Skip to content

Commit 46730e8

Browse files
wkliaowkliao
committedJan 12, 2025
alltoallw.c
1 parent 88ebabd commit 46730e8

File tree

1 file changed

+83
-83
lines changed

1 file changed

+83
-83
lines changed
 

‎MPI/alltoallw.c

+83-83
Original file line numberDiff line numberDiff line change
@@ -14,6 +14,7 @@
1414
* Usage: ./alltoallw [OPTION]
1515
* [-h] Print this help message
1616
* [-v] Verbose mode (default: no)
17+
* [-d] debug mode to check receive buffer contents (default: no)
1718
* [-n num] number of iterations (default: 1)
1819
* [-r num] every ratio processes is a receiver (default: 1)
1920
* [-l num] receive amount per iteration (default: 8 MB)
@@ -22,10 +23,10 @@
2223
* Example run command and output on screen:
2324
* % mpiexec -n 2048 ./alltoallw -n 253 -r 32
2425
*
25-
* nprocs = 2048
26-
* ntimes = 253
27-
* num_recvers = 64
28-
* individual message len = 4096 bytes
26+
* number of MPI processes = 2048
27+
* number of iterations = 253
28+
* numbe of receivers = 64
29+
* individual message length = 4096 bytes
2930
* send/recv buffer gap = 4 int(s)
3031
* Recv amount per iteration = 8388608 bytes
3132
* Time for using MPI_alltoallw = 53.60 sec
@@ -42,6 +43,7 @@
4243
#include <mpi.h>
4344

4445
static int verbose;
46+
static int debug;
4547

4648
#define ERR \
4749
if (err != MPI_SUCCESS) { \
@@ -52,6 +54,56 @@ static int verbose;
5254
goto err_out; \
5355
}
5456

57+
void initialize_bufs(int num_recvers,
58+
int len,
59+
int gap,
60+
int *sendBuf,
61+
int *recvBuf)
62+
{
63+
int i, j, k, nprocs, rank;
64+
65+
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
66+
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
67+
68+
for (i=0; i<(len + gap)*num_recvers; i++)
69+
sendBuf[i] = -2;
70+
for (i=0; i<(len + gap)*nprocs; i++)
71+
recvBuf[i] = -3;
72+
k = 0;
73+
for (i=0; i<num_recvers; i++) {
74+
for (j=0; j<len; j++) {
75+
sendBuf[k++] = rank;
76+
}
77+
k += gap;
78+
}
79+
}
80+
81+
int check_recv_buf(char *comm_op,
82+
int len,
83+
int gap,
84+
int *recvBuf)
85+
{
86+
int i, j, k, expect, err=0, nprocs, rank;
87+
88+
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
89+
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
90+
91+
k = 0;
92+
for (i=0; i<nprocs; i++) {
93+
for (j=0; j<len+gap; j++) {
94+
expect = (i == rank) ? -3 : ((j < len) ? i : -3);
95+
if (recvBuf[k] != expect) {
96+
printf("Error(%s): rank %d i=%d j=%d expect %d but got %d\n",
97+
comm_op, rank, i, j, expect, recvBuf[k]);
98+
goto err_out;
99+
}
100+
k++;
101+
}
102+
}
103+
err_out:
104+
return err;
105+
}
106+
55107
void run_alltoallw(int ntimes,
56108
int ratio,
57109
int is_receiver,
@@ -60,7 +112,6 @@ void run_alltoallw(int ntimes,
60112
int *sendBuf,
61113
int *recvBuf)
62114
{
63-
int *sendPtr, *recvPtr;
64115
int i, j, err, nprocs, rank, num_recvers;
65116
int *sendCounts, *recvCounts, *sendDisps, *recvDisps;
66117
MPI_Datatype *sendTypes, *recvTypes;
@@ -82,9 +133,6 @@ void run_alltoallw(int ntimes,
82133
sendDisps = (int*) calloc(nprocs * 2, sizeof(int));
83134
recvDisps = sendDisps + nprocs;
84135

85-
sendPtr = sendBuf;
86-
recvPtr = recvBuf;
87-
88136
/* Only receivers has non-zero data to receive */
89137
if (is_receiver) {
90138
j = 0;
@@ -113,11 +161,15 @@ void run_alltoallw(int ntimes,
113161
}
114162

115163
for (i=0; i<ntimes; i++) {
116-
err = MPI_Alltoallw(sendPtr, sendCounts, sendDisps, sendTypes,
117-
recvPtr, recvCounts, recvDisps, recvTypes,
164+
if (debug)
165+
initialize_bufs(num_recvers, len, gap, sendBuf, recvBuf);
166+
167+
err = MPI_Alltoallw(sendBuf, sendCounts, sendDisps, sendTypes,
168+
recvBuf, recvCounts, recvDisps, recvTypes,
118169
MPI_COMM_WORLD); ERR
119-
sendPtr += num_recvers * (len + gap);
120-
recvPtr += nprocs * (len + gap);
170+
171+
if (debug && is_receiver)
172+
check_recv_buf("alltoallw", len, gap, recvBuf);
121173
}
122174

123175
err_out:
@@ -156,11 +208,13 @@ void run_async_send_recv(int ntimes,
156208
reqs = (MPI_Request*) malloc(sizeof(MPI_Request) * (nprocs + num_recvers));
157209
st = (MPI_Status*) malloc(sizeof(MPI_Status) * (nprocs + num_recvers));
158210

159-
sendPtr = sendBuf;
160-
recvPtr = recvBuf;
161-
162211
for (i=0; i<ntimes; i++) {
212+
if (debug)
213+
initialize_bufs(num_recvers, len, gap, sendBuf, recvBuf);
214+
163215
nreqs = 0;
216+
sendPtr = sendBuf;
217+
recvPtr = recvBuf;
164218

165219
/* Only receivers post recv requests */
166220
if (is_receiver) {
@@ -186,6 +240,9 @@ void run_async_send_recv(int ntimes,
186240
}
187241

188242
err = MPI_Waitall(nreqs, reqs, st); ERR
243+
244+
if (debug && is_receiver)
245+
check_recv_buf("isend/irecv", len, gap, recvBuf);
189246
}
190247

191248
err_out:
@@ -198,67 +255,12 @@ void run_async_send_recv(int ntimes,
198255
printf("Time for using MPI_Issend/Irecv = %.2f sec\n", maxt);
199256
}
200257

201-
void initialize_bufs(int ntimes,
202-
int num_recvers,
203-
int len,
204-
int gap,
205-
int *sendBuf,
206-
int *recvBuf)
207-
{
208-
int i, j, k, m, nprocs, rank;
209-
210-
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
211-
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
212-
213-
for (i=0; i<(len + gap)*ntimes*num_recvers; i++)
214-
sendBuf[i] = -2;
215-
for (i=0; i<(len + gap)*ntimes*nprocs; i++)
216-
recvBuf[i] = -3;
217-
m = 0;
218-
for (i=0; i<ntimes; i++) {
219-
for (j=0; j<num_recvers; j++) {
220-
for (k=0; k<len; k++) {
221-
sendBuf[m++] = rank;
222-
}
223-
m += gap;
224-
}
225-
}
226-
}
227-
228-
int check_recv_buf(char *comm_op,
229-
int ntimes,
230-
int len,
231-
int gap,
232-
int *recvBuf)
233-
{
234-
int i, j, k, m, expect, err=0, nprocs, rank;
235-
236-
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
237-
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
238-
239-
m = 0;
240-
for (i=0; i<ntimes; i++) {
241-
for (j=0; j<nprocs; j++) {
242-
for (k=0; k<len+gap; k++) {
243-
expect = (j == rank) ? -3 : ((k < len) ? j : -3);
244-
if (recvBuf[m] != expect) {
245-
printf("Error(%s): rank %d i=%d j=%d k=%d expect %d but got %d\n",
246-
comm_op, rank, i, j, k, expect, recvBuf[m]);
247-
goto err_out;
248-
}
249-
m++;
250-
}
251-
}
252-
}
253-
err_out:
254-
return err;
255-
}
256-
257258
/*----< usage() >------------------------------------------------------------*/
258259
static void usage (char *argv0) {
259260
char *help = "Usage: %s [OPTION]\n\
260261
[-h] Print this help message\n\
261262
[-v] Verbose mode (default: no)\n\
263+
[-d] debug mode to check receive buffer contents (default: no)\n\
262264
[-n num] number of iterations (default: 1)\n\
263265
[-r num] every ratio processes is a receiver (default: 1)\n\
264266
[-l num] receive amount per iteration (default: 8 MB)\n\
@@ -279,17 +281,21 @@ int main(int argc, char **argv) {
279281
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
280282

281283
verbose = 0;
284+
debug = 0;
282285
ntimes = 1;
283286
ratio = 1;
284287
block_len = 8 * 1024 * 1024;
285288
gap = 4;
286289

287290
/* command-line arguments */
288-
while ((i = getopt (argc, argv, "hvn:r:l:g:")) != EOF)
291+
while ((i = getopt (argc, argv, "hdvn:r:l:g:")) != EOF)
289292
switch (i) {
290293
case 'v':
291294
verbose = 1;
292295
break;
296+
case 'd':
297+
debug = 1;
298+
break;
293299
case 'n':
294300
ntimes = atoi(optarg);
295301
break;
@@ -328,7 +334,7 @@ int main(int argc, char **argv) {
328334
if (verbose) fflush(stdout);
329335

330336
if (rank == 0) {
331-
printf("nprocs = %d\n", nprocs);
337+
printf("number of MPI processes = %d\n", nprocs);
332338
printf("number of iterations = %d\n", ntimes);
333339
printf("numbe of receivers = %d\n", num_recvers);
334340
printf("individual message length = %zd bytes\n",len*sizeof(int));
@@ -337,25 +343,19 @@ int main(int argc, char **argv) {
337343
}
338344

339345
/* allocate and initialize send and recevive buffer */
340-
sendBuf = (int*) malloc(sizeof(int) * (len + gap) * ntimes * num_recvers);
341-
recvBuf = (int*) malloc(sizeof(int) * (len + gap) * ntimes * nprocs);
346+
sendBuf = (int*) malloc(sizeof(int) * (len + gap) * num_recvers);
347+
recvBuf = (int*) malloc(sizeof(int) * (len + gap) * nprocs);
342348

343-
initialize_bufs(ntimes, num_recvers, len, gap, sendBuf, recvBuf);
349+
initialize_bufs(num_recvers, len, gap, sendBuf, recvBuf);
344350

345351
MPI_Barrier(MPI_COMM_WORLD);
346352
run_alltoallw(ntimes, ratio, is_receiver, len, gap, sendBuf, recvBuf);
347353

348-
if (is_receiver)
349-
check_recv_buf("alltoallw", ntimes, len, gap, recvBuf);
350-
351-
initialize_bufs(ntimes, num_recvers, len, gap, sendBuf, recvBuf);
354+
initialize_bufs(num_recvers, len, gap, sendBuf, recvBuf);
352355

353356
MPI_Barrier(MPI_COMM_WORLD);
354357
run_async_send_recv(ntimes, ratio, is_receiver, len, gap, sendBuf, recvBuf);
355358

356-
if (is_receiver)
357-
check_recv_buf("isend/irecv", ntimes, len, gap, recvBuf);
358-
359359
free(recvBuf);
360360
free(sendBuf);
361361

0 commit comments

Comments
 (0)