[源码解析] TensorFlow 分布式环境(5) --- Session

[源码解析] TensorFlow 分布式环境(5) --- Session

目录

• [源码解析] TensorFlow 分布式环境(5) --- Session
  • 1. 概述
    • 1.1 Session 分类
    • 1.2 会话流程
      • 1.2.1 MasterSession 生命周期
      • 1.2.2 WorkerSession 生命周期
  • 2. GrpcSession
    • 2.1 定义
    • 2.2 注册&工厂类
    • 2.3 创建GrpcSession
    • 2.4 创建MasterSession
      • 2.4.1 GrpcRemoteMaster::CreateSession
      • 2.4.2 GrpcMasterService::CreateSessionHandler
      • 2.4.3 Master::CreateSession
  • 3. MasterSession
    • 3.1 定义
    • 3.2 创建
      • 3.2.1 创建计算图
      • 3.2.2 创建 WorkerSession
        • GrpcRemoteWorker
        • GrpcWorkerService
  • 4. WorkerSession
    • 4.1 SessionMgr
      • 4.1.1 定义
      • 4.1.2 建立 Session
      • 4.1.3 注册图
    • 4.2 WorkerSession
      • 4.2.1 定义
  • 0xFF 参考

会话机制是TensorFlow 分布式运行时的核心，我们接下来按照从 Client 到 worker 的流程，把 Session 机制从前到后走一边。

本系列其他文章是：

[翻译] TensorFlow 分布式之论文篇 "TensorFlow : Large-Scale Machine Learning on Heterogeneous Distributed Systems"

[翻译] TensorFlow 分布式之论文篇 "Implementation of Control Flow in TensorFlow"

[源码解析] TensorFlow 分布式环境(1) --- 总体架构

[源码解析] TensorFlow 分布式环境(2)---Master 静态逻辑

[源码解析] TensorFlow 分布式环境(3)--- Worker 静态逻辑

[源码解析] TensorFlow 分布式环境(4) --- WorkerCache

1. 概述

1.1 Session 分类

分布式模式由如下 sessions 彼此协作完成了会话控制，其中：

• GrpcSession 位于 Client 之上，控制 Client 的会话生命周期；
• MasterSession 位于 Master 之上，可能存在多个 Client 同时接入到同一个 Master，Master 会为每个 Client 构建一个 MasterSession。MasterSession 控制 Master 的会话生命周 期；
• WorkerSession 位于 Worker 之上，可能存在多个 Master 接入到同一个 Worker，Worker 会为每个 Master 创建一个 WorkerSession。WorkerSession 控制 Worker 的会话生命周期；

如下图所示，这里 Master 和 Worker 都是一个 Server，每个 Server 之上运行一个 MasterService，一个 WorkerService，每个 Server 可能会扮演不同角色，具体取决于用户如何配置计算图和集群。因为存在这种两层一对多关系，为了区别这种不同的数据流和控制关系，有逻辑关系的这三个 session 绑定在同一个 session_handle 之上，每个 session_handle 标示一条完整的数据流。

图 1 Session 关系

1.2 会话流程

我们从 GrpcSession 入手，其基本功能如下：

• 创建会话
  • 获取远端设备集;
  • 在 Master 之上创建 MasterSession;
  • 在各个 Worker 之上创建 WorkerSession;
• 迭代执行
  • 启动执行;
  • 图分裂;
  • 注册子图;
  • 运行子图;
• 关闭会话
  • 关闭 MasterSession
  • 关闭 WorkerSession;

1.2.1 MasterSession 生命周期

在分布式模式下，Master 运行时被 MasterSession 控制，其生命周期如下图所示。

图 2 MasterSession 生命周期

1.2.2 WorkerSession 生命周期

在分布式模式下，Worker 运行时由 WorkerSession 控制，其生命周期如下图所示。

图 3 WorkerSession 生命周期

2. GrpcSession

GrpcSession 是 tensorflow::grpc::MasterService 的简单封装。其使用远程设备集作为计算资源，使用 grpc 作为远端调用机制，让调用者在远端设备上对 TensorFlow 图进行计算。

2.1 定义

我们依然只给出成员变量定义和部分重要函数，其就是利用 master_ 对 tensorflow::grpc::MasterService 进行调用。

class GrpcSession : public Session {
  // 有多种创建方式
  Status Create(const GraphDef& graph) override;
  Status Create(const RunOptions& run_options, const GraphDef& graph) override;
  Status Create(GraphDef&& graph) override;
  Status Create(const RunOptions& run_options, GraphDef&& graph) override;  
  
 private:
  const SessionOptions options_;
  std::unique_ptr<MasterInterface> master_;
  mutex mu_;

  // handle_ returned by the master to identify this session.
  string handle_ TF_GUARDED_BY(mu_);

  // The current version of the graph.
  int64_t current_graph_version_ TF_GUARDED_BY(mu_);

  bool is_local_ = false;
};

2.2 注册&工厂类

GrpcSession 的使用是通过工厂类完成，比如：

Status NewSession(const SessionOptions& options, Session** out_session) {
  SessionFactory* factory;
  Status s = SessionFactory::GetFactory(options, &factory);
  if (!s.ok()) {
    *out_session = nullptr;
    return s;
  }
  // Starts exporting metrics through a platform-specific monitoring API (if
  // provided). For builds using "tensorflow/core/platform/default", this is
  // currently a no-op.
  session_created->GetCell()->Set(true);
  s = factory->NewSession(options, out_session);
  if (!s.ok()) {
    *out_session = nullptr;
  }
  return s;
}

GrpcSession 由 GrpcSessionFactory 来多态创建，如果 protocal 使用了"grpc://"，就会产生 GrpcSession。而 GrpcSessionFactory 会实现注册到系统之上。

const char* const kSchemePrefix = "grpc://";
const size_t kSchemePrefixLength = strlen(kSchemePrefix);

class GrpcSessionFactory : public SessionFactory {
 public:
  bool AcceptsOptions(const SessionOptions& options) override {
    return absl::StartsWith(options.target, kSchemePrefix);
  }

  Status NewSession(const SessionOptions& options,
                    Session** out_session) override {
    std::unique_ptr<GrpcSession> session;
    TF_RETURN_IF_ERROR(GrpcSession::Create(options, &session));
    *out_session = session.release();
    return Status::OK();
  }

  // Invokes the session specific static method to reset containers.
  Status Reset(const SessionOptions& options,
               const std::vector<string>& containers) override {
    return GrpcSession::Reset(options, containers);
  }
};

class GrpcSessionRegistrar {
 public:
  GrpcSessionRegistrar() {
    SessionFactory::Register("GRPC_SESSION", new GrpcSessionFactory());
  }
};
static GrpcSessionRegistrar registrar;

2.3 创建GrpcSession

GrpcSession::Create 方法完成了获取工作。Client 通过 GrpcSession 调用 Master Service，但是具体如何与 Master Service 交互？则通过 MasterInterface。

所以说，这里最重要的就是如何构建 MasterInterface 实例。我们前文提到过，MasterInterface有两种实现，都是用来和 Master service 进行通信，分别对应了不同的应用场景。

• LocalMaster 用于进程间的直接通信，此时 Client 和 Master 在同一个进程。
• GrpcRemoteMaster 则使用 Grpc 来和 Master service 进行通信，此时Client 和 Master 分别部署在两个不同进程。GrpcRemoteMaster 其实就实现了 gRPC 客户端，它通过 Stub 访问远端 Master 上的 MasterService 服务。

图上两个矩形封装的 Master 代表实际的 Master 类，此类实现了具体 Master 功能。

图 1 Master 逻辑关系

从下面代码可以看到，GrpcSession 会依据 options.target 来决定如何创建，options.target 一般就是"grpc://"，如果通过 LocalMaster::Lookup 方法得到 LocalMaster 类，就直接使用，如果没有找到，就使用 NewGrpcMaster 来生成一个 GrpcRemoteMaster。

/* static */
Status GrpcSession::Create(const SessionOptions& options,
                           std::unique_ptr<GrpcSession>* out_session) {
  std::unique_ptr<GrpcSession> session(new GrpcSession(options));
  std::unique_ptr<MasterInterface> master;
  // For testing, we enable the client to disable the use of the local
  // master registry, so that the RPC stack is exercised.
  if (!options.config.rpc_options().use_rpc_for_inprocess_master()) {
    master = LocalMaster::Lookup(options.target);
  }
  if (!master) {
    SharedGrpcChannelPtr master_channel;
    TF_RETURN_IF_ERROR(
        NewHostPortGrpcChannel(options.target.substr(kSchemePrefixLength),
                               &options.config.rpc_options(), &master_channel));
    master.reset(NewGrpcMaster(master_channel));
  } else {
    session->is_local_ = true;
  }
  session->SetRemoteMaster(std::move(master));
  *out_session = std::move(session);
  return Status::OK();
}

2.4 创建MasterSession

在 GrpcSession 创建之后，系统会接着创建 MasterSession，这是通过 GrpcSession::Create(graph_def) 完成的。GrpcSession::Create(graph_def) 会构建 CreateSessionRequst 消息，然后通过 GrpcRemoteMaster 把初始计算图发给 Master。Master 收到 CreateSessionRequst 消息之后就构建相应的 MasterSession，然后返回 CreateSessionResponse 给 GrpcSession，消息包括。

• 该 MasterSession 的 session_handle。用于标识 Master 侧的 MasterSession 实例
• 初始计算图的版本号 graph_version。用于后续发起 ExtendSession 操作，比如往原始的计算图中追加新的节点。

图 2 创建MasterSession

具体代码如下，首先是两个 create 方法，其最终调用到 CreateImpl。

Status GrpcSession::Create(const RunOptions& run_options,
                           const GraphDef& graph) {
  return Create(run_options, GraphDef(graph));
}

Status GrpcSession::Create(GraphDef&& graph) {
  CallOptions call_options;
  call_options.SetTimeout(options_.config.operation_timeout_in_ms());
  return CreateImpl(&call_options, std::move(graph));
}

CreateImpl 方法如下：

Status GrpcSession::CreateImpl(CallOptions* call_options, GraphDef graph) {
  {
    mutex_lock l(mu_);
    if (!handle_.empty()) {
      return errors::InvalidArgument("A session is alive.");
    }
  }
  CreateSessionRequest req;
  *req.mutable_config() = options_.config;
  req.mutable_graph_def()->Swap(&graph);
  req.set_target(options_.target);
  ReEncodeConsts(req.mutable_graph_def());
  CreateSessionResponse resp;
  Status s = master_->CreateSession(call_options, &req, &resp);
  if (s.ok()) {
    SetHandleAndGraphVersion(resp.session_handle(), resp.graph_version());
  }
  return s;
}

2.4.1 GrpcRemoteMaster::CreateSession

GrpcRemoteMaster 是位于 Client 的 gRPC 客户端实现，它的 CreateSession 方法只是通过 gRPC stub 来调用 远端服务 MasterService 的 CreateSession 接口，其实就是发送一个 CreateSessionRequest 请求。

Status CreateSession(CallOptions* call_options,
                     const CreateSessionRequest* request,
                     CreateSessionResponse* response) override {
  return CallWithRetry(call_options, request, response,
                       &MasterServiceStub::CreateSession);
}

2.4.2 GrpcMasterService::CreateSessionHandler

GrpcMasterService 是 Master 提供的 gRPC 服务，收到 CreateSessionRequest 消息之后， 服务调用 GrpcMasterService::CreateSessionHandler 来处理消息，而真正业务处理是由 master_impl_（Master 类的实例）来完成，就是调用了 Master::CreateSession。

当 master_impl_ 处理完成后，会向 Client 返回 CreateSessionResponse 响应。

// RPC handler for creating a session.
void CreateSessionHandler(
    MasterCall<CreateSessionRequest, CreateSessionResponse>* call) {
  CreateSessionRequest* rewritten_req = new CreateSessionRequest;
  rewritten_req->mutable_config()->MergeFrom(default_session_config_);
  rewritten_req->MergeFrom(call->request);
  master_impl_->CreateSession(rewritten_req, &call->response,
                              [call, rewritten_req](const Status& status) {
                                call->SendResponse(ToGrpcStatus(status));
                                delete rewritten_req;
                              });
  ENQUEUE_REQUEST(CreateSession, true);
}

2.4.3 Master::CreateSession

Master::CreateSession 会从线程池之中拿到一个线程，在线程之中会做如下处理：

• 如果定义了 clust_spec，则按照配置寻找所有的 worker。
• 获取远端设备。
• 获取远端worker。
• 通过factory 建立 MasterSession。
• 利用 worker_cache_factory，让 MasterSession 建立 WorkerSession 会话。
• 通过 sessions_.insert 在 Master 内部的 <session_handle, MasterSession> 二元组之中保存对应关系，这样后续 Master 就可以通过 session_handle 得到对应的 MasterSession。

void Master::CreateSession(const CreateSessionRequest* req,
                           CreateSessionResponse* resp, MyClosure done) {
  SchedClosure([this, req, resp, done]() {
    Status status;
    WorkerCacheFactoryOptions worker_cache_factory_options;
    string grpc_protocol("grpc");
    worker_cache_factory_options.protocol = &grpc_protocol;
    auto call_done = gtl::MakeCleanup([&status, &done] { done(status); });
    status = ValidateExternalGraphDefSyntax(req->graph_def());
    if (!status.ok()) return;

    // The following 4 variables are set differently, depending on whether this
    // session uses a client-provided clusterspec or not.
    WorkerCacheInterface* worker_cache = nullptr;
    // Note: worker_cache_ptr will be null except if this session is using a
    // client-supplied ClusterDef (ClusterSpec propagation).
    std::unique_ptr<WorkerCacheInterface> worker_cache_ptr;
    std::unique_ptr<DeviceSet> device_set;
    // TODO(saeta): Convert to std::make_unique when available.
    std::unique_ptr<std::vector<std::unique_ptr<Device>>> remote_devices(
        new std::vector<std::unique_ptr<Device>>());

    if (req->config().has_cluster_def()) { // 如果定义了集群
      worker_cache_factory_options.cluster_def = &req->config().cluster_def();

      // Set the server_def's job_name and task_index fields.
      string normalized_string;
      string grpc_protocol(kGrpcProtocol);
      if (req->target().compare(0, grpc_protocol.length(), grpc_protocol) ==
          0) {
        normalized_string =
            req->target().substr(grpc_protocol.length(), string::npos);
      } else {
        normalized_string = req->target();
      }
      for (auto&& job : req->config().cluster_def().job()) {
        for (auto&& task : job.tasks()) {
          if (task.second == normalized_string) {
            if (worker_cache_factory_options.job_name != nullptr) {
              return;
            }
            if (env_->local_devices[0]->parsed_name().job == job.name() &&
                env_->local_devices[0]->parsed_name().task == task.first) {
              return;
            }
            worker_cache_factory_options.job_name = &job.name();
            worker_cache_factory_options.task_index = task.first;
          }
        }
      }
      worker_cache_factory_options.rpc_options = &req->config().rpc_options();
      // Create the worker cache from the computed server_def.
      status = env_->worker_cache_factory(worker_cache_factory_options,
                                          &worker_cache);
      if (!status.ok()) return;
      worker_cache_ptr = std::unique_ptr<WorkerCacheInterface>(worker_cache);
      // Ping all the workers and build the list of devices that the
      // session will use.
      // 获取设备
      status =
          DeviceFinder::GetRemoteDevices(req->config().device_filters(), env_,
                                         worker_cache, remote_devices.get());
      if (!status.ok()) return;
      device_set.reset(new DeviceSet);
      for (auto&& d : *remote_devices) {
        device_set->AddDevice(d.get());
        DeviceNameUtils::ParsedName name = d->parsed_name();
        if (name.job == *worker_cache_factory_options.job_name &&
            name.task == worker_cache_factory_options.task_index &&
            name.type == "CPU" && name.id == 0) {
          device_set->set_client_device(d.get());
        }
      }
    } else { // 没有集群
      worker_cache = env_->worker_cache;
      // Ping all the workers and build the list of devices that the
      // session will use.
      // 获取远端设备
      status =
          DeviceFinder::GetRemoteDevices(req->config().device_filters(), env_,
                                         worker_cache, remote_devices.get());
      if (!status.ok()) return;
      device_set.reset(new DeviceSet);
      for (auto&& d : *remote_devices) {
        device_set->AddDevice(d.get());
      }
      int num_local_devices = 0;
      for (Device* d : env_->local_devices) {
        device_set->AddDevice(d);
        if (num_local_devices == 0) {
          // Uses the first local device as the client device.
          device_set->set_client_device(d);
        }
        num_local_devices++;
      }
    }

    SessionOptions options;
    options.config = req->config();

    // 获取远端worker
    std::vector<string> filtered_worker_list;
    DeviceFinder::GetRemoteWorkers(req->config().device_filters(), env_,
                                   worker_cache, &filtered_worker_list);

    // 通过factory找到会话
    MasterSession* session = env_->master_session_factory(
        options, env_, std::move(remote_devices), std::move(worker_cache_ptr),
        std::move(device_set), std::move(filtered_worker_list));

    GraphDef* gdef =
        const_cast<CreateSessionRequest*>(req)->mutable_graph_def();

    // 建立会话，把图传给会话
    status = session->Create(std::move(*gdef), worker_cache_factory_options);
    if (!status.ok()) {
      session->Close().IgnoreError();
      session->Unref();
      return;
    }
    resp->set_session_handle(session->handle());
    // Insert into the session map, which takes ownership of the session.
    {
      mutex_lock l(mu_);
      CHECK(sessions_.insert({session->handle(), session}).second);
    }
  });
}

3. MasterSession

MasterSession 位于 Master 之上，可能存在多个 Client 同时接入到同一个 Master，Master 会为每个 Client 构建一个 MasterSession。MasterSession 控制 Master 的会话生命周 期。

3.1 定义

MasterSession 的定义如下。

// MasterSession wraps ClientGraph in a reference counted object.
// This way, MasterSession can clear up the cache mapping Run requests to
// compiled graphs while the compiled graph is still being used.
class MasterSession::ReffedClientGraph : public core::RefCounted {
 public:
  ReffedClientGraph(const string& handle, const BuildGraphOptions& bopts,
                    std::unique_ptr<ClientGraph> client_graph,
                    const SessionOptions& session_opts,
                    const StatsPublisherFactory& stats_publisher_factory,
                    bool is_partial, WorkerCacheInterface* worker_cache,
                    bool should_deregister)
      : session_handle_(handle),
        bg_opts_(bopts),
        client_graph_before_register_(std::move(client_graph)),
        session_opts_(session_opts),
        is_partial_(is_partial),
        callable_opts_(bopts.callable_options),
        worker_cache_(worker_cache),
        should_deregister_(should_deregister),
        collective_graph_key_(
            client_graph_before_register_->collective_graph_key) {
    VLOG(1) << "Created ReffedClientGraph for node with "
            << client_graph_before_register_->graph.num_node_ids();

    stats_publisher_ = stats_publisher_factory(handle, bopts, session_opts);

    // Initialize a name to node map for processing device stats.
    for (Node* n : client_graph_before_register_->graph.nodes()) {
      name_to_node_details_.emplace(
          n->name(),
          NodeDetails(n->type_string(),
                      strings::StrCat(
                          "(", absl::StrJoin(n->requested_inputs(), ", "))));
    }
  }

  ~ReffedClientGraph() override {
    if (should_deregister_) {
      DeregisterPartitions();
    } else {
      for (Part& part : partitions_) {
        worker_cache_->ReleaseWorker(part.name, part.worker);
      }
    }
  }

 private:
  const string session_handle_;
  const BuildGraphOptions bg_opts_;

  // NOTE(mrry): This pointer will be null after `RegisterPartitions()` returns.
  std::unique_ptr<ClientGraph> client_graph_before_register_ TF_GUARDED_BY(mu_);
  const SessionOptions session_opts_;
  const bool is_partial_;
  const CallableOptions callable_opts_;
  WorkerCacheInterface* const worker_cache_;  // Not owned.

  struct NodeDetails {
    explicit NodeDetails(string type_string, string detail_text)
        : type_string(std::move(type_string)),
          detail_text(std::move(detail_text)) {}
    const string type_string;
    const string detail_text;
  };
  std::unordered_map<string, NodeDetails> name_to_node_details_;

  const bool should_deregister_;
  const int64_t collective_graph_key_;
  std::atomic<int64_t> execution_count_ = {0};

  // Graph partitioned into per-location subgraphs.
  struct Part {
    // Worker name.
    string name;

    // Maps feed names to rendezvous keys. Empty most of the time.
    std::unordered_map<string, string> feed_key;

    // Maps rendezvous keys to fetch names. Empty most of the time.
    std::unordered_map<string, string> key_fetch;

    // The interface to the worker. Owned.
    WorkerInterface* worker = nullptr;

    // After registration with the worker, graph_handle identifies
    // this partition on the worker.
    string graph_handle;

    Part() : feed_key(3), key_fetch(3) {}
  };

  // partitions_ is immutable after RegisterPartitions() call
  // finishes.  RunPartitions() can access partitions_ safely without
  // acquiring locks.
  std::vector<Part> partitions_;

  mutable mutex mu_;

  // Partition initialization and registration only needs to happen
  // once. `!client_graph_before_register_ && !init_done_.HasBeenNotified()`
  // indicates the initialization is ongoing.
  Notification init_done_;

  // init_result_ remembers the initialization error if any.
  Status init_result_ TF_GUARDED_BY(mu_);

  std::unique_ptr<StatsPublisherInterface> stats_publisher_;
};

3.2 创建

MasterSession::Create(graph_def) 的工作如下：

• 调用 MakeForBaseGraph 来初始化计算图，并生成 SimpleGraphExecutionState 实例;
• 调用 CreateWorkerSessions，如果动态配置集群，则广播通知给所有 Worker，让其创建对应的 WorkerSession。

Status MasterSession::Create(GraphDef&& graph_def,
                             const WorkerCacheFactoryOptions& options) {
  if (session_opts_.config.use_per_session_threads() ||
      session_opts_.config.session_inter_op_thread_pool_size() > 0) {
    return errors::InvalidArgument(
        "Distributed session does not support session thread pool options.");
  }
  if (session_opts_.config.graph_options().place_pruned_graph()) {
    session_opts_.config.mutable_graph_options()->set_place_pruned_graph(false);
  }

  GraphExecutionStateOptions execution_options;
  execution_options.device_set = devices_.get();
  execution_options.session_options = &session_opts_;
  {
    mutex_lock l(mu_);
    TF_RETURN_IF_ERROR(GraphExecutionState::MakeForBaseGraph(
        std::move(graph_def), execution_options, &execution_state_));
  }
  should_delete_worker_sessions_ = true;
  return CreateWorkerSessions(options);
}

3.2.1 创建计算图

这里会构建 GraphExecutionState，依据 GraphDef 构建对应的 FullGraph。

GraphDef 是原始图结构，ConvertGraphDefToGraph 完成从 GraphDef 到 Graph 的格式转换，GraphDef 包含了图的元数据，Graph 则包含图结构的其他信息，被运行时系统所使用。

/* static */ Status GraphExecutionState::MakeForBaseGraph(
    GraphDef&& graph_def, const GraphExecutionStateOptions& options,
    std::unique_ptr<GraphExecutionState>* out_state) {

  auto flib_def = absl::make_unique<FunctionLibraryDefinition>(
      OpRegistry::Global(), graph_def.library());

  TF_RETURN_IF_ERROR(AddDefaultAttrsToGraphDef(&graph_def, *flib_def, 0));

  if (options.session_options->config.graph_options().place_pruned_graph() ||
      !options.session_options->config.experimental()
           .optimize_for_static_graph()) {
    auto ret = absl::WrapUnique(new GraphExecutionState(
        absl::make_unique<GraphDef>(std::move(graph_def)), std::move(flib_def),
        options));

    // When place_pruned_graph is true, a different Graph* will be initialized
    // each time we prune the original graph, so there is no need to
    // construct a Graph* in this case.
    if (!options.session_options->config.graph_options().place_pruned_graph()) {
      auto base_graph = absl::make_unique<Graph>(OpRegistry::Global());
      TF_RETURN_IF_ERROR(ConvertGraphDefToGraph({}, *ret->original_graph_def_,
                                                base_graph.get()));
      TF_RETURN_IF_ERROR(ret->InitBaseGraph(std::move(base_graph)));
    }
    *out_state = std::move(ret);
  } else {
    auto ret = absl::WrapUnique(
        new GraphExecutionState(nullptr, std::move(flib_def), options));
    auto base_graph = absl::make_unique<Graph>(OpRegistry::Global());
    TF_RETURN_IF_ERROR(
        ConvertGraphDefToGraph({}, std::move(graph_def), base_graph.get()));
    TF_RETURN_IF_ERROR(ret->InitBaseGraph(std::move(base_graph)));
    *out_state = std::move(ret);
  }
  return Status::OK();
}

InitBaseGraph 会调用 Placer.run 完成算子编排。就是把计算图之中的算子放到最适合的设备上计算，这样可以最大化效率。Placer 会对 Graph 做分析，并且结合用户的要求对每个Node如何放置进行微调，具体原则有如下四种：

• 尽量满足用户的要求。用户可以通过 device 信息或者 loc 来制定设备，尽量优先满足。
• 尽量使用快速设备。TF 系统之中每个设备都有优先级，级别越高计算性能越好，优先选择级别高的设备。
• 尽量保证程序可运行。如果某个 Node 指定了在某种设备上执行，但是系统之中没有，则会选择一个可用的设备来重写 Placement。
• 尽量考虑近邻性。比如尽量让 Consumer 和 Producer 在同一个设备上，避免无意义的跨设备拷贝。

Status GraphExecutionState::InitBaseGraph(std::unique_ptr<Graph>&& new_graph) {
  // Save stateful placements before placing.
  RestoreStatefulNodes(new_graph.get());

  GraphOptimizationPassOptions optimization_options;
  optimization_options.session_handle = session_handle_;
  optimization_options.session_options = session_options_;
  optimization_options.graph = &new_graph;
  optimization_options.flib_def = flib_def_.get();
  optimization_options.device_set = device_set_;

  TF_RETURN_IF_ERROR(OptimizationPassRegistry::Global()->RunGrouping(
      OptimizationPassRegistry::PRE_PLACEMENT, optimization_options));

  Placer placer(new_graph.get(), "", flib_def_.get(), device_set_,
                /* default_local_device= */ nullptr,
                session_options_ == nullptr ||
                    session_options_->config.allow_soft_placement(),
                session_options_ != nullptr &&
                    session_options_->config.log_device_placement());
  TF_RETURN_IF_ERROR(placer.Run());

  TF_RETURN_IF_ERROR(OptimizationPassRegistry::Global()->RunGrouping(
      OptimizationPassRegistry::POST_PLACEMENT, optimization_options));

  for (const Node* n : new_graph->nodes()) {
    node_name_to_cost_id_map_[n->name()] = n->cost_id();
  }

  SaveStatefulNodes(new_graph.get());
  graph_ = new_graph.release();
  return Status::OK();
}

3.2.2 创建 WorkerSession

当 MasterSession 创建成功后，如果没有动态配置集群 (默认的分布式配置环境)， 则不会广播所有 Worker 动态地创建 WorkerSession。事实上，每个 Worker 都存在一个 SessionMgr 实例，它持有一个名为 legacy_session_ 的 WorkerSession 实例。因此，每个 Worker 存在一个全局唯一的 WorkerSession 实例。

图 3 创建 WorkerSession

逻辑如下：

• 首先，调用 ReleaseWorker 来释放已有的 workers。
• 其次，调用 GetOrCreateWorker 重新在缓存之中获取 Worker，如果没有，缓存自会构建。
• 最后，遍历 Workers，调用 CreateWorkerSessionAsync 来让每个 Worker 各自创建一个 WorkerSession，每个请求都会用 set_session_handle(handle_) 来把 MasterSession 的 session_handle 设置进入，这样每个 WorkerSession 都和 MasterSession 共享同样的 session_handle，它们都隶属于同一个 MasterSession。

为了收集全部 Workers 返回的消息，这里使用了计数器 BlockingCounter 来等待，其会把初始数值设置为 Worker 数目，当收集全部 Workers 的 CreateWorkerSessionResponse 响应消息之后，计数器会减少为 0，则 BlockingCounter 会被唤醒。

Status MasterSession::CreateWorkerSessions(
    const WorkerCacheFactoryOptions& options) {
  const std::vector<string> worker_names = filtered_worker_list_;
  WorkerCacheInterface* worker_cache = get_worker_cache();

  struct WorkerGroup {
    // The worker name. (Not owned.)
    const string* name;

    // The worker referenced by name. (Not owned.)
    WorkerInterface* worker = nullptr;

    // Request and responses used for a given worker.
    CreateWorkerSessionRequest request;
    CreateWorkerSessionResponse response;
    Status status = Status::OK();
  };
  BlockingCounter done(worker_names.size());
  std::vector<WorkerGroup> workers(worker_names.size());

  // Release the workers.
  auto cleanup = gtl::MakeCleanup([&workers, worker_cache] {
    for (auto&& worker_group : workers) {
      if (worker_group.worker != nullptr) {
        worker_cache->ReleaseWorker(*worker_group.name, worker_group.worker);
      }
    }
  });

  string task_name;
  string local_device_name;
  DeviceNameUtils::SplitDeviceName(devices_->client_device()->name(),
                                   &task_name, &local_device_name);
  const int64_t client_device_incarnation =
      devices_->client_device()->attributes().incarnation();

  Status status = Status::OK();
  // Create all the workers & kick off the computations.
  for (size_t i = 0; i < worker_names.size(); ++i) {
    workers[i].name = &worker_names[i];
    workers[i].worker = worker_cache->GetOrCreateWorker(worker_names[i]);
    workers[i].request.set_session_handle(handle_);
    workers[i].request.set_master_task(task_name);
    workers[i].request.set_master_incarnation(client_device_incarnation);
    if (session_opts_.config.share_cluster_devices_in_session() ||
        session_opts_.config.experimental()
            .share_cluster_devices_in_session()) {
      for (const auto& remote_dev : devices_->devices()) {
        *workers[i].request.add_cluster_device_attributes() =
            remote_dev->attributes();
      }

      if (!session_opts_.config.share_cluster_devices_in_session() &&
          session_opts_.config.experimental()
              .share_cluster_devices_in_session()) {
      }
    }

    DeviceNameUtils::ParsedName name;
    if (!DeviceNameUtils::ParseFullName(worker_names[i], &name)) {
      status = errors::Internal("Could not parse name ", worker_names[i]);
      return status;
    }
    if (!name.has_job || !name.has_task) {
      status = errors::Internal("Incomplete worker name ", worker_names[i]);
      return status;
    }

    if (options.cluster_def) {
      *workers[i].request.mutable_server_def()->mutable_cluster() =
          *options.cluster_def;
      workers[i].request.mutable_server_def()->set_protocol(*options.protocol);
      workers[i].request.mutable_server_def()->set_job_name(name.job);
      workers[i].request.mutable_server_def()->set_task_index(name.task);
      // Session state is always isolated when ClusterSpec propagation
      // is in use.
      workers[i].request.set_isolate_session_state(true);
    } else {
      // NOTE(mrry): Do not set any component of the ServerDef,
      // because the worker will use its local configuration.
      workers[i].request.set_isolate_session_state(
          session_opts_.config.isolate_session_state());
    }
    if (session_opts_.config.experimental()
            .share_session_state_in_clusterspec_propagation()) {
      // In a dynamic cluster, the ClusterSpec info is usually propagated by
      // master sessions. However, in data parallel training with multiple
      // masters
      // ("between-graph replication"), we need to disable isolation for
      // different worker sessions to update the same variables in PS tasks.
      workers[i].request.set_isolate_session_state(false);
    }
  }

  for (size_t i = 0; i < worker_names.size(); ++i) {
    auto cb = [i, &workers, &done](const Status& s) {
      workers[i].status = s;
      done.DecrementCount();
    };
    workers[i].worker->CreateWorkerSessionAsync(&workers[i].request,
                                                &workers[i].response, cb);
  }

  done.Wait();
  for (size_t i = 0; i < workers.size(); ++i) {
    status.Update(workers[i].status);
  }
  return status;
}

GrpcRemoteWorker

GrpcRemoteWorker 是 gRPC 的客户端，通过 stub 调用远端 WorkerService 相应的服务接口。

void CreateWorkerSessionAsync(const CreateWorkerSessionRequest* request,
                              CreateWorkerSessionResponse* response,
                              StatusCallback done) override {
  IssueRequest(request, response, createworkersession_, std::move(done));
}

GrpcWorkerService

远端 Worker 之中，接收到消息是在 GrpcWorkerService 之中，当收到 CreateWorkerSessionRequest 消息，将 由 CreateWorkerSessionHandler 回调处理，CreateWorkerSessionHandler 是一个宏，其在线程池中启动一个可运行的线程，触发 Worker（就是GrpcWorker） 的 CreateWorkerSession 方法来动态创建 WorkerSession 实例。

#define HANDLE_CALL(method, may_block_on_compute_pool)                        \
  void method##Handler(WorkerCall<method##Request, method##Response>* call) { \
    auto closure = [this, call]() {                                           \
      Status s = worker_->method(&call->request, &call->response);            \
      if (!s.ok()) {                                                          \
        VLOG(3) << "Bad response from " << #method << ": " << s;              \
      }                                                                       \
      call->SendResponse(ToGrpcStatus(s));                                    \
    };                                                                        \
    if ((may_block_on_compute_pool)) {                                        \
      worker_->env()->env->SchedClosure(std::move(closure));                  \
    } else {                                                                  \
      worker_->env()->compute_pool->Schedule(std::move(closure));             \
    }                                                                         \
    ENQUEUE_REQUEST(method, false);                                           \
  }

  HANDLE_CALL(CreateWorkerSession, false);

4. WorkerSession

其实，GrpcWorker 最终调用的是 WorkerInterface.CreateWorkerSession 方法。

Status CreateWorkerSession(const CreateWorkerSessionRequest* request,
                           CreateWorkerSessionResponse* response) {
  return CallAndWait(&ME::CreateWorkerSessionAsync, request, response);
}

CreateWorkerSessionRequest 消息之中携带了 MasterSession 分配的 session_handle，GrpcWorker 将据此创建一个 WorkerSession，session_handle 在这个 Worker 之内唯一标识这个 WorkerSession。

在 GrpcWorker 的 WorkerEnv 上下文之中有一个 SessionMgr，SessionMgr 负责统一管理和维护所有的 WorkerSession 生命周期。SessionMgr 与 WorkerSession 是一对多的关系，每个 WorkerSession 实例使用 session_handle 标识。

void Worker::CreateWorkerSessionAsync(const CreateWorkerSessionRequest* request,
                                      CreateWorkerSessionResponse* response,
                                      StatusCallback done) {
  Status s = env_->session_mgr->CreateSession(
      request->session_handle(), request->server_def(),
      request->cluster_device_attributes(), request->isolate_session_state(),
      request->master_task(), request->master_incarnation());
  done(s);
}

4.1 SessionMgr

4.1.1 定义

重点是如下，维护了 session_handle 和 WorkerSession 之间的对应关系，每个 WorkerSession 由 session_handle 来标识。

• std::map<string, std::shared_ptr> sessions_ ：维护了对应关系。

• std::shared_ptr legacy_session_ ：本地 WorkerSession 实例。

图 4 SessionMgr

class SessionMgr {
 public:
  typedef std::function<Status(const ServerDef&, WorkerCacheInterface**)>
      WorkerCacheFactory;

  explicit SessionMgr(
      WorkerEnv* worker_env, const string& default_worker_name,
      std::unique_ptr<WorkerCacheInterface> default_worker_cache,
      WorkerCacheFactory worker_cache_factory);
  ~SessionMgr() {}

  // Allocates state for a new session.
  Status CreateSession(const string& session, const ServerDef& server_def,
                       bool isolate_session_state);
  Status CreateSession(
      const string& session, const ServerDef& server_def,
      const protobuf::RepeatedPtrField<DeviceAttributes>& device_attributes,
      bool isolate_session_state);

  // Create WorkerSession from the master with the given `master_task` and
  // `master_incarnation`. We first look for existing WorkerSessions associated
  // with the specified master task. If there are sessions created by the same
  // master but with a different incarnation, it indicates that the remote
  // master has restarted before deleting the sessions on worker. When it
  // happens, old sessions associated with the master will be automatically
  // removed before the new session is created.
  Status CreateSession(
      const string& session, const ServerDef& server_def,
      const protobuf::RepeatedPtrField<DeviceAttributes>& device_attributes,
      bool isolate_session_state, string master_task,
      int64_t master_incarnation);

  void ResetDefaultWorkerCache(WorkerCacheInterface* worker_cache);

  // Updates state (worker cache, devices) of worker session identified by
  // session name (`session`) based on a new server_def and set of devices.
  Status UpdateSession(const string& session, const ServerDef& server_def,
                       const protobuf::RepeatedPtrField<DeviceAttributes>&
                           cluster_device_attributes,
                       bool isolate_session_state);

  // Locates the worker session for a given session handle
  Status WorkerSessionForSession(const string& session_handle,
                                 std::shared_ptr<WorkerSession>* out_session);
  std::shared_ptr<WorkerSession> LegacySession();

  Status DeleteSession(const string& session);

  static string WorkerNameFromServerDef(const ServerDef& server_def);

  void SetLogging(bool active);

  void RetrieveLogs(int64_t step_id, LoggingResponse* response);

  void ClearLogs();

 private:
  WorkerEnv* const worker_env_;  // Not owned.

  // A note about destruction:
  // We must delete graph_mgr before device_mgr, due to shared
  // ownership of OpKernels in the executors. (The graph_mgr will
  // free all stateless OpKernels, and pass over borrowed stateful
  // OpKernels, which are also held in their respective devices'
  // OpSegments.)
  //
  // legacy_session_ owns the worker_env_.device_mgr, and so we must ensure
  // that sessions_'s WorkerSessions are deleted (which do not own the
  // underlying devices, but instead own RenamedDevices) before
  // legacy_session_ is deleted. Further, we must ensure that WorkerSession's
  // device_mgr is deleted after WorkerSession's graph_mgr.

  std::unique_ptr<WorkerCacheInterface> default_worker_cache_;
  std::shared_ptr<WorkerSession> legacy_session_;

  bool is_logging_active_ = false;

  const WorkerCacheFactory worker_cache_factory_;

  Status WorkerSessionForSessionLocked(
      const string& session_handle, std::shared_ptr<WorkerSession>* out_session)
      TF_EXCLUSIVE_LOCKS_REQUIRED(mu_);

  mutex mu_;
  // A map from session identifier to internal session structure.
  std::map<string, std::shared_ptr<WorkerSession>> sessions_ TF_GUARDED_BY(mu_);

  // Incarnation and WorkerSession handle associated with a master task.
  struct MasterAssociatedSession {
    const int64_t master_incarnation;
    const string session_handle;
  };
  // A map from master task name to its associated worker sessions.
  std::unordered_multimap<string, MasterAssociatedSession>
      master_to_associated_sessions_ TF_GUARDED_BY(mu_);
};

4.1.2 建立 Session

CreateSession 方法会创建 WorkerSession 和 GraphMgr。

Status SessionMgr::CreateSession(
    const string& session, const ServerDef& server_def,
    const protobuf::RepeatedPtrField<DeviceAttributes>&
        cluster_device_attributes,
    bool isolate_session_state, string master_task,
    int64_t master_incarnation) {
  mutex_lock l(mu_);
  if (session.empty()) {
    return errors::InvalidArgument("Session must be non-empty.");
  }

  // For given master task name, check if one or more `WorkerSession`s have been
  // created previously on this worker, and if so garbage collect the expired
  // `WorkerSession`s. This happens when the master fails before sending
  // `DeleteSession` requests, which can cause `WorkerSession`s to be leaked.
  if (!master_task.empty()) {
    auto it_range = master_to_associated_sessions_.equal_range(master_task);
    if (it_range.first != it_range.second &&
        it_range.first->second.master_incarnation != master_incarnation) {
      auto it = it_range.first;
      while (it != it_range.second) {
        auto session_it = sessions_.find(it->second.session_handle);
        if (session_it != sessions_.end()) {
          sessions_.erase(session_it);
        }
        it = master_to_associated_sessions_.erase(it);
      }
    }
  }

  WorkerCacheInterface* worker_cache = nullptr;
  string worker_name;
  if (server_def.cluster().job().empty()) {
    worker_cache = new WorkerCacheWrapper(default_worker_cache_.get());
    worker_name = legacy_session_->worker_name();
  } else {
    TF_RETURN_IF_ERROR(worker_cache_factory_(server_def, &worker_cache));
    worker_name = WorkerNameFromServerDef(server_def);
  }

  if (worker_cache != nullptr && default_worker_cache_ != nullptr) {
    worker_cache->SetLogging(this->is_logging_active_);
  }

  std::shared_ptr<WorkerSession> worker_session;
  std::vector<std::unique_ptr<Device>> cluster_devices;

  if (isolate_session_state || server_def.cluster().job_size()) {

    // Create a private copy of the DeviceMgr for the WorkerSession.
    std::vector<std::unique_ptr<Device>> renamed_devices;
    for (Device* d : worker_env_->local_devices) {
      renamed_devices.push_back(RenamedDevice::NewRenamedDevice(
          worker_name, d, false, isolate_session_state));
    }
    auto device_mgr = MakeUnique<StaticDeviceMgr>(std::move(renamed_devices));
    LookupLocalDevice cb = [&device_mgr](StringPiece name, Device** device) {
      return device_mgr->LookupDevice(name, device);
    };
    AsRemoteDevices(worker_env_->env, cluster_device_attributes, cb,
                    &cluster_devices);
    std::unique_ptr<DynamicDeviceMgr> remote_devices;
    if (!cluster_device_attributes.empty()) {
      remote_devices = MakeUnique<DynamicDeviceMgr>();
      TF_RETURN_IF_ERROR(
          remote_devices->AddDevices(std::move(cluster_devices)));
    }

    auto graph_mgr = MakeUnique<GraphMgr>(worker_env_, device_mgr.get());
    worker_session.reset(
        new WorkerSession(session, worker_name,
                          std::unique_ptr<WorkerCacheInterface>(worker_cache),
                          std::move(device_mgr), std::move(graph_mgr),
                          std::move(remote_devices)));
  } else {
    AsRemoteDevices(worker_env_->env, cluster_device_attributes, nullptr,
                    &cluster_devices);
    std::unique_ptr<DynamicDeviceMgr> remote_devices;
    if (!cluster_device_attributes.empty()) {
      remote_devices = MakeUnique<DynamicDeviceMgr>();
      TF_RETURN_IF_ERROR(
          remote_devices->AddDevices(std::move(cluster_devices)));
    }
    // Borrow the WorkerEnv's DeviceMgr for the WorkerSession, so
    // that resources using it can use its devices after the
    // WorkerSession has been deleted.
    auto graph_mgr = MakeUnique<GraphMgr>(worker_env_, worker_env_->device_mgr);
    worker_session = WorkerSession::CreateWithBorrowedDeviceMgr(
        session, worker_name,
        std::unique_ptr<WorkerCacheInterface>(worker_cache),
        worker_env_->device_mgr, std::move(graph_mgr),
        std::move(remote_devices));
  }

  sessions_.insert(std::make_pair(session, std::move(worker_session)));
  if (!master_task.empty()) {
    MasterAssociatedSession s{master_incarnation, session};
    master_to_associated_sessions_.emplace(master_task, s);
  }
  return Status::OK();
}

4.1.3 注册图

我们用 RegisterGraphAsync 为例来看看 worker 内部功能。可以看到其使用 GraphMgr 完成了基础功能。

void Worker::RegisterGraphAsync(const RegisterGraphRequest* request,
                                RegisterGraphResponse* response,
                                StatusCallback done) {
  std::shared_ptr<WorkerSession> session;
  Status s;
  if (request->create_worker_session_called()) {
    s = env_->session_mgr->WorkerSessionForSession(request->session_handle(),
                                                   &session);
  } else {
    session = env_->session_mgr->LegacySession();
  }
  if (s.ok()) {
    s = session->graph_mgr()->Register(
        request->session_handle(), request->graph_def(), session.get(),
        request->graph_options(), request->debug_options(),
        request->config_proto(), request->collective_graph_key(),
        session->cluster_flr(), response->mutable_graph_handle());
  }
  done(s);
}

4.2 WorkerSession

4.2.1 定义

WorkerSession 之中比较重要的几个成员变量包括几个管理类 GraphMgr，DeviceMgr，DynamicDeviceMgr：

• string session_name_ ：Session 名称。

• string worker_name_ ：Worker 名称，比如 /job:mnist/replica:0/task:1。

• std::shared_ptr worker_cache_ ：Worker 缓存。

• std::unique_ptr graph_mgr_ ：本 session 注册的计算图，每个 Worker 可以注册和运行多个计算图，每个计算图使用 graph）handle 标识。

• std::unique_ptr device_mgr_ ：本地计算设备集合信息。

图 5 WorkerSession 概念

// WorkerSession encapsulates all of the state relating to a given session.
class WorkerSession {
 public:
  // Collection of local devices. These devices are typically
  // RenamedDevices in all except the SessionMgr.legacy_session_ and
  // sessions created with `isolate_session_state == false`. In the
  // those cases, this method returns a pointer to a borrowed
  // DeviceMgr (typically the `worker_env.device_mgr`).
  DeviceMgr* device_mgr() {
    return device_mgr_ ? device_mgr_.get() : borrowed_device_mgr_;
  }

  DynamicDeviceMgr* remote_device_mgr() { return remote_device_mgr_.get(); }

  const string& session_name() const { return session_name_; }
  const string& worker_name() const { return worker_name_; }

  WorkerCacheInterface* worker_cache() const {
    tf_shared_lock l(worker_session_state_mu_);
    return worker_cache_.get();
  }
  GraphMgr* graph_mgr() const { return graph_mgr_.get(); }

  ClusterFunctionLibraryRuntime* cluster_flr() const {
    return cluster_flr_.get();
  }

  WorkerSession(const string& session_name, const string& worker_name,
                std::unique_ptr<WorkerCacheInterface> worker_cache,
                std::unique_ptr<DeviceMgr> device_mgr,
                std::unique_ptr<GraphMgr> graph_mgr,
                std::unique_ptr<DynamicDeviceMgr> remote_device_mgr);

  static std::shared_ptr<WorkerSession> CreateWithBorrowedDeviceMgr(
      const string& session_name, const string& worker_name,
      std::unique_ptr<WorkerCacheInterface> worker_cache,
      DeviceMgr* borrowed_device_mgr, std::unique_ptr<GraphMgr> graph_mgr,
      std::unique_ptr<DynamicDeviceMgr> remote_device_mgr);

  // In the eager runtime we allow WorkerSession to be updated, where the
  // worker cache will be recreated. If WorkerSession upate is expected and a
  // worker in the cache is used in RPCs, the caller should hold a shared
  // pointer to avoid the workers getting deleted.
  std::shared_ptr<WorkerCacheInterface> GetSharedWorkerCache() {
    tf_shared_lock l(worker_session_state_mu_);
    return worker_cache_;
  }

  // Update an existing worker session with new set of remote workers and
  // devices. Added devices will be owned by the worker session, and removed
  // devices will be freed by their names.
  Status UpdateWorkerCacheAndDevices(
      std::unique_ptr<WorkerCacheInterface> new_worker_cache,
      std::vector<std::unique_ptr<Device>> added_remote_devices,
      const std::vector<Device*>& removed_remote_devices);

  ~WorkerSession();

 private:
  WorkerSession(const string& session_name, const string& worker_name,
                std::unique_ptr<WorkerCacheInterface> worker_cache,
                DeviceMgr* borrowed_device_mgr,
                std::unique_ptr<GraphMgr> graph_mgr,
                std::unique_ptr<DynamicDeviceMgr> remote_device_mgr);

  // The name of the session.
  const string session_name_;

  // The name of the worker. E.g., /job:mnist/replica:0/task:1.
  const string worker_name_;

  mutable mutex worker_session_state_mu_;
  // Object from which WorkerInterface instances can be obtained.
  std::shared_ptr<WorkerCacheInterface> worker_cache_
      TF_GUARDED_BY(worker_session_state_mu_);

  // graph_mgr keeps track of the registered graphs of this session.
  //
  // Note: graph_mgr must be deleted before rendezvous_mgr!
  // Note: graph_mgr must be deleted before device_mgr!
  const std::unique_ptr<GraphMgr> graph_mgr_;

  std::unique_ptr<ClusterFunctionLibraryRuntime> cluster_flr_;

  const std::unique_ptr<DeviceMgr> device_mgr_;
  DeviceMgr* const borrowed_device_mgr_;  // Not owned.
  std::unique_ptr<DynamicDeviceMgr> remote_device_mgr_;
};

至此，session 基本流程我们梳理完成，下面就会对业务进行详细分析。

0xFF 参考

TensorFlow中的Placement启发式算法模块——Placer