• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2016 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_DISTRIBUTED_RUNTIME_SCHEDULER_H_
17 #define TENSORFLOW_CORE_DISTRIBUTED_RUNTIME_SCHEDULER_H_
18 
19 #include <deque>
20 #include <functional>
21 #include <map>
22 #include <unordered_map>
23 #include <vector>
24 
25 #include "tensorflow/core/common_runtime/device.h"
26 #include "tensorflow/core/common_runtime/device_set.h"
27 #include "tensorflow/core/graph/costmodel.h"
28 
29 namespace tensorflow {
30 
31 class SlackAnalysis {
32  public:
33   SlackAnalysis(const Graph* g, const CostModel* cost_model);
34 
~SlackAnalysis()35   ~SlackAnalysis() {}
36 
37   // Compute the earliest possible start time for each node, based on
38   // a given cost model. 'asap_time' is indexed by node id.
39   Microseconds ComputeAsap(std::vector<Microseconds>* asap_times);
40 
41   // Compute the latest possible start time for each node, based on
42   // a given cost model. 'alap_time' is indexed by node id.
43   Microseconds ComputeAlap(std::vector<Microseconds>* alap_times);
44 
45   // Compute the "slack" of each node. 'slacks' is indexed by node id.
46   void ComputeSlack(std::vector<int64>* slacks);
47 
48  private:
49   const Graph* graph_;
50   const CostModel* cost_model_;
51 
52   TF_DISALLOW_COPY_AND_ASSIGN(SlackAnalysis);
53 };
54 
55 class GreedyScheduler {
56  public:
57   struct Sim {
58     int degree_parallelism;
59     int num_running;
60     std::vector<const Node*> ready_nodes;
61   };
62 
63   struct Event {
64     const Node* node;
65     Microseconds time;
66     bool is_completion;
67 
68     bool operator<(const Event& other) const { return time < other.time; }
69   };
70 
71   GreedyScheduler(const DeviceSet* devices, const CostModel* cost_model,
72                   const Graph* g, std::vector<int64>* priority);
73 
74   ~GreedyScheduler();
75 
76   // Computes the start time of each node given the priorities of
77   // the nodes.
78   Microseconds ComputeSchedule(std::vector<Microseconds>* start_times);
79 
80  private:
81   // Returns the ready node with the highest priority for a sim.
82   const Node* GetNodeWithHighestPriority(const std::vector<const Node*>& nodes);
83 
84   const DeviceSet* devices_;
85   const CostModel* cost_model_;
86   const Graph* graph_;
87   std::vector<int64>* priority_;
88   std::unordered_map<string, Sim*> device_states_;
89 
90   TF_DISALLOW_COPY_AND_ASSIGN(GreedyScheduler);
91 };
92 
93 class PriorityScheduler {
94  public:
95   PriorityScheduler(const DeviceSet* devices, const CostModel* cost_model,
96                     const Graph* g);
97 
~PriorityScheduler()98   ~PriorityScheduler() {}
99 
100   // Computes a schedule of the ideal start time for each node.
101   // Returns the makespan (the total running time).
102   Microseconds ComputeSchedule(std::vector<Microseconds>* start_times);
103 
104   // Computes a schedule and assigns priorities to the nodes based on
105   // the schedule. Returns the makespan.
106   Microseconds AssignPriorities(std::vector<int64>* priorities);
107 
108  private:
109   const DeviceSet* devices_;
110   const CostModel* cost_model_;
111   const Graph* graph_;
112 
113   TF_DISALLOW_COPY_AND_ASSIGN(PriorityScheduler);
114 };
115 
116 }  // namespace tensorflow
117 
118 #endif  // TENSORFLOW_CORE_DISTRIBUTED_RUNTIME_SCHEDULER_H_
119