15], [1024], [10976, 10976], [8160], [8000],
+                          [35937, 35937, 35937, 35937, 35937, 35937]]
+
+        self.check_split_output(shapes, expected_sizes, 8192)
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -93,30 +93,33 @@ def same_or_split_var(p_name, var_name):
     return p_name == var_name or p_name.startswith(var_name + ".block")
 
 
-def split_dense_variable(var_list,
+def split_dense_variable(var_list, service_count, min_block_size=8192):
-                         pserver_count,
-                         min_block_size=1024,
-                         max_block_size=1048576):
     """
-        We may need to split dense tensor to one or more blocks and put
+    We may need to split dense tensor to one or more blocks and put
-        them equally onto parameter server. One block is a sub-tensor
+    them equally onto parameter server. One block is a sub-tensor
-        aligned by dim[0] of the tensor.
+    aligned by dim[0] of the tensor.
-
+
-        We need to have a minimal block size so that the calculations in
+    We need to have a minimal block size so that the calculations in
-        the parameter server side can gain better performance. By default
+    the parameter server side can gain better performance. By default
-        minimum block size is 1024. The max block size is used to prevent
+    minimum block size 8K elements (maybe 16bit or 32bit or 64bit). 
-        very large blocks that may cause send error.
+
-        :return: A list of VarBlocks. Each VarBlock specifies a shard of
+    Args:
-           the var.
+        var_list (list): List of variables.
+        service_count (int): Numel of pserver services. A pserver may have two
+            or more listening ports.
+        min_block_size (int): Minimum splitted block size.
+    Returns:
+        blocks (list[(varname, block_id, current_block_size)]): A list 
+            of VarBlocks. Each VarBlock specifies a shard of the var.
     """
     blocks = []
     for var in var_list:
-        split_count = pserver_count
+        split_count = service_count
         var_numel = reduce(lambda x, y: x * y, var.shape)
         max_pserver_count = int(math.floor(var_numel / float(min_block_size)))
         if max_pserver_count == 0:
             max_pserver_count = 1
-        if max_pserver_count < pserver_count:
+        if max_pserver_count < service_count:
             split_count = max_pserver_count
         block_size = int(math.ceil(var_numel / float(split_count)))
 
@@ -270,6 +273,7 @@ class DistributeTranspiler:
         grad_var_mapping = self._append_split_op(program, grad_blocks)
         param_var_mapping = self._create_vars_from_blocklist(program,
                                                              param_blocks)
+
         # step3: Add gradients as send op inputs and parameters as send
         # op outputs.
         send_inputs = []
@@ -277,9 +281,11 @@ class DistributeTranspiler:
         for b in grad_blocks:  # append by order
             varname, block_id, _ = b.split(":")
             send_inputs.append(grad_var_mapping[varname][int(block_id)])
+
         for b in param_blocks:
             varname, block_id, _ = b.split(":")
             send_outputs.append(param_var_mapping[varname][int(block_id)])
+
         # let send_op know which endpoint to send which var to, eplist has the same
         # order as send_inputs.
         eplist = split_method(send_inputs, pserver_endpoints)
@@ -751,9 +757,18 @@ class DistributeTranspiler:
         Create vars for each split.
         NOTE: only grads need to be named for different trainers, use
               add_trainer_suffix to rename the grad vars.
-        :return: A dict mapping from original var name to each var split.
+        Args:
+            program (ProgramDesc): ProgramDesc which gradients blong.
+            block_list (list[(varname, block_id, block_size)]): List of gradient blocks.
+            add_trainer_suffix (Bool): Add trainer suffix to new variable's name if set True.
+        Returns: 
+            var_mapping (dict(varname->[new_varname_variable])):A dict mapping 
+                from original var name to each var split.
         """
+
+        # varname->[(block_id, current_block_size)]
         block_map = dict()
+
         var_mapping = dict()
         for block_str in block_list:
             varname, offset, size = block_str.split(":")
@@ -824,7 +839,16 @@ class DistributeTranspiler:
             persistable=persistable)
 
     def _append_split_op(self, program, gradblocks):
-        # Split variables that need to be split and append respective ops
+        """
+        Split variables that need to be split and append respective ops
+        Args:
+            program (ProgramDesc): ProgramDesc that gradients blong.
+            gradblocks (list[(varname, block_id, block_size)]): List of gradient blocks.
+        Returns:
+            var_mapping (dict(varname->[new_splitted_variable])):A dict mapping 
+                from original var name to each var split.
+        """
+
         add_suffix = False
         if self.trainer_num > 1:
             add_suffix = True
@@ -1148,6 +1172,12 @@ class DistributeTranspiler:
         return lr_ops
 
     def _get_optimize_pass(self):
+        """
+        Get optimizer operators, paramters and gradients from origin_program
+        Returns:
+            opt_ops (list): optimize operators.
+            params_grads (dict): paramter->gradient.
+        """
         block = self.origin_program.global_block()
         opt_ops = []
         params_grads = []