0001
0002
0003
0004
0005
0006
0007
0008
0009 #include <stdio.h>
0010 #include "bench.h"
0011 #include "../util/debug.h"
0012 #include "../util/session.h"
0013 #include "../util/stat.h"
0014 #include "../util/synthetic-events.h"
0015 #include "../util/target.h"
0016 #include "../util/thread_map.h"
0017 #include "../util/tool.h"
0018 #include "../util/util.h"
0019 #include <linux/atomic.h>
0020 #include <linux/err.h>
0021 #include <linux/time64.h>
0022 #include <subcmd/parse-options.h>
0023
0024 static unsigned int min_threads = 1;
0025 static unsigned int max_threads = UINT_MAX;
0026 static unsigned int single_iterations = 10000;
0027 static unsigned int multi_iterations = 10;
0028 static bool run_st;
0029 static bool run_mt;
0030
0031 static const struct option options[] = {
0032 OPT_BOOLEAN('s', "st", &run_st, "Run single threaded benchmark"),
0033 OPT_BOOLEAN('t', "mt", &run_mt, "Run multi-threaded benchmark"),
0034 OPT_UINTEGER('m', "min-threads", &min_threads,
0035 "Minimum number of threads in multithreaded bench"),
0036 OPT_UINTEGER('M', "max-threads", &max_threads,
0037 "Maximum number of threads in multithreaded bench"),
0038 OPT_UINTEGER('i', "single-iterations", &single_iterations,
0039 "Number of iterations used to compute single-threaded average"),
0040 OPT_UINTEGER('I', "multi-iterations", &multi_iterations,
0041 "Number of iterations used to compute multi-threaded average"),
0042 OPT_END()
0043 };
0044
0045 static const char *const bench_usage[] = {
0046 "perf bench internals synthesize <options>",
0047 NULL
0048 };
0049
0050 static atomic_t event_count;
0051
0052 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
0053 union perf_event *event __maybe_unused,
0054 struct perf_sample *sample __maybe_unused,
0055 struct machine *machine __maybe_unused)
0056 {
0057 atomic_inc(&event_count);
0058 return 0;
0059 }
0060
0061 static int do_run_single_threaded(struct perf_session *session,
0062 struct perf_thread_map *threads,
0063 struct target *target, bool data_mmap)
0064 {
0065 const unsigned int nr_threads_synthesize = 1;
0066 struct timeval start, end, diff;
0067 u64 runtime_us;
0068 unsigned int i;
0069 double time_average, time_stddev, event_average, event_stddev;
0070 int err;
0071 struct stats time_stats, event_stats;
0072
0073 init_stats(&time_stats);
0074 init_stats(&event_stats);
0075
0076 for (i = 0; i < single_iterations; i++) {
0077 atomic_set(&event_count, 0);
0078 gettimeofday(&start, NULL);
0079 err = __machine__synthesize_threads(&session->machines.host,
0080 NULL,
0081 target, threads,
0082 process_synthesized_event,
0083 true, data_mmap,
0084 nr_threads_synthesize);
0085 if (err)
0086 return err;
0087
0088 gettimeofday(&end, NULL);
0089 timersub(&end, &start, &diff);
0090 runtime_us = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
0091 update_stats(&time_stats, runtime_us);
0092 update_stats(&event_stats, atomic_read(&event_count));
0093 }
0094
0095 time_average = avg_stats(&time_stats);
0096 time_stddev = stddev_stats(&time_stats);
0097 printf(" Average %ssynthesis took: %.3f usec (+- %.3f usec)\n",
0098 data_mmap ? "data " : "", time_average, time_stddev);
0099
0100 event_average = avg_stats(&event_stats);
0101 event_stddev = stddev_stats(&event_stats);
0102 printf(" Average num. events: %.3f (+- %.3f)\n",
0103 event_average, event_stddev);
0104
0105 printf(" Average time per event %.3f usec\n",
0106 time_average / event_average);
0107 return 0;
0108 }
0109
0110 static int run_single_threaded(void)
0111 {
0112 struct perf_session *session;
0113 struct target target = {
0114 .pid = "self",
0115 };
0116 struct perf_thread_map *threads;
0117 int err;
0118
0119 perf_set_singlethreaded();
0120 session = perf_session__new(NULL, NULL);
0121 if (IS_ERR(session)) {
0122 pr_err("Session creation failed.\n");
0123 return PTR_ERR(session);
0124 }
0125 threads = thread_map__new_by_pid(getpid());
0126 if (!threads) {
0127 pr_err("Thread map creation failed.\n");
0128 err = -ENOMEM;
0129 goto err_out;
0130 }
0131
0132 puts(
0133 "Computing performance of single threaded perf event synthesis by\n"
0134 "synthesizing events on the perf process itself:");
0135
0136 err = do_run_single_threaded(session, threads, &target, false);
0137 if (err)
0138 goto err_out;
0139
0140 err = do_run_single_threaded(session, threads, &target, true);
0141
0142 err_out:
0143 if (threads)
0144 perf_thread_map__put(threads);
0145
0146 perf_session__delete(session);
0147 return err;
0148 }
0149
0150 static int do_run_multi_threaded(struct target *target,
0151 unsigned int nr_threads_synthesize)
0152 {
0153 struct timeval start, end, diff;
0154 u64 runtime_us;
0155 unsigned int i;
0156 double time_average, time_stddev, event_average, event_stddev;
0157 int err;
0158 struct stats time_stats, event_stats;
0159 struct perf_session *session;
0160
0161 init_stats(&time_stats);
0162 init_stats(&event_stats);
0163 for (i = 0; i < multi_iterations; i++) {
0164 session = perf_session__new(NULL, NULL);
0165 if (IS_ERR(session))
0166 return PTR_ERR(session);
0167
0168 atomic_set(&event_count, 0);
0169 gettimeofday(&start, NULL);
0170 err = __machine__synthesize_threads(&session->machines.host,
0171 NULL,
0172 target, NULL,
0173 process_synthesized_event,
0174 true, false,
0175 nr_threads_synthesize);
0176 if (err) {
0177 perf_session__delete(session);
0178 return err;
0179 }
0180
0181 gettimeofday(&end, NULL);
0182 timersub(&end, &start, &diff);
0183 runtime_us = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
0184 update_stats(&time_stats, runtime_us);
0185 update_stats(&event_stats, atomic_read(&event_count));
0186 perf_session__delete(session);
0187 }
0188
0189 time_average = avg_stats(&time_stats);
0190 time_stddev = stddev_stats(&time_stats);
0191 printf(" Average synthesis took: %.3f usec (+- %.3f usec)\n",
0192 time_average, time_stddev);
0193
0194 event_average = avg_stats(&event_stats);
0195 event_stddev = stddev_stats(&event_stats);
0196 printf(" Average num. events: %.3f (+- %.3f)\n",
0197 event_average, event_stddev);
0198
0199 printf(" Average time per event %.3f usec\n",
0200 time_average / event_average);
0201 return 0;
0202 }
0203
0204 static int run_multi_threaded(void)
0205 {
0206 struct target target = {
0207 .cpu_list = "0"
0208 };
0209 unsigned int nr_threads_synthesize;
0210 int err;
0211
0212 if (max_threads == UINT_MAX)
0213 max_threads = sysconf(_SC_NPROCESSORS_ONLN);
0214
0215 puts(
0216 "Computing performance of multi threaded perf event synthesis by\n"
0217 "synthesizing events on CPU 0:");
0218
0219 for (nr_threads_synthesize = min_threads;
0220 nr_threads_synthesize <= max_threads;
0221 nr_threads_synthesize++) {
0222 if (nr_threads_synthesize == 1)
0223 perf_set_singlethreaded();
0224 else
0225 perf_set_multithreaded();
0226
0227 printf(" Number of synthesis threads: %u\n",
0228 nr_threads_synthesize);
0229
0230 err = do_run_multi_threaded(&target, nr_threads_synthesize);
0231 if (err)
0232 return err;
0233 }
0234 perf_set_singlethreaded();
0235 return 0;
0236 }
0237
0238 int bench_synthesize(int argc, const char **argv)
0239 {
0240 int err = 0;
0241
0242 argc = parse_options(argc, argv, options, bench_usage, 0);
0243 if (argc) {
0244 usage_with_options(bench_usage, options);
0245 exit(EXIT_FAILURE);
0246 }
0247
0248
0249
0250
0251
0252 if (!run_st && !run_mt)
0253 run_st = true;
0254
0255 if (run_st)
0256 err = run_single_threaded();
0257
0258 if (!err && run_mt)
0259 err = run_multi_threaded();
0260
0261 return err;
0262 }