1 /* Copyright 2015 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_COMMON_RUNTIME_PLACER_H_
17 #define TENSORFLOW_CORE_COMMON_RUNTIME_PLACER_H_
18 
19 #include <string>
20 
21 #include "tensorflow/core/common_runtime/device_set.h"
22 #include "tensorflow/core/framework/function.h"
23 #include "tensorflow/core/graph/graph.h"
24 #include "tensorflow/core/lib/core/status.h"
25 #include "tensorflow/core/platform/macros.h"
26 #include "tensorflow/core/platform/types.h"
27 #include "tensorflow/core/public/session_options.h"
28 
29 namespace tensorflow {
30 
31 // A placement algorithm that assigns the nodes of the given Graph to
32 // devices the given DeviceSet, respecting the following constraints:
33 //
34 // 1. Existing device assignments remain unchanged.
35 // 2. Requested (partial or complete) device specifications given by device name
36 //    for each node are granted.
37 // 3. Nodes connected by edges of a reference type are colocated on
38 //    the same device.
39 // 4. Given nodes "A" and "B", if node "B" has a colocation group
40 //    "@loc:A", nodes "A" and "B" will be colocated on the same device.
41 //
42 // The implementation builds a constraint graph with the same set of
43 // nodes, and edges that represent colocation constraints between
44 // nodes.  Each connected component in the resulting constraint graph
45 // is then assigned to a set of valid devices.
46 //
47 // Run() will finally assign the device to each node given the list of
48 // possible devices.
49 //
50 // TODO(mrry): "Soft" constraints, such as "place node 'x' as close as
51 // possible to node 'y' while respecting the other constraints"?
52 // TODO(mrry): Create a common interface for this and the other
53 // placement algorithms so that they may be injected into the graph
54 // builder.
55 class Placer {
56  public:
57   // Creates an instance of the Placer algorithm for the given
58   // Graph "graph" (nodes in which may or may not be assigned) on the
59   // given DeviceSet "devices".
60   // "function_name" should be set to the name of the function whose body is
61   // represented by "graph". If "graph" is not representing a function body,
62   // "function_name" should be empty.
63   //
64   // If non-null, default_local_device is used where possible as a placement for
65   // nodes which do not have a device specified, ahead of other devices which
66   // would otherwise be higher priority. default_local_device should be on the
67   // local host so that its FLR is directly accessible by the current process.
68   //
69   // The "graph", "devices", and "default_local_device" pointer arguments are
70   // borrowed by this Placer, and must outlive it.
71   Placer(Graph* graph, const string& function_name,
72          const FunctionLibraryDefinition* flib_def, const DeviceSet* devices,
73          const Device* default_local_device, bool allow_soft_placement,
74          bool log_device_placement);
75 
76   Placer(Graph* graph, const string& function_name, const DeviceSet* devices,
77          const Device* default_local_device);
78 
79   Placer(Graph* graph, const string& function_name, const DeviceSet* devices);
80 
81   ~Placer();
82 
83   // Assigns each node in this Placer's graph to a device in its
84   // set of devices.
85   //
86   // This method is not thread-safe.
87   // Run() may be invoked at most once.
88   Status Run();
89 
90  private:
91   // Returns true if the device type of 'candidate_device_name' is
92   // found in 'devices'.
93   bool CanAssignToDevice(const string& candidate_device_name,
94                          const std::vector<Device*>& devices) const;
95 
96   Graph* const graph_;  // Not owned.
97   const string function_name_;
98   const FunctionLibraryDefinition* const flib_def_;  // Not owned.
99   const DeviceSet* const devices_;                   // Not owned.
100   const Device* default_local_device_;               // Not owned.
101   const bool allow_soft_placement_;
102   const bool log_device_placement_;
103 
104   TF_DISALLOW_COPY_AND_ASSIGN(Placer);
105 };
106 
107 }  // namespace tensorflow
108 
109 #endif  // TENSORFLOW_CORE_COMMON_RUNTIME_PLACER_H_
110