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
1 ////////////////////////////////////////////////////////////////////////////////
2 //
3 // A SimplePlacerTest method has three phases:
4 //
5 // 1. Build a TensorFlow graph, with no (or partial) device assignments.
6 // 2. Attempt to compute a placement using the SimplePlacer.
7 // 3. EITHER: test that the constraints implied by the graph are respected;
8 // or that an appropriate error was reported.
9 //
10 ////////////////////////////////////////////////////////////////////////////////
11 class SimplePlacerTest : public ::testing::Test {
12 protected:
13 SimplePlacerTest() {
14 // Build a set of 10 GPU and 10 CPU devices.
15 // NOTE: this->local_devices_ owns the device objects;
16 // this->devices_ contains borrowed pointers to the device
17 // objects.
18 for (int i = 0; i < 10; ++i) { // 添加了10 cpu和10 gpu的fake devices
19 local_devices_.emplace_back(FakeDevice::MakeCPU(
20 strings::StrCat("/job:a/replica:0/task:0/cpu:", i)));
21 devices_.AddDevice(local_devices_.back().get());
22 // Insert the GPUs in reverse order.
23 local_devices_.emplace_back(FakeDevice::MakeGPU(
24 strings::StrCat("/job:a/replica:0/task:0/gpu:", 9 - i)));
25 devices_.AddDevice(local_devices_.back().get());
26 }
27 }
28 ...
29 }
30 ...
31 // Test that a graph with no constraints will successfully assign nodes to the
32 // "best available" device (i.e. prefer GPU over CPU).
33 TEST_F(SimplePlacerTest, TestNoConstraints) {
34 Graph g(OpRegistry::Global());
35 { // Scope for temporary variables used to construct g. // 用GraphDefBuilder构建graph的结构
36 GraphDefBuilder b(GraphDefBuilder::kFailImmediately);
37 Node* input = ops::SourceOp("TestInput", b.opts().WithName("in"));
38 ops::UnaryOp("TestRelu", ops::NodeOut(input, 0), b.opts().WithName("n1"));
39 ops::UnaryOp("TestRelu", ops::NodeOut(input, 1), b.opts().WithName("n2"));
40 TF_EXPECT_OK(BuildGraph(b, &g)); // BuildGraph函数将GraphDefBuilder的图写入到Graph中
41 }
42
43 TF_EXPECT_OK(Place(&g)); // Place函数将graph中的node布放到设备列表中
44 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
45 EXPECT_DEVICE_TYPE(g, "n1", DEVICE_GPU);
46 EXPECT_DEVICE_TYPE(g, "n2", DEVICE_GPU);
47 }
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
1 // Builds the given graph, and (if successful) indexes the node
2 // names for use in placement, and later lookup.
3 Status BuildGraph(const GraphDefBuilder& builder, Graph* out_graph) {
4 TF_RETURN_IF_ERROR(builder.ToGraph(out_graph));
5 nodes_by_name_.clear();
6 for (Node* node : out_graph->nodes()) {
7 nodes_by_name_[node->name()] = node->id();
8 }
9 return Status::OK();
10 }
11 // Invokes the SimplePlacer on "graph". If no DeviceSet is specified, the
12 // placement will use the default DeviceSet (of 10 CPU and 10 GPU devices).
13 //
14 // REQUIRES: "*graph" was produced by the most recent call to BuildGraph.
15 Status Place(Graph* graph, DeviceSet* devices, SessionOptions* options) {
16 SimplePlacer placer(graph, devices, options);
17 return placer.Run();
18 }
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 // 1. First add all of the nodes. Note that steps (1) and (2)
2 // requires two passes over the nodes because the graph (and hence
3 // the constraints) may not be acyclic. 这里graph可能是有环的？
4 for (Node* node : graph_->nodes()) {
5 // Skip the source and sink nodes.
6 if (!node->IsOp()) { continue; }
7 status = colocation_graph.AddNode(*node);
8 if (!status.ok()) return AttachDef(status, node->def());
9 }
10 // 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，
11 ...
1 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. 2 The implementation uses the union- find algorithm to maintain the connected components efficiently and incrementally as edges (implied by ColocationGraph::ColocateNodes() invocations) are added. 3 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。
1 3. For each node, assign a device based on the constraints in thedisjoint node set.
2 std::vector<Device*> devices;
3 std::vector<Node*> second_pass;
4 for (Node* node : graph_->nodes()) {
5 // Skip the source and sink nodes.
6 if (!node->IsOp()) {
7 continue;
8 }
9 // Skip nodes that already have an assigned name.
10 if (!node->assigned_device_name().empty()) {
11 continue;
12 }
13 // Heuristic A: prefer to place "generators" with their only
14 // consumers.
15 //
16 // If this is a node with no inputs and a single (non-ref)
17 // consumer, we save this for a second pass, so that the
18 // consumer‘s placement is chosen.
19 if (IsGeneratorNode(node)) { // generator node: no input, one output, not a reference-type node
20 second_pass.push_back(node);
21 continue;
22 }
23 status = colocation_graph.GetDevicesForNode(node, &devices);
24 ...
25 // Returns the first device in sorted devices list so we will always
26 // choose the same device.
27 //
28 // TODO(vrv): Factor this assignment out into a pluggable
29 // algorithm, so that SimplePlacer is responsible for enforcing
30 // preconditions and we can experiment with other algorithms when
31 // given a choice of devices. Once we have a better idea of the
32 // types of heuristics we want to use and the information needed
33 // to perform good placement we can add an interface for this.
34 string assigned_device = devices[0]->name();
35 // Heuristic B: If the node only operates on metadata, not data,
36 // then it is desirable to place that metadata node with its
37 // input.
38 if (IsMetadataNode(node)) {
39 // Make sure that the input device type is in the list of supported
40 // device types for this node.
41 const Node* input = (*node->in_edges().begin())->src();
42 // TODO(vrv): if the input is empty, consider postponing this
43 // node‘s assignment to the second pass, so that we handle the
44 // case where a metadata node‘s input comes from a backedge
45 // of a loop.
46 const string& input_device_name = input->assigned_device_name();
47 if (CanAssignToDevice(input_device_name, devices)) {
48 assigned_device = input_device_name;
49 }
50 }
51 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
52 }
1 bool IsGeneratorNode(const Node* node) {
2 return node->num_inputs() == 0 && node->num_outputs() == 1 && node->out_edges().size() == 1 && !IsRefType(node->output_type(0));
3 }
1 bool IsMetadataNode(const Node* node) {
2 const string& node_type = node->type_string();
3 return (node_type == "Size" || node_type == "Shape" || node_type == "Rank");
4 }
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“
时间： 2024-07-31 12:18:41