forked from rapidsai/rapidsmpf
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathmpi.cpp
More file actions
356 lines (320 loc) · 11.5 KB
/
mpi.cpp
File metadata and controls
356 lines (320 loc) · 11.5 KB
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
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
/**
* SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES.
* SPDX-License-Identifier: Apache-2.0
*/
#include <algorithm>
#include <array>
#include <memory>
#include <utility>
#include <mpi.h>
#include <rmm/cuda_stream_view.hpp>
#include <rapidsmpf/communicator/mpi.hpp>
#include <rapidsmpf/error.hpp>
namespace rapidsmpf {
namespace mpi {
void init(int* argc, char*** argv) {
if (!is_initialized()) {
int provided;
// Initialize MPI with the desired level of thread support
RAPIDSMPF_MPI(MPI_Init_thread(argc, argv, MPI_THREAD_MULTIPLE, &provided));
RAPIDSMPF_EXPECTS(
provided == MPI_THREAD_MULTIPLE,
"didn't get the requested thread level support: MPI_THREAD_MULTIPLE"
);
}
// Check if max MPI TAG can accommodate the OpID + TagPrefixT
int flag;
int32_t* max_tag;
MPI_Comm_get_attr(MPI_COMM_WORLD, MPI_TAG_UB, &max_tag, &flag);
RAPIDSMPF_EXPECTS(flag, "Unable to get the MPI_TAG_UB attr");
RAPIDSMPF_EXPECTS(
(*max_tag) >= Tag::max_value(),
"MPI_TAG_UB(" + std::to_string(*max_tag)
+ ") is unable to accommodate the required max tag("
+ std::to_string(Tag::max_value()) + ")"
);
}
bool is_initialized() {
int flag;
RAPIDSMPF_MPI(MPI_Initialized(&flag));
return flag;
}
void detail::check_mpi_error(int error_code, const char* file, int line) {
if (error_code != MPI_SUCCESS) {
std::array<char, MPI_MAX_ERROR_STRING> error_string;
int error_length;
MPI_Error_string(error_code, error_string.data(), &error_length);
std::cerr << "MPI error at " << file << ":" << line << ": "
<< std::string(
error_string.data(), static_cast<std::size_t>(error_length)
)
<< std::endl;
MPI_Abort(MPI_COMM_WORLD, error_code);
}
}
} // namespace mpi
namespace {
void check_mpi_thread_support() {
int level;
RAPIDSMPF_MPI(MPI_Query_thread(&level));
std::string level_str;
switch (level) {
case MPI_THREAD_SINGLE:
level_str = "MPI_THREAD_SINGLE";
break;
case MPI_THREAD_FUNNELED:
level_str = "MPI_THREAD_FUNNELED";
break;
case MPI_THREAD_SERIALIZED:
level_str = "MPI_THREAD_SERIALIZED";
break;
case MPI_THREAD_MULTIPLE:
level_str = "MPI_THREAD_MULTIPLE";
break;
default:
throw std::logic_error("MPI_Query_thread(): unknown thread level support");
}
RAPIDSMPF_EXPECTS(
level == MPI_THREAD_MULTIPLE,
"MPI thread level support " + level_str
+ " isn't sufficient, need MPI_THREAD_MULTIPLE"
);
}
} // namespace
MPI::MPI(MPI_Comm comm, config::Options options)
: comm_{comm}, logger_{this, std::move(options)} {
int rank;
int nranks;
RAPIDSMPF_MPI(MPI_Comm_rank(comm_, &rank));
RAPIDSMPF_MPI(MPI_Comm_size(comm_, &nranks));
rank_ = rank;
nranks_ = nranks;
check_mpi_thread_support();
}
std::unique_ptr<Communicator::Future> MPI::send(
std::unique_ptr<std::vector<uint8_t>> msg, Rank rank, Tag tag
) {
RAPIDSMPF_EXPECTS(
msg->size() <= std::numeric_limits<int>::max(),
"send buffer size exceeds MPI max count"
);
MPI_Request req;
RAPIDSMPF_MPI(
MPI_Isend(msg->data(), msg->size(), MPI_UINT8_T, rank, tag, comm_, &req)
);
return std::make_unique<Future>(req, std::move(msg));
}
std::unique_ptr<Communicator::Future> MPI::send(
std::unique_ptr<Buffer> msg, Rank rank, Tag tag
) {
RAPIDSMPF_EXPECTS(msg->is_latest_write_done(), "msg must be ready");
RAPIDSMPF_EXPECTS(
msg->size <= std::numeric_limits<int>::max(),
"send buffer size exceeds MPI max count"
);
MPI_Request req;
RAPIDSMPF_MPI(MPI_Isend(msg->data(), msg->size, MPI_UINT8_T, rank, tag, comm_, &req));
return std::make_unique<Future>(req, std::move(msg));
}
namespace {
void mpi_recv_impl(
Rank rank, Tag tag, auto* data, size_t size, MPI_Comm comm, MPI_Request* req
) {
RAPIDSMPF_EXPECTS(
size <= std::numeric_limits<int>::max(), "recv buffer size exceeds MPI max count"
);
RAPIDSMPF_MPI(MPI_Irecv(data, size, MPI_UINT8_T, rank, tag, comm, req));
}
} // namespace
std::unique_ptr<Communicator::Future> MPI::recv(
Rank rank, Tag tag, std::unique_ptr<Buffer> recv_buffer
) {
RAPIDSMPF_EXPECTS(
recv_buffer != nullptr, "recv buffer cannot be null", std::invalid_argument
);
RAPIDSMPF_EXPECTS(recv_buffer->is_latest_write_done(), "msg must be ready");
MPI_Request req;
mpi_recv_impl(
rank, tag, recv_buffer->exclusive_data_access(), recv_buffer->size, comm_, &req
);
return std::make_unique<Future>(req, std::move(recv_buffer));
}
std::unique_ptr<Communicator::Future> MPI::recv_sync_host_data(
Rank rank, Tag tag, std::unique_ptr<std::vector<uint8_t>> synced_buffer
) {
RAPIDSMPF_EXPECTS(
synced_buffer != nullptr,
"synced host buffer cannot be null",
std::invalid_argument
);
MPI_Request req;
mpi_recv_impl(rank, tag, synced_buffer->data(), synced_buffer->size(), comm_, &req);
return std::make_unique<Future>(req, std::move(synced_buffer));
}
std::pair<std::unique_ptr<std::vector<uint8_t>>, Rank> MPI::recv_any(Tag tag) {
int msg_available;
MPI_Status probe_status;
MPI_Message matched_msg;
RAPIDSMPF_MPI(MPI_Improbe(
MPI_ANY_SOURCE, tag, comm_, &msg_available, &matched_msg, &probe_status
));
if (!msg_available) {
return {nullptr, 0};
}
RAPIDSMPF_EXPECTS(
tag == probe_status.MPI_TAG || tag == MPI_ANY_TAG, "corrupt mpi tag"
);
MPI_Count size;
RAPIDSMPF_MPI(MPI_Get_elements_x(&probe_status, MPI_UINT8_T, &size));
RAPIDSMPF_EXPECTS(
size <= std::numeric_limits<int>::max(), "recv buffer size exceeds MPI max count"
);
auto msg = std::make_unique<std::vector<uint8_t>>(size); // TODO: uninitialize
MPI_Status msg_status;
RAPIDSMPF_MPI(
MPI_Mrecv(msg->data(), msg->size(), MPI_UINT8_T, &matched_msg, &msg_status)
);
RAPIDSMPF_MPI(MPI_Get_elements_x(&msg_status, MPI_UINT8_T, &size));
RAPIDSMPF_EXPECTS(
static_cast<std::size_t>(size) == msg->size(),
"incorrect size of the MPI_Recv message"
);
return {std::move(msg), probe_status.MPI_SOURCE};
}
std::unique_ptr<std::vector<uint8_t>> MPI::recv_from(Rank src, Tag tag) {
int msg_available;
MPI_Status probe_status;
MPI_Message matched_msg;
RAPIDSMPF_MPI(
MPI_Improbe(src, tag, comm_, &msg_available, &matched_msg, &probe_status)
);
if (!msg_available) {
return nullptr;
}
RAPIDSMPF_EXPECTS(tag == probe_status.MPI_TAG, "corrupt mpi tag");
MPI_Count size;
RAPIDSMPF_MPI(MPI_Get_elements_x(&probe_status, MPI_UINT8_T, &size));
RAPIDSMPF_EXPECTS(
size <= std::numeric_limits<int>::max(), "recv buffer size exceeds MPI max count"
);
auto msg = std::make_unique<std::vector<uint8_t>>(size); // TODO: uninitialize
MPI_Status msg_status;
RAPIDSMPF_MPI(
MPI_Mrecv(msg->data(), msg->size(), MPI_UINT8_T, &matched_msg, &msg_status)
);
RAPIDSMPF_MPI(MPI_Get_elements_x(&msg_status, MPI_UINT8_T, &size));
RAPIDSMPF_EXPECTS(
static_cast<std::size_t>(size) == msg->size(),
"incorrect size of the MPI_Recv message"
);
return msg;
}
std::pair<std::vector<std::unique_ptr<Communicator::Future>>, std::vector<std::size_t>>
MPI::test_some(std::vector<std::unique_ptr<Communicator::Future>>& future_vector) {
if (future_vector.empty()) {
return {};
}
std::vector<MPI_Request> reqs;
reqs.reserve(future_vector.size());
for (auto const& future : future_vector) {
auto mpi_future = dynamic_cast<Future const*>(future.get());
RAPIDSMPF_EXPECTS(mpi_future != nullptr, "future isn't a MPI::Future");
reqs.push_back(mpi_future->req_);
}
// Get completed requests as indices into `future_vector` (and `reqs`).
std::vector<int> indices(reqs.size());
int num_completed{0};
RAPIDSMPF_MPI(MPI_Testsome(
reqs.size(), reqs.data(), &num_completed, indices.data(), MPI_STATUSES_IGNORE
));
RAPIDSMPF_EXPECTS(
num_completed != MPI_UNDEFINED, "Expected at least one active handle."
);
if (num_completed == 0) {
return {};
}
std::vector<std::unique_ptr<Communicator::Future>> completed;
completed.reserve(static_cast<std::size_t>(num_completed));
std::ranges::transform(
indices.begin(),
indices.begin() + num_completed,
std::back_inserter(completed),
[&](std::size_t i) { return std::move(future_vector[i]); }
);
std::erase(future_vector, nullptr);
return {
std::move(completed),
std::vector<std::size_t>(indices.begin(), indices.begin() + num_completed)
};
}
std::vector<std::size_t> MPI::test_some(
std::unordered_map<std::size_t, std::unique_ptr<Communicator::Future>> const&
future_map
) {
std::vector<MPI_Request> reqs;
std::vector<std::size_t> key_reqs;
reqs.reserve(future_map.size());
key_reqs.reserve(future_map.size());
for (auto const& [key, future] : future_map) {
auto mpi_future = dynamic_cast<Future const*>(future.get());
RAPIDSMPF_EXPECTS(mpi_future != nullptr, "future isn't a MPI::Future");
reqs.push_back(mpi_future->req_);
key_reqs.push_back(key);
}
// Get completed requests as indices into `key_reqs` (and `reqs`).
std::vector<int> completed(reqs.size());
{
int num_completed{0};
RAPIDSMPF_MPI(MPI_Testsome(
reqs.size(),
reqs.data(),
&num_completed,
completed.data(),
MPI_STATUSES_IGNORE
));
completed.resize(static_cast<std::size_t>(num_completed));
}
std::vector<std::size_t> ret;
ret.reserve(completed.size());
for (int i : completed) {
ret.push_back(key_reqs.at(static_cast<std::size_t>(i)));
}
return ret;
}
std::unique_ptr<Buffer> MPI::wait(std::unique_ptr<Communicator::Future> future) {
auto mpi_future = dynamic_cast<Future*>(future.get());
RAPIDSMPF_EXPECTS(mpi_future != nullptr, "future isn't a MPI::Future");
RAPIDSMPF_MPI(MPI_Wait(&mpi_future->req_, MPI_STATUS_IGNORE));
mpi_future->data_buffer_->unlock();
return std::move(mpi_future->data_buffer_);
}
std::unique_ptr<Buffer> MPI::release_data(std::unique_ptr<Communicator::Future> future) {
auto mpi_future = dynamic_cast<Future*>(future.get());
RAPIDSMPF_EXPECTS(mpi_future != nullptr, "future isn't a MPI::Future");
RAPIDSMPF_EXPECTS(
mpi_future->data_buffer_ != nullptr, "future has no data", std::invalid_argument
);
mpi_future->data_buffer_->unlock();
return std::move(mpi_future->data_buffer_);
}
std::unique_ptr<std::vector<uint8_t>> MPI::release_sync_host_data(
std::unique_ptr<Communicator::Future> future
) {
auto mpi_future = dynamic_cast<Future*>(future.get());
RAPIDSMPF_EXPECTS(mpi_future != nullptr, "future isn't a MPI::Future");
RAPIDSMPF_EXPECTS(
mpi_future->synced_host_data_ != nullptr,
"future has no synced host data",
std::invalid_argument
);
return std::move(mpi_future->synced_host_data_);
}
std::string MPI::str() const {
int version, subversion;
RAPIDSMPF_MPI(MPI_Get_version(&version, &subversion));
std::stringstream ss;
ss << "MPI(rank=" << rank_ << ", nranks: " << nranks_ << ", mpi-version=" << version
<< "." << subversion << ")";
return ss.str();
}
} // namespace rapidsmpf