1 /* Copyright 2016 The TensorFlow Authors. All Rights Reserved. 2 3 Licensed under the Apache License, Version 2.0 (the "License"); 4 you may not use this file except in compliance with the License. 5 You may obtain a copy of the License at 6 7 http://www.apache.org/licenses/LICENSE-2.0 8 9 Unless required by applicable law or agreed to in writing, software 10 distributed under the License is distributed on an "AS IS" BASIS, 11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 See the License for the specific language governing permissions and 13 limitations under the License. 14 ==============================================================================*/ 15 16 #ifndef TENSORFLOW_CORE_DISTRIBUTED_RUNTIME_RPC_GRPC_SERVER_LIB_H_ 17 #define TENSORFLOW_CORE_DISTRIBUTED_RUNTIME_RPC_GRPC_SERVER_LIB_H_ 18 19 // GrpcServer manages the lifecycle of an Eager, Worker and Master service. 20 21 #include <memory> 22 23 #include "grpcpp/grpcpp.h" 24 #include "grpcpp/security/credentials.h" 25 #include "tensorflow/core/common_runtime/eager/context.h" 26 #include "tensorflow/core/common_runtime/process_util.h" 27 #include "tensorflow/core/common_runtime/stats_publisher_interface.h" 28 #include "tensorflow/core/distributed_runtime/master_env.h" 29 #include "tensorflow/core/distributed_runtime/rpc/async_service_interface.h" 30 #include "tensorflow/core/distributed_runtime/rpc/grpc_channel.h" 31 #include "tensorflow/core/distributed_runtime/rpc/grpc_worker_cache.h" 32 #include "tensorflow/core/distributed_runtime/rpc/grpc_worker_service.h" 33 #include "tensorflow/core/distributed_runtime/server_lib.h" 34 #include "tensorflow/core/distributed_runtime/session_mgr.h" 35 #include "tensorflow/core/distributed_runtime/worker_env.h" 36 #include "tensorflow/core/framework/collective.h" 37 #include "tensorflow/core/framework/op.h" 38 #include "tensorflow/core/platform/env.h" 39 #include "tensorflow/core/profiler/profiler_service.grpc.pb.h" 40 41 namespace tensorflow { 42 43 class GrpcWorker; 44 class Master; 45 46 // function that creates a RendezvousMgr. 47 typedef std::function<RendezvousMgrInterface*(const WorkerEnv*)> 48 RendezvousMgrCreationFunction; 49 50 // function that creates a CollectiveExecutorMgr. 51 typedef std::function<CollectiveExecutorMgrInterface*( 52 const ConfigProto&, const WorkerEnv*, WorkerCacheInterface*)> 53 CollectiveMgrCreationFunction; 54 55 // function that registers a service to the server. The service needs to 56 // be registered before builder.BuildAndStart(). 57 typedef std::function<void(const WorkerEnv*, ::grpc::ServerBuilder*)> 58 ServiceInitFunction; 59 60 // function that creates a grpc based worker implementation. 61 typedef std::function<std::unique_ptr<GrpcWorker>(WorkerEnv*, 62 const ConfigProto& config)> 63 WorkerCreationFunction; 64 65 struct GrpcServerOptions { 66 ServiceInitFunction service_func = nullptr; 67 RendezvousMgrCreationFunction rendezvous_mgr_func = nullptr; 68 CollectiveMgrCreationFunction collective_mgr_func = nullptr; 69 WorkerCreationFunction worker_func = nullptr; 70 StatsPublisherFactory stats_factory = CreateNoOpStatsPublisher; 71 GrpcWorkerServiceOptions worker_service_options; 72 DeviceMgr* local_device_mgr = nullptr; 73 }; 74 75 class GrpcServer : public ServerInterface { 76 protected: 77 GrpcServer(const ServerDef& server_def, Env* env); 78 GrpcServer(const ServerDef& server_def, DeviceMgr* local_device_mgr, 79 Env* env); 80 // Allow children classes to override this and provide custom args to the 81 // server before it is constructed. Default behavior is to do nothing. 82 // requested_port provides the port requested by caller as bound_port() is 83 // not available till BuildAndStart has been called. MaybeMutateBuilder(::grpc::ServerBuilder * builder,int requested_port)84 virtual void MaybeMutateBuilder(::grpc::ServerBuilder* builder, 85 int requested_port) {} 86 87 public: 88 static Status Create(const ServerDef& server_def, Env* env, 89 std::unique_ptr<ServerInterface>* out_server); 90 static Status Create(const ServerDef& server_def, Env* env, 91 std::unique_ptr<GrpcServer>* out_server); 92 // Reuse the local_device_mgr. 93 static Status Create(const ServerDef& server_def, Env* env, 94 DeviceMgr* local_device_mgr, 95 std::unique_ptr<ServerInterface>* out_server); 96 97 // Destruction is only supported in the factory method. Clean 98 // shutdown is not currently implemented for this server type. 99 virtual ~GrpcServer(); 100 101 // Implementations of ServerInterface methods. 102 Status Start() override; 103 Status Stop() override; 104 Status Join() override; 105 const string target() const override; 106 worker_env()107 WorkerEnv* worker_env() override { return &worker_env_; } master_env()108 MasterEnv* master_env() override { return &master_env_; } 109 110 // Add master eager context to local eager service in order to handle enqueue 111 // requests from remote workers. 112 Status AddMasterEagerContextToEagerService( 113 const tensorflow::uint64 context_id, 114 tensorflow::EagerContext* context) override; 115 // Update the set of workers that can be reached by the GRPC server 116 Status UpdateServerDef(const ServerDef& server_def) override; 117 // Pass coordination service agent instance to server's RPC handler 118 Status SetCoordinationServiceAgentInstance( 119 CoordinationServiceAgent* agent) override; 120 // TODO(hanyangtay): Remove this method once gRPC server clean shutdown is 121 // supported. 122 Status StopCoordinationService() override; 123 124 protected: 125 virtual Status GetHostAndPort(const ServerDef& server_def, string* host_name, 126 int* port) const; 127 Status Init(const GrpcServerOptions& opts = GrpcServerOptions()); 128 129 // A subclass can override this method to support secure credentials. 130 virtual std::shared_ptr<::grpc::ServerCredentials> GetServerCredentials( 131 const ServerDef& server_def) const; 132 133 virtual ChannelCreationFunction GetChannelCreationFunction() const; 134 135 virtual std::unique_ptr<Master> CreateMaster(MasterEnv* master_env); 136 137 // Creates a WorkerCacheInterface for a session. 138 virtual Status WorkerCacheFactory(const WorkerCacheFactoryOptions& options, 139 WorkerCacheInterface** worker_cache); 140 141 // Override to return extra services to be brought up and managed along with 142 // the standard {master, worker, eager} services. The map key is an aribtrary 143 // string and the value is a pointer to the service to be brought up. 144 // Ownership of the pointer is transferred to GrpcServer after this call 145 // returns, and the service will be destroyed during the destruction of 146 // GrpcServer. Each service will have its HandleRPCsLoop called in a separate 147 // thread. An example usage would be to add a RDMA based partial worker 148 // service to offload tensor and data buffer transfers. ExtraServices(::grpc::ServerBuilder *)149 virtual std::map<std::string, AsyncServiceInterface*> ExtraServices( 150 ::grpc::ServerBuilder*) { 151 return {}; 152 } 153 GetExtraServices()154 virtual std::map<std::string, AsyncServiceInterface*> GetExtraServices() { 155 return extra_services_; 156 } 157 158 // Parses a WorkerCacheFactoryOptions into a GrpcChannelSpec. 159 Status ParseChannelSpec(const WorkerCacheFactoryOptions& options, 160 GrpcChannelSpec* channel_spec); 161 162 // Returns the port to which this server is bound. 163 // This method may only be called after `this->Init()` returns successfully. bound_port()164 int bound_port() const { return bound_port_; } 165 166 // Returns hostname. host_name()167 const string& host_name() const { return host_name_; } 168 server_def()169 const ServerDef& server_def() const { return server_def_; } worker_impl()170 GrpcWorker* worker_impl() const { return worker_impl_.get(); } grpc_worker_env()171 GrpcWorkerEnv* grpc_worker_env() const { return grpc_worker_env_.get(); } 172 173 private: 174 Env* env_; 175 176 // The port to which this server is bound. 177 int bound_port_ = 0; 178 179 // The host name of this server 180 string host_name_; 181 182 // Guards server configuration, server, and state. 183 mutex mu_; 184 185 // Represents the current state of the server, which changes as follows: 186 // 187 // Join() Join() 188 // ___ ___ 189 // Start() \ / Stop() \ / 190 // NEW ---------> STARTED --------> STOPPED 191 // \ / 192 // \________________________/ 193 // Stop(), Join() 194 enum State { NEW, STARTED, STOPPED }; 195 State state_ TF_GUARDED_BY(mu_); 196 197 // Implementation of a TensorFlow master, and RPC polling thread. 198 MasterEnv master_env_; 199 std::unique_ptr<Master> master_impl_; 200 AsyncServiceInterface* master_service_ = nullptr; 201 std::unique_ptr<Thread> master_thread_ TF_GUARDED_BY(mu_); 202 203 std::map<std::string, AsyncServiceInterface*> extra_services_; 204 std::vector<std::unique_ptr<Thread>> extra_service_threads_ 205 TF_GUARDED_BY(mu_); 206 207 // Implementation of a TensorFlow worker, and RPC polling thread. 208 WorkerEnv worker_env_; 209 std::unique_ptr<const DeviceMgr> owned_device_manager_; 210 std::unique_ptr<GrpcWorker> worker_impl_; 211 AsyncServiceInterface* worker_service_ = nullptr; 212 std::unique_ptr<Thread> worker_thread_ TF_GUARDED_BY(mu_); 213 std::unique_ptr<GrpcWorkerEnv> grpc_worker_env_; 214 215 // TensorFlow Eager implementation, and RPC polling thread. 216 AsyncServiceInterface* eager_service_ = nullptr; 217 std::unique_ptr<Thread> eager_thread_ TF_GUARDED_BY(mu_); 218 std::shared_ptr<WorkerSession> worker_session_; 219 220 // Experimental coordination service implementation, and RPC polling thread. 221 AsyncServiceInterface* coordination_service_ = nullptr; 222 std::unique_ptr<Thread> coordination_thread_ TF_GUARDED_BY(mu_); 223 224 // TensorFlow profiler service implementation. 225 std::unique_ptr<grpc::ProfilerService::Service> profiler_service_ = nullptr; 226 227 // The overall server configuration. 228 ServerDef server_def_ TF_GUARDED_BY(mu_); 229 230 std::unique_ptr<::grpc::Server> server_ TF_GUARDED_BY(mu_); 231 }; 232 233 } // namespace tensorflow 234 235 #endif // TENSORFLOW_CORE_DISTRIBUTED_RUNTIME_RPC_GRPC_SERVER_LIB_H_ 236