1 // SPDX-License-Identifier: GPL-2.0
2 #ifdef HAVE_EVENTFD_SUPPORT
4 * Copyright (C) 2018 Davidlohr Bueso.
6 * This program benchmarks concurrent epoll_wait(2) monitoring multiple
7 * file descriptors under one or two load balancing models. The first,
8 * and default, is the single/combined queueing (which refers to a single
9 * epoll instance for N worker threads):
13 * [combined queue] .---> [worker C]
17 * While the second model, enabled via --multiq option, uses multiple
18 * queueing (which refers to one epoll instance per worker). For example,
19 * short lived tcp connections in a high throughput httpd server will
20 * distribute the accept()'ing connections across CPUs. In this case each
21 * worker does a limited amount of processing.
23 * [queue A] ---> [worker]
24 * [queue B] ---> [worker]
25 * [queue C] ---> [worker]
26 * [queue D] ---> [worker]
27 * [queue E] ---> [worker]
29 * Naturally, the single queue will enforce more concurrency on the epoll
30 * instance, and can therefore scale poorly compared to multiple queues.
31 * However, this is a benchmark raw data and must be taken with a grain of
32 * salt when choosing how to make use of sys_epoll.
34 * Each thread has a number of private, nonblocking file descriptors,
35 * referred to as fdmap. A writer thread will constantly be writing to
36 * the fdmaps of all threads, minimizing each threads's chances of
37 * epoll_wait not finding any ready read events and blocking as this
38 * is not what we want to stress. The size of the fdmap can be adjusted
39 * by the user; enlarging the value will increase the chances of
40 * epoll_wait(2) blocking as the lineal writer thread will take "longer",
41 * at least at a high level.
43 * Note that because fds are private to each thread, this workload does
44 * not stress scenarios where multiple tasks are awoken per ready IO; ie:
45 * EPOLLEXCLUSIVE semantics.
47 * The end result/metric is throughput: number of ops/second where an
48 * operation consists of:
50 * epoll_wait(2) + [others]
52 * ... where [others] is the cost of re-adding the fd (EPOLLET),
53 * or rearming it (EPOLLONESHOT).
56 * The purpose of this is program is that it be useful for measuring
57 * kernel related changes to the sys_epoll, and not comparing different
58 * IO polling methods, for example. Hence everything is very adhoc and
59 * outputs raw microbenchmark numbers. Also this uses eventfd, similar
60 * tools tend to use pipes or sockets, but the result is the same.
63 /* For the CLR_() macros */
72 #include <linux/compiler.h>
73 #include <linux/kernel.h>
75 #include <sys/resource.h>
76 #include <sys/epoll.h>
77 #include <sys/eventfd.h>
78 #include <sys/types.h>
79 #include <perf/cpumap.h>
81 #include "../util/stat.h"
82 #include <subcmd/parse-options.h>
87 #define printinfo(fmt, arg...) \
88 do { if (__verbose) { printf(fmt, ## arg); fflush(stdout); } } while (0)
90 static unsigned int nthreads = 0;
91 static unsigned int nsecs = 8;
92 static bool wdone, done, __verbose, randomize, nonblocking;
95 * epoll related shared variables.
98 /* Maximum number of nesting allowed inside epoll sets */
99 #define EPOLL_MAXNESTS 4
102 static int *epollfdp;
103 static bool noaffinity;
104 static unsigned int nested = 0;
105 static bool et; /* edge-trigger */
107 static bool multiq; /* use an epoll instance per thread */
109 /* amount of fds to monitor, per thread */
110 static unsigned int nfds = 64;
112 static pthread_mutex_t thread_lock;
113 static unsigned int threads_starting;
114 static struct stats throughput_stats;
115 static pthread_cond_t thread_parent, thread_worker;
119 int epollfd; /* for --multiq */
125 static const struct option options[] = {
126 /* general benchmark options */
127 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
128 OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"),
129 OPT_UINTEGER('f', "nfds", &nfds, "Specify amount of file descriptors to monitor for each thread"),
130 OPT_BOOLEAN( 'n', "noaffinity", &noaffinity, "Disables CPU affinity"),
131 OPT_BOOLEAN('R', "randomize", &randomize, "Enable random write behaviour (default is lineal)"),
132 OPT_BOOLEAN( 'v', "verbose", &__verbose, "Verbose mode"),
134 /* epoll specific options */
135 OPT_BOOLEAN( 'm', "multiq", &multiq, "Use multiple epoll instances (one per thread)"),
136 OPT_BOOLEAN( 'B', "nonblocking", &nonblocking, "Nonblocking epoll_wait(2) behaviour"),
137 OPT_UINTEGER( 'N', "nested", &nested, "Nesting level epoll hierarchy (default is 0, no nesting)"),
138 OPT_BOOLEAN( 'S', "oneshot", &oneshot, "Use EPOLLONESHOT semantics"),
139 OPT_BOOLEAN( 'E', "edge", &et, "Use Edge-triggered interface (default is LT)"),
144 static const char * const bench_epoll_wait_usage[] = {
145 "perf bench epoll wait <options>",
151 * Arrange the N elements of ARRAY in random order.
152 * Only effective if N is much smaller than RAND_MAX;
153 * if this may not be the case, use a better random
154 * number generator. -- Ben Pfaff.
156 static void shuffle(void *array, size_t n, size_t size)
158 char *carray = array;
165 aux = calloc(1, size);
167 err(EXIT_FAILURE, "calloc");
169 for (i = 1; i < n; ++i) {
170 size_t j = i + rand() / (RAND_MAX / (n - i) + 1);
173 memcpy(aux, &carray[j], size);
174 memcpy(&carray[j], &carray[i*size], size);
175 memcpy(&carray[i*size], aux, size);
182 static void *workerfn(void *arg)
185 struct worker *w = (struct worker *) arg;
186 unsigned long ops = w->ops;
187 struct epoll_event ev;
189 int to = nonblocking? 0 : -1;
190 int efd = multiq ? w->epollfd : epollfd;
192 pthread_mutex_lock(&thread_lock);
194 if (!threads_starting)
195 pthread_cond_signal(&thread_parent);
196 pthread_cond_wait(&thread_worker, &thread_lock);
197 pthread_mutex_unlock(&thread_lock);
201 * Block indefinitely waiting for the IN event.
202 * In order to stress the epoll_wait(2) syscall,
203 * call it event per event, instead of a larger
207 ret = epoll_wait(efd, &ev, 1, to);
208 } while (ret < 0 && errno == EINTR);
210 err(EXIT_FAILURE, "epoll_wait");
215 r = read(fd, &val, sizeof(val));
216 } while (!done && (r < 0 && errno == EAGAIN));
219 ev.events = EPOLLIN | EPOLLET;
220 ret = epoll_ctl(efd, EPOLL_CTL_ADD, fd, &ev);
224 /* rearm the file descriptor with a new event mask */
225 ev.events |= EPOLLIN | EPOLLONESHOT;
226 ret = epoll_ctl(efd, EPOLL_CTL_MOD, fd, &ev);
239 static void nest_epollfd(struct worker *w)
242 struct epoll_event ev;
243 int efd = multiq ? w->epollfd : epollfd;
245 if (nested > EPOLL_MAXNESTS)
246 nested = EPOLL_MAXNESTS;
248 epollfdp = calloc(nested, sizeof(*epollfdp));
250 err(EXIT_FAILURE, "calloc");
252 for (i = 0; i < nested; i++) {
253 epollfdp[i] = epoll_create(1);
255 err(EXIT_FAILURE, "epoll_create");
258 ev.events = EPOLLHUP; /* anything */
259 ev.data.u64 = i; /* any number */
261 for (i = nested - 1; i; i--) {
262 if (epoll_ctl(epollfdp[i - 1], EPOLL_CTL_ADD,
263 epollfdp[i], &ev) < 0)
264 err(EXIT_FAILURE, "epoll_ctl");
267 if (epoll_ctl(efd, EPOLL_CTL_ADD, *epollfdp, &ev) < 0)
268 err(EXIT_FAILURE, "epoll_ctl");
271 static void toggle_done(int sig __maybe_unused,
272 siginfo_t *info __maybe_unused,
273 void *uc __maybe_unused)
275 /* inform all threads that we're done for the day */
277 gettimeofday(&bench__end, NULL);
278 timersub(&bench__end, &bench__start, &bench__runtime);
281 static void print_summary(void)
283 unsigned long avg = avg_stats(&throughput_stats);
284 double stddev = stddev_stats(&throughput_stats);
286 printf("\nAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
287 avg, rel_stddev_stats(stddev, avg),
288 (int)bench__runtime.tv_sec);
291 static int do_threads(struct worker *worker, struct perf_cpu_map *cpu)
293 pthread_attr_t thread_attr, *attrp = NULL;
296 int ret = 0, events = EPOLLIN;
301 events |= EPOLLONESHOT;
305 printinfo("starting worker/consumer %sthreads%s\n",
306 noaffinity ? "":"CPU affinity ",
307 nonblocking ? " (nonblocking)":"");
309 pthread_attr_init(&thread_attr);
311 nrcpus = perf_cpu_map__nr(cpu);
312 cpuset = CPU_ALLOC(nrcpus);
314 size = CPU_ALLOC_SIZE(nrcpus);
316 for (i = 0; i < nthreads; i++) {
317 struct worker *w = &worker[i];
320 w->epollfd = epoll_create(1);
322 err(EXIT_FAILURE, "epoll_create");
329 w->fdmap = calloc(nfds, sizeof(int));
333 for (j = 0; j < nfds; j++) {
334 int efd = multiq ? w->epollfd : epollfd;
335 struct epoll_event ev;
337 w->fdmap[j] = eventfd(0, EFD_NONBLOCK);
339 err(EXIT_FAILURE, "eventfd");
341 ev.data.fd = w->fdmap[j];
344 ret = epoll_ctl(efd, EPOLL_CTL_ADD,
347 err(EXIT_FAILURE, "epoll_ctl");
351 CPU_ZERO_S(size, cpuset);
352 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu,
355 ret = pthread_attr_setaffinity_np(&thread_attr, size, cpuset);
358 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
361 attrp = &thread_attr;
364 ret = pthread_create(&w->thread, attrp, workerfn,
365 (void *)(struct worker *) w);
368 err(EXIT_FAILURE, "pthread_create");
374 pthread_attr_destroy(&thread_attr);
379 static void *writerfn(void *p)
381 struct worker *worker = p;
383 const uint64_t val = 1;
385 struct timespec ts = { .tv_sec = 0,
388 printinfo("starting writer-thread: doing %s writes ...\n",
389 randomize? "random":"lineal");
391 for (iter = 0; !wdone; iter++) {
393 shuffle((void *)worker, nthreads, sizeof(*worker));
396 for (i = 0; i < nthreads; i++) {
397 struct worker *w = &worker[i];
400 shuffle((void *)w->fdmap, nfds, sizeof(int));
403 for (j = 0; j < nfds; j++) {
405 sz = write(w->fdmap[j], &val, sizeof(val));
406 } while (!wdone && (sz < 0 && errno == EAGAIN));
410 nanosleep(&ts, NULL);
413 printinfo("exiting writer-thread (total full-loops: %zd)\n", iter);
417 static int cmpworker(const void *p1, const void *p2)
420 struct worker *w1 = (struct worker *) p1;
421 struct worker *w2 = (struct worker *) p2;
422 return w1->tid > w2->tid;
425 int bench_epoll_wait(int argc, const char **argv)
428 struct sigaction act;
430 struct worker *worker = NULL;
431 struct perf_cpu_map *cpu;
433 struct rlimit rl, prevrl;
435 argc = parse_options(argc, argv, options, bench_epoll_wait_usage, 0);
437 usage_with_options(bench_epoll_wait_usage, options);
441 memset(&act, 0, sizeof(act));
442 sigfillset(&act.sa_mask);
443 act.sa_sigaction = toggle_done;
444 sigaction(SIGINT, &act, NULL);
446 cpu = perf_cpu_map__new(NULL);
450 /* a single, main epoll instance */
452 epollfd = epoll_create(1);
454 err(EXIT_FAILURE, "epoll_create");
457 * Deal with nested epolls, if any.
463 printinfo("Using %s queue model\n", multiq ? "multi" : "single");
464 printinfo("Nesting level(s): %d\n", nested);
466 /* default to the number of CPUs and leave one for the writer pthread */
468 nthreads = perf_cpu_map__nr(cpu) - 1;
470 worker = calloc(nthreads, sizeof(*worker));
475 if (getrlimit(RLIMIT_NOFILE, &prevrl))
476 err(EXIT_FAILURE, "getrlimit");
477 rl.rlim_cur = rl.rlim_max = nfds * nthreads * 2 + 50;
478 printinfo("Setting RLIMIT_NOFILE rlimit from %" PRIu64 " to: %" PRIu64 "\n",
479 (uint64_t)prevrl.rlim_max, (uint64_t)rl.rlim_max);
480 if (setrlimit(RLIMIT_NOFILE, &rl) < 0)
481 err(EXIT_FAILURE, "setrlimit");
483 printf("Run summary [PID %d]: %d threads monitoring%s on "
484 "%d file-descriptors for %d secs.\n\n",
485 getpid(), nthreads, oneshot ? " (EPOLLONESHOT semantics)": "", nfds, nsecs);
487 init_stats(&throughput_stats);
488 pthread_mutex_init(&thread_lock, NULL);
489 pthread_cond_init(&thread_parent, NULL);
490 pthread_cond_init(&thread_worker, NULL);
492 threads_starting = nthreads;
494 gettimeofday(&bench__start, NULL);
496 do_threads(worker, cpu);
498 pthread_mutex_lock(&thread_lock);
499 while (threads_starting)
500 pthread_cond_wait(&thread_parent, &thread_lock);
501 pthread_cond_broadcast(&thread_worker);
502 pthread_mutex_unlock(&thread_lock);
505 * At this point the workers should be blocked waiting for read events
506 * to become ready. Launch the writer which will constantly be writing
507 * to each thread's fdmap.
509 ret = pthread_create(&wthread, NULL, writerfn,
510 (void *)(struct worker *) worker);
512 err(EXIT_FAILURE, "pthread_create");
515 toggle_done(0, NULL, NULL);
516 printinfo("main thread: toggling done\n");
520 ret = pthread_join(wthread, NULL);
522 err(EXIT_FAILURE, "pthread_join");
524 /* cleanup & report results */
525 pthread_cond_destroy(&thread_parent);
526 pthread_cond_destroy(&thread_worker);
527 pthread_mutex_destroy(&thread_lock);
529 /* sort the array back before reporting */
531 qsort(worker, nthreads, sizeof(struct worker), cmpworker);
533 for (i = 0; i < nthreads; i++) {
534 unsigned long t = bench__runtime.tv_sec > 0 ?
535 worker[i].ops / bench__runtime.tv_sec : 0;
537 update_stats(&throughput_stats, t);
540 printf("[thread %2d] fdmap: %p [ %04ld ops/sec ]\n",
541 worker[i].tid, &worker[i].fdmap[0], t);
543 printf("[thread %2d] fdmap: %p ... %p [ %04ld ops/sec ]\n",
544 worker[i].tid, &worker[i].fdmap[0],
545 &worker[i].fdmap[nfds-1], t);
553 err(EXIT_FAILURE, "calloc");
555 #endif // HAVE_EVENTFD_SUPPORT