1/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
2
3Licensed under the Apache License, Version 2.0 (the "License");
4you may not use this file except in compliance with the License.
5You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9Unless required by applicable law or agreed to in writing, software
10distributed under the License is distributed on an "AS IS" BASIS,
11WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12See the License for the specific language governing permissions and
13limitations under the License.
14==============================================================================*/
15
16#ifndef TENSORFLOW_CORE_DISTRIBUTED_RUNTIME_GRAPH_MGR_H_
17#define TENSORFLOW_CORE_DISTRIBUTED_RUNTIME_GRAPH_MGR_H_
18
19#include <unordered_map>
20#include <vector>
21
22#include "tensorflow/core/common_runtime/costmodel_manager.h"
23#include "tensorflow/core/common_runtime/executor.h"
24#include "tensorflow/core/common_runtime/process_function_library_runtime.h"
25#include "tensorflow/core/distributed_runtime/message_wrappers.h"
26#include "tensorflow/core/distributed_runtime/worker_env.h"
27#include "tensorflow/core/framework/cancellation.h"
28#include "tensorflow/core/framework/collective.h"
29#include "tensorflow/core/framework/cost_graph.pb.h"
30#include "tensorflow/core/framework/function.h"
31#include "tensorflow/core/lib/core/refcount.h"
32#include "tensorflow/core/platform/logging.h"
33#include "tensorflow/core/platform/macros.h"
34#include "tensorflow/core/platform/mutex.h"
35#include "tensorflow/core/platform/types.h"
36#include "tensorflow/core/protobuf/config.pb.h"
37#include "tensorflow/core/protobuf/debug.pb.h"
38#include "tensorflow/core/protobuf/worker.pb.h"
39
40namespace tensorflow {
41
42class ExecutorOpts;
43class StepStatsCollector;
44class RendezvousMgrInterface;
45class DeviceMgr;
46class WorkerSession;
47class CoordinationServiceAgent;
48
49// GraphMgr keeps track of a set of graphs that are registered with a
50// TensorFlow worker. Each registered graph is identified by a handle
51// that is generated by GraphMgr and returned to the caller.
52//
53// After a successful registration, the caller executes a graph using
54// the graph handle. Each execution is distinguished from others by a
55// caller generated global unique id "step_id". Multiple executions
56// can use the same graph concurrently and independently as long as
57// "step_id" used are different.
58//
59// Multiple threads can call GraphMgr methods concurrently.
60//
61// E.g.,
62// GraphMgr gmgr(worker_env);
63// string handle;
64// TF_CHECK_OK(gmgr.Register("session", { graph computes c = a + b },
65// &handle));
66// GraphMgr::NamedTensors in = { { "a", Tensor({1, 2}) },
67// { "b", Tensor({3, 4}) } };
68// GraphMgr::NamedTensors out = { { "c", Tensor() } };
69// TF_CHECK_OK(gmgr.Execute(handle, 0x0001, in, &out));
70// EXPECT_EQ(out["c"], Tensor({4, 6}));
71class GraphMgr {
72 public:
73 explicit GraphMgr(const WorkerEnv* worker_env, const DeviceMgr* device_mgr);
74 ~GraphMgr();
75
76 // Registers a graph. Fills in "handle". The registered graph retains a
77 // reference to cluster_flr to do cross process function calls.
78 Status Register(const string& handle, const GraphDef& gdef,
79 const GraphOptions& graph_options,
80 const DebugOptions& debug_options,
81 const ConfigProto& config_proto, int64_t collective_graph_key,
82 WorkerSession* session,
83 DistributedFunctionLibraryRuntime* cluster_flr,
84 string* graph_handle);
85
86 // Executes one step of a registered graph "handle".
87 //
88 // If "out" is not nullptr, "out" specifies all keys the execution
89 // should receive upon finish.
90 typedef std::map<string, Tensor> NamedTensors;
91 typedef std::function<void(const Status&)> StatusCallback;
92 void ExecuteAsync(const string& handle, const int64_t step_id,
93 const ExecutorOpts& opts, const NamedTensors& in,
94 WorkerSession* session, StepStatsCollector* collector,
95 MutableRunGraphResponseWrapper* response,
96 CancellationManager* cancellation_manager,
97 CoordinationServiceAgent* coordination_service_agent,
98 StatusCallback done);
99
100 Status SendInputs(const int64_t step_id, const NamedTensors& in);
101 Status RecvOutputs(const int64_t step_id, NamedTensors* out);
102 void RecvOutputsAsync(const int64_t step_id, NamedTensors* out,
103 StatusCallback done);
104
105 // Deregisters a graph.
106 Status Deregister(const string& handle);
107
108 // Deregister all graphs.
109 Status DeregisterAll();
110
111 private:
112 typedef GraphMgr ME;
113
114 struct ExecutionUnit {
115 std::unique_ptr<Graph> graph = nullptr;
116 Device* device = nullptr; // not owned.
117 Executor* root = nullptr; // not owned.
118 FunctionLibraryRuntime* lib = nullptr; // not owned.
119 // Build the cost model if this value is strictly positive.
120 int64_t build_cost_model = 0;
121 };
122
123 struct Item : public core::RefCounted {
124 // TODO(zhifengc): Keeps a copy of the original graph if the need arises.
125 // TODO(zhifengc): Stats, updated by multiple runs potentially.
126 // TODO(zhifengc): Dup-detection. Ensure step_id only run once.
127 ~Item() override;
128
129 // Session handle.
130 string session;
131
132 // Graph handle.
133 string handle;
134
135 std::unique_ptr<FunctionLibraryDefinition> lib_def;
136 // Owns the FunctionLibraryRuntime objects needed to execute functions, one
137 // per device.
138 std::unique_ptr<ProcessFunctionLibraryRuntime> proc_flr;
139 // A graph is partitioned over multiple devices. Each partition
140 // has a root executor which may call into the runtime library.
141 std::vector<ExecutionUnit> units;
142
143 // Used to deregister a cost model when cost model is required in graph
144 // manager.
145 GraphMgr* graph_mgr;
146
147 int64_t collective_graph_key;
148 };
149
150 const WorkerEnv* worker_env_; // Not owned.
151 const DeviceMgr* device_mgr_;
152
153 CostModelManager cost_model_manager_;
154
155 // Owned.
156 mutex mu_;
157 int64_t next_id_ TF_GUARDED_BY(mu_) = 0;
158
159 // If true, blocks until device has finished all queued operations in a step.
160 bool sync_on_finish_ = true;
161
162 // Table mapping graph handles to registered graphs.
163 //
164 // TODO(zhifengc): If the client does not call Deregister, we'll
165 // lose memory over time. We should implement a timeout-based
166 // mechanism to gc these graphs.
167 std::unordered_map<string, Item*> table_;
168
169 void StartParallelExecutors(
170 const string& handle, int64_t step_id, Item* item, Rendezvous* rendezvous,
171 CollectiveExecutor::Handle* ce_handle, StepStatsCollector* collector,
172 CostGraphDef* cost_graph, CancellationManager* cancellation_manager,
173 WorkerSession* session, int64_t start_time_usecs,
174 CoordinationServiceAgent* coordination_service_agent,
175 StatusCallback done);
176
177 // Don't attempt to process cost models unless explicitly requested for at
178 // least one of the items.
179 bool skip_cost_models_ = true;
180
181 void BuildCostModel(Item* item, StepStatsCollector* collector,
182 CostGraphDef* cost_graph);
183
184 Status InitItem(const string& handle, const GraphDef& gdef,
185 const GraphOptions& graph_options,
186 const DebugOptions& debug_options,
187 const ConfigProto& config_proto, int64_t collective_graph_key,
188 WorkerSession* session,
189 DistributedFunctionLibraryRuntime* cluster_flr, Item* item);
190
191 Status DecorateAndPublishGraphForDebug(const DebugOptions& debug_options,
192 Graph* graph, Device* device);
193
194 TF_DISALLOW_COPY_AND_ASSIGN(GraphMgr);
195};
196
197} // end namespace tensorflow
198
199#endif // TENSORFLOW_CORE_DISTRIBUTED_RUNTIME_GRAPH_MGR_H_
200