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
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
|
/*
* vim: ai ts=4 sts=4 sw=4 cinoptions=>4 expandtab
*/
#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <err.h>
#include <sys/types.h>
#include <sys/time.h>
#include <time.h>
#include <string.h>
#include <unistd.h>
#ifdef MULTITHREADED
#include <pthread.h>
#include <semaphore.h>
#endif
#ifdef NUMA
#include <numaif.h>
#include <numa.h>
#endif
/* how many runs to average by default */
#define DEFAULT_NR_LOOPS 40
/* we have 3 tests at the moment */
#define MAX_TESTS 3
/* default block size for test 2, in bytes */
#define DEFAULT_BLOCK_SIZE 262144
/* test types */
#define TEST_MEMCPY 0
#define TEST_DUMB 1
#define TEST_MCBLOCK 2
/* version number */
#define VERSION "1.5+smaug"
/*
* MBW memory bandwidth benchmark
*
* 2006, 2012 Andras.Horvath@gmail.com
* 2013 j.m.slocum@gmail.com
* (Special thanks to Stephen Pasich)
*
* http://github.com/raas/mbw
*
* compile with:
* gcc -O -o mbw mbw.c
*
* run with eg.:
*
* ./mbw 300
*
* or './mbw -h' for help
*
* watch out for swap usage (or turn off swap)
*/
#ifdef MULTITHREADED
unsigned long num_threads = 1;
volatile unsigned int done = 0;
pthread_t *threads;
sem_t start_sem, stop_sem, sync_sem;
#endif
long *arr_a, *arr_b; /* the two arrays to be copied from/to */
unsigned long long arr_size=0; /* array size (elements in array) */
unsigned int test_type;
/* fixed memcpy block size for -t2 */
unsigned long long block_size=DEFAULT_BLOCK_SIZE;
#ifdef NUMA
void* mp_pages[1];
int mp_status[1];
int mp_nodes[1];
int numa_node_a = -1;
int numa_node_b = -1;
int numa_node_cpu = -1;
struct bitmask* bitmask_a = NULL;
struct bitmask* bitmask_b = NULL;
#endif
void usage()
{
printf("mbw memory benchmark v%s, https://github.com/raas/mbw\n", VERSION);
printf("Usage: mbw [options] array_size_in_MiB\n");
printf("Options:\n");
printf(" -n: number of runs per test (0 to run forever)\n");
printf(" -a: Don't display average\n");
printf(" -t%d: memcpy test\n", TEST_MEMCPY);
printf(" -t%d: dumb (b[i]=a[i] style) test\n", TEST_DUMB);
printf(" -t%d: memcpy test with fixed block size\n", TEST_MCBLOCK);
printf(" -b <size>: block size in bytes for -t2 (default: %d)\n", DEFAULT_BLOCK_SIZE);
printf(" -q: quiet (print statistics only)\n");
#ifdef NUMA
printf(" -a <node>: allocate source array on NUMA node\n");
printf(" -b <node>: allocate target array on NUMA node\n");
printf(" -c <node>: schedule task/threads on NUME node\n");
#endif
printf("(will then use two arrays, watch out for swapping)\n");
printf("'Bandwidth' is amount of data copied over the time this operation took.\n");
printf("\nThe default is to run all tests available.\n");
}
/* ------------------------------------------------------ */
/* allocate a test array and fill it with data
* so as to force Linux to _really_ allocate it */
long *make_array()
{
unsigned long long t;
unsigned int long_size=sizeof(long);
long *a;
a=calloc(arr_size, long_size);
if(NULL==a) {
perror("Error allocating memory");
exit(1);
}
/* make sure both arrays are allocated, fill with pattern */
for(t=0; t<arr_size; t++) {
a[t]=0xaa;
}
return a;
}
#ifdef MULTITHREADED
void *thread_worker(void *arg)
{
unsigned long thread_id = (unsigned long)arg;
unsigned int long_size=sizeof(long);
unsigned long long array_bytes=arr_size*long_size;
unsigned long long t;
unsigned long long const dumb_start = thread_id * (arr_size / num_threads);
unsigned long long const dumb_stop = (thread_id + 1) * (arr_size / num_threads);
while (!done) {
if (sem_wait(&start_sem) != 0) {
err(1, "sem_wait(start_sem)");
}
if (done) {
return NULL;
}
if(test_type==TEST_MEMCPY) { /* memcpy test */
memcpy(arr_b + (thread_id * (arr_size / num_threads)), arr_a + (thread_id * (arr_size / num_threads)), array_bytes / num_threads);
} else if(test_type==TEST_MCBLOCK) { /* memcpy block test */
char* src = (char*)(arr_a + (thread_id * (arr_size / num_threads)));
char* dst = (char*)(arr_b + (thread_id * (arr_size / num_threads)));
for (t=array_bytes / num_threads; t >= block_size; t-=block_size, src+=block_size){
dst=(char *) memcpy(dst, src, block_size) + block_size;
}
if(t) {
dst=(char *) memcpy(dst, src, t) + t;
}
} else if(test_type==TEST_DUMB) { /* dumb test */
for(t=dumb_start; t<dumb_stop; t++) {
arr_b[t]=arr_a[t];
}
}
if (sem_post(&stop_sem) != 0) {
err(1, "sem_post(stop_sem)");
}
if (sem_wait(&sync_sem) != 0) {
err(1, "sem_wait(sync_sem)");
}
}
return NULL;
}
void start_threads()
{
for (unsigned int i = 0 ; i < num_threads; i++) {
sem_post(&start_sem);
}
}
void await_threads()
{
for (unsigned int i = 0 ; i < num_threads; i++) {
sem_wait(&stop_sem);
}
}
void sync_threads()
{
for (unsigned int i = 0 ; i < num_threads; i++) {
sem_post(&sync_sem);
}
}
#endif
/* actual benchmark */
/* arr_size: number of type 'long' elements in test arrays
* long_size: sizeof(long) cached
* test_type: 0=use memcpy, 1=use dumb copy loop (whatever GCC thinks best)
*
* return value: elapsed time in seconds
*/
double worker()
{
struct timespec starttime, endtime;
double te;
/* array size in bytes */
#ifdef MULTITHREADED
clock_gettime(CLOCK_MONOTONIC, &starttime);
start_threads();
await_threads();
clock_gettime(CLOCK_MONOTONIC, &endtime);
sync_threads();
#else
unsigned int long_size=sizeof(long);
unsigned long long array_bytes=arr_size*long_size;
unsigned long long t;
if(test_type==TEST_MEMCPY) { /* memcpy test */
clock_gettime(CLOCK_MONOTONIC, &starttime);
memcpy(arr_b, arr_a, array_bytes);
clock_gettime(CLOCK_MONOTONIC, &endtime);
} else if(test_type==TEST_MCBLOCK) { /* memcpy block test */
char* src = (char*)arr_a;
char* dst = (char*)arr_b;
clock_gettime(CLOCK_MONOTONIC, &starttime);
for (t=array_bytes; t >= block_size; t-=block_size, src+=block_size){
dst=(char *) memcpy(dst, src, block_size) + block_size;
}
if(t) {
dst=(char *) memcpy(dst, src, t) + t;
}
clock_gettime(CLOCK_MONOTONIC, &endtime);
} else if(test_type==TEST_DUMB) { /* dumb test */
clock_gettime(CLOCK_MONOTONIC, &starttime);
for(t=0; t<arr_size; t++) {
arr_b[t]=arr_a[t];
}
clock_gettime(CLOCK_MONOTONIC, &endtime);
}
#endif
te=((double)(endtime.tv_sec*1000000000-starttime.tv_sec*1000000000+endtime.tv_nsec-starttime.tv_nsec))/1000000000;
return te;
}
/* ------------------------------------------------------ */
/* pretty print worker's output in human-readable terms */
/* te: elapsed time in seconds
* mt: amount of transferred data in MiB
* test_type: see 'worker' above
*
* return value: -
*/
void printout(double te, double mt)
{
switch(test_type) {
case TEST_MEMCPY:
printf("e_method=MEMCPY ");
break;
case TEST_DUMB:
printf("e_method=DUMB ");
break;
case TEST_MCBLOCK:
printf("e_method=MCBLOCK ");
break;
}
printf("| data_MiB=%f time_s=%f throughput_MiBps=%f\n", mt, te, mt/te);
return;
}
/* ------------------------------------------------------ */
int main(int argc, char **argv)
{
unsigned int long_size=0;
double te, te_sum; /* time elapsed */
unsigned int i;
int o; /* getopt options */
unsigned long testno;
/* options */
/* how many runs to average? */
unsigned int nr_loops=DEFAULT_NR_LOOPS;
/* what tests to run (-t x) */
int tests[MAX_TESTS];
double mt=0; /* MiBytes transferred == array size in MiB */
int quiet=0; /* suppress extra messages */
tests[0]=0;
tests[1]=0;
tests[2]=0;
while((o=getopt(argc, argv, "ha:b:c:qn:N:t:B:")) != EOF) {
switch(o) {
case 'h':
usage();
exit(1);
break;
#ifdef NUMA
case 'a': /* NUMA node */
bitmask_a = numa_parse_nodestring(optarg);
break;
case 'b': /* NUMA node */
bitmask_b = numa_parse_nodestring(optarg);
break;
case 'c': /* NUMA node */
numa_node_cpu = strtoul(optarg, (char **)NULL, 10);
break;
#endif
case 'n': /* no. loops */
nr_loops=strtoul(optarg, (char **)NULL, 10);
break;
#ifdef MULTITHREADED
case 'N': /* no. threads */
num_threads=strtoul(optarg, (char **)NULL, 10);
break;
#endif
case 't': /* test to run */
testno=strtoul(optarg, (char **)NULL, 10);
if(testno>MAX_TESTS-1) {
printf("Error: test number must be between 0 and %d\n", MAX_TESTS-1);
exit(1);
}
tests[testno]=1;
break;
case 'B': /* block size in bytes*/
block_size=strtoull(optarg, (char **)NULL, 10);
if(0>=block_size) {
printf("Error: what block size do you mean?\n");
exit(1);
}
break;
case 'q': /* quiet */
quiet=1;
break;
default:
break;
}
}
/* default is to run all tests if no specific tests were requested */
if( (tests[0]+tests[1]+tests[2]) == 0) {
tests[0]=1;
tests[1]=1;
tests[2]=1;
}
if( nr_loops==0 && ((tests[0]+tests[1]+tests[2]) != 1) ) {
printf("Error: nr_loops can be zero if only one test selected!\n");
exit(1);
}
if(optind<argc) {
mt=strtoul(argv[optind++], (char **)NULL, 10);
} else {
printf("Error: no array size given!\n");
exit(1);
}
if(0>=mt) {
printf("Error: array size wrong!\n");
exit(1);
}
/* ------------------------------------------------------ */
long_size=sizeof(long); /* the size of long on this platform */
arr_size=1024*1024/long_size*mt; /* how many longs then in one array? */
if(arr_size*long_size < block_size) {
printf("Error: array size larger than block size (%llu bytes)!\n", block_size);
exit(1);
}
if(!quiet) {
printf("Long uses %d bytes. ", long_size);
printf("Allocating 2*%lld elements = %lld bytes of memory.\n", arr_size, 2*arr_size*long_size);
if(tests[2]) {
printf("Using %lld bytes as blocks for memcpy block copy test.\n", block_size);
}
}
#ifdef NUMA
struct bitmask *bitmask_all = numa_allocate_nodemask();
numa_bitmask_setall(bitmask_all);
if (bitmask_a) {
numa_set_membind(bitmask_a);
numa_free_nodemask(bitmask_a);
}
#endif
arr_a=make_array();
#ifdef NUMA
if (bitmask_b) {
numa_set_membind(bitmask_b);
numa_free_nodemask(bitmask_b);
}
#endif
arr_b=make_array();
#ifdef NUMA
numa_set_membind(bitmask_all);
numa_free_nodemask(bitmask_all);
#endif
#ifdef NUMA
mp_pages[0] = arr_a;
if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
perror("move_pages(arr_a)");
}
else if (mp_status[0] < 0) {
printf("move_pages error: %d", mp_status[0]);
}
else {
numa_node_a = mp_status[0];
}
mp_pages[0] = arr_b;
if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
perror("move_pages(arr_b)");
}
else if (mp_status[0] < 0) {
printf("move_pages error: %d", mp_status[0]);
}
else {
numa_node_b = mp_status[0];
}
if (numa_node_cpu != -1) {
if (numa_run_on_node(numa_node_cpu) == -1) {
perror("numa_run_on_node");
numa_node_cpu = -1;
}
}
#endif
/* ------------------------------------------------------ */
if(!quiet) {
printf("Getting down to business... Doing %d runs per test.\n", nr_loops);
}
#ifdef MULTITHREADED
if (sem_init(&start_sem, 0, 0) != 0) {
err(1, "sem_init");
}
if (sem_init(&stop_sem, 0, 0) != 0) {
err(1, "sem_init");
}
if (sem_init(&sync_sem, 0, 0) != 0) {
err(1, "sem_init");
}
threads = calloc(num_threads, sizeof(pthread_t));
for (i=0; i < num_threads; i++) {
if (pthread_create(&threads[i], NULL, thread_worker, (void*)(unsigned long)i) != 0) {
err(1, "pthread_create");
}
}
#endif
/* run all tests requested, the proper number of times */
for(test_type=0; test_type<MAX_TESTS; test_type++) {
te_sum=0;
if(tests[test_type]) {
for (i=0; nr_loops==0 || i<nr_loops; i++) {
te=worker();
te_sum+=te;
if (test_type == TEST_MEMCPY) {
printf("[::] memcpy");
} else if (test_type == TEST_DUMB) {
printf("[::] copy");
} else if (test_type == TEST_MCBLOCK) {
printf("[::] mcblock");
}
printf(" | block_size_B=%llu array_size_B=%llu ", block_size, arr_size*long_size);
#ifdef MULTITHREADED
printf("n_threads=%ld ", num_threads);
#else
printf("n_threads=1 ");
#endif
#ifdef NUMA
printf("from_numa_node=%d to_numa_node=%d cpu_numa_node=%d numa_distance_ram_ram=%d numa_distance_ram_cpu=%d numa_distance_cpu_ram=%d ", numa_node_a, numa_node_b, numa_node_cpu, numa_distance(numa_node_a, numa_node_b), numa_distance(numa_node_a, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_b));
#else
printf("from_numa_node=X to_numa_node=X cpu_numa_node=X numa_distance_ram_ram=X numa_distance_ram_cpu=X numa_distance_cpu_ram=X ");
#endif
printout(te, mt);
}
}
}
#ifdef MULTITHREADED
done = 1;
start_threads();
for (i=0; i < num_threads; i++) {
if (pthread_join(threads[i], NULL) != 0) {
err(1, "pthread_join");
}
}
#endif
free(arr_a);
free(arr_b);
return 0;
}
|