app/test/test_ring_perf.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2013 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  *
  33  */
  34
  35
  36 #include <stdio.h>
  37 #include <inttypes.h>
  38 #include <rte_ring.h>
  39 #include <rte_cycles.h>
  40 #include <rte_launch.h>
  41
  42 #include <cmdline_parse.h>
  43
  44 #include "test.h"
  45
  46 /*
  47  * Ring
  48  * ====
  49  *
  50  * Measures performance of various operations using rdtsc
  51  *  * Empty ring dequeue
  52  *  * Enqueue/dequeue of bursts in 1 threads
  53  *  * Enqueue/dequeue of bursts in 2 threads
  54  */
  55
  56 #define RING_NAME "RING_PERF"
  57 #define RING_SIZE 4096
  58 #define MAX_BURST 32
  59
  60 /*
  61  * the sizes to enqueue and dequeue in testing
  62  * (marked volatile so they won't be seen as compile-time constants)
  63  */
  64 static const volatile unsigned bulk_sizes[] = { 8, 32 };
  65
  66 /* The ring structure used for tests */
  67 static struct rte_ring *r;
  68
  69 struct lcore_pair {
  70         unsigned c1, c2;
  71 };
  72
  73 static volatile unsigned lcore_count = 0;
  74
  75 /**** Functions to analyse our core mask to get cores for different tests ***/
  76
  77 static int
  78 get_two_hyperthreads(struct lcore_pair *lcp)
  79 {
  80         unsigned id1, id2;
  81         unsigned c1, c2, s1, s2;
  82         RTE_LCORE_FOREACH(id1) {
  83                 /* inner loop just re-reads all id's. We could skip the first few
  84                  * elements, but since number of cores is small there is little point
  85                  */
  86                 RTE_LCORE_FOREACH(id2) {
  87                         if (id1 == id2)
  88                                 continue;
  89                         c1 = lcore_config[id1].core_id;
  90                         c2 = lcore_config[id2].core_id;
  91                         s1 = lcore_config[id1].socket_id;
  92                         s2 = lcore_config[id2].socket_id;
  93                         if ((c1 == c2) && (s1 == s2)){
  94                                 lcp->c1 = id1;
  95                                 lcp->c2 = id2;
  96                                 return 0;
  97                         }
  98                 }
  99         }
 100         return 1;
 101 }
 102
 103 static int
 104 get_two_cores(struct lcore_pair *lcp)
 105 {
 106         unsigned id1, id2;
 107         unsigned c1, c2, s1, s2;
 108         RTE_LCORE_FOREACH(id1) {
 109                 RTE_LCORE_FOREACH(id2) {
 110                         if (id1 == id2)
 111                                 continue;
 112                         c1 = lcore_config[id1].core_id;
 113                         c2 = lcore_config[id2].core_id;
 114                         s1 = lcore_config[id1].socket_id;
 115                         s2 = lcore_config[id2].socket_id;
 116                         if ((c1 != c2) && (s1 == s2)){
 117                                 lcp->c1 = id1;
 118                                 lcp->c2 = id2;
 119                                 return 0;
 120                         }
 121                 }
 122         }
 123         return 1;
 124 }
 125
 126 static int
 127 get_two_sockets(struct lcore_pair *lcp)
 128 {
 129         unsigned id1, id2;
 130         unsigned s1, s2;
 131         RTE_LCORE_FOREACH(id1) {
 132                 RTE_LCORE_FOREACH(id2) {
 133                         if (id1 == id2)
 134                                 continue;
 135                         s1 = lcore_config[id1].socket_id;
 136                         s2 = lcore_config[id2].socket_id;
 137                         if (s1 != s2){
 138                                 lcp->c1 = id1;
 139                                 lcp->c2 = id2;
 140                                 return 0;
 141                         }
 142                 }
 143         }
 144         return 1;
 145 }
 146
 147 /* Get cycle counts for dequeuing from an empty ring. Should be 2 or 3 cycles */
 148 static void
 149 test_empty_dequeue(void)
 150 {
 151         const unsigned iter_shift = 26;
 152         const unsigned iterations = 1<<iter_shift;
 153         unsigned i = 0;
 154         void *burst[MAX_BURST];
 155
 156         const uint64_t sc_start = rte_rdtsc();
 157         for (i = 0; i < iterations; i++)
 158                 rte_ring_sc_dequeue_bulk(r, burst, bulk_sizes[0]);
 159         const uint64_t sc_end = rte_rdtsc();
 160
 161         const uint64_t mc_start = rte_rdtsc();
 162         for (i = 0; i < iterations; i++)
 163                 rte_ring_mc_dequeue_bulk(r, burst, bulk_sizes[0]);
 164         const uint64_t mc_end = rte_rdtsc();
 165
 166         printf("SC empty dequeue: %.2F\n",
 167                         (double)(sc_end-sc_start) / iterations);
 168         printf("MC empty dequeue: %.2F\n",
 169                         (double)(mc_end-mc_start) / iterations);
 170 }
 171
 172 /*
 173  * for the separate enqueue and dequeue threads they take in one param
 174  * and return two. Input = burst size, output = cycle average for sp/sc & mp/mc
 175  */
 176 struct thread_params {
 177         unsigned size;        /* input value, the burst size */
 178         double spsc, mpmc;    /* output value, the single or multi timings */
 179 };
 180
 181 /*
 182  * Function that uses rdtsc to measure timing for ring enqueue. Needs pair
 183  * thread running dequeue_bulk function
 184  */
 185 static int
 186 enqueue_bulk(void *p)
 187 {
 188         const unsigned iter_shift = 23;
 189         const unsigned iterations = 1<<iter_shift;
 190         struct thread_params *params = p;
 191         const unsigned size = params->size;
 192         unsigned i;
 193         void *burst[MAX_BURST] = {0};
 194
 195         if ( __sync_add_and_fetch(&lcore_count, 1) != 2 )
 196                 while(lcore_count != 2)
 197                         rte_pause();
 198
 199         const uint64_t sp_start = rte_rdtsc();
 200         for (i = 0; i < iterations; i++)
 201                 while (rte_ring_sp_enqueue_bulk(r, burst, size) != 0)
 202                         rte_pause();
 203         const uint64_t sp_end = rte_rdtsc();
 204
 205         const uint64_t mp_start = rte_rdtsc();
 206         for (i = 0; i < iterations; i++)
 207                 while (rte_ring_mp_enqueue_bulk(r, burst, size) != 0)
 208                         rte_pause();
 209         const uint64_t mp_end = rte_rdtsc();
 210
 211         params->spsc = ((double)(sp_end - sp_start))/(iterations*size);
 212         params->mpmc = ((double)(mp_end - mp_start))/(iterations*size);
 213         return 0;
 214 }
 215
 216 /*
 217  * Function that uses rdtsc to measure timing for ring dequeue. Needs pair
 218  * thread running enqueue_bulk function
 219  */
 220 static int
 221 dequeue_bulk(void *p)
 222 {
 223         const unsigned iter_shift = 23;
 224         const unsigned iterations = 1<<iter_shift;
 225         struct thread_params *params = p;
 226         const unsigned size = params->size;
 227         unsigned i;
 228         void *burst[MAX_BURST] = {0};
 229
 230         if ( __sync_add_and_fetch(&lcore_count, 1) != 2 )
 231                 while(lcore_count != 2)
 232                         rte_pause();
 233
 234         const uint64_t sc_start = rte_rdtsc();
 235         for (i = 0; i < iterations; i++)
 236                 while (rte_ring_sc_dequeue_bulk(r, burst, size) != 0)
 237                         rte_pause();
 238         const uint64_t sc_end = rte_rdtsc();
 239
 240         const uint64_t mc_start = rte_rdtsc();
 241         for (i = 0; i < iterations; i++)
 242                 while (rte_ring_mc_dequeue_bulk(r, burst, size) != 0)
 243                         rte_pause();
 244         const uint64_t mc_end = rte_rdtsc();
 245
 246         params->spsc = ((double)(sc_end - sc_start))/(iterations*size);
 247         params->mpmc = ((double)(mc_end - mc_start))/(iterations*size);
 248         return 0;
 249 }
 250
 251 /*
 252  * Function that calls the enqueue and dequeue bulk functions on pairs of cores.
 253  * used to measure ring perf between hyperthreads, cores and sockets.
 254  */
 255 static void
 256 run_on_core_pair(struct lcore_pair *cores,
 257                 lcore_function_t f1, lcore_function_t f2)
 258 {
 259         struct thread_params param1 = {.size = 0}, param2 = {.size = 0};
 260         unsigned i;
 261         for (i = 0; i < sizeof(bulk_sizes)/sizeof(bulk_sizes[0]); i++) {
 262                 lcore_count = 0;
 263                 param1.size = param2.size = bulk_sizes[i];
 264                 if (cores->c1 == rte_get_master_lcore()) {
 265                         rte_eal_remote_launch(f2, &param2, cores->c2);
 266                         f1(&param1);
 267                         rte_eal_wait_lcore(cores->c2);
 268                 } else {
 269                         rte_eal_remote_launch(f1, &param1, cores->c1);
 270                         rte_eal_remote_launch(f2, &param2, cores->c2);
 271                         rte_eal_wait_lcore(cores->c1);
 272                         rte_eal_wait_lcore(cores->c2);
 273                 }
 274                 printf("SP/SC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[i],
 275                                 param1.spsc + param2.spsc);
 276                 printf("MP/MC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[i],
 277                                 param1.mpmc + param2.mpmc);
 278         }
 279 }
 280
 281 /*
 282  * Test function that determines how long an enqueue + dequeue of a single item
 283  * takes on a single lcore. Result is for comparison with the bulk enq+deq.
 284  */
 285 static void
 286 test_single_enqueue_dequeue(void)
 287 {
 288         const unsigned iter_shift = 24;
 289         const unsigned iterations = 1<<iter_shift;
 290         unsigned i = 0;
 291         void *burst = NULL;
 292
 293         const uint64_t sc_start = rte_rdtsc();
 294         for (i = 0; i < iterations; i++) {
 295                 rte_ring_sp_enqueue(r, burst);
 296                 rte_ring_sc_dequeue(r, &burst);
 297         }
 298         const uint64_t sc_end = rte_rdtsc();
 299
 300         const uint64_t mc_start = rte_rdtsc();
 301         for (i = 0; i < iterations; i++) {
 302                 rte_ring_mp_enqueue(r, burst);
 303                 rte_ring_mc_dequeue(r, &burst);
 304         }
 305         const uint64_t mc_end = rte_rdtsc();
 306
 307         printf("SP/SC single enq/dequeue: %"PRIu64"\n",
 308                         (sc_end-sc_start) >> iter_shift);
 309         printf("MP/MC single enq/dequeue: %"PRIu64"\n",
 310                         (mc_end-mc_start) >> iter_shift);
 311 }
 312
 313 /*
 314  * Test that does both enqueue and dequeue on a core using the burst() API calls
 315  * instead of the bulk() calls used in other tests. Results should be the same
 316  * as for the bulk function called on a single lcore.
 317  */
 318 static void
 319 test_burst_enqueue_dequeue(void)
 320 {
 321         const unsigned iter_shift = 23;
 322         const unsigned iterations = 1<<iter_shift;
 323         unsigned sz, i = 0;
 324         void *burst[MAX_BURST] = {0};
 325
 326         for (sz = 0; sz < sizeof(bulk_sizes)/sizeof(bulk_sizes[0]); sz++) {
 327                 const uint64_t sc_start = rte_rdtsc();
 328                 for (i = 0; i < iterations; i++) {
 329                         rte_ring_sp_enqueue_burst(r, burst, bulk_sizes[sz]);
 330                         rte_ring_sc_dequeue_burst(r, burst, bulk_sizes[sz]);
 331                 }
 332                 const uint64_t sc_end = rte_rdtsc();
 333
 334                 const uint64_t mc_start = rte_rdtsc();
 335                 for (i = 0; i < iterations; i++) {
 336                         rte_ring_mp_enqueue_burst(r, burst, bulk_sizes[sz]);
 337                         rte_ring_mc_dequeue_burst(r, burst, bulk_sizes[sz]);
 338                 }
 339                 const uint64_t mc_end = rte_rdtsc();
 340
 341                 uint64_t mc_avg = ((mc_end-mc_start) >> iter_shift) / bulk_sizes[sz];
 342                 uint64_t sc_avg = ((sc_end-sc_start) >> iter_shift) / bulk_sizes[sz];
 343
 344                 printf("SP/SC burst enq/dequeue (size: %u): %"PRIu64"\n", bulk_sizes[sz],
 345                                 sc_avg);
 346                 printf("MP/MC burst enq/dequeue (size: %u): %"PRIu64"\n", bulk_sizes[sz],
 347                                 mc_avg);
 348         }
 349 }
 350
 351 /* Times enqueue and dequeue on a single lcore */
 352 static void
 353 test_bulk_enqueue_dequeue(void)
 354 {
 355         const unsigned iter_shift = 23;
 356         const unsigned iterations = 1<<iter_shift;
 357         unsigned sz, i = 0;
 358         void *burst[MAX_BURST] = {0};
 359
 360         for (sz = 0; sz < sizeof(bulk_sizes)/sizeof(bulk_sizes[0]); sz++) {
 361                 const uint64_t sc_start = rte_rdtsc();
 362                 for (i = 0; i < iterations; i++) {
 363                         rte_ring_sp_enqueue_bulk(r, burst, bulk_sizes[sz]);
 364                         rte_ring_sc_dequeue_bulk(r, burst, bulk_sizes[sz]);
 365                 }
 366                 const uint64_t sc_end = rte_rdtsc();
 367
 368                 const uint64_t mc_start = rte_rdtsc();
 369                 for (i = 0; i < iterations; i++) {
 370                         rte_ring_mp_enqueue_bulk(r, burst, bulk_sizes[sz]);
 371                         rte_ring_mc_dequeue_bulk(r, burst, bulk_sizes[sz]);
 372                 }
 373                 const uint64_t mc_end = rte_rdtsc();
 374
 375                 double sc_avg = ((double)(sc_end-sc_start) /
 376                                 (iterations * bulk_sizes[sz]));
 377                 double mc_avg = ((double)(mc_end-mc_start) /
 378                                 (iterations * bulk_sizes[sz]));
 379
 380                 printf("SP/SC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[sz],
 381                                 sc_avg);
 382                 printf("MP/MC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[sz],
 383                                 mc_avg);
 384         }
 385 }
 386
 387 int
 388 test_ring_perf(void)
 389 {
 390         struct lcore_pair cores;
 391         r = rte_ring_create(RING_NAME, RING_SIZE, rte_socket_id(), 0);
 392         if (r == NULL && (r = rte_ring_lookup(RING_NAME)) == NULL)
 393                 return -1;
 394
 395         printf("### Testing single element and burst enq/deq ###\n");
 396         test_single_enqueue_dequeue();
 397         test_burst_enqueue_dequeue();
 398
 399         printf("\n### Testing empty dequeue ###\n");
 400         test_empty_dequeue();
 401
 402         printf("\n### Testing using a single lcore ###\n");
 403         test_bulk_enqueue_dequeue();
 404
 405         if (get_two_hyperthreads(&cores) == 0) {
 406                 printf("\n### Testing using two hyperthreads ###\n");
 407                 run_on_core_pair(&cores, enqueue_bulk, dequeue_bulk);
 408         }
 409         if (get_two_cores(&cores) == 0) {
 410                 printf("\n### Testing using two physical cores ###\n");
 411                 run_on_core_pair(&cores, enqueue_bulk, dequeue_bulk);
 412         }
 413         if (get_two_sockets(&cores) == 0) {
 414                 printf("\n### Testing using two NUMA nodes ###\n");
 415                 run_on_core_pair(&cores, enqueue_bulk, dequeue_bulk);
 416         }
 417         return 0;
 418 }