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
|
/*
* Copyright (C) by Argonne National Laboratory
* See COPYRIGHT in top-level directory
*/
#include "mpi.h"
#include "stdio.h"
#include "stdlib.h"
#include "mpitest.h"
/*
* Tests for lock contention, including special cases within the MPICH code
* (any MPI implementation should pass these tests; in the MPICH case, our
* coverage analysis showed that the lockcontention.c test was not covering
* all cases, and in fact, this test revealed a bug in the code).
*
* In all of these tests, each process writes (or accesses) the values
* rank + i*size_of_world for NELM times.
*
* This test strives to avoid operations not strictly permitted by MPI RMA,
* for example, it doesn't target the same locations with multiple put/get
* calls in the same access epoch.
*/
#define NELM 200
#define NBLOCK 10
#define MAX_ERRS_REPORT 10
/*
* Each process writes data into the rmabuf on the process with target rank
* trank. The final result in rmabuf are the consecutive integers starting
* from 0. Each process, however, does not write a consecutive block.
* Instead, they write these locations:
*
* for i=0,...,NELM-1
* for j=0,...,NBLOCK-1
* j + NBLOCK * (rank + i * wsize)
*
* The value written is the location.
*
* In many cases, multiple RMA operations are needed. Where these must not
* overlap, the above pattern is replicated at NBLOCK*NELM*wsize.
* (NBLOCK is either 1 or NBLOCK in the code below, depending on use)
*/
static int toterrs = 0;
int testValues(int, int, int, int *, const char *);
int main(int argc, char *argv[])
{
int rank, wsize, i, j, cnt;
int *rmabuf, *localbuf, *localbuf2, *vals;
MPI_Win win;
int trank = 0;
int windowsize;
MTest_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &wsize);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
if (wsize < 2) {
fprintf(stderr, "Run this program with at least 2 processes\n");
MPI_Abort(MPI_COMM_WORLD, 1);
}
windowsize = (2 * NBLOCK + 2) * NELM * wsize;
rmabuf = (int *) malloc(windowsize * sizeof(int));
localbuf = (int *) malloc(NELM * sizeof(int));
localbuf2 = (int *) malloc(NELM * NBLOCK * sizeof(int));
vals = (int *) malloc(NELM * sizeof(int));
/*
* Initialize the buffers
*/
for (i = 0; i < NELM; i++) {
localbuf[i] = rank + i * wsize;
}
cnt = 0;
for (i = 0; i < NELM; i++) {
for (j = 0; j < NBLOCK; j++) {
localbuf2[cnt++] = j + NBLOCK * (rank + i * wsize);
}
}
for (i = 0; i < windowsize; i++) {
rmabuf[i] = -1;
}
/* Create the window */
MPI_Win_create(rmabuf, windowsize * sizeof(int), sizeof(int), MPI_INFO_NULL,
MPI_COMM_WORLD, &win);
/* Multiple puts, with contention at trank */
MPI_Barrier(MPI_COMM_WORLD);
for (i = 0; i < NELM; i++) {
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, trank, 0, win);
MPI_Put(&localbuf[i], 1, MPI_INT, trank, rank + i * wsize, 1, MPI_INT, win);
MPI_Put(&localbuf[i], 1, MPI_INT, trank, rank + (i + NELM) * wsize, 1, MPI_INT, win);
MPI_Win_unlock(trank, win);
}
MPI_Barrier(MPI_COMM_WORLD);
if (rank == trank) {
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, trank, 0, win);
toterrs += testValues(1, NELM, wsize, rmabuf, "Multiple puts (1)");
toterrs += testValues(1, NELM, wsize, rmabuf + wsize * NELM, "Multiple puts (2)");
MPI_Win_unlock(trank, win);
}
MPI_Barrier(MPI_COMM_WORLD);
/* Reinit the rmabuf */
for (i = 0; i < windowsize; i++) {
rmabuf[i] = -1;
}
MPI_Barrier(MPI_COMM_WORLD);
/* Single put with contention */
trank = 0;
for (i = 0; i < NELM; i++) {
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, trank, 0, win);
MPI_Put(&localbuf[i], 1, MPI_INT, trank, rank + i * wsize, 1, MPI_INT, win);
MPI_Win_unlock(trank, win);
}
MPI_Barrier(MPI_COMM_WORLD);
if (rank == trank) {
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, trank, 0, win);
toterrs += testValues(1, NELM, wsize, rmabuf, "Single put");
MPI_Win_unlock(trank, win);
}
/* Reinit the rmabuf */
for (i = 0; i < windowsize; i++) {
rmabuf[i] = -1;
}
/* Longer puts with contention at trank */
MPI_Barrier(MPI_COMM_WORLD);
for (i = 0; i < NELM; i++) {
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, trank, 0, win);
if (rank != trank) {
MPI_Put(&localbuf2[i * NBLOCK], NBLOCK, MPI_INT, trank,
NELM * wsize + NBLOCK * (rank + i * wsize), NBLOCK, MPI_INT, win);
MPI_Put(&localbuf2[i * NBLOCK], NBLOCK, MPI_INT, trank,
NELM * wsize + NBLOCK * (rank + (i + NELM) * wsize), NBLOCK, MPI_INT, win);
}
MPI_Put(&localbuf[i], 1, MPI_INT, trank, rank + i * wsize, 1, MPI_INT, win);
MPI_Win_unlock(trank, win);
}
MPI_Barrier(MPI_COMM_WORLD);
if (rank == trank) {
/* For simplicity in testing, set the values that rank==trank
* would have set. */
for (i = 0; i < NELM; i++) {
for (j = 0; j < NBLOCK; j++) {
rmabuf[NELM * wsize + NBLOCK * (trank + i * wsize) + j] =
j + NBLOCK * (trank + i * wsize);
rmabuf[NELM * wsize + NBLOCK * (trank + (i + NELM) * wsize) + j] =
j + NBLOCK * (trank + i * wsize);
}
}
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, trank, 0, win);
toterrs += testValues(1, NELM, wsize, rmabuf, "Long puts (1)");
toterrs += testValues(NBLOCK, NELM, wsize, rmabuf + NELM * wsize, "Long puts(2)");
toterrs += testValues(NBLOCK, NELM, wsize,
rmabuf + NELM * wsize * (1 + NBLOCK), "Long puts(3)");
MPI_Win_unlock(trank, win);
}
/* Reinit the rmabuf */
for (i = 0; i < windowsize; i++) {
rmabuf[i] = -1;
}
for (i = 0; i < NELM; i++)
vals[i] = -2;
/* Put mixed with Get */
MPI_Barrier(MPI_COMM_WORLD);
for (i = 0; i < NELM; i++) {
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, trank, 0, win);
if (rank != trank) {
MPI_Put(&localbuf2[i], NBLOCK, MPI_INT, trank,
NELM * wsize + NBLOCK * (rank + i * wsize), NBLOCK, MPI_INT, win);
MPI_Put(&localbuf[i], 1, MPI_INT, trank, rank + i * wsize, 1, MPI_INT, win);
} else {
MPI_Get(&vals[i], 1, MPI_INT, trank, i, 1, MPI_INT, win);
}
MPI_Win_unlock(trank, win);
}
MPI_Barrier(MPI_COMM_WORLD);
if (rank == trank) {
/* Just test the Get */
for (i = 0; i < wsize; i++) {
if (i == trank) {
if (vals[i] != -1) {
toterrs++;
if (toterrs < MAX_ERRS_REPORT) {
printf("put/get: vals[%d] = %d, expected -1\n", i, vals[i]);
}
}
} else if (vals[i] != i && vals[i] != -1) {
toterrs++;
if (toterrs < MAX_ERRS_REPORT) {
printf("put/get: vals[%d] = %d, expected -1 or %d\n", i, vals[i], i);
}
}
}
}
/* Contention only with get */
for (i = 0; i < windowsize; i++) {
rmabuf[i] = -i;
}
for (i = 0; i < NELM; i++)
vals[i] = -2;
MPI_Barrier(MPI_COMM_WORLD);
for (i = 0; i < NELM; i++) {
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, trank, 0, win);
MPI_Get(&vals[i], 1, MPI_INT, trank, i, 1, MPI_INT, win);
MPI_Win_unlock(trank, win);
}
MPI_Barrier(MPI_COMM_WORLD);
if (rank == trank) {
for (i = 0; i < NELM; i++) {
if (vals[i] != -i) {
toterrs++;
if (toterrs < MAX_ERRS_REPORT) {
printf("single get: vals[%d] = %d, expected %d\n", i, vals[i], -i);
}
}
}
}
/* Contention with accumulate */
MPI_Barrier(MPI_COMM_WORLD);
for (i = 0; i < NELM * wsize; i++) {
rmabuf[i] = 0;
}
MPI_Barrier(MPI_COMM_WORLD);
for (i = 0; i < NELM; i++) {
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, trank, 0, win);
MPI_Accumulate(&localbuf[i], 1, MPI_INT, trank, rank + i * wsize, 1, MPI_INT, MPI_SUM, win);
MPI_Accumulate(&localbuf[i], 1, MPI_INT, trank, rank + i * wsize, 1, MPI_INT, MPI_SUM, win);
MPI_Win_unlock(trank, win);
}
MPI_Barrier(MPI_COMM_WORLD);
if (rank == trank) {
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, trank, 0, win);
for (i = 0; i < NELM * wsize; i++) {
if (rmabuf[i] != 2 * i) {
toterrs++;
if (toterrs < MAX_ERRS_REPORT) {
printf("2 accumulate: rmabuf[%d] = %d, expected %d\n", i, rmabuf[i], 2 * i);
}
}
}
MPI_Win_unlock(trank, win);
}
MPI_Win_free(&win);
free(rmabuf);
free(localbuf);
free(localbuf2);
free(vals);
MTest_Finalize(toterrs);
return MTestReturnValue(toterrs);
}
/* Test the values in the rmabuf against the expected values. Return the
number of errors */
int testValues(int nb, int nelm, int wsize, int *rmabuf, const char *msg)
{
int i, errs = 0;
for (i = 0; i < nb * nelm * wsize; i++) {
if (rmabuf[i] != i) {
if (toterrs + errs < MAX_ERRS_REPORT) {
printf("%s:rmabuf[%d] = %d expected %d\n", msg, i, rmabuf[i], i);
}
errs++;
}
}
return errs;
}
|
