Add brpc serialization support.

benchmark/fluid/fluid_benchmark.py

@@ -81,9 +81,11 @@ def dist_transpile(trainer_id, args, train_prog, startup_prog):
     # the role, should be either PSERVER or TRAINER
     training_role = os.getenv("PADDLE_TRAINING_ROLE")
 
-    config = distribute_transpiler.DistributeTranspilerConfig()
+    config = fluid.DistributeTranspilerConfig()
     config.slice_var_up = not args.no_split_var
+    config.min_block_size = 1048576
     t = distribute_transpiler.DistributeTranspiler(config=config)
+
     t.transpile(
         trainer_id,
         # NOTE: *MUST* use train_prog, for we are using with guard to

cmake/external/brpc.cmake

@@ -14,14 +14,16 @@
 
 INCLUDE(ExternalProject)
 
-find_library(SSL_LIBRARY NAMES ssl)
+find_package(OpenSSL REQUIRED) 
+
+message(STATUS "ssl:" ${OPENSSL_SSL_LIBRARY})
+message(STATUS "crypto:" ${OPENSSL_CRYPTO_LIBRARY})
+
 ADD_LIBRARY(ssl SHARED IMPORTED GLOBAL)
-SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${SSL_LIBRARY})
+SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${OPENSSL_SSL_LIBRARY})
 
-find_library(CRYPTO_LIBRARY NAMES crypto)
 ADD_LIBRARY(crypto SHARED IMPORTED GLOBAL)
-SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${CRYPTO_LIBRARY})
-
+SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${OPENSSL_CRYPTO_LIBRARY})
 
 SET(BRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/brpc)
 SET(BRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/brpc)
@@ -31,14 +33,15 @@ SET(BRPC_LIBRARIES "${BRPC_INSTALL_DIR}/lib/libbrpc.a" CACHE FILEPATH "brpc libr
 INCLUDE_DIRECTORIES(${BRPC_INCLUDE_DIR})
 
 # Reference https://stackoverflow.com/questions/45414507/pass-a-list-of-prefix-paths-to-externalproject-add-in-cmake-args
-set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib")
+set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib|${THIRD_PARTY_PATH}/install/glog")
 
 # If minimal .a is need, you can set  WITH_DEBUG_SYMBOLS=OFF
 ExternalProject_Add(
     extern_brpc
     ${EXTERNAL_PROJECT_LOG_ARGS}
+    # TODO(gongwb): change to de newst repo when they changed.
     GIT_REPOSITORY  "https://github.com/gongweibao/brpc"
-    GIT_TAG         "7dc04defad1fd4173aae170c3fcbde131b65155a"
+    GIT_TAG         "e9b67ec1b7458f2af5fae76451afe1e27e01b4b4"
     PREFIX          ${BRPC_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
@@ -50,7 +53,7 @@ ExternalProject_Add(
                     -DCMAKE_POSITION_INDEPENDENT_CODE=ON
                     -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
                     -DCMAKE_PREFIX_PATH=${prefix_path}
-                    -DBRPC_WITH_GLOG=ON
+                    -DWITH_GLOG=ON
                     -DIOBUF_WITH_HUGE_BLOCK=ON
                     -DBRPC_WITH_RDMA=${WITH_BRPC_RDMA}
                     ${EXTERNAL_OPTIONAL_ARGS}
@@ -65,5 +68,6 @@ ADD_LIBRARY(brpc STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES})
 ADD_DEPENDENCIES(brpc extern_brpc)
 
+add_definitions(-DBRPC_WITH_GLOG)
 
 LIST(APPEND external_project_dependencies brpc)

cmake/external/gtest.cmake

@@ -12,8 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-IF(WITH_TESTING)
-    ENABLE_TESTING()
+#FIXME:(gongwb) Move brpc's gtest dependency.
+IF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
+    IF(WITH_TESTING)
+        ENABLE_TESTING()
+    ENDIF(WITH_TESTING)
+
     INCLUDE(ExternalProject)
 
     SET(GTEST_SOURCES_DIR ${THIRD_PARTY_PATH}/gtest)
@@ -76,4 +80,4 @@ IF(WITH_TESTING)
     ADD_DEPENDENCIES(gtest_main extern_gtest)
 
     LIST(APPEND external_project_dependencies gtest gtest_main)
-ENDIF(WITH_TESTING)
+ENDIF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))

cmake/external/leveldb.cmake

@@ -24,8 +24,8 @@ ExternalProject_Add(
     extern_leveldb
     ${EXTERNAL_PROJECT_LOG_ARGS}
     PREFIX ${LEVELDB_SOURCES_DIR}
-    URL "https://github.com/google/leveldb/archive/v1.18.tar.gz"
-    URL_MD5 "73770de34a2a5ab34498d2e05b2b7fa0"
+    GIT_REPOSITORY "https://github.com/google/leveldb"
+    GIT_TAG v1.18
     CONFIGURE_COMMAND ""
     BUILD_COMMAND CXXFLAGS=-fPIC make -j ${NUM_OF_PROCESSOR} libleveldb.a
     INSTALL_COMMAND mkdir -p ${LEVELDB_INSTALL_DIR}/lib/ 

paddle/fluid/framework/CMakeLists.txt

@@ -169,9 +169,12 @@ cc_library(variable_helper SRCS variable_helper.cc DEPS lod_tensor)
 cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
 
 if(WITH_DISTRIBUTE)
-  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass variable_helper)
-  set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
-  set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog
+        lod_rank_table feed_fetch_method sendrecvop_rpc  ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper)
+
+   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
+   set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+
 else()
   if(WITH_NGRAPH)
     if(NOT WIN32)

paddle/fluid/framework/details/CMakeLists.txt

@@ -12,12 +12,19 @@ cc_library(multi_devices_graph_check_pass SRCS multi_devices_graph_check_pass.cc
 
 cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)
 
+if(WITH_DISTRIBUTE)
+    if(NOT WITH_GRPC)
+        set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
+        set_source_files_properties(reduce_op_handle.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    endif()
+endif()
+
 if(WITH_GPU)
     nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
             dynload_cuda variable_visitor)
     if(WITH_DISTRIBUTE)
         nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
-            ddim dynload_cuda selected_rows_functor sendrecvop_grpc)
+            ddim dynload_cuda selected_rows_functor sendrecvop_rpc)
     else()
         nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
             ddim dynload_cuda selected_rows_functor)
@@ -30,7 +37,7 @@ else()
              variable_visitor)
     if(WITH_DISTRIBUTE)
         cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
-            ddim selected_rows_functor sendrecvop_grpc)
+            ddim selected_rows_functor sendrecvop_rpc)
     else()
         cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
             ddim selected_rows_functor)

paddle/fluid/framework/executor.cc

@@ -157,9 +157,9 @@ void Executor::Close() {
 #ifdef PADDLE_WITH_DISTRIBUTE
   // TODO(typhoonzero): complete message will need to use real trainer_id,
   // except 0.
-  ::paddle::operators::distributed::RPCClient::GetInstance<
-      ::paddle::operators::distributed::GRPCClient>(0)
-      ->SendComplete();
+  auto client =
+      paddle::operators::distributed::RPCClient::GetInstance<RPCCLIENT_T>(0);
+  client->SendComplete();
 #endif
 }
 

paddle/fluid/operators/distributed/CMakeLists.txt

@@ -12,44 +12,51 @@ configure_file(send_recv.proto.in ${CMAKE_CURRENT_SOURCE_DIR}/send_recv.proto @O
 set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
 
 if(WITH_GRPC)
-  grpc_library(sendrecvop_grpc SRCS grpc_bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc
+  grpc_library(sendrecvop_rpc SRCS grpc_bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc
         request_handler_impl.cc rpc_client.cc rpc_server.cc grpc_server.cc variable_response.cc grpc_variable_response.cc grpc_serde.cc collective_client.cc collective_server.cc
       PROTO send_recv.proto 
       DEPS lod_tensor selected_rows_functor memory)
 
   set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 
   cc_test(grpc_serde_test SRCS grpc_serde_test.cc 
-    DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL)
+    DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_rpc scope profiler math_function SERIAL)
 
   cc_test(rpc_server_test SRCS rpc_server_test.cc
-    DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor  proto_desc lookup_sparse_table_op SERIAL)
+    DEPS sendrecvop_rpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor  proto_desc lookup_sparse_table_op SERIAL)
 
   cc_test(varhandle_test SRCS varhandle_test.cc DEPS profiler)
 
   if(WITH_GPU)
   cc_test(collective_server_test SRCS collective_server_test.cc 
-      DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor
+      DEPS sendrecvop_rpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor
       selected_rows_functor  scope math_function SERIAL)
   endif()
 
-  cc_library(parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_grpc memory)
+  cc_library(parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_rpc memory)
 else()
-  set_source_files_properties(brpc_server.cc brpc_client.cc rpc_server_test.cc brpc_serde_test.cc
-      brpc_variable_response.cc brpc_sendrecvop_utils.cc brpc_rdma_pool.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+  set_source_files_properties(brpc_server.cc parameter_prefetch.cc brpc_client.cc rpc_server_test.cc brpc_serde_test.cc
+      brpc_variable_response.cc brpc_sendrecvop_utils.cc brpc_rdma_pool.cc collective_server.cc collective_server_test.cc
+      collective_client.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 
-  brpc_library(sendrecvop_brpc SRCS brpc_client.cc brpc_server.cc rpc_server.cc rpc_client.cc request_handler_impl.cc brpc_sendrecvop_utils.cc
-      brpc_variable_response.cc variable_response.cc sendrecvop_utils.cc brpc_rdma_pool.cc
+  brpc_library(sendrecvop_rpc SRCS brpc_client.cc brpc_server.cc rpc_server.cc rpc_client.cc request_handler_impl.cc brpc_sendrecvop_utils.cc
+      brpc_variable_response.cc variable_response.cc sendrecvop_utils.cc brpc_rdma_pool.cc collective_client.cc collective_server.cc
     PROTO send_recv.proto
     DEPS lod_tensor selected_rows memory)
 
-  cc_library(parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_brpc memory)
+  cc_library(parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_rpc memory)
 
-  set(brpc_test_depends sendrecvop_brpc brpc ssl crypto protobuf leveldb gflags glog executor proto_desc lookup_table_op snappystream snappy)
+  set(brpc_test_depends sendrecvop_rpc brpc ssl crypto protobuf leveldb gflags glog executor
+      proto_desc lookup_sparse_table_op snappystream snappy zlib)
 
-  cc_test(brpc_server_test SRCS rpc_server_test.cc
+  cc_test(rpc_server_test SRCS rpc_server_test.cc
       DEPS ${brpc_test_depends} SERIAL)
 
   cc_test(brpc_serde_test SRCS brpc_serde_test.cc
       DEPS ${brpc_test_depends} SERIAL)
+
+  if(WITH_GPU)
+  cc_test(collective_server_test SRCS collective_server_test.cc 
+      DEPS ${brpc_test_depends} selected_rows_functor  scope math_function SERIAL)
+  endif()
 endif()