tests/qht-bench.c

   1 /*
   2  * Copyright (C) 2016, Emilio G. Cota <cota@braap.org>
   3  *
   4  * License: GNU GPL, version 2 or later.
   5  *   See the COPYING file in the top-level directory.
   6  */
   7 #include "qemu/osdep.h"
   8 #include "qemu/processor.h"
   9 #include "qemu/atomic.h"
  10 #include "qemu/qht.h"
  11 #include "qemu/rcu.h"
  12 #include "exec/tb-hash-xx.h"
  13
  14 struct thread_stats {
  15     size_t rd;
  16     size_t not_rd;
  17     size_t in;
  18     size_t not_in;
  19     size_t rm;
  20     size_t not_rm;
  21     size_t rz;
  22     size_t not_rz;
  23 };
  24
  25 struct thread_info {
  26     void (*func)(struct thread_info *);
  27     struct thread_stats stats;
  28     uint64_t r;
  29     bool write_op; /* writes alternate between insertions and removals */
  30     bool resize_down;
  31 } QEMU_ALIGNED(64); /* avoid false sharing among threads */
  32
  33 static struct qht ht;
  34 static QemuThread *rw_threads;
  35
  36 #define DEFAULT_RANGE (4096)
  37 #define DEFAULT_QHT_N_ELEMS DEFAULT_RANGE
  38
  39 static unsigned int duration = 1;
  40 static unsigned int n_rw_threads = 1;
  41 static unsigned long lookup_range = DEFAULT_RANGE;
  42 static unsigned long update_range = DEFAULT_RANGE;
  43 static size_t init_range = DEFAULT_RANGE;
  44 static size_t init_size = DEFAULT_RANGE;
  45 static size_t n_ready_threads;
  46 static long populate_offset;
  47 static long *keys;
  48
  49 static size_t resize_min;
  50 static size_t resize_max;
  51 static struct thread_info *rz_info;
  52 static unsigned long resize_delay = 1000;
  53 static double resize_rate; /* 0.0 to 1.0 */
  54 static unsigned int n_rz_threads = 1;
  55 static QemuThread *rz_threads;
  56 static bool precompute_hash;
  57
  58 static double update_rate; /* 0.0 to 1.0 */
  59 static uint64_t update_threshold;
  60 static uint64_t resize_threshold;
  61
  62 static size_t qht_n_elems = DEFAULT_QHT_N_ELEMS;
  63 static int qht_mode;
  64
  65 static bool test_start;
  66 static bool test_stop;
  67
  68 static struct thread_info *rw_info;
  69
  70 static const char commands_string[] =
  71     " -d = duration, in seconds\n"
  72     " -n = number of threads\n"
  73     "\n"
  74     " -o = offset at which keys start\n"
  75     "\n"
  76     " -g = set -s,-k,-K,-l,-r to the same value\n"
  77     " -s = initial size hint\n"
  78     " -k = initial number of keys\n"
  79     " -K = initial range of keys (will be rounded up to pow2)\n"
  80     " -l = lookup range of keys (will be rounded up to pow2)\n"
  81     " -r = update range of keys (will be rounded up to pow2)\n"
  82     "\n"
  83     " -u = update rate (0.0 to 100.0), 50/50 split of insertions/removals\n"
  84     "\n"
  85     " -R = enable auto-resize\n"
  86     " -S = resize rate (0.0 to 100.0)\n"
  87     " -D = delay (in us) between potential resizes\n"
  88     " -N = number of resize threads";
  89
  90 static void usage_complete(int argc, char *argv[])
  91 {
  92     fprintf(stderr, "Usage: %s [options]\n", argv[0]);
  93     fprintf(stderr, "options:\n%s\n", commands_string);
  94     exit(-1);
  95 }
  96
  97 static bool is_equal(const void *ap, const void *bp)
  98 {
  99     const long *a = ap;
 100     const long *b = bp;
 101
 102     return *a == *b;
 103 }
 104
 105 static uint32_t h(unsigned long v)
 106 {
 107     return tb_hash_func7(v, 0, 0, 0, 0);
 108 }
 109
 110 static uint32_t hval(unsigned long v)
 111 {
 112     return v;
 113 }
 114
 115 static uint32_t (*hfunc)(unsigned long v) = h;
 116
 117 /*
 118  * From: https://en.wikipedia.org/wiki/Xorshift
 119  * This is faster than rand_r(), and gives us a wider range (RAND_MAX is only
 120  * guaranteed to be >= INT_MAX).
 121  */
 122 static uint64_t xorshift64star(uint64_t x)
 123 {
 124     x ^= x >> 12; /* a */
 125     x ^= x << 25; /* b */
 126     x ^= x >> 27; /* c */
 127     return x * UINT64_C(2685821657736338717);
 128 }
 129
 130 static void do_rz(struct thread_info *info)
 131 {
 132     struct thread_stats *stats = &info->stats;
 133
 134     if (info->r < resize_threshold) {
 135         size_t size = info->resize_down ? resize_min : resize_max;
 136         bool resized;
 137
 138         resized = qht_resize(&ht, size);
 139         info->resize_down = !info->resize_down;
 140
 141         if (resized) {
 142             stats->rz++;
 143         } else {
 144             stats->not_rz++;
 145         }
 146     }
 147     g_usleep(resize_delay);
 148 }
 149
 150 static void do_rw(struct thread_info *info)
 151 {
 152     struct thread_stats *stats = &info->stats;
 153     uint32_t hash;
 154     long *p;
 155
 156     if (info->r >= update_threshold) {
 157         bool read;
 158
 159         p = &keys[info->r & (lookup_range - 1)];
 160         hash = hfunc(*p);
 161         read = qht_lookup(&ht, p, hash);
 162         if (read) {
 163             stats->rd++;
 164         } else {
 165             stats->not_rd++;
 166         }
 167     } else {
 168         p = &keys[info->r & (update_range - 1)];
 169         hash = hfunc(*p);
 170         if (info->write_op) {
 171             bool written = false;
 172
 173             if (qht_lookup(&ht, p, hash) == NULL) {
 174                 written = qht_insert(&ht, p, hash, NULL);
 175             }
 176             if (written) {
 177                 stats->in++;
 178             } else {
 179                 stats->not_in++;
 180             }
 181         } else {
 182             bool removed = false;
 183
 184             if (qht_lookup(&ht, p, hash)) {
 185                 removed = qht_remove(&ht, p, hash);
 186             }
 187             if (removed) {
 188                 stats->rm++;
 189             } else {
 190                 stats->not_rm++;
 191             }
 192         }
 193         info->write_op = !info->write_op;
 194     }
 195 }
 196
 197 static void *thread_func(void *p)
 198 {
 199     struct thread_info *info = p;
 200
 201     rcu_register_thread();
 202
 203     atomic_inc(&n_ready_threads);
 204     while (!atomic_read(&test_start)) {
 205         cpu_relax();
 206     }
 207
 208     rcu_read_lock();
 209     while (!atomic_read(&test_stop)) {
 210         info->r = xorshift64star(info->r);
 211         info->func(info);
 212     }
 213     rcu_read_unlock();
 214
 215     rcu_unregister_thread();
 216     return NULL;
 217 }
 218
 219 /* sets everything except info->func */
 220 static void prepare_thread_info(struct thread_info *info, int i)
 221 {
 222     /* seed for the RNG; each thread should have a different one */
 223     info->r = (i + 1) ^ time(NULL);
 224     /* the first update will be a write */
 225     info->write_op = true;
 226     /* the first resize will be down */
 227     info->resize_down = true;
 228
 229     memset(&info->stats, 0, sizeof(info->stats));
 230 }
 231
 232 static void
 233 th_create_n(QemuThread **threads, struct thread_info **infos, const char *name,
 234             void (*func)(struct thread_info *), int offset, int n)
 235 {
 236     struct thread_info *info;
 237     QemuThread *th;
 238     int i;
 239
 240     th = g_malloc(sizeof(*th) * n);
 241     *threads = th;
 242
 243     info = qemu_memalign(64, sizeof(*info) * n);
 244     *infos = info;
 245
 246     for (i = 0; i < n; i++) {
 247         prepare_thread_info(&info[i], offset + i);
 248         info[i].func = func;
 249         qemu_thread_create(&th[i], name, thread_func, &info[i],
 250                            QEMU_THREAD_JOINABLE);
 251     }
 252 }
 253
 254 static void create_threads(void)
 255 {
 256     th_create_n(&rw_threads, &rw_info, "rw", do_rw, 0, n_rw_threads);
 257     th_create_n(&rz_threads, &rz_info, "rz", do_rz, n_rw_threads, n_rz_threads);
 258 }
 259
 260 static void pr_params(void)
 261 {
 262     printf("Parameters:\n");
 263     printf(" duration:          %d s\n", duration);
 264     printf(" # of threads:      %u\n", n_rw_threads);
 265     printf(" initial # of keys: %zu\n", init_size);
 266     printf(" initial size hint: %zu\n", qht_n_elems);
 267     printf(" auto-resize:       %s\n",
 268            qht_mode & QHT_MODE_AUTO_RESIZE ? "on" : "off");
 269     if (resize_rate) {
 270         printf(" resize_rate:       %f%%\n", resize_rate * 100.0);
 271         printf(" resize range:      %zu-%zu\n", resize_min, resize_max);
 272         printf(" # resize threads   %u\n", n_rz_threads);
 273     }
 274     printf(" update rate:       %f%%\n", update_rate * 100.0);
 275     printf(" offset:            %ld\n", populate_offset);
 276     printf(" initial key range: %zu\n", init_range);
 277     printf(" lookup range:      %lu\n", lookup_range);
 278     printf(" update range:      %lu\n", update_range);
 279 }
 280
 281 static void do_threshold(double rate, uint64_t *threshold)
 282 {
 283     if (rate == 1.0) {
 284         *threshold = UINT64_MAX;
 285     } else {
 286         *threshold = rate * UINT64_MAX;
 287     }
 288 }
 289
 290 static void htable_init(void)
 291 {
 292     unsigned long n = MAX(init_range, update_range);
 293     uint64_t r = time(NULL);
 294     size_t retries = 0;
 295     size_t i;
 296
 297     /* avoid allocating memory later by allocating all the keys now */
 298     keys = g_malloc(sizeof(*keys) * n);
 299     for (i = 0; i < n; i++) {
 300         long val = populate_offset + i;
 301
 302         keys[i] = precompute_hash ? h(val) : hval(val);
 303     }
 304
 305     /* some sanity checks */
 306     g_assert_cmpuint(lookup_range, <=, n);
 307
 308     /* compute thresholds */
 309     do_threshold(update_rate, &update_threshold);
 310     do_threshold(resize_rate, &resize_threshold);
 311
 312     if (resize_rate) {
 313         resize_min = n / 2;
 314         resize_max = n;
 315         assert(resize_min < resize_max);
 316     } else {
 317         n_rz_threads = 0;
 318     }
 319
 320     /* initialize the hash table */
 321     qht_init(&ht, is_equal, qht_n_elems, qht_mode);
 322     assert(init_size <= init_range);
 323
 324     pr_params();
 325
 326     fprintf(stderr, "Initialization: populating %zu items...", init_size);
 327     for (i = 0; i < init_size; i++) {
 328         for (;;) {
 329             uint32_t hash;
 330             long *p;
 331
 332             r = xorshift64star(r);
 333             p = &keys[r & (init_range - 1)];
 334             hash = hfunc(*p);
 335             if (qht_insert(&ht, p, hash, NULL)) {
 336                 break;
 337             }
 338             retries++;
 339         }
 340     }
 341     fprintf(stderr, " populated after %zu retries\n", retries);
 342 }
 343
 344 static void add_stats(struct thread_stats *s, struct thread_info *info, int n)
 345 {
 346     int i;
 347
 348     for (i = 0; i < n; i++) {
 349         struct thread_stats *stats = &info[i].stats;
 350
 351         s->rd += stats->rd;
 352         s->not_rd += stats->not_rd;
 353
 354         s->in += stats->in;
 355         s->not_in += stats->not_in;
 356
 357         s->rm += stats->rm;
 358         s->not_rm += stats->not_rm;
 359
 360         s->rz += stats->rz;
 361         s->not_rz += stats->not_rz;
 362     }
 363 }
 364
 365 static void pr_stats(void)
 366 {
 367     struct thread_stats s = {};
 368     double tx;
 369
 370     add_stats(&s, rw_info, n_rw_threads);
 371     add_stats(&s, rz_info, n_rz_threads);
 372
 373     printf("Results:\n");
 374
 375     if (resize_rate) {
 376         printf(" Resizes:           %zu (%.2f%% of %zu)\n",
 377                s.rz, (double)s.rz / (s.rz + s.not_rz) * 100, s.rz + s.not_rz);
 378     }
 379
 380     printf(" Read:              %.2f M (%.2f%% of %.2fM)\n",
 381            (double)s.rd / 1e6,
 382            (double)s.rd / (s.rd + s.not_rd) * 100,
 383            (double)(s.rd + s.not_rd) / 1e6);
 384     printf(" Inserted:          %.2f M (%.2f%% of %.2fM)\n",
 385            (double)s.in / 1e6,
 386            (double)s.in / (s.in + s.not_in) * 100,
 387            (double)(s.in + s.not_in) / 1e6);
 388     printf(" Removed:           %.2f M (%.2f%% of %.2fM)\n",
 389            (double)s.rm / 1e6,
 390            (double)s.rm / (s.rm + s.not_rm) * 100,
 391            (double)(s.rm + s.not_rm) / 1e6);
 392
 393     tx = (s.rd + s.not_rd + s.in + s.not_in + s.rm + s.not_rm) / 1e6 / duration;
 394     printf(" Throughput:        %.2f MT/s\n", tx);
 395     printf(" Throughput/thread: %.2f MT/s/thread\n", tx / n_rw_threads);
 396 }
 397
 398 static void run_test(void)
 399 {
 400     unsigned int remaining;
 401     int i;
 402
 403     while (atomic_read(&n_ready_threads) != n_rw_threads + n_rz_threads) {
 404         cpu_relax();
 405     }
 406     atomic_set(&test_start, true);
 407     do {
 408         remaining = sleep(duration);
 409     } while (remaining);
 410     atomic_set(&test_stop, true);
 411
 412     for (i = 0; i < n_rw_threads; i++) {
 413         qemu_thread_join(&rw_threads[i]);
 414     }
 415     for (i = 0; i < n_rz_threads; i++) {
 416         qemu_thread_join(&rz_threads[i]);
 417     }
 418 }
 419
 420 static void parse_args(int argc, char *argv[])
 421 {
 422     int c;
 423
 424     for (;;) {
 425         c = getopt(argc, argv, "d:D:g:k:K:l:hn:N:o:pr:Rs:S:u:");
 426         if (c < 0) {
 427             break;
 428         }
 429         switch (c) {
 430         case 'd':
 431             duration = atoi(optarg);
 432             break;
 433         case 'D':
 434             resize_delay = atol(optarg);
 435             break;
 436         case 'g':
 437             init_range = pow2ceil(atol(optarg));
 438             lookup_range = pow2ceil(atol(optarg));
 439             update_range = pow2ceil(atol(optarg));
 440             qht_n_elems = atol(optarg);
 441             init_size = atol(optarg);
 442             break;
 443         case 'h':
 444             usage_complete(argc, argv);
 445             exit(0);
 446         case 'k':
 447             init_size = atol(optarg);
 448             break;
 449         case 'K':
 450             init_range = pow2ceil(atol(optarg));
 451             break;
 452         case 'l':
 453             lookup_range = pow2ceil(atol(optarg));
 454             break;
 455         case 'n':
 456             n_rw_threads = atoi(optarg);
 457             break;
 458         case 'N':
 459             n_rz_threads = atoi(optarg);
 460             break;
 461         case 'o':
 462             populate_offset = atol(optarg);
 463             break;
 464         case 'p':
 465             precompute_hash = true;
 466             hfunc = hval;
 467             break;
 468         case 'r':
 469             update_range = pow2ceil(atol(optarg));
 470             break;
 471         case 'R':
 472             qht_mode |= QHT_MODE_AUTO_RESIZE;
 473             break;
 474         case 's':
 475             qht_n_elems = atol(optarg);
 476             break;
 477         case 'S':
 478             resize_rate = atof(optarg) / 100.0;
 479             if (resize_rate > 1.0) {
 480                 resize_rate = 1.0;
 481             }
 482             break;
 483         case 'u':
 484             update_rate = atof(optarg) / 100.0;
 485             if (update_rate > 1.0) {
 486                 update_rate = 1.0;
 487             }
 488             break;
 489         }
 490     }
 491 }
 492
 493 int main(int argc, char *argv[])
 494 {
 495     parse_args(argc, argv);
 496     htable_init();
 497     create_threads();
 498     run_test();
 499     pr_stats();
 500     return 0;
 501 }