MXNet源码分析 | KVStore进程间通信

本文主要基于MXNet1.6.0版本进行分析。

在上一篇文章中，我们分析了MXNet中KVStore的进程内通信机制。在这篇文章中，我们主要分析KVStore如何进行多节点分布式通信。

在KVStore的实现中，KVStoreDist和KVStoreDistServer分别对应参数服务器中的worker节点与server节点。KVStoreDist继承自KVStoreLocal，通过封装PS-Lite中的KVWorker实现了Push和Pull等接口，从而向server发送各类请求；而KVStoreDistServer则封装了PS-Lite中的KVServer，用来处理并响应worker发来的各类请求。

worker端执行逻辑

worker创建

KVStoreDist的构造函数为每个worker节点创建一个ps::KVWorker<char>类型的对象。如果当前worker节点不是一个recovery的节点，那么就阻塞到所有的worker和server启动。

explicit KVStoreDist(bool use_device_comm)
    : KVStoreLocal(use_device_comm), ps_worker_(nullptr), server_(nullptr) {
  if (IsWorkerNode()) {
    int new_customer_id = GetNewCustomerId();
    ps_worker_ = new ps::KVWorker<char>(0, new_customer_id);
    ps::StartAsync(new_customer_id, "mxnet\0");
    if (!ps::Postoffice::Get()->is_recovery()) { 
      ps::Postoffice::Get()->Barrier(
        new_customer_id,
        ps::kWorkerGroup + ps::kServerGroup + ps::kScheduler);
    }
  }
  bigarray_bound_ = dmlc::GetEnv("MXNET_KVSTORE_BIGARRAY_BOUND", 1000 * 1000);
  log_verbose_ = dmlc::GetEnv("MXNET_KVSTORE_DIST_ROW_SPARSE_VERBOSE", false);
}

worker的初始化过程

在初始化时，每个worker首先检查key的唯一性，随后调用comm_->Init为每个key初始化进行本地通信的资源。本地初始化完成后，worker0把自己本地的权重发送给所有的server。worker0在其push操作完成后，会将数据写入到comm_buf_和compr_buf_这两个缓冲区中。

void InitImpl(const std::vector<int>& keys,
              const std::vector<NDArray>& values) override {
  CheckUnique(keys);
  for (size_t i = 0; i < keys.size(); ++i) {
    comm_->Init(keys[i], values[i].storage_type(), values[i].shape(), values[i].dtype());
  }
  if (get_rank() == 0 && this->ps_worker_->get_customer()->customer_id() == 0) {
    Push_(keys, values, 0, false);
    // wait until the push is finished
    for (const int key : keys) {
      comm_buf_[key].WaitToWrite();
      compr_buf_[key].WaitToWrite();
    }
  } else {
    // do nothing
  }
  if (!ps::Postoffice::Get()->is_recovery()) {
    Barrier();
  }
}

worker发送控制消息

worker端通过SendCommandToServers函数向server端发送控制消息。例如，在KVStoreDist的析构函数中有如下代码，用来从worker0节点向所有server节点发送一个终止的命令。

if (get_rank() == 0 && ps_worker_->get_customer()->customer_id() == 0) {
   // stop the executor at servers
  SendCommandToServers(static_cast<int>(CommandType::kStopServer), "");
}

worker发送数据消息

worker会调用Push_函数向server发送数据请求，它的核心逻辑如下所示（省略部分代码）。与之前提到的本地通信类似，在向server节点发送数据之前，会先调用GroupPairsPush把具有相同key的value汇总到一个vector中。对于每个key，先在本地进行一次Reduce操作聚合所有设备上的梯度，并将结果存放到comm_buf中。随后，通过EncodeDefaultKey把key和value编码成PS-Lite支持的数据结构，再调用PushDefault把对应的数据发送出去。

void KVStoreDist::Push_(const std::vector<int>& keys,
                        const std::vector<NDArray>& values,
                        int priority,
                        bool do_merge) {
  std::vector<int> uniq_keys;
  std::vector<std::vector<NDArray>> grouped_val;
  GroupKVPairsPush(keys, values, &uniq_keys, &grouped_val, false);

  for (size_t i = 0; i < uniq_keys.size(); ++i) {
    int key = uniq_keys[i];
    const auto& vals = grouped_vals[i];
    NDArray merged = do_merge ? comm_->Reduce(key, vals, priority) : vals[0];

    auto &comm_buf = comm_buf_[key];
    if (merged.ctx().dev_mask() == cpu::kDevMask) {
      // Start of a push doesn't guarantee that the previous pushes are completed.
      // This shouldn't affect training of networks though because training involves
      // a sequence of push, pull, then push. This imposes ordering that the
      // second push happens after the first pull, and the pull happens after first push.
      comm_buf = merged;  // avoid memory copy
    } else {
      if (comm_buf.is_none()) {
        comm_buf = NDArray(merged.shape(), pinned_ctx_, true, merged.dtype());
      }
      CopyFromTo(merged, &comm_buf);
    }
    const int dtype = merged.dtype();
    const int num_bytes = mshadow::mshadow_sizeof(dtype);
    PSKV& pskv = EncodeDefaultKey(key, comm_buf.shape().Size(), num_bytes);
    PushDefault(key, comm_buf, pskv, priority);
  }
}

PushDefault会调用ps_worker_->ZPush来完成梯度的发送，梯度发送以及发送之前的一些准备操作都被封装到一个lambda表达式中，这个lambda表达式随后被压入到MXNet后端的依赖引擎中等待执行。

void PushDefault(int key, const NDArray &send_buf, const PSKV& pskv, int priority) {
  auto push_to_servers =
      [this, key, pskv, send_buf](RunContext rctx, Engine::CallbackOnComplete cb) {
        const int dtype = send_buf.dtype();
        // convert to ps keys
        const size_t size = send_buf.shape().Size() * mshadow::mshadow_sizeof(dtype);
        char* data = static_cast<char *>(send_buf.data().dptr_);
        // do push. false means no delete
        ps::SArray<char> vals(data, size, false);
        int cmd = GetCommandType(RequestType::kDefaultPushPull, dtype);
        CHECK_NOTNULL(ps_worker_)->ZPush(
            pskv.keys, vals, pskv.lens,
            cmd, [cb]() { cb(); });
      };
  Engine::Get()->PushAsync(
      push_to_servers,
      pinned_ctx_,
      {send_buf.var()},
      {},
      FnProperty::kNormal,
      priority,
      "KVStoreDistDefaultPush");
}

Pull操作的过程如下所示。在准备工作完成后，调用ps_server_->ZPull完成权重的拉取，最后在本地执行Broadcast把从server端拉回的权重广播到所有设备上。

void PullImpl(const std::vector<int>& keys,
              const std::vector<NDArray*>& values,
              int priority, bool ignore_sparse) override {
  CHECK(ignore_sparse) << "dist kvstore pull doesn't support ignore_sparse=False";
  std::vector<int> uniq_keys;
  std::vector<std::vector<NDArray*> > grouped_vals;
  GroupKVPairsPull(keys, values, &uniq_keys, &grouped_vals, true);

  for (size_t i = 0; i < uniq_keys.size(); ++i) {
    int key = uniq_keys[i];
    // use the same array for merging to guarantee that pull always happens
    // after the previous push on this key
    auto& recv_buf = comm_buf_[key];
    const auto storage_type = grouped_vals[i][0]->storage_type();
    CHECK_EQ(storage_type, kDefaultStorage)
             << "Expected stype of value to be kDefaultStorage";
    if (recv_buf.is_none()) {
      // it may happen for the first time a no-rank-0 worker pull the weight.
      recv_buf = NDArray(grouped_vals[i][0]->shape(), pinned_ctx_,
                         true, grouped_vals[i][0]->dtype());
    }
    auto pull_from_servers = [this, key, recv_buf](
        RunContext rctx, Engine::CallbackOnComplete cb) {
      // convert to ps keys
      size_t size = recv_buf.shape().Size();
      const int dtype = recv_buf.dtype();
      const int num_bytes = mshadow::mshadow_sizeof(dtype);
      PSKV& pskv = EncodeDefaultKey(key, size, num_bytes) :
      char* data = static_cast<char*> (recv_buf.data().dptr_);
      // false means not to delete data when SArray is deleted
      auto vals = new ps::SArray<char>(data, size * num_bytes, false);
      // issue pull
      RequestType mode = RequestType::kDefaultPushPull;
      const int cmd = GetCommandType(mode, dtype);
      CHECK_NOTNULL(ps_worker_)->ZPull(
       pskv.keys, vals, &pskv.lens, cmd, [vals, cb](){ delete vals; cb(); });
    };

    CHECK_NOTNULL(Engine::Get())->PushAsync(
        pull_from_servers,
        pinned_ctx_,
        {},
        {recv_buf.var()},
        FnProperty::kNormal,
        priority,
        "KVStoreDistDefaultStoragePull");

    comm_->Broadcast(key, recv_buf, grouped_vals[i], priority);
  }
}

server端执行逻辑

server的创建以及启动

首先在KVStoreDistServer的构造函数中为ps_server_绑定处理命令请求的CommandHandle以及处理数据请求的DataHandleEx。注意到在绑定CommandHandle时，ps_server_被向上转型成ps::SimpleApp*类型。这是因为ps::SimpleApp中实现的set_request_handle只能接收包含两个形参的函数对象，而ps::KVServer继承了ps::SimpleApp并且重载了set_request_handle，使之可以接收包含三个形参的函数对象。这样一来，就完成了对控制请求和数据请求的分开处理。

KVStoreDistServer() {
  using namespace std::placeholders;
  ps_server_ = new ps::KVServer<char>(0);
  static_cast<ps::SimpleApp*>(ps_server_)->set_request_handle(
      std::bind(&KVStoreDistServer::CommandHandle, this, _1, _2));
  ps_server_->set_request_handle(
      std::bind(&KVStoreDistServer::DataHandleEx, this, _1, _2, _3));
  sync_mode_ = false;
  gradient_compression_ = std::make_shared<GradientCompression>();
  log_verbose_ = dmlc::GetEnv("MXNET_KVSTORE_DIST_ROW_SPARSE_VERBOSE", false);
}

处理控制请求

server接收到worker0发来的命令后，会根据命令的类型，执行不同的操作。例如，当worker发来StopServer的命令后，server就会被停止。相应的命令执行完毕后，server会发送一个响应给worker0。注意这里负责发送响应的不是ps::KVWorker<char>类型的对象，而是ps::SimpleApp类型的对象。

void CommandHandle(const ps::SimpleData& recved, ps::SimpleApp* app) {
  CommandType recved_type = static_cast<CommandType>(recved.head);
  switch (recved_type) {
    case CommandType::kStopServer:
      exec_.Stop();
      break;
    case CommandType::kSyncMode:
      sync_mode_ = true;
      break;
    case CommandType::kSetGradientCompression:
      gradient_compression_->DecodeParams(recved.body);
      break;
    case CommandType::kSetProfilerParams:
      // last char is the type of profiler command
      ProcessServerProfilerCommands(static_cast<KVStoreServerProfilerCommand>
                                                (recved.body.back() - '0'),
                                    recved.body);
      break;
    case CommandType::kSetMultiPrecision:
      // uses value 1 for message id from frontend
      if (!multi_precision_) {
        multi_precision_ = true;
        CreateMultiPrecisionCopies();
      }
      break;
    case CommandType::kController:
      // this uses value 0 for message id from frontend
      // let the main thread to execute ctrl, which is necessary for python
      exec_.Exec([this, recved]() {
          CHECK(controller_);
          controller_(recved.head, recved.body);
        });
      break;
  }
  app->Response(recved);
}

处理数据请求

前面提到，DataHandleEx被注册为处理数据请求的函数，它会根据数据请求类型去调用不同的处理函数。默认情况下会调用DataHandleDefalut，该函数会对worker发来的push和pull请求分开处理。当worker节点push梯度到server时，如果某个key是第一次被push，那么server会为相应的key申请内存空间；否则会根据sync_mode_的值分别进行处理。在sync_mode_ == true（即同步训练模式）的情况下，所有worker上的梯度会被聚合到update_buf_[key].merged中；而在异步训练模式下，server把从某个worker接收的梯度放在update_buf_[key].temp_array中。随后，worker发来的push请求信息会被记录到update_buf_[key].request中。待上面的工作完成后，会调用ApplyUpdates函数去更新key对应的模型参数。当worker节点向server节点发送pull请求时，server会直接调用DefaultStorageResponse把server节点最新的模型参数发送给worker。

void DataHandleDefault(const DataHandleType type, const ps::KVMeta& req_meta,
                       const ps::KVPairs<char>& req_data, ps::KVServer<char>* server) {
  int key = DecodeKey(req_data.keys[0]);
  auto& stored = store_[key];
  if (req_meta.push) { // push operation
    size_t ds[] = {(size_t) req_data.lens[0] / mshadow::mshadow_sizeof(type.dtype)};
    mxnet::TShape dshape(ds, ds + 1);
    TBlob recv_blob;
    MSHADOW_REAL_TYPE_SWITCH(type.dtype, DType, {
      recv_blob = TBlob(reinterpret_cast<DType*>(req_data.vals.data()), dshape, cpu::kDevMask);
    })
    NDArray recved = NDArray(recv_blob, 0);
    if (stored.is_none()) { // the first push request
      // initialization
      stored = NDArray(dshape, Context(), false, type.dtype);
      CopyFromTo(recved, &stored, 0);
      server->Response(req_meta);
      stored.WaitToRead();
    } else {
      auto& updates = update_buf_[key];
      if (sync_mode_ && updates.merged.is_none() {
        updates.merged = NDArray(dshape, Context(), false, type.dtype);
      }
      if (updates.request.empty()) { // the first 
        if (sync_mode_) {
          CopyFromTo(recvd, updates.merged);
        } else { // async training
          updates.temp_array = recved;
        }
      } else {
        updates.merged += recved;
      }
      updates.request.push_back(req_meta);
      ApplyUpdates(type, key, req_data, &updates, server);
  } else { // pull operation
    DefaultStorageResponse(type, key, req_meta, req_data, server);
  }
}

函数ApplyUpdates实现了模型权重更新的核心逻辑。如果是异步训练模式，或者当前的update_buf中的push请求数量等于worker的数量（意味着server收到了所有worker上的梯度），那么就会执行参数的更新过程；否则就不进行更新，直接调用server->Response给worker发一个不带任何数据的响应消息，表示收到了相关的数据。如果server端设置了更新器updater_，那么就会在server端执行更新操作；否则，server只对梯度进行聚合。如下代码的7~16行描述了这一过程，更新或聚合的结果会被存放到store_[key]中。由于update_buf_[key].request中保存的请求既有可能是push，也有可能是pushpull（唯独不可能是pull，因为我们只在req_meta.push==true时才把req_meta加入到update_buf_[key].request中），因此我们还要额外处理pushpull这类请求。对于update_buf_[key].request中的每个请求，如果该请求req.pull==true，那么就调用DefaultStorageResponse把模型权重传输给worker。在更新过程完成后，update_buf_[key].request就会被清空，以等待下一次更新。

inline void ApplyUpdates(const DataHandleType type, const int key,
                         const ps::KVPairs<char>& req_data, UpdateBuf *update_buf,
                         ps::KVServer<char>* server) {
  if (!sync_mode_ || update_buf->request.size() == (size_t) ps::NumWorkers()) {
    // let the main thread to execute updater_, which is necessary for python
    auto& stored = store_[key];
    auto& update =  sync_mode_ ? update_buf->merged : update_buf->temp_array;
    if (updater_) { // update_on_kvstore == True
      exec_.Exec([this, key, &update, &stored](){
        CHECK(updater_);
        updater_(key, update, &stored);
      });
    } else { // update_on_kvstore == False, only support for sync mode
      CHECK(sync_mode_) << "Updater needs to be set for async mode";
      // if no updater, just copy
      CopyFromTo(update_buf->merged, &stored);
    }
    /**
     * Request can be for either push or pushpull
     * If pull flag is set, respond immediately with the updated values
     * Otherwise, only send the notification
     */
    bool has_pull = false;
    for (const auto& req : update_buf->request) {
      has_pull = has_pull || req.pull;
    }
    if (has_pull) {
      // if there is a pull request, perform WaitToRead() once before DefaultStorageResponse
      stored.WaitToRead();
      for (const auto& req : update_buf->request) {
        if (req.pull) {
          DefaultStorageResponse(type, key, req, req_data, server);
        }
      }
      update_buf->request.clear();
    } else {
      // otherwise, send response directly
      for (const auto& req : update_buf->request) {
        server->Response(req);
      }
      update_buf->request.clear();
      stored.WaitToRead();
    }
  } else { // donot perform update operation
    update_buf->merged.WaitToRead();
  }
}

DefaultStorageResponse会根据传入的req_meta和req_data这两个参数针对worker的push请求构建出对应的带数据的响应消息。响应是一个ps::KVPairs<char>类型的对象，其中的数据部分拷贝自store_[key]。响应对象构建完成后，同样会调用server->Response将消息发回对应的worker。

void DefaultStorageResponse(const DataHandleType type,
                            const int key,
                            const ps::KVMeta& req_meta,
                            const ps::KVPairs<char> &req_data,
                            ps::KVServer<char>* server) {
  ps::KVPairs<char> response;
  const NDArray& stored = store_[key];
  CHECK(!stored.is_none()) << "init " << key << " first";

  auto len = stored.shape().Size() * mshadow::mshadow_sizeof(stored.dtype());
  response.keys = req_data.keys;
  response.lens = {len};
  // TODO(mli) try to remove this CopyFrom
  response.vals.CopyFrom(static_cast<const char*>(stored.data().dptr_), len);
  server->Response(req_meta, response);
}