aboutsummaryrefslogtreecommitdiffstats
path: root/tools/perf/bench
diff options
context:
space:
mode:
authorAndrea Bastoni <bastoni@cs.unc.edu>2010-05-30 19:16:45 -0400
committerAndrea Bastoni <bastoni@cs.unc.edu>2010-05-30 19:16:45 -0400
commitada47b5fe13d89735805b566185f4885f5a3f750 (patch)
tree644b88f8a71896307d71438e9b3af49126ffb22b /tools/perf/bench
parent43e98717ad40a4ae64545b5ba047c7b86aa44f4f (diff)
parent3280f21d43ee541f97f8cda5792150d2dbec20d5 (diff)
Merge branch 'wip-2.6.34' into old-private-masterarchived-private-master
Diffstat (limited to 'tools/perf/bench')
-rw-r--r--tools/perf/bench/bench.h17
-rw-r--r--tools/perf/bench/mem-memcpy.c193
-rw-r--r--tools/perf/bench/sched-messaging.c338
-rw-r--r--tools/perf/bench/sched-pipe.c127
4 files changed, 675 insertions, 0 deletions
diff --git a/tools/perf/bench/bench.h b/tools/perf/bench/bench.h
new file mode 100644
index 000000000000..f7781c6267c0
--- /dev/null
+++ b/tools/perf/bench/bench.h
@@ -0,0 +1,17 @@
1#ifndef BENCH_H
2#define BENCH_H
3
4extern int bench_sched_messaging(int argc, const char **argv, const char *prefix);
5extern int bench_sched_pipe(int argc, const char **argv, const char *prefix);
6extern int bench_mem_memcpy(int argc, const char **argv, const char *prefix __used);
7
8#define BENCH_FORMAT_DEFAULT_STR "default"
9#define BENCH_FORMAT_DEFAULT 0
10#define BENCH_FORMAT_SIMPLE_STR "simple"
11#define BENCH_FORMAT_SIMPLE 1
12
13#define BENCH_FORMAT_UNKNOWN -1
14
15extern int bench_format;
16
17#endif
diff --git a/tools/perf/bench/mem-memcpy.c b/tools/perf/bench/mem-memcpy.c
new file mode 100644
index 000000000000..89773178e894
--- /dev/null
+++ b/tools/perf/bench/mem-memcpy.c
@@ -0,0 +1,193 @@
1/*
2 * mem-memcpy.c
3 *
4 * memcpy: Simple memory copy in various ways
5 *
6 * Written by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
7 */
8#include <ctype.h>
9
10#include "../perf.h"
11#include "../util/util.h"
12#include "../util/parse-options.h"
13#include "../util/string.h"
14#include "../util/header.h"
15#include "bench.h"
16
17#include <stdio.h>
18#include <stdlib.h>
19#include <string.h>
20#include <sys/time.h>
21#include <errno.h>
22
23#define K 1024
24
25static const char *length_str = "1MB";
26static const char *routine = "default";
27static int use_clock = 0;
28static int clock_fd;
29
30static const struct option options[] = {
31 OPT_STRING('l', "length", &length_str, "1MB",
32 "Specify length of memory to copy. "
33 "available unit: B, MB, GB (upper and lower)"),
34 OPT_STRING('r', "routine", &routine, "default",
35 "Specify routine to copy"),
36 OPT_BOOLEAN('c', "clock", &use_clock,
37 "Use CPU clock for measuring"),
38 OPT_END()
39};
40
41struct routine {
42 const char *name;
43 const char *desc;
44 void * (*fn)(void *dst, const void *src, size_t len);
45};
46
47struct routine routines[] = {
48 { "default",
49 "Default memcpy() provided by glibc",
50 memcpy },
51 { NULL,
52 NULL,
53 NULL }
54};
55
56static const char * const bench_mem_memcpy_usage[] = {
57 "perf bench mem memcpy <options>",
58 NULL
59};
60
61static struct perf_event_attr clock_attr = {
62 .type = PERF_TYPE_HARDWARE,
63 .config = PERF_COUNT_HW_CPU_CYCLES
64};
65
66static void init_clock(void)
67{
68 clock_fd = sys_perf_event_open(&clock_attr, getpid(), -1, -1, 0);
69
70 if (clock_fd < 0 && errno == ENOSYS)
71 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
72 else
73 BUG_ON(clock_fd < 0);
74}
75
76static u64 get_clock(void)
77{
78 int ret;
79 u64 clk;
80
81 ret = read(clock_fd, &clk, sizeof(u64));
82 BUG_ON(ret != sizeof(u64));
83
84 return clk;
85}
86
87static double timeval2double(struct timeval *ts)
88{
89 return (double)ts->tv_sec +
90 (double)ts->tv_usec / (double)1000000;
91}
92
93int bench_mem_memcpy(int argc, const char **argv,
94 const char *prefix __used)
95{
96 int i;
97 void *dst, *src;
98 size_t length;
99 double bps = 0.0;
100 struct timeval tv_start, tv_end, tv_diff;
101 u64 clock_start, clock_end, clock_diff;
102
103 clock_start = clock_end = clock_diff = 0ULL;
104 argc = parse_options(argc, argv, options,
105 bench_mem_memcpy_usage, 0);
106
107 tv_diff.tv_sec = 0;
108 tv_diff.tv_usec = 0;
109 length = (size_t)perf_atoll((char *)length_str);
110
111 if ((s64)length <= 0) {
112 fprintf(stderr, "Invalid length:%s\n", length_str);
113 return 1;
114 }
115
116 for (i = 0; routines[i].name; i++) {
117 if (!strcmp(routines[i].name, routine))
118 break;
119 }
120 if (!routines[i].name) {
121 printf("Unknown routine:%s\n", routine);
122 printf("Available routines...\n");
123 for (i = 0; routines[i].name; i++) {
124 printf("\t%s ... %s\n",
125 routines[i].name, routines[i].desc);
126 }
127 return 1;
128 }
129
130 dst = zalloc(length);
131 if (!dst)
132 die("memory allocation failed - maybe length is too large?\n");
133
134 src = zalloc(length);
135 if (!src)
136 die("memory allocation failed - maybe length is too large?\n");
137
138 if (bench_format == BENCH_FORMAT_DEFAULT) {
139 printf("# Copying %s Bytes from %p to %p ...\n\n",
140 length_str, src, dst);
141 }
142
143 if (use_clock) {
144 init_clock();
145 clock_start = get_clock();
146 } else {
147 BUG_ON(gettimeofday(&tv_start, NULL));
148 }
149
150 routines[i].fn(dst, src, length);
151
152 if (use_clock) {
153 clock_end = get_clock();
154 clock_diff = clock_end - clock_start;
155 } else {
156 BUG_ON(gettimeofday(&tv_end, NULL));
157 timersub(&tv_end, &tv_start, &tv_diff);
158 bps = (double)((double)length / timeval2double(&tv_diff));
159 }
160
161 switch (bench_format) {
162 case BENCH_FORMAT_DEFAULT:
163 if (use_clock) {
164 printf(" %14lf Clock/Byte\n",
165 (double)clock_diff / (double)length);
166 } else {
167 if (bps < K)
168 printf(" %14lf B/Sec\n", bps);
169 else if (bps < K * K)
170 printf(" %14lfd KB/Sec\n", bps / 1024);
171 else if (bps < K * K * K)
172 printf(" %14lf MB/Sec\n", bps / 1024 / 1024);
173 else {
174 printf(" %14lf GB/Sec\n",
175 bps / 1024 / 1024 / 1024);
176 }
177 }
178 break;
179 case BENCH_FORMAT_SIMPLE:
180 if (use_clock) {
181 printf("%14lf\n",
182 (double)clock_diff / (double)length);
183 } else
184 printf("%lf\n", bps);
185 break;
186 default:
187 /* reaching this means there's some disaster: */
188 die("unknown format: %d\n", bench_format);
189 break;
190 }
191
192 return 0;
193}
diff --git a/tools/perf/bench/sched-messaging.c b/tools/perf/bench/sched-messaging.c
new file mode 100644
index 000000000000..81cee78181fa
--- /dev/null
+++ b/tools/perf/bench/sched-messaging.c
@@ -0,0 +1,338 @@
1/*
2 *
3 * sched-messaging.c
4 *
5 * messaging: Benchmark for scheduler and IPC mechanisms
6 *
7 * Based on hackbench by Rusty Russell <rusty@rustcorp.com.au>
8 * Ported to perf by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
9 *
10 */
11
12#include "../perf.h"
13#include "../util/util.h"
14#include "../util/parse-options.h"
15#include "../builtin.h"
16#include "bench.h"
17
18/* Test groups of 20 processes spraying to 20 receivers */
19#include <pthread.h>
20#include <stdio.h>
21#include <stdlib.h>
22#include <string.h>
23#include <errno.h>
24#include <unistd.h>
25#include <sys/types.h>
26#include <sys/socket.h>
27#include <sys/wait.h>
28#include <sys/time.h>
29#include <sys/poll.h>
30#include <limits.h>
31
32#define DATASIZE 100
33
34static int use_pipes = 0;
35static unsigned int loops = 100;
36static unsigned int thread_mode = 0;
37static unsigned int num_groups = 10;
38
39struct sender_context {
40 unsigned int num_fds;
41 int ready_out;
42 int wakefd;
43 int out_fds[0];
44};
45
46struct receiver_context {
47 unsigned int num_packets;
48 int in_fds[2];
49 int ready_out;
50 int wakefd;
51};
52
53static void barf(const char *msg)
54{
55 fprintf(stderr, "%s (error: %s)\n", msg, strerror(errno));
56 exit(1);
57}
58
59static void fdpair(int fds[2])
60{
61 if (use_pipes) {
62 if (pipe(fds) == 0)
63 return;
64 } else {
65 if (socketpair(AF_UNIX, SOCK_STREAM, 0, fds) == 0)
66 return;
67 }
68
69 barf(use_pipes ? "pipe()" : "socketpair()");
70}
71
72/* Block until we're ready to go */
73static void ready(int ready_out, int wakefd)
74{
75 char dummy;
76 struct pollfd pollfd = { .fd = wakefd, .events = POLLIN };
77
78 /* Tell them we're ready. */
79 if (write(ready_out, &dummy, 1) != 1)
80 barf("CLIENT: ready write");
81
82 /* Wait for "GO" signal */
83 if (poll(&pollfd, 1, -1) != 1)
84 barf("poll");
85}
86
87/* Sender sprays loops messages down each file descriptor */
88static void *sender(struct sender_context *ctx)
89{
90 char data[DATASIZE];
91 unsigned int i, j;
92
93 ready(ctx->ready_out, ctx->wakefd);
94
95 /* Now pump to every receiver. */
96 for (i = 0; i < loops; i++) {
97 for (j = 0; j < ctx->num_fds; j++) {
98 int ret, done = 0;
99
100again:
101 ret = write(ctx->out_fds[j], data + done,
102 sizeof(data)-done);
103 if (ret < 0)
104 barf("SENDER: write");
105 done += ret;
106 if (done < DATASIZE)
107 goto again;
108 }
109 }
110
111 return NULL;
112}
113
114
115/* One receiver per fd */
116static void *receiver(struct receiver_context* ctx)
117{
118 unsigned int i;
119
120 if (!thread_mode)
121 close(ctx->in_fds[1]);
122
123 /* Wait for start... */
124 ready(ctx->ready_out, ctx->wakefd);
125
126 /* Receive them all */
127 for (i = 0; i < ctx->num_packets; i++) {
128 char data[DATASIZE];
129 int ret, done = 0;
130
131again:
132 ret = read(ctx->in_fds[0], data + done, DATASIZE - done);
133 if (ret < 0)
134 barf("SERVER: read");
135 done += ret;
136 if (done < DATASIZE)
137 goto again;
138 }
139
140 return NULL;
141}
142
143static pthread_t create_worker(void *ctx, void *(*func)(void *))
144{
145 pthread_attr_t attr;
146 pthread_t childid;
147 int err;
148
149 if (!thread_mode) {
150 /* process mode */
151 /* Fork the receiver. */
152 switch (fork()) {
153 case -1:
154 barf("fork()");
155 break;
156 case 0:
157 (*func) (ctx);
158 exit(0);
159 break;
160 default:
161 break;
162 }
163
164 return (pthread_t)0;
165 }
166
167 if (pthread_attr_init(&attr) != 0)
168 barf("pthread_attr_init:");
169
170#ifndef __ia64__
171 if (pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN) != 0)
172 barf("pthread_attr_setstacksize");
173#endif
174
175 err = pthread_create(&childid, &attr, func, ctx);
176 if (err != 0) {
177 fprintf(stderr, "pthread_create failed: %s (%d)\n",
178 strerror(err), err);
179 exit(-1);
180 }
181 return childid;
182}
183
184static void reap_worker(pthread_t id)
185{
186 int proc_status;
187 void *thread_status;
188
189 if (!thread_mode) {
190 /* process mode */
191 wait(&proc_status);
192 if (!WIFEXITED(proc_status))
193 exit(1);
194 } else {
195 pthread_join(id, &thread_status);
196 }
197}
198
199/* One group of senders and receivers */
200static unsigned int group(pthread_t *pth,
201 unsigned int num_fds,
202 int ready_out,
203 int wakefd)
204{
205 unsigned int i;
206 struct sender_context *snd_ctx = malloc(sizeof(struct sender_context)
207 + num_fds * sizeof(int));
208
209 if (!snd_ctx)
210 barf("malloc()");
211
212 for (i = 0; i < num_fds; i++) {
213 int fds[2];
214 struct receiver_context *ctx = malloc(sizeof(*ctx));
215
216 if (!ctx)
217 barf("malloc()");
218
219
220 /* Create the pipe between client and server */
221 fdpair(fds);
222
223 ctx->num_packets = num_fds * loops;
224 ctx->in_fds[0] = fds[0];
225 ctx->in_fds[1] = fds[1];
226 ctx->ready_out = ready_out;
227 ctx->wakefd = wakefd;
228
229 pth[i] = create_worker(ctx, (void *)receiver);
230
231 snd_ctx->out_fds[i] = fds[1];
232 if (!thread_mode)
233 close(fds[0]);
234 }
235
236 /* Now we have all the fds, fork the senders */
237 for (i = 0; i < num_fds; i++) {
238 snd_ctx->ready_out = ready_out;
239 snd_ctx->wakefd = wakefd;
240 snd_ctx->num_fds = num_fds;
241
242 pth[num_fds+i] = create_worker(snd_ctx, (void *)sender);
243 }
244
245 /* Close the fds we have left */
246 if (!thread_mode)
247 for (i = 0; i < num_fds; i++)
248 close(snd_ctx->out_fds[i]);
249
250 /* Return number of children to reap */
251 return num_fds * 2;
252}
253
254static const struct option options[] = {
255 OPT_BOOLEAN('p', "pipe", &use_pipes,
256 "Use pipe() instead of socketpair()"),
257 OPT_BOOLEAN('t', "thread", &thread_mode,
258 "Be multi thread instead of multi process"),
259 OPT_INTEGER('g', "group", &num_groups,
260 "Specify number of groups"),
261 OPT_INTEGER('l', "loop", &loops,
262 "Specify number of loops"),
263 OPT_END()
264};
265
266static const char * const bench_sched_message_usage[] = {
267 "perf bench sched messaging <options>",
268 NULL
269};
270
271int bench_sched_messaging(int argc, const char **argv,
272 const char *prefix __used)
273{
274 unsigned int i, total_children;
275 struct timeval start, stop, diff;
276 unsigned int num_fds = 20;
277 int readyfds[2], wakefds[2];
278 char dummy;
279 pthread_t *pth_tab;
280
281 argc = parse_options(argc, argv, options,
282 bench_sched_message_usage, 0);
283
284 pth_tab = malloc(num_fds * 2 * num_groups * sizeof(pthread_t));
285 if (!pth_tab)
286 barf("main:malloc()");
287
288 fdpair(readyfds);
289 fdpair(wakefds);
290
291 total_children = 0;
292 for (i = 0; i < num_groups; i++)
293 total_children += group(pth_tab+total_children, num_fds,
294 readyfds[1], wakefds[0]);
295
296 /* Wait for everyone to be ready */
297 for (i = 0; i < total_children; i++)
298 if (read(readyfds[0], &dummy, 1) != 1)
299 barf("Reading for readyfds");
300
301 gettimeofday(&start, NULL);
302
303 /* Kick them off */
304 if (write(wakefds[1], &dummy, 1) != 1)
305 barf("Writing to start them");
306
307 /* Reap them all */
308 for (i = 0; i < total_children; i++)
309 reap_worker(pth_tab[i]);
310
311 gettimeofday(&stop, NULL);
312
313 timersub(&stop, &start, &diff);
314
315 switch (bench_format) {
316 case BENCH_FORMAT_DEFAULT:
317 printf("# %d sender and receiver %s per group\n",
318 num_fds, thread_mode ? "threads" : "processes");
319 printf("# %d groups == %d %s run\n\n",
320 num_groups, num_groups * 2 * num_fds,
321 thread_mode ? "threads" : "processes");
322 printf(" %14s: %lu.%03lu [sec]\n", "Total time",
323 diff.tv_sec,
324 (unsigned long) (diff.tv_usec/1000));
325 break;
326 case BENCH_FORMAT_SIMPLE:
327 printf("%lu.%03lu\n", diff.tv_sec,
328 (unsigned long) (diff.tv_usec/1000));
329 break;
330 default:
331 /* reaching here is something disaster */
332 fprintf(stderr, "Unknown format:%d\n", bench_format);
333 exit(1);
334 break;
335 }
336
337 return 0;
338}
diff --git a/tools/perf/bench/sched-pipe.c b/tools/perf/bench/sched-pipe.c
new file mode 100644
index 000000000000..4f77c7c27640
--- /dev/null
+++ b/tools/perf/bench/sched-pipe.c
@@ -0,0 +1,127 @@
1/*
2 *
3 * sched-pipe.c
4 *
5 * pipe: Benchmark for pipe()
6 *
7 * Based on pipe-test-1m.c by Ingo Molnar <mingo@redhat.com>
8 * http://people.redhat.com/mingo/cfs-scheduler/tools/pipe-test-1m.c
9 * Ported to perf by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
10 *
11 */
12
13#include "../perf.h"
14#include "../util/util.h"
15#include "../util/parse-options.h"
16#include "../builtin.h"
17#include "bench.h"
18
19#include <unistd.h>
20#include <stdio.h>
21#include <stdlib.h>
22#include <signal.h>
23#include <sys/wait.h>
24#include <linux/unistd.h>
25#include <string.h>
26#include <errno.h>
27#include <assert.h>
28#include <sys/time.h>
29#include <sys/types.h>
30
31#define LOOPS_DEFAULT 1000000
32static int loops = LOOPS_DEFAULT;
33
34static const struct option options[] = {
35 OPT_INTEGER('l', "loop", &loops,
36 "Specify number of loops"),
37 OPT_END()
38};
39
40static const char * const bench_sched_pipe_usage[] = {
41 "perf bench sched pipe <options>",
42 NULL
43};
44
45int bench_sched_pipe(int argc, const char **argv,
46 const char *prefix __used)
47{
48 int pipe_1[2], pipe_2[2];
49 int m = 0, i;
50 struct timeval start, stop, diff;
51 unsigned long long result_usec = 0;
52
53 /*
54 * why does "ret" exist?
55 * discarding returned value of read(), write()
56 * causes error in building environment for perf
57 */
58 int ret, wait_stat;
59 pid_t pid, retpid;
60
61 argc = parse_options(argc, argv, options,
62 bench_sched_pipe_usage, 0);
63
64 assert(!pipe(pipe_1));
65 assert(!pipe(pipe_2));
66
67 pid = fork();
68 assert(pid >= 0);
69
70 gettimeofday(&start, NULL);
71
72 if (!pid) {
73 for (i = 0; i < loops; i++) {
74 ret = read(pipe_1[0], &m, sizeof(int));
75 ret = write(pipe_2[1], &m, sizeof(int));
76 }
77 } else {
78 for (i = 0; i < loops; i++) {
79 ret = write(pipe_1[1], &m, sizeof(int));
80 ret = read(pipe_2[0], &m, sizeof(int));
81 }
82 }
83
84 gettimeofday(&stop, NULL);
85 timersub(&stop, &start, &diff);
86
87 if (pid) {
88 retpid = waitpid(pid, &wait_stat, 0);
89 assert((retpid == pid) && WIFEXITED(wait_stat));
90 } else {
91 exit(0);
92 }
93
94 switch (bench_format) {
95 case BENCH_FORMAT_DEFAULT:
96 printf("# Extecuted %d pipe operations between two tasks\n\n",
97 loops);
98
99 result_usec = diff.tv_sec * 1000000;
100 result_usec += diff.tv_usec;
101
102 printf(" %14s: %lu.%03lu [sec]\n\n", "Total time",
103 diff.tv_sec,
104 (unsigned long) (diff.tv_usec/1000));
105
106 printf(" %14lf usecs/op\n",
107 (double)result_usec / (double)loops);
108 printf(" %14d ops/sec\n",
109 (int)((double)loops /
110 ((double)result_usec / (double)1000000)));
111 break;
112
113 case BENCH_FORMAT_SIMPLE:
114 printf("%lu.%03lu\n",
115 diff.tv_sec,
116 (unsigned long) (diff.tv_usec / 1000));
117 break;
118
119 default:
120 /* reaching here is something disaster */
121 fprintf(stderr, "Unknown format:%d\n", bench_format);
122 exit(1);
123 break;
124 }
125
126 return 0;
127}
generated by cgit v1.2.2 (git 2.25.0) at 2025-04-08 10:15:34 -0400