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
|
/*
* Copyright (C) by Argonne National Laboratory
* See COPYRIGHT in top-level directory
*/
/* This program provides a simple test of send-receive performance between
two (or more) processes. This sometimes called head-to-head or
ping-ping test, as both processes send at the same time.
*/
#include "mpi.h"
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include "mpitest.h"
#include "mpithreadtest.h"
/* Alignment prevents either false-sharing on the CPU or serialization in the
* NIC's parallel TLB engine. At least it need be cacheline size. Using page
* size for optimum results. */
#define BUFFER_ALIGNMENT 4096
#define MESSAGE_SIZE 8
#define NUM_MESSAGES 64000
#define WINDOW_SIZE 64
#define ERROR_MARGIN 0.05 /* FIXME: a better margin? */
MPI_Comm *thread_comms;
double *t_elapsed;
MTEST_THREAD_RETURN_TYPE thread_fn(void *arg);
MTEST_THREAD_RETURN_TYPE thread_fn(void *arg)
{
int error;
int tid;
MPI_Comm my_comm;
int rank;
int win_i, win_post_i, win_posts;
void *buf;
int sync_buf;
MPI_Request requests[WINDOW_SIZE];
MPI_Status statuses[WINDOW_SIZE];
double t_start, t_end;
tid = (int) (long) arg;
my_comm = thread_comms[tid];
MPI_Comm_rank(my_comm, &rank);
win_posts = NUM_MESSAGES / WINDOW_SIZE;
assert(win_posts * WINDOW_SIZE == NUM_MESSAGES);
error = posix_memalign(&buf, BUFFER_ALIGNMENT, MESSAGE_SIZE * sizeof(char));
if (error) {
fprintf(stderr, "Thread %d: Error in allocating send buffer\n", tid);
}
/* Benchmark */
t_start = MPI_Wtime();
for (win_post_i = 0; win_post_i < win_posts; win_post_i++) {
for (win_i = 0; win_i < WINDOW_SIZE; win_i++) {
if (rank == 0) {
MPI_Isend(buf, MESSAGE_SIZE, MPI_CHAR, 1, tid, my_comm, &requests[win_i]);
} else {
MPI_Irecv(buf, MESSAGE_SIZE, MPI_CHAR, 0, tid, my_comm, &requests[win_i]);
}
}
MPI_Waitall(WINDOW_SIZE, requests, statuses);
}
/* Sync */
if (rank == 0) {
MPI_Recv(&sync_buf, 1, MPI_INT, 1, tid, my_comm, MPI_STATUS_IGNORE);
} else {
MPI_Send(&sync_buf, 1, MPI_INT, 0, tid, my_comm);
}
if (rank == 0) {
t_end = MPI_Wtime();
t_elapsed[tid] = t_end - t_start;
}
free(buf);
return 0;
}
int main(int argc, char *argv[])
{
int rank, size;
int provided;
int num_threads;
double onethread_msg_rate, multithread_msg_rate;
int errors;
MPI_Info info;
if (argc > 2) {
fprintf(stderr, "Can support at most only the -nthreads argument.\n");
MPI_Abort(MPI_COMM_WORLD, 1);
}
MTest_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
if (provided != MPI_THREAD_MULTIPLE) {
fprintf(stderr, "MPI_THREAD_MULTIPLE required for this test.\n");
MPI_Abort(MPI_COMM_WORLD, 1);
}
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (size != 2) {
fprintf(stderr, "please run with exactly two processes.\n");
MPI_Abort(MPI_COMM_WORLD, 1);
}
errors = MTest_thread_barrier_init();
if (errors) {
fprintf(stderr, "Could not create thread barrier\n");
MPI_Abort(MPI_COMM_WORLD, 1);
}
MTestArgList *head = MTestArgListCreate(argc, argv);
num_threads = MTestArgListGetInt(head, "nthreads");
MTestArgListDestroy(head);
thread_comms = (MPI_Comm *) malloc(sizeof(MPI_Comm) * num_threads);
t_elapsed = calloc(num_threads, sizeof(double));
/* Create a communicator per thread */
MPI_Info_create(&info);
MPI_Info_set(info, "mpi_assert_new_vci", "true");
for (int i = 0; i < num_threads; i++) {
MPI_Comm_dup_with_info(MPI_COMM_WORLD, info, &thread_comms[i]);
}
/* Run test with 1 thread */
thread_fn((void *) 0);
onethread_msg_rate = ((double) NUM_MESSAGES / t_elapsed[0]) / 1e6;
/* Run test with multiple threads */
for (int i = 1; i < num_threads; i++) {
MTest_Start_thread(thread_fn, (void *) (long) i);
}
thread_fn((void *) 0);
MTest_Join_threads();
MTest_thread_barrier_free();
/* Calculate message rate with multiple threads */
if (rank == 0) {
MTestPrintfMsg(1, "Number of messages: %d\n", NUM_MESSAGES);
MTestPrintfMsg(1, "Message size: %d\n", MESSAGE_SIZE);
MTestPrintfMsg(1, "Window size: %d\n", WINDOW_SIZE);
MTestPrintfMsg(1, "Mmsgs/s with one thread: %-10.2f\n\n", onethread_msg_rate);
MTestPrintfMsg(1, "%-10s\t%-10s\t%-10s\n", "Thread", "Mmsgs/s", "Error");
multithread_msg_rate = 0;
errors = 0;
for (int tid = 0; tid < num_threads; tid++) {
double my_msg_rate = ((double) NUM_MESSAGES / t_elapsed[tid]) / 1e6;
int my_error = 0;
if ((1 - (my_msg_rate / onethread_msg_rate)) > ERROR_MARGIN) {
/* Erroneous */
errors++;
my_error = 1;
fprintf(stderr,
"Thread %d message rate below threshold: %.2f / %.2f = %.2f (threshold = %.2f)\n",
tid, my_msg_rate, onethread_msg_rate, (my_msg_rate / onethread_msg_rate),
ERROR_MARGIN);
}
MTestPrintfMsg(1, "%-10d\t%-10.2f\t%-10d\n", tid, my_msg_rate, my_error);
multithread_msg_rate += my_msg_rate;
}
MTestPrintfMsg(1, "\n%-10s\t%-10s\t%-10s\t%-10s\n", "Size", "Threads", "Mmsgs/s", "Errors");
MTestPrintfMsg(1, "%-10d\t%-10d\t%-10.2f\t%-10d\n", MESSAGE_SIZE, num_threads,
multithread_msg_rate, errors);
}
for (int i = 0; i < num_threads; i++) {
MPI_Comm_free(&thread_comms[i]);
}
MPI_Info_free(&info);
free(thread_comms);
free(t_elapsed);
MTest_Finalize(errors);
return 0;
}
|
