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
|
/*
* Copyright (C) by Argonne National Laboratory
* See COPYRIGHT in top-level directory
*/
#include "mpitest.h"
#include "mpi.h"
#include <stdio.h>
#include <stdlib.h>
#include "mtest_dtp.h"
#include "dtpools.h"
#include <assert.h>
/* Extended from src/pt2pt/pingping.c with partitioned APIs. */
/*
static char MTEST_Descrip[] = "Pingping test with partitioned point-to-point routines";
*/
#define MAX_TOTAL_MSG_SIZE (32 * 1024 * 1024)
#define MAXMSG (4096)
/* Used on both sender and receiver sides. If the total count of datatype elements
* cannot be completely divided, we set local partitions to 1. */
#define PARTITIONS 4
/* Utility functions */
static void set_partitions_count(int total_count, int *partitions, MPI_Count * count)
{
if (total_count % PARTITIONS == 0) {
*partitions = PARTITIONS;
*count = total_count / PARTITIONS; /* count per partition */
} else {
*partitions = 1;
*count = total_count; /* count per partition */
}
}
/* Test subroutines */
static void send_test(void *sbuf, int partitions, MPI_Count count, MPI_Datatype stype,
int dest, MPI_Comm comm)
{
MPI_Request req = MPI_REQUEST_NULL;
MPI_Psend_init(sbuf, partitions, count, stype, dest, 0, comm, MPI_INFO_NULL, &req);
MPI_Start(&req);
/* Set partition ready separately to test partial data transfer */
for (int i = 0; i < partitions; i++)
MPI_Pready(i, req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
MPI_Request_free(&req);
}
static void recv_test(void *rbuf, int partitions, MPI_Count count, MPI_Datatype rtype,
int source, MPI_Comm comm, MPI_Status * p_status)
{
MPI_Request req = MPI_REQUEST_NULL;
MPI_Precv_init(rbuf, partitions, count, rtype, source, 0, comm, MPI_INFO_NULL, &req);
MPI_Start(&req);
MPI_Wait(&req, p_status);
MPI_Request_free(&req);
}
int main(int argc, char *argv[])
{
int errs = 0, err;
int rank, size, source, dest;
int minsize = 2, nmsg, maxmsg;
int seed, testsize;
MPI_Aint sendcnt, recvcnt;
MPI_Aint maxbufsize;
MPI_Comm comm;
DTP_pool_s dtp;
struct mtest_obj send, recv;
mtest_mem_type_e sendmem, recvmem;
char *basic_type;
MTest_Init(&argc, &argv);
MTestArgList *head = MTestArgListCreate(argc, argv);
seed = MTestArgListGetInt(head, "seed");
testsize = MTestArgListGetInt(head, "testsize");
sendcnt = MTestArgListGetLong(head, "sendcnt");
recvcnt = MTestArgListGetLong(head, "recvcnt");
basic_type = MTestArgListGetString(head, "type");
sendmem = MTestArgListGetMemType(head, "sendmem");
recvmem = MTestArgListGetMemType(head, "recvmem");
maxbufsize = MTestDefaultMaxBufferSize();
err = DTP_pool_create(basic_type, sendcnt, seed, &dtp);
if (err != DTP_SUCCESS) {
fprintf(stderr, "Error while creating send pool (%s,%ld)\n", basic_type, sendcnt);
fflush(stderr);
}
MTestArgListDestroy(head);
MTest_dtp_obj_start(&send, "send", dtp, sendmem, 0, false);
MTest_dtp_obj_start(&recv, "recv", dtp, recvmem, 0, false);
int nbytes;
MPI_Type_size(dtp.DTP_base_type, &nbytes);
nbytes *= sendcnt;
maxmsg = MAX_TOTAL_MSG_SIZE / nbytes;
if (maxmsg > MAXMSG)
maxmsg = MAXMSG;
/* The following illustrates the use of the routines to
* run through a selection of communicators and datatypes.
* Use subsets of these for tests that do not involve combinations
* of communicators, datatypes, and counts of datatypes */
while (MTestGetIntracommGeneral(&comm, minsize, 1)) {
if (comm == MPI_COMM_NULL) {
/* for NULL comms, make sure these processes create the
* same number of objects, so the target knows what
* datatype layout to check for */
errs += MTEST_CREATE_AND_FREE_DTP_OBJS(dtp, testsize);
errs += MTEST_CREATE_AND_FREE_DTP_OBJS(dtp, testsize);
continue;
}
/* Determine the sender and receiver */
MPI_Comm_rank(comm, &rank);
MPI_Comm_size(comm, &size);
source = 0;
dest = size - 1;
DTP_pool_update_count(dtp, rank == source ? sendcnt : recvcnt);
/* To improve reporting of problems about operations, we
* change the error handler to errors return */
MPI_Comm_set_errhandler(comm, MPI_ERRORS_RETURN);
for (int i = 0; i < testsize; i++) {
errs += MTest_dtp_create(&send, rank == source);
errs += MTest_dtp_create(&recv, rank == dest);
int partitions = 0;
MPI_Count partition_count = 0;
MPI_Datatype dtype = MPI_DATATYPE_NULL;
if (rank == source) {
MTest_dtp_init(&send, 0, 1, sendcnt);
dtype = send.dtp_obj.DTP_datatype;
set_partitions_count(send.dtp_obj.DTP_type_count, &partitions, &partition_count);
MTestPrintfMsg(1,
"Sending partitions = %d, count = %ld (total size %d bytes) of datatype %s",
partitions, partition_count, nbytes,
DTP_obj_get_description(send.dtp_obj));
for (nmsg = 1; nmsg < maxmsg; nmsg++)
send_test((char *) send.buf + send.dtp_obj.DTP_buf_offset, partitions,
partition_count, dtype, dest, comm);
} else if (rank == dest) {
dtype = recv.dtp_obj.DTP_datatype;
set_partitions_count(recv.dtp_obj.DTP_type_count, &partitions, &partition_count);
MTestPrintfMsg(1, "Receiving partitions = %d, count = %d of datatype %s\n",
partitions, partition_count, DTP_obj_get_description(recv.dtp_obj));
for (nmsg = 1; nmsg < maxmsg; nmsg++) {
MTest_dtp_init(&recv, -1, -1, recvcnt);
MPI_Status status;
recv_test((char *) recv.buf + recv.dtp_obj.DTP_buf_offset, partitions,
partition_count, dtype, source, comm, &status);
errs += MTestCheckStatus(&status, dtp.DTP_base_type, sendcnt, source, 0,
errs < 10);
errs += MTest_dtp_check(&recv, 0, 1, sendcnt, &send, errs < 10);
}
}
MTest_dtp_destroy(&send);
MTest_dtp_destroy(&recv);
#ifdef USE_BARRIER
/* NOTE: Without MPI_Barrier, recv side can easily accumulate large unexpected queue
* across multiple batches, especially in an async test. Currently, both libfabric and ucx
* netmod does not handle large message queue well, resulting in exponential slow-downs.
* Adding barrier let the current tests pass.
*/
/* FIXME: fix netmod issues then remove the barrier (and corresponding tests). */
MPI_Barrier(comm);
#endif
}
MTestFreeComm(&comm);
}
MTest_dtp_obj_finish(&send);
MTest_dtp_obj_finish(&recv);
DTP_pool_free(dtp);
MTest_Finalize(errs);
return MTestReturnValue(errs);
}
|
