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
|
/*
* Copyright (C) by Argonne National Laboratory
* See COPYRIGHT in top-level directory
*/
/* This test is a simplification of the one in perf/manyrma.c that tests
for correct handling of the case where many RMA operations occur between
synchronization events.
This is one of the ways that RMA may be used, and is used in the
reference implementation of the graph500 benchmark.
*/
#include "mpi.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include "mpitest.h"
#define MAX_COUNT 65536*4/16
#define MAX_RMA_SIZE 2 /* 16 in manyrma performance test */
#define MAX_RUNS 8
#define MAX_ITER_TIME 5.0 /* seconds */
typedef enum { SYNC_NONE = 0,
SYNC_ALL = -1, SYNC_FENCE = 1, SYNC_LOCK = 2, SYNC_PSCW = 4
} sync_t;
typedef enum { RMA_NONE = 0, RMA_ALL = -1, RMA_PUT = 1, RMA_ACC = 2, RMA_GET = 4 } rma_t;
/* Note GET not yet implemented */
/* By default, run only a subset of the available tests, to keep the
total runtime reasonably short. Command line arguments may be used
to run other tests. */
sync_t syncChoice = SYNC_FENCE;
rma_t rmaChoice = RMA_ACC;
static int verbose = 0;
void RunAccFence(MPI_Win win, int destRank, int cnt, int sz);
void RunAccLock(MPI_Win win, int destRank, int cnt, int sz);
void RunPutFence(MPI_Win win, int destRank, int cnt, int sz);
void RunPutLock(MPI_Win win, int destRank, int cnt, int sz);
void RunAccPSCW(MPI_Win win, int destRank, int cnt, int sz,
MPI_Group exposureGroup, MPI_Group accessGroup);
void RunPutPSCW(MPI_Win win, int destRank, int cnt, int sz,
MPI_Group exposureGroup, MPI_Group accessGroup);
int main(int argc, char *argv[])
{
int arraysize, i, cnt, sz, maxCount = MAX_COUNT, *arraybuffer;
int wrank, wsize, destRank, srcRank;
MPI_Win win;
MPI_Group wgroup, accessGroup, exposureGroup;
int maxSz = MAX_RMA_SIZE;
double start, end;
MTest_Init(&argc, &argv);
for (i = 1; i < argc; i++) {
if (strcmp(argv[i], "-put") == 0) {
if (rmaChoice == RMA_ALL)
rmaChoice = RMA_NONE;
rmaChoice |= RMA_PUT;
} else if (strcmp(argv[i], "-acc") == 0) {
if (rmaChoice == RMA_ALL)
rmaChoice = RMA_NONE;
rmaChoice |= RMA_ACC;
} else if (strcmp(argv[i], "-fence") == 0) {
if (syncChoice == SYNC_ALL)
syncChoice = SYNC_NONE;
syncChoice |= SYNC_FENCE;
} else if (strcmp(argv[i], "-lock") == 0) {
if (syncChoice == SYNC_ALL)
syncChoice = SYNC_NONE;
syncChoice |= SYNC_LOCK;
} else if (strcmp(argv[i], "-pscw") == 0) {
if (syncChoice == SYNC_ALL)
syncChoice = SYNC_NONE;
syncChoice |= SYNC_PSCW;
} else if (strcmp(argv[i], "-maxsz") == 0) {
i++;
maxSz = atoi(argv[i]);
} else if (strcmp(argv[i], "-maxcount") == 0) {
i++;
maxCount = atoi(argv[i]);
} else {
fprintf(stderr, "Unrecognized argument %s\n", argv[i]);
fprintf(stderr,
"%s [ -put ] [ -acc ] [ -lock ] [ -fence ] [ -pscw ] [ -maxsz msgsize ]\n",
argv[0]);
MPI_Abort(MPI_COMM_WORLD, 1);
}
}
MPI_Comm_rank(MPI_COMM_WORLD, &wrank);
MPI_Comm_size(MPI_COMM_WORLD, &wsize);
destRank = wrank + 1;
while (destRank >= wsize)
destRank = destRank - wsize;
srcRank = wrank - 1;
if (srcRank < 0)
srcRank += wsize;
/* Create groups for PSCW */
MPI_Comm_group(MPI_COMM_WORLD, &wgroup);
MPI_Group_incl(wgroup, 1, &destRank, &accessGroup);
MPI_Group_incl(wgroup, 1, &srcRank, &exposureGroup);
MPI_Group_free(&wgroup);
arraysize = maxSz * MAX_COUNT;
#ifdef USE_WIN_ALLOCATE
MPI_Win_allocate(arraysize * sizeof(int), (int) sizeof(int), MPI_INFO_NULL,
MPI_COMM_WORLD, &arraybuffer, &win);
if (!arraybuffer) {
fprintf(stderr, "Unable to allocate %d words\n", arraysize);
MPI_Abort(MPI_COMM_WORLD, 1);
}
#else
arraybuffer = (int *) malloc(arraysize * sizeof(int));
if (!arraybuffer) {
fprintf(stderr, "Unable to allocate %d words\n", arraysize);
MPI_Abort(MPI_COMM_WORLD, 1);
}
MPI_Win_create(arraybuffer, arraysize * sizeof(int), (int) sizeof(int),
MPI_INFO_NULL, MPI_COMM_WORLD, &win);
#endif
if (maxCount > MAX_COUNT) {
fprintf(stderr, "MaxCount must not exceed %d\n", MAX_COUNT);
MPI_Abort(MPI_COMM_WORLD, 1);
}
if ((syncChoice & SYNC_FENCE) && (rmaChoice & RMA_ACC)) {
for (sz = 1; sz <= maxSz; sz = sz + sz) {
if (wrank == 0 && verbose)
printf("Accumulate with fence, %d elements\n", sz);
for (cnt = 1; cnt <= maxCount; cnt *= 2) {
start = MPI_Wtime();
RunAccFence(win, destRank, cnt, sz);
end = MPI_Wtime();
if (end - start > MAX_ITER_TIME)
break;
}
}
}
if ((syncChoice & SYNC_LOCK) && (rmaChoice & RMA_ACC)) {
for (sz = 1; sz <= maxSz; sz = sz + sz) {
if (wrank == 0 && verbose)
printf("Accumulate with lock, %d elements\n", sz);
for (cnt = 1; cnt <= maxCount; cnt *= 2) {
start = MPI_Wtime();
RunAccLock(win, destRank, cnt, sz);
end = MPI_Wtime();
if (end - start > MAX_ITER_TIME)
break;
}
}
}
if ((syncChoice & SYNC_FENCE) && (rmaChoice & RMA_PUT)) {
for (sz = 1; sz <= maxSz; sz = sz + sz) {
if (wrank == 0 && verbose)
printf("Put with fence, %d elements\n", sz);
for (cnt = 1; cnt <= maxCount; cnt *= 2) {
start = MPI_Wtime();
RunPutFence(win, destRank, cnt, sz);
end = MPI_Wtime();
if (end - start > MAX_ITER_TIME)
break;
}
}
}
if ((syncChoice & SYNC_LOCK) && (rmaChoice & RMA_PUT)) {
for (sz = 1; sz <= maxSz; sz = sz + sz) {
if (wrank == 0 && verbose)
printf("Put with lock, %d elements\n", sz);
for (cnt = 1; cnt <= maxCount; cnt *= 2) {
start = MPI_Wtime();
RunPutLock(win, destRank, cnt, sz);
end = MPI_Wtime();
if (end - start > MAX_ITER_TIME)
break;
}
}
}
if ((syncChoice & SYNC_PSCW) && (rmaChoice & RMA_PUT)) {
for (sz = 1; sz <= maxSz; sz = sz + sz) {
if (wrank == 0 && verbose)
printf("Put with pscw, %d elements\n", sz);
for (cnt = 1; cnt <= maxCount; cnt *= 2) {
start = MPI_Wtime();
RunPutPSCW(win, destRank, cnt, sz, exposureGroup, accessGroup);
end = MPI_Wtime();
if (end - start > MAX_ITER_TIME)
break;
}
}
}
if ((syncChoice & SYNC_PSCW) && (rmaChoice & RMA_ACC)) {
for (sz = 1; sz <= maxSz; sz = sz + sz) {
if (wrank == 0 && verbose)
printf("Accumulate with pscw, %d elements\n", sz);
for (cnt = 1; cnt <= maxCount; cnt *= 2) {
start = MPI_Wtime();
RunAccPSCW(win, destRank, cnt, sz, exposureGroup, accessGroup);
end = MPI_Wtime();
if (end - start > MAX_ITER_TIME)
break;
}
}
}
MPI_Win_free(&win);
#ifndef USE_WIN_ALLOCATE
free(arraybuffer);
#endif
MPI_Group_free(&accessGroup);
MPI_Group_free(&exposureGroup);
MTest_Finalize(0);
return 0;
}
void RunAccFence(MPI_Win win, int destRank, int cnt, int sz)
{
int k, i, j;
int *buf = malloc(sz * sizeof(int));
for (i = 0; i < sz; i++) {
buf[i] = i;
}
for (k = 0; k < MAX_RUNS; k++) {
MPI_Barrier(MPI_COMM_WORLD);
MPI_Win_fence(0, win);
j = 0;
for (i = 0; i < cnt; i++) {
MPI_Accumulate(buf, sz, MPI_INT, destRank, j, sz, MPI_INT, MPI_SUM, win);
j += sz;
}
MPI_Win_fence(0, win);
}
free(buf);
}
void RunAccLock(MPI_Win win, int destRank, int cnt, int sz)
{
int k, i, j;
int *buf = malloc(sz * sizeof(int));
for (i = 0; i < sz; i++) {
buf[i] = i;
}
for (k = 0; k < MAX_RUNS; k++) {
MPI_Barrier(MPI_COMM_WORLD);
MPI_Win_lock(MPI_LOCK_SHARED, destRank, 0, win);
j = 0;
for (i = 0; i < cnt; i++) {
MPI_Accumulate(buf, sz, MPI_INT, destRank, j, sz, MPI_INT, MPI_SUM, win);
j += sz;
}
MPI_Win_unlock(destRank, win);
}
free(buf);
}
void RunPutFence(MPI_Win win, int destRank, int cnt, int sz)
{
int k, i, j;
int *buf = malloc(sz * sizeof(int));
for (k = 0; k < MAX_RUNS; k++) {
MPI_Barrier(MPI_COMM_WORLD);
MPI_Win_fence(0, win);
j = 0;
for (i = 0; i < cnt; i++) {
MPI_Put(buf, sz, MPI_INT, destRank, j, sz, MPI_INT, win);
j += sz;
}
MPI_Win_fence(0, win);
}
free(buf);
}
void RunPutLock(MPI_Win win, int destRank, int cnt, int sz)
{
int k, i, j;
int *buf = malloc(sz * sizeof(int));
for (k = 0; k < MAX_RUNS; k++) {
MPI_Barrier(MPI_COMM_WORLD);
MPI_Win_lock(MPI_LOCK_SHARED, destRank, 0, win);
j = 0;
for (i = 0; i < cnt; i++) {
MPI_Put(buf, sz, MPI_INT, destRank, j, sz, MPI_INT, win);
j += sz;
}
MPI_Win_unlock(destRank, win);
}
free(buf);
}
void RunPutPSCW(MPI_Win win, int destRank, int cnt, int sz,
MPI_Group exposureGroup, MPI_Group accessGroup)
{
int k, i, j;
int *buf = malloc(sz * sizeof(int));
for (k = 0; k < MAX_RUNS; k++) {
MPI_Barrier(MPI_COMM_WORLD);
MPI_Win_post(exposureGroup, 0, win);
MPI_Win_start(accessGroup, 0, win);
j = 0;
for (i = 0; i < cnt; i++) {
MPI_Put(buf, sz, MPI_INT, destRank, j, sz, MPI_INT, win);
j += sz;
}
MPI_Win_complete(win);
MPI_Win_wait(win);
}
free(buf);
}
void RunAccPSCW(MPI_Win win, int destRank, int cnt, int sz,
MPI_Group exposureGroup, MPI_Group accessGroup)
{
int k, i, j;
int *buf = malloc(sz * sizeof(int));
for (k = 0; k < MAX_RUNS; k++) {
MPI_Barrier(MPI_COMM_WORLD);
MPI_Win_post(exposureGroup, 0, win);
MPI_Win_start(accessGroup, 0, win);
j = 0;
for (i = 0; i < cnt; i++) {
MPI_Accumulate(buf, sz, MPI_INT, destRank, j, sz, MPI_INT, MPI_SUM, win);
j += sz;
}
MPI_Win_complete(win);
MPI_Win_wait(win);
}
free(buf);
}
|
