1 /* Measure mutex_lock for different threads and critical sections.
2 Copyright (C) 2022 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
9
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
14
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, see
17 <https://www.gnu.org/licenses/>. */
18
19 #define TEST_MAIN
20 #define TEST_NAME "pthread-mutex-locks"
21 #define TIMEOUT (20 * 60)
22
23 #include <stdio.h>
24 #include <stdlib.h>
25 #include <string.h>
26 #include <unistd.h>
27 #include <math.h>
28 #include <pthread.h>
29 #include <sys/time.h>
30 #include <sys/sysinfo.h>
31 #include "bench-timing.h"
32 #include "json-lib.h"
33
34 static pthread_mutex_t lock;
35 static pthread_mutexattr_t attr;
36 static pthread_barrier_t barrier;
37
38 #define START_ITERS 1000
39
40 #pragma GCC push_options
41 #pragma GCC optimize(1)
42
fibonacci(int i)43 static int __attribute__ ((noinline)) fibonacci (int i)
44 {
45 asm("");
46 if (i > 2)
47 return fibonacci (i - 1) + fibonacci (i - 2);
48 return 10 + i;
49 }
50
51 static void
do_filler(void)52 do_filler (void)
53 {
54 char buf1[512], buf2[512];
55 int f = fibonacci (4);
56 memcpy (buf1, buf2, f);
57 }
58
59 static void
do_filler_shared(void)60 do_filler_shared (void)
61 {
62 static char buf1[512], buf2[512];
63 int f = fibonacci (4);
64 memcpy (buf1, buf2, f);
65 }
66
67 #pragma GCC pop_options
68
69 #define UNIT_WORK_CRT do_filler_shared ()
70 #define UNIT_WORK_NON_CRT do_filler ()
71
72 static inline void
critical_section(int length)73 critical_section (int length)
74 {
75 for (int i = length; i >= 0; i--)
76 UNIT_WORK_CRT;
77 }
78
79 static inline void
non_critical_section(int length)80 non_critical_section (int length)
81 {
82 for (int i = length; i >= 0; i--)
83 UNIT_WORK_NON_CRT;
84 }
85
86 typedef struct Worker_Params
87 {
88 long iters;
89 int crt_len;
90 int non_crt_len;
91 timing_t duration;
92 } Worker_Params;
93
94 static void *
worker(void * v)95 worker (void *v)
96 {
97 timing_t start, stop;
98 Worker_Params *p = (Worker_Params *) v;
99 long iters = p->iters;
100 int crt_len = p->crt_len;
101 int non_crt_len = p->non_crt_len;
102
103 pthread_barrier_wait (&barrier);
104 TIMING_NOW (start);
105 while (iters--)
106 {
107 pthread_mutex_lock (&lock);
108 critical_section (crt_len);
109 pthread_mutex_unlock (&lock);
110 non_critical_section (non_crt_len);
111 }
112 TIMING_NOW (stop);
113
114 TIMING_DIFF (p->duration, start, stop);
115 return NULL;
116 }
117
118 static double
do_one_test(int num_threads,int crt_len,int non_crt_len,long iters)119 do_one_test (int num_threads, int crt_len, int non_crt_len, long iters)
120 {
121 int i;
122 timing_t mean;
123 Worker_Params *p, params[num_threads];
124 pthread_t threads[num_threads];
125
126 pthread_mutex_init (&lock, &attr);
127 pthread_barrier_init (&barrier, NULL, num_threads);
128
129 for (i = 0; i < num_threads; i++)
130 {
131 p = ¶ms[i];
132 p->iters = iters;
133 p->crt_len = crt_len;
134 p->non_crt_len = non_crt_len;
135 pthread_create (&threads[i], NULL, worker, (void *) p);
136 }
137 for (i = 0; i < num_threads; i++)
138 pthread_join (threads[i], NULL);
139
140 pthread_mutex_destroy (&lock);
141 pthread_barrier_destroy (&barrier);
142
143 mean = 0;
144 for (i = 0; i < num_threads; i++)
145 mean += params[i].duration;
146 mean /= num_threads;
147 return mean;
148 }
149
150 #define RUN_COUNT 10
151 #define MIN_TEST_SEC 0.01
152
153 static void
do_bench_one(const char * name,int num_threads,int crt_len,int non_crt_len,json_ctx_t * js)154 do_bench_one (const char *name, int num_threads, int crt_len, int non_crt_len,
155 json_ctx_t *js)
156 {
157 timing_t cur;
158 struct timeval ts, te;
159 double tsd, ted, td;
160 long iters, iters_limit, total_iters;
161 timing_t curs[RUN_COUNT + 2];
162 int i, j;
163 double mean, stdev;
164
165 iters = START_ITERS;
166 iters_limit = LONG_MAX / 100;
167
168 while (1)
169 {
170 gettimeofday (&ts, NULL);
171 cur = do_one_test (num_threads, crt_len, non_crt_len, iters);
172 gettimeofday (&te, NULL);
173 /* Make sure the test to run at least MIN_TEST_SEC. */
174 tsd = ts.tv_sec + ts.tv_usec / 1000000.0;
175 ted = te.tv_sec + te.tv_usec / 1000000.0;
176 td = ted - tsd;
177 if (td >= MIN_TEST_SEC || iters >= iters_limit)
178 break;
179
180 iters *= 10;
181 }
182
183 curs[0] = cur;
184 for (i = 1; i < RUN_COUNT + 2; i++)
185 curs[i] = do_one_test (num_threads, crt_len, non_crt_len, iters);
186
187 /* Sort the results so we can discard the fastest and slowest
188 times as outliers. */
189 for (i = 0; i < RUN_COUNT + 1; i++)
190 for (j = i + 1; j < RUN_COUNT + 2; j++)
191 if (curs[i] > curs[j])
192 {
193 timing_t temp = curs[i];
194 curs[i] = curs[j];
195 curs[j] = temp;
196 }
197
198 /* Calculate mean and standard deviation. */
199 mean = 0.0;
200 total_iters = iters * num_threads;
201 for (i = 1; i < RUN_COUNT + 1; i++)
202 mean += (double) curs[i] / (double) total_iters;
203 mean /= RUN_COUNT;
204
205 stdev = 0.0;
206 for (i = 1; i < RUN_COUNT + 1; i++)
207 {
208 double s = (double) curs[i] / (double) total_iters - mean;
209 stdev += s * s;
210 }
211 stdev = sqrt (stdev / (RUN_COUNT - 1));
212
213 char buf[256];
214 snprintf (buf, sizeof buf, "%s,non_crt_len=%d,crt_len=%d,threads=%d", name,
215 non_crt_len, crt_len, num_threads);
216
217 json_attr_object_begin (js, buf);
218
219 json_attr_double (js, "duration", (double) cur);
220 json_attr_double (js, "iterations", (double) total_iters);
221 json_attr_double (js, "mean", mean);
222 json_attr_double (js, "stdev", stdev);
223 json_attr_double (js, "min-outlier",
224 (double) curs[0] / (double) total_iters);
225 json_attr_double (js, "min", (double) curs[1] / (double) total_iters);
226 json_attr_double (js, "max",
227 (double) curs[RUN_COUNT] / (double) total_iters);
228 json_attr_double (js, "max-outlier",
229 (double) curs[RUN_COUNT + 1] / (double) total_iters);
230
231 json_attr_object_end (js);
232 }
233
234 #define TH_CONF_MAX 10
235
236 int
do_bench(void)237 do_bench (void)
238 {
239 int rv = 0;
240 json_ctx_t json_ctx;
241 int i, j, k;
242 int th_num, th_conf, nprocs;
243 int threads[TH_CONF_MAX];
244 int crt_lens[] = { 0, 1, 2, 4, 8, 16, 32, 64, 128 };
245 int non_crt_lens[] = { 1, 32, 128 };
246 char name[128];
247
248 json_init (&json_ctx, 2, stdout);
249 json_attr_object_begin (&json_ctx, "pthread_mutex_locks");
250
251 /* The thread config begins from 1, and increases by 2x until nprocs.
252 We also wants to test over-saturation case (1.25*nprocs). */
253 nprocs = get_nprocs ();
254 th_num = 1;
255 for (th_conf = 0; th_conf < (TH_CONF_MAX - 2) && th_num < nprocs; th_conf++)
256 {
257 threads[th_conf] = th_num;
258 th_num <<= 1;
259 }
260 threads[th_conf++] = nprocs;
261 threads[th_conf++] = nprocs + nprocs / 4;
262
263 pthread_mutexattr_init (&attr);
264 pthread_mutexattr_settype (&attr, PTHREAD_MUTEX_ADAPTIVE_NP);
265 snprintf (name, sizeof name, "type=adaptive");
266
267 for (k = 0; k < (sizeof (non_crt_lens) / sizeof (int)); k++)
268 {
269 int non_crt_len = non_crt_lens[k];
270 for (j = 0; j < (sizeof (crt_lens) / sizeof (int)); j++)
271 {
272 int crt_len = crt_lens[j];
273 for (i = 0; i < th_conf; i++)
274 {
275 th_num = threads[i];
276 do_bench_one (name, th_num, crt_len, non_crt_len, &json_ctx);
277 }
278 }
279 }
280
281 json_attr_object_end (&json_ctx);
282
283 return rv;
284 }
285
286 #define TEST_FUNCTION do_bench ()
287
288 #include "../test-skeleton.c"
289