app/test/test_lpm_perf.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 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 #include <stdio.h>
  35 #include <stdint.h>
  36 #include <stdlib.h>
  37
  38 #include <rte_cycles.h>
  39 #include <rte_random.h>
  40 #include <rte_branch_prediction.h>
  41 #include <rte_lpm.h>
  42
  43 #include "test.h"
  44 #include "test_lpm_routes.h"
  45 #include "test_xmmt_ops.h"
  46
  47 #define TEST_LPM_ASSERT(cond) do {                                            \
  48         if (!(cond)) {                                                        \
  49                 printf("Error at line %d: \n", __LINE__);                     \
  50                 return -1;                                                    \
  51         }                                                                     \
  52 } while(0)
  53
  54 #define ITERATIONS (1 << 10)
  55 #define BATCH_SIZE (1 << 12)
  56 #define BULK_SIZE 32
  57
  58 static void
  59 print_route_distribution(const struct route_rule *table, uint32_t n)
  60 {
  61         unsigned i, j;
  62
  63         printf("Route distribution per prefix width: \n");
  64         printf("DEPTH    QUANTITY (PERCENT)\n");
  65         printf("--------------------------- \n");
  66
  67         /* Count depths. */
  68         for (i = 1; i <= 32; i++) {
  69                 unsigned depth_counter = 0;
  70                 double percent_hits;
  71
  72                 for (j = 0; j < n; j++)
  73                         if (table[j].depth == (uint8_t) i)
  74                                 depth_counter++;
  75
  76                 percent_hits = ((double)depth_counter)/((double)n) * 100;
  77                 printf("%.2u%15u (%.2f)\n", i, depth_counter, percent_hits);
  78         }
  79         printf("\n");
  80 }
  81
  82 static int
  83 test_lpm_perf(void)
  84 {
  85         struct rte_lpm *lpm = NULL;
  86         struct rte_lpm_config config;
  87
  88         config.max_rules = 1000000;
  89         config.number_tbl8s = 256;
  90         config.flags = 0;
  91         uint64_t begin, total_time, lpm_used_entries = 0;
  92         unsigned i, j;
  93         uint32_t next_hop_add = 0xAA, next_hop_return = 0;
  94         int status = 0;
  95         uint64_t cache_line_counter = 0;
  96         int64_t count = 0;
  97
  98         rte_srand(rte_rdtsc());
  99
 100         printf("No. routes = %u\n", (unsigned) NUM_ROUTE_ENTRIES);
 101
 102         print_route_distribution(large_route_table, (uint32_t) NUM_ROUTE_ENTRIES);
 103
 104         lpm = rte_lpm_create(__func__, SOCKET_ID_ANY, &config);
 105         TEST_LPM_ASSERT(lpm != NULL);
 106
 107         /* Measue add. */
 108         begin = rte_rdtsc();
 109
 110         for (i = 0; i < NUM_ROUTE_ENTRIES; i++) {
 111                 if (rte_lpm_add(lpm, large_route_table[i].ip,
 112                                 large_route_table[i].depth, next_hop_add) == 0)
 113                         status++;
 114         }
 115         /* End Timer. */
 116         total_time = rte_rdtsc() - begin;
 117
 118         printf("Unique added entries = %d\n", status);
 119         /* Obtain add statistics. */
 120         for (i = 0; i < RTE_LPM_TBL24_NUM_ENTRIES; i++) {
 121                 if (lpm->tbl24[i].valid)
 122                         lpm_used_entries++;
 123
 124                 if (i % 32 == 0) {
 125                         if ((uint64_t)count < lpm_used_entries) {
 126                                 cache_line_counter++;
 127                                 count = lpm_used_entries;
 128                         }
 129                 }
 130         }
 131
 132         printf("Used table 24 entries = %u (%g%%)\n",
 133                         (unsigned) lpm_used_entries,
 134                         (lpm_used_entries * 100.0) / RTE_LPM_TBL24_NUM_ENTRIES);
 135         printf("64 byte Cache entries used = %u (%u bytes)\n",
 136                         (unsigned) cache_line_counter, (unsigned) cache_line_counter * 64);
 137
 138         printf("Average LPM Add: %g cycles\n",
 139                         (double)total_time / NUM_ROUTE_ENTRIES);
 140
 141         /* Measure single Lookup */
 142         total_time = 0;
 143         count = 0;
 144
 145         for (i = 0; i < ITERATIONS; i++) {
 146                 static uint32_t ip_batch[BATCH_SIZE];
 147
 148                 for (j = 0; j < BATCH_SIZE; j++)
 149                         ip_batch[j] = rte_rand();
 150
 151                 /* Lookup per batch */
 152                 begin = rte_rdtsc();
 153
 154                 for (j = 0; j < BATCH_SIZE; j++) {
 155                         if (rte_lpm_lookup(lpm, ip_batch[j], &next_hop_return) != 0)
 156                                 count++;
 157                 }
 158
 159                 total_time += rte_rdtsc() - begin;
 160
 161         }
 162         printf("Average LPM Lookup: %.1f cycles (fails = %.1f%%)\n",
 163                         (double)total_time / ((double)ITERATIONS * BATCH_SIZE),
 164                         (count * 100.0) / (double)(ITERATIONS * BATCH_SIZE));
 165
 166         /* Measure bulk Lookup */
 167         total_time = 0;
 168         count = 0;
 169         for (i = 0; i < ITERATIONS; i++) {
 170                 static uint32_t ip_batch[BATCH_SIZE];
 171                 uint32_t next_hops[BULK_SIZE];
 172
 173                 /* Create array of random IP addresses */
 174                 for (j = 0; j < BATCH_SIZE; j++)
 175                         ip_batch[j] = rte_rand();
 176
 177                 /* Lookup per batch */
 178                 begin = rte_rdtsc();
 179                 for (j = 0; j < BATCH_SIZE; j += BULK_SIZE) {
 180                         unsigned k;
 181                         rte_lpm_lookup_bulk(lpm, &ip_batch[j], next_hops, BULK_SIZE);
 182                         for (k = 0; k < BULK_SIZE; k++)
 183                                 if (unlikely(!(next_hops[k] & RTE_LPM_LOOKUP_SUCCESS)))
 184                                         count++;
 185                 }
 186
 187                 total_time += rte_rdtsc() - begin;
 188         }
 189         printf("BULK LPM Lookup: %.1f cycles (fails = %.1f%%)\n",
 190                         (double)total_time / ((double)ITERATIONS * BATCH_SIZE),
 191                         (count * 100.0) / (double)(ITERATIONS * BATCH_SIZE));
 192
 193         /* Measure LookupX4 */
 194         total_time = 0;
 195         count = 0;
 196         for (i = 0; i < ITERATIONS; i++) {
 197                 static uint32_t ip_batch[BATCH_SIZE];
 198                 uint32_t next_hops[4];
 199
 200                 /* Create array of random IP addresses */
 201                 for (j = 0; j < BATCH_SIZE; j++)
 202                         ip_batch[j] = rte_rand();
 203
 204                 /* Lookup per batch */
 205                 begin = rte_rdtsc();
 206                 for (j = 0; j < BATCH_SIZE; j += RTE_DIM(next_hops)) {
 207                         unsigned k;
 208                         xmm_t ipx4;
 209
 210                         ipx4 = vect_loadu_sil128((xmm_t *)(ip_batch + j));
 211                         ipx4 = *(xmm_t *)(ip_batch + j);
 212                         rte_lpm_lookupx4(lpm, ipx4, next_hops, UINT32_MAX);
 213                         for (k = 0; k < RTE_DIM(next_hops); k++)
 214                                 if (unlikely(next_hops[k] == UINT32_MAX))
 215                                         count++;
 216                 }
 217
 218                 total_time += rte_rdtsc() - begin;
 219         }
 220         printf("LPM LookupX4: %.1f cycles (fails = %.1f%%)\n",
 221                         (double)total_time / ((double)ITERATIONS * BATCH_SIZE),
 222                         (count * 100.0) / (double)(ITERATIONS * BATCH_SIZE));
 223
 224         /* Delete */
 225         status = 0;
 226         begin = rte_rdtsc();
 227
 228         for (i = 0; i < NUM_ROUTE_ENTRIES; i++) {
 229                 /* rte_lpm_delete(lpm, ip, depth) */
 230                 status += rte_lpm_delete(lpm, large_route_table[i].ip,
 231                                 large_route_table[i].depth);
 232         }
 233
 234         total_time += rte_rdtsc() - begin;
 235
 236         printf("Average LPM Delete: %g cycles\n",
 237                         (double)total_time / NUM_ROUTE_ENTRIES);
 238
 239         rte_lpm_delete_all(lpm);
 240         rte_lpm_free(lpm);
 241
 242         return 0;
 243 }
 244
 245 REGISTER_TEST_COMMAND(lpm_perf_autotest, test_lpm_perf);