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
|
/**
* @file app.c
* @brief Template for a Host Application Source File.
*
*/
#include <stdio.h>
#include <stdlib.h>
#include <stdbool.h>
#include <string.h>
#include <unistd.h>
#include <getopt.h>
#include <assert.h>
#include <stdint.h>
#include <omp.h>
#include "../../support/timer.h"
#if NUMA
#include <numaif.h>
#include <numa.h>
void* mp_pages[1];
int mp_status[1];
int mp_nodes[1];
int numa_node_in = -1;
int numa_node_out = -1;
int numa_node_cpu = -1;
#endif
#define XSTR(x) STR(x)
#define STR(x) #x
#ifndef T
#define T int32_t
#endif
static T *A;
static T *B;
static T *C;
#if NUMA_MEMCPY
int numa_node_in_is_local = 0;
static T *A_local;
static T *B_local;
#endif
/**
* @brief compute output in the host
*/
static void vector_addition_host(unsigned int nr_elements, int t) {
omp_set_num_threads(t);
#pragma omp parallel for
for (int i = 0; i < nr_elements; i++) {
#if NUMA_MEMCPY
C[i] = A_local[i] + B_local[i];
#else
C[i] = A[i] + B[i];
#endif
}
}
// Params ---------------------------------------------------------------------
typedef struct Params {
int input_size;
int n_warmup;
int n_reps;
int exp;
int n_threads;
#if NUMA
struct bitmask* bitmask_in;
struct bitmask* bitmask_out;
int numa_node_cpu;
#endif
#if NUMA_MEMCPY
struct bitmask* bitmask_cpu;
#endif
}Params;
void usage() {
fprintf(stderr,
"\nUsage: ./program [options]"
"\n"
"\nGeneral options:"
"\n -h help"
"\n -t <T> # of threads (default=8)"
"\n -w <W> # of untimed warmup iterations (default=1)"
"\n -e <E> # of timed repetition iterations (default=3)"
"\n -x <X> Weak (0) or strong (1) scaling (default=0)"
"\n"
"\nBenchmark-specific options:"
"\n -i <I> input size (default=8M elements)"
"\n");
}
struct Params input_params(int argc, char **argv) {
struct Params p;
p.input_size = 16777216;
p.n_warmup = 1;
p.n_reps = 3;
p.exp = 1;
p.n_threads = 5;
#if NUMA
p.bitmask_in = NULL;
p.bitmask_out = NULL;
p.numa_node_cpu = -1;
#endif
#if NUMA_MEMCPY
p.bitmask_cpu = NULL;
#endif
int opt;
while((opt = getopt(argc, argv, "hi:w:e:x:t:a:b:c:")) >= 0) {
switch(opt) {
case 'h':
usage();
exit(0);
break;
case 'i': p.input_size = atoi(optarg); break;
case 'w': p.n_warmup = atoi(optarg); break;
case 'e': p.n_reps = atoi(optarg); break;
case 'x': p.exp = atoi(optarg); break;
case 't': p.n_threads = atoi(optarg); break;
#if NUMA
case 'a': p.bitmask_in = numa_parse_nodestring(optarg); break;
case 'b': p.bitmask_out = numa_parse_nodestring(optarg); break;
#if NUMA_MEMCPY
case 'c': p.numa_node_cpu = atoi(optarg);
p.bitmask_cpu = numa_parse_nodestring(optarg); break;
#else
case 'c': p.numa_node_cpu = atoi(optarg); break;
#endif // NUMA_MEMCPY
#endif // NUMA
default:
fprintf(stderr, "\nUnrecognized option!\n");
usage();
exit(0);
}
}
assert(p.n_threads > 0 && "Invalid # of ranks!");
return p;
}
/**
* @brief Main of the Host Application.
*/
int main(int argc, char **argv) {
struct Params p = input_params(argc, argv);
const unsigned int input_size = p.exp == 0 ? p.input_size * p.n_threads : p.input_size;
// Create an input file with arbitrary data.
/**
* @brief creates a "test file" by filling a buffer of 64MB with pseudo-random values
* @param nr_elements how many 32-bit elements we want the file to be
* @return the buffer address
*/
srand(0);
#if NUMA
if (p.bitmask_in) {
numa_set_membind(p.bitmask_in);
numa_free_nodemask(p.bitmask_in);
}
A = (T*) numa_alloc(input_size * sizeof(T));
B = (T*) numa_alloc(input_size * sizeof(T));
#else
A = (T*) malloc(input_size * sizeof(T));
B = (T*) malloc(input_size * sizeof(T));
#endif
#if NUMA
if (p.bitmask_out) {
numa_set_membind(p.bitmask_out);
numa_free_nodemask(p.bitmask_out);
}
C = (T*) numa_alloc(input_size * sizeof(T));
#else
C = (T*) malloc(input_size * sizeof(T));
#endif
for (unsigned int i = 0; i < input_size; i++) {
A[i] = (T) (rand());
B[i] = (T) (rand());
}
#if NUMA
#if NUMA_MEMCPY
if (p.bitmask_cpu) {
numa_set_membind(p.bitmask_cpu);
numa_free_nodemask(p.bitmask_cpu);
}
#else
struct bitmask *bitmask_all = numa_allocate_nodemask();
numa_bitmask_setall(bitmask_all);
numa_set_membind(bitmask_all);
numa_free_nodemask(bitmask_all);
#endif // NUMA_MEMCPY
#endif // NUMA
#if NUMA
mp_pages[0] = A;
if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
perror("move_pages(A)");
}
else if (mp_status[0] < 0) {
printf("move_pages error: %d", mp_status[0]);
}
else {
numa_node_in = mp_status[0];
}
mp_pages[0] = C;
if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
perror("move_pages(C)");
}
else if (mp_status[0] < 0) {
printf("move_pages error: %d", mp_status[0]);
}
else {
numa_node_out = mp_status[0];
}
numa_node_cpu = p.numa_node_cpu;
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 NUMA_MEMCPY
numa_node_in_is_local = ((numa_node_cpu == numa_node_in) || (numa_node_cpu + 8 == numa_node_in)) * 1;
#endif
Timer timer;
for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
#if NUMA_MEMCPY
start(&timer, 1, 0);
if (!numa_node_in_is_local) {
A_local = (T*) numa_alloc(input_size * sizeof(T));
B_local = (T*) numa_alloc(input_size * sizeof(T));
}
stop(&timer, 1);
start(&timer, 2, 0);
if (!numa_node_in_is_local) {
memcpy(A_local, A, input_size * sizeof(T));
memcpy(B_local, B, input_size * sizeof(T));
} else {
A_local = A;
B_local = B;
}
stop(&timer, 2);
#endif
start(&timer, 0, 0);
vector_addition_host(input_size, p.n_threads);
stop(&timer, 0);
#if NUMA_MEMCPY
start(&timer, 3, 0);
if (!numa_node_in_is_local) {
numa_free(A_local, input_size * sizeof(T));
numa_free(B_local, input_size * sizeof(T));
}
stop(&timer, 3);
#endif
unsigned int nr_threads = 0;
#pragma omp parallel
#pragma omp atomic
nr_threads++;
if (rep >= p.n_warmup) {
#if NUMA_MEMCPY
printf("[::] VA-CPU-MEMCPY | n_threads=%d e_type=%s n_elements=%d"
" numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
" | throughput_MBps=%f",
nr_threads, XSTR(T), input_size,
numa_node_in, numa_node_out, numa_node_cpu, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
input_size * 3 * sizeof(T) / timer.time[0]);
printf(" throughput_MOpps=%f",
input_size / timer.time[0]);
printf(" latency_kernel_us=%f latency_alloc_us=%f latency_memcpy_us=%f latency_free_us=%f latency_total_us=%f\n",
timer.time[0], timer.time[1], timer.time[2], timer.time[3],
timer.time[0] + timer.time[1] + timer.time[2] + timer.time[3]);
#else
printf("[::] VA-CPU | n_threads=%d e_type=%s n_elements=%d"
#if NUMA
" numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
#endif
" | throughput_MBps=%f",
nr_threads, XSTR(T), input_size,
#if NUMA
numa_node_in, numa_node_out, numa_node_cpu, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
#endif
input_size * 3 * sizeof(T) / timer.time[0]);
printf(" throughput_MOpps=%f",
input_size / timer.time[0]);
printf(" latency_us=%f\n",
timer.time[0]);
#endif // NUMA_MEMCPY
}
}
#if NUMA
numa_free(A, input_size * sizeof(T));
numa_free(B, input_size * sizeof(T));
numa_free(C, input_size * sizeof(T));
#else
free(A);
free(B);
free(C);
#endif
return 0;
}
|
