tensorflow节点布放（device assignment of node）算法：simpler_placer

tensorflow v0.9中目前在用的devcie assignment算法是simple placer算法，相比于白皮书中cost model算法实现简单。simpler placer算法优先选择/gpu:0设备， 但不支持 multi gpu assignment。

白皮书提到的cost model可以根据设备资源代价、数据传输代价平衡分配设备，在v0.9版本中有部分实现，但还未开放使用，见 core/graph/costmodel.cc 

 

simple_placer的实现代码在文件python/core/common_runtime/simple_placer.cc，其中包含device_assignment的核心功能。

core/common_runtime/simple_placer_test.cc测试片段如下

////////////////////////////////////////////////////////////////////////////////
//
// A SimplePlacerTest method has three phases:
//
// 1. Build a TensorFlow graph, with no (or partial) device assignments.
// 2. Attempt to compute a placement using the SimplePlacer.
// 3. EITHER: test that the constraints implied by the graph are respected;
//    or that an appropriate error was reported.
//
////////////////////////////////////////////////////////////////////////////////
class SimplePlacerTest : public ::testing::Test {
 protected:
  SimplePlacerTest() {
    // Build a set of 10 GPU and 10 CPU devices.
    // NOTE: this->local_devices_ owns the device objects;
    // this->devices_ contains borrowed pointers to the device
    // objects.
    for (int i = 0; i < 10; ++i) {    // 添加了10 cpu和10 gpu的fake devices
      local_devices_.emplace_back(FakeDevice::MakeCPU(
          strings::StrCat("/job:a/replica:0/task:0/cpu:", i)));
      devices_.AddDevice(local_devices_.back().get());
      // Insert the GPUs in reverse order.
      local_devices_.emplace_back(FakeDevice::MakeGPU(
          strings::StrCat("/job:a/replica:0/task:0/gpu:", 9 - i)));
      devices_.AddDevice(local_devices_.back().get());
    }
  }
  ...
}
...
// Test that a graph with no constraints will successfully assign nodes to the
// "best available" device (i.e. prefer GPU over CPU).
TEST_F(SimplePlacerTest, TestNoConstraints) {
  Graph g(OpRegistry::Global());
  {  // Scope for temporary variables used to construct g.   // 用GraphDefBuilder构建graph的结构
    GraphDefBuilder b(GraphDefBuilder::kFailImmediately);
    Node* input = ops::SourceOp("TestInput", b.opts().WithName("in"));    
    ops::UnaryOp("TestRelu", ops::NodeOut(input, 0), b.opts().WithName("n1"));
    ops::UnaryOp("TestRelu", ops::NodeOut(input, 1), b.opts().WithName("n2"));
    TF_EXPECT_OK(BuildGraph(b, &g));   //  BuildGraph函数将GraphDefBuilder的图写入到Graph中
  }
 
  TF_EXPECT_OK(Place(&g));   // Place函数将graph中的node布放到设备列表中
  EXPECT_DEVICE_TYPE(g, "in", DEVICE_CPU);   // 期望：input节点在CPU中，n1节点在GPU中，n2节点在GPU中，故而GPU优先级大于CPU
  EXPECT_DEVICE_TYPE(g, "n1", DEVICE_GPU);
  EXPECT_DEVICE_TYPE(g, "n2", DEVICE_GPU);
}

其中BuildGraph函数将GraphDefBuilder 对象中的graph 结构定义写入到Graph中。Place函数将graph中的node布放到设备列表中，其中device assignment算法的核心在SimplePlacer::Run函数中

 // Builds the given graph, and (if successful) indexes the node
  // names for use in placement, and later lookup.
  Status BuildGraph(const GraphDefBuilder& builder, Graph* out_graph) {
    TF_RETURN_IF_ERROR(builder.ToGraph(out_graph));
    nodes_by_name_.clear();
    for (Node* node : out_graph->nodes()) {
      nodes_by_name_[node->name()] = node->id();
    }
    return Status::OK();
  }
  // Invokes the SimplePlacer on "graph". If no DeviceSet is specified, the
  // placement will use the default DeviceSet (of 10 CPU and 10 GPU devices).
  //
  // REQUIRES: "*graph" was produced by the most recent call to BuildGraph.
  Status Place(Graph* graph, DeviceSet* devices, SessionOptions* options) {
    SimplePlacer placer(graph, devices, options);
    return placer.Run();
  }

SimplePlacer::Run()在core/common_runtime/simple_placer.cc文件中，具体实现分为4个步骤： 

步骤1和2： 遍历graph的node，将node加入到ColocationGraph对象中（不包含source和sink节点）。

// 1. First add all of the nodes. Note that steps (1) and (2)
// requires two passes over the nodes because the graph (and hence
// the constraints) may not be acyclic.  这里graph可能是有环的？
for (Node* node : graph_->nodes()) {
    // Skip the source and sink nodes.
    if (!node->IsOp()) { continue; }
    status = colocation_graph.AddNode(*node);
    if (!status.ok()) return AttachDef(status, node->def());
  }
// 2. Enumerate the constraint edges, and use them to update the disjoint node set.         // disjoint set（并查集，即不相交的节点集合），一种树型数据结构，
...

ColocationGraph maintains the connected components of a colocation constraint graph, and uses this information to assign a satisfying device placement to the nodes of the graph. 
The implementation uses the union- find algorithm to maintain the connected components efficiently and incrementally as edges (implied by ColocationGraph::ColocateNodes() invocations) are added.  
参考：并查集wiki

 

步骤3：如下图和code所示，source和sink节点分配在cpu上，已指定device的节点不再重新分配。分配方式有方面，见Heuristic A和Heuristic B。

 3. For each node, assign a device based on the constraints in thedisjoint node set.
  std::vector<Device*> devices;
  std::vector<Node*> second_pass;
  for (Node* node : graph_->nodes()) {
    // Skip the source and sink nodes.
    if (!node->IsOp()) {
      continue;
    }
    // Skip nodes that already have an assigned name.
    if (!node->assigned_device_name().empty()) {
      continue;
    }
    // Heuristic A: prefer to place "generators" with their only
    // consumers.
    //
    // If this is a node with no inputs and a single (non-ref)
    // consumer, we save this for a second pass, so that the
    // consumer's placement is chosen.
    if (IsGeneratorNode(node)) {    // generator node: no input, one output, not a reference-type node
      second_pass.push_back(node);
      continue;
    }
    status = colocation_graph.GetDevicesForNode(node, &devices);
    ...
    // Returns the first device in sorted devices list so we will always
    // choose the same device.
    //
    // TODO(vrv): Factor this assignment out into a pluggable
    // algorithm, so that SimplePlacer is responsible for enforcing
    // preconditions and we can experiment with other algorithms when
    // given a choice of devices. Once we have a better idea of the
    // types of heuristics we want to use and the information needed
    // to perform good placement we can add an interface for this.
    string assigned_device = devices[0]->name();
    // Heuristic B: If the node only operates on metadata, not data,
    // then it is desirable to place that metadata node with its
    // input.
    if (IsMetadataNode(node)) {   
      // Make sure that the input device type is in the list of supported
      // device types for this node.
      const Node* input = (*node->in_edges().begin())->src();
      // TODO(vrv): if the input is empty, consider postponing this
      // node's assignment to the second pass, so that we handle the
      // case where a metadata node's input comes from a backedge
      // of a loop.
      const string& input_device_name = input->assigned_device_name();
      if (CanAssignToDevice(input_device_name, devices)) {
        assigned_device = input_device_name;
      }
    }
    AssignAndLog(assigned_device, node);   // 将assigned_device分配个node节点，在步骤3中没有对符合Heuristic A的GeneratorNode分配设备，而是在步骤4中完成的
  }

bool IsGeneratorNode(const Node* node) {
  return node->num_inputs() == 0 && node->num_outputs() == 1 && node->out_edges().size() == 1 && !IsRefType(node->output_type(0));
}

bool IsMetadataNode(const Node* node) {
  const string& node_type = node->type_string();
  return (node_type == "Size" || node_type == "Shape" || node_type == "Rank");
}

步骤4：给步骤3中的Generator Node分配device。

// 4. Perform a second pass assignment for those nodes explicitly skipped during the first pass.
... 

部分参考：

http://bettercstomorrow.com/2016/07/14/distributed-tensorflow-internal-architecture-summary/

http://bettercstomorrow.com/2016/07/06/distributed-tensorflow-internal-architecture-6/  （韩文的-_-）

”tensorflow:  large-scale machine learning on heterogeneous distributed systems“

 

 

 

 

 

 

 

 

 

 

 

 

来自为知笔记(Wiz)