1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
|
/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "tensorflow/contrib/gdr/gdr_worker.h"
#include "tensorflow/core/common_runtime/device.h"
#include "tensorflow/core/common_runtime/device_mgr.h"
#include "tensorflow/core/common_runtime/dma_helper.h"
#if GOOGLE_CUDA
#include "tensorflow/core/common_runtime/gpu/gpu_util.h"
#endif // GOOGLE_CUDA
#include "tensorflow/core/common_runtime/process_util.h"
#include "tensorflow/core/common_runtime/step_stats_collector.h"
#include "tensorflow/core/distributed_runtime/graph_mgr.h"
#include "tensorflow/core/distributed_runtime/rendezvous_mgr_interface.h"
#include "tensorflow/core/distributed_runtime/rpc/grpc_call.h"
#include "tensorflow/core/distributed_runtime/rpc/grpc_tensor_coding.h"
#include "tensorflow/core/distributed_runtime/rpc/grpc_util.h"
#include "tensorflow/core/distributed_runtime/worker.h"
#include "tensorflow/core/distributed_runtime/worker_cache.h"
#include "tensorflow/core/distributed_runtime/worker_session.h"
#include "tensorflow/core/framework/cancellation.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/tracing.h"
namespace tensorflow {
GdrWorker::GdrWorker(WorkerEnv* worker_env,
RemoteMemoryManager* remote_memory_manager)
: GrpcWorker(worker_env), remote_memory_manager_(remote_memory_manager) {}
void GdrWorker::GrpcRecvTensorAsync(CallOptions* opts,
const RecvTensorRequest* request,
::grpc::ByteBuffer* response,
StatusCallback done) {
const int64 step_id = request->step_id();
const string& key = request->rendezvous_key();
TRACEPRINTF("RecvTensor: %lld %s", step_id, key.c_str());
Rendezvous::ParsedKey parsed;
Status s = Rendezvous::ParseKey(key, &parsed);
Device* src_dev = nullptr;
if (s.ok()) {
s = PrepareRecvTensor(parsed, &src_dev);
}
if (!s.ok()) {
done(s);
return;
}
// Request the tensor associated with the rendezvous key. Any time
// while waiting for the tensor to be produced, up until the start
// of execution of the callback lambda body below, an RPC
// cancellation should abort the rendezvous.
opts->SetCancelCallback([this, step_id]() { AbortStep(step_id); });
const bool dma_ok = request->dma_ok();
env_->rendezvous_mgr->RecvLocalAsync(
step_id, parsed,
[this, opts, response, done, src_dev, dma_ok](
const Status& status, const Rendezvous::Args& send_args,
const Rendezvous::Args&, const Tensor& val, const bool is_dead) {
opts->ClearCancelCallback();
if (status.ok()) {
// DMA can only be used for Tensors that do not fall into
// the following three odd edge cases: 1) a zero-size
// buffer, 2) a dead tensor which has an uninit value, and
// 3) the tensor has the on_host allocation attribute,
// i.e. it's in CPU RAM *independent of its assigned
// device type*.
const bool on_host =
(src_dev->tensorflow_gpu_device_info() == nullptr) ||
send_args.alloc_attrs.on_host();
if (val.TotalBytes() > 0 && (!is_dead) &&
DMAHelper::CanUseDMA(&val) && dma_ok) {
// DMA cases.
RecvTensorResponse* proto = new RecvTensorResponse;
proto->set_is_dead(is_dead);
proto->set_send_start_micros(Env::Default()->NowMicros());
TensorProto* tensor_proto = proto->mutable_tensor();
tensor_proto->set_dtype(val.dtype());
val.shape().AsProto(tensor_proto->mutable_tensor_shape());
auto transport_options = proto->mutable_transport_options();
remote_memory_manager_->TransportOptionsFromTensor(
transport_options, val, src_dev, send_args.device_context,
on_host, [proto, done, response](const Status& s) {
if (s.ok()) {
grpc::EncodeRecvTensorResponseToByteBuffer(*proto,
response);
done(Status::OK());
} else {
done(s);
}
delete proto;
});
} else {
// Non-DMA cases.
if (src_dev->tensorflow_gpu_device_info() && (!on_host)) {
#if GOOGLE_CUDA
const DeviceContext* send_dev_context = send_args.device_context;
AllocatorAttributes alloc_attrs;
alloc_attrs.set_gpu_compatible(true);
alloc_attrs.set_on_host(true);
Allocator* alloc = src_dev->GetAllocator(alloc_attrs);
Tensor* copy = new Tensor(alloc, val.dtype(), val.shape());
CHECK(send_dev_context)
<< "send dev name: " << src_dev->name()
<< " gpu_info: " << src_dev->tensorflow_gpu_device_info();
// "val" is on a GPU. Uses GPUUtil to fill the response proto.
StatusCallback copy_ready = [response, done, copy,
is_dead](const Status& s) {
// The value is now ready to be returned on the wire.
grpc::EncodeTensorToByteBuffer(is_dead, *copy, response);
done(s);
delete copy;
};
GPUUtil::CopyGPUTensorToCPU(src_dev, send_dev_context, &val, copy,
copy_ready);
#else
done(errors::Internal("No GPU device in process"));
#endif // GOOGLE_CUDA
} else {
grpc::EncodeTensorToByteBuffer(is_dead, val, response);
done(Status::OK());
}
}
} else {
// !s.ok()
done(status);
}
});
}
} // namespace tensorflow
|
