ceph/src/spdk/dpdk/app/test-compress-perf/comp_perf_test_benchmark.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2018 Intel Corporation
   3  */
   4
   5 #include <rte_malloc.h>
   6 #include <rte_eal.h>
   7 #include <rte_log.h>
   8 #include <rte_cycles.h>
   9 #include <rte_compressdev.h>
  10
  11 #include "comp_perf_test_benchmark.h"
  12
  13 static int
  14 main_loop(struct comp_test_data *test_data, uint8_t level,
  15                         enum rte_comp_xform_type type)
  16 {
  17         uint8_t dev_id = test_data->cdev_id;
  18         uint32_t i, iter, num_iter;
  19         struct rte_comp_op **ops, **deq_ops;
  20         void *priv_xform = NULL;
  21         struct rte_comp_xform xform;
  22         struct rte_mbuf **input_bufs, **output_bufs;
  23         int res = 0;
  24         int allocated = 0;
  25         uint32_t out_seg_sz;
  26
  27         if (test_data == NULL || !test_data->burst_sz) {
  28                 RTE_LOG(ERR, USER1,
  29                         "Unknown burst size\n");
  30                 return -1;
  31         }
  32
  33         ops = rte_zmalloc_socket(NULL,
  34                 2 * test_data->total_bufs * sizeof(struct rte_comp_op *),
  35                 0, rte_socket_id());
  36
  37         if (ops == NULL) {
  38                 RTE_LOG(ERR, USER1,
  39                         "Can't allocate memory for ops strucures\n");
  40                 return -1;
  41         }
  42
  43         deq_ops = &ops[test_data->total_bufs];
  44
  45         if (type == RTE_COMP_COMPRESS) {
  46                 xform = (struct rte_comp_xform) {
  47                         .type = RTE_COMP_COMPRESS,
  48                         .compress = {
  49                                 .algo = RTE_COMP_ALGO_DEFLATE,
  50                                 .deflate.huffman = test_data->huffman_enc,
  51                                 .level = level,
  52                                 .window_size = test_data->window_sz,
  53                                 .chksum = RTE_COMP_CHECKSUM_NONE,
  54                                 .hash_algo = RTE_COMP_HASH_ALGO_NONE
  55                         }
  56                 };
  57                 input_bufs = test_data->decomp_bufs;
  58                 output_bufs = test_data->comp_bufs;
  59                 out_seg_sz = test_data->out_seg_sz;
  60         } else {
  61                 xform = (struct rte_comp_xform) {
  62                         .type = RTE_COMP_DECOMPRESS,
  63                         .decompress = {
  64                                 .algo = RTE_COMP_ALGO_DEFLATE,
  65                                 .chksum = RTE_COMP_CHECKSUM_NONE,
  66                                 .window_size = test_data->window_sz,
  67                                 .hash_algo = RTE_COMP_HASH_ALGO_NONE
  68                         }
  69                 };
  70                 input_bufs = test_data->comp_bufs;
  71                 output_bufs = test_data->decomp_bufs;
  72                 out_seg_sz = test_data->seg_sz;
  73         }
  74
  75         /* Create private xform */
  76         if (rte_compressdev_private_xform_create(dev_id, &xform,
  77                         &priv_xform) < 0) {
  78                 RTE_LOG(ERR, USER1, "Private xform could not be created\n");
  79                 res = -1;
  80                 goto end;
  81         }
  82
  83         uint64_t tsc_start, tsc_end, tsc_duration;
  84
  85         tsc_start = tsc_end = tsc_duration = 0;
  86         tsc_start = rte_rdtsc();
  87         num_iter = test_data->num_iter;
  88
  89         for (iter = 0; iter < num_iter; iter++) {
  90                 uint32_t total_ops = test_data->total_bufs;
  91                 uint32_t remaining_ops = test_data->total_bufs;
  92                 uint32_t total_deq_ops = 0;
  93                 uint32_t total_enq_ops = 0;
  94                 uint16_t ops_unused = 0;
  95                 uint16_t num_enq = 0;
  96                 uint16_t num_deq = 0;
  97
  98                 while (remaining_ops > 0) {
  99                         uint16_t num_ops = RTE_MIN(remaining_ops,
 100                                                    test_data->burst_sz);
 101                         uint16_t ops_needed = num_ops - ops_unused;
 102
 103                         /*
 104                          * Move the unused operations from the previous
 105                          * enqueue_burst call to the front, to maintain order
 106                          */
 107                         if ((ops_unused > 0) && (num_enq > 0)) {
 108                                 size_t nb_b_to_mov =
 109                                       ops_unused * sizeof(struct rte_comp_op *);
 110
 111                                 memmove(ops, &ops[num_enq], nb_b_to_mov);
 112                         }
 113
 114                         /* Allocate compression operations */
 115                         if (ops_needed && !rte_comp_op_bulk_alloc(
 116                                                 test_data->op_pool,
 117                                                 &ops[ops_unused],
 118                                                 ops_needed)) {
 119                                 RTE_LOG(ERR, USER1,
 120                                       "Could not allocate enough operations\n");
 121                                 res = -1;
 122                                 goto end;
 123                         }
 124                         allocated += ops_needed;
 125
 126                         for (i = 0; i < ops_needed; i++) {
 127                                 /*
 128                                  * Calculate next buffer to attach to operation
 129                                  */
 130                                 uint32_t buf_id = total_enq_ops + i +
 131                                                 ops_unused;
 132                                 uint16_t op_id = ops_unused + i;
 133                                 /* Reset all data in output buffers */
 134                                 struct rte_mbuf *m = output_bufs[buf_id];
 135
 136                                 m->pkt_len = out_seg_sz * m->nb_segs;
 137                                 while (m) {
 138                                         m->data_len = m->buf_len - m->data_off;
 139                                         m = m->next;
 140                                 }
 141                                 ops[op_id]->m_src = input_bufs[buf_id];
 142                                 ops[op_id]->m_dst = output_bufs[buf_id];
 143                                 ops[op_id]->src.offset = 0;
 144                                 ops[op_id]->src.length =
 145                                         rte_pktmbuf_pkt_len(input_bufs[buf_id]);
 146                                 ops[op_id]->dst.offset = 0;
 147                                 ops[op_id]->flush_flag = RTE_COMP_FLUSH_FINAL;
 148                                 ops[op_id]->input_chksum = buf_id;
 149                                 ops[op_id]->private_xform = priv_xform;
 150                         }
 151
 152                         num_enq = rte_compressdev_enqueue_burst(dev_id, 0, ops,
 153                                                                 num_ops);
 154                         if (num_enq == 0) {
 155                                 struct rte_compressdev_stats stats;
 156
 157                                 rte_compressdev_stats_get(dev_id, &stats);
 158                                 if (stats.enqueue_err_count) {
 159                                         res = -1;
 160                                         goto end;
 161                                 }
 162                         }
 163
 164                         ops_unused = num_ops - num_enq;
 165                         remaining_ops -= num_enq;
 166                         total_enq_ops += num_enq;
 167
 168                         num_deq = rte_compressdev_dequeue_burst(dev_id, 0,
 169                                                            deq_ops,
 170                                                            test_data->burst_sz);
 171                         total_deq_ops += num_deq;
 172
 173                         if (iter == num_iter - 1) {
 174                                 for (i = 0; i < num_deq; i++) {
 175                                         struct rte_comp_op *op = deq_ops[i];
 176
 177                                         if (op->status !=
 178                                                 RTE_COMP_OP_STATUS_SUCCESS) {
 179                                                 RTE_LOG(ERR, USER1,
 180                                                         "Some operations were not successful\n");
 181                                                 goto end;
 182                                         }
 183
 184                                         struct rte_mbuf *m = op->m_dst;
 185
 186                                         m->pkt_len = op->produced;
 187                                         uint32_t remaining_data = op->produced;
 188                                         uint16_t data_to_append;
 189
 190                                         while (remaining_data > 0) {
 191                                                 data_to_append =
 192                                                         RTE_MIN(remaining_data,
 193                                                              out_seg_sz);
 194                                                 m->data_len = data_to_append;
 195                                                 remaining_data -=
 196                                                                 data_to_append;
 197                                                 m = m->next;
 198                                         }
 199                                 }
 200                         }
 201                         rte_mempool_put_bulk(test_data->op_pool,
 202                                              (void **)deq_ops, num_deq);
 203                         allocated -= num_deq;
 204                 }
 205
 206                 /* Dequeue the last operations */
 207                 while (total_deq_ops < total_ops) {
 208                         num_deq = rte_compressdev_dequeue_burst(dev_id, 0,
 209                                                 deq_ops, test_data->burst_sz);
 210                         if (num_deq == 0) {
 211                                 struct rte_compressdev_stats stats;
 212
 213                                 rte_compressdev_stats_get(dev_id, &stats);
 214                                 if (stats.dequeue_err_count) {
 215                                         res = -1;
 216                                         goto end;
 217                                 }
 218                         }
 219
 220                         total_deq_ops += num_deq;
 221
 222                         if (iter == num_iter - 1) {
 223                                 for (i = 0; i < num_deq; i++) {
 224                                         struct rte_comp_op *op = deq_ops[i];
 225
 226                                         if (op->status !=
 227                                                 RTE_COMP_OP_STATUS_SUCCESS) {
 228                                                 RTE_LOG(ERR, USER1,
 229                                                         "Some operations were not successful\n");
 230                                                 goto end;
 231                                         }
 232
 233                                         struct rte_mbuf *m = op->m_dst;
 234
 235                                         m->pkt_len = op->produced;
 236                                         uint32_t remaining_data = op->produced;
 237                                         uint16_t data_to_append;
 238
 239                                         while (remaining_data > 0) {
 240                                                 data_to_append =
 241                                                 RTE_MIN(remaining_data,
 242                                                         out_seg_sz);
 243                                                 m->data_len = data_to_append;
 244                                                 remaining_data -=
 245                                                                 data_to_append;
 246                                                 m = m->next;
 247                                         }
 248                                 }
 249                         }
 250                         rte_mempool_put_bulk(test_data->op_pool,
 251                                              (void **)deq_ops, num_deq);
 252                         allocated -= num_deq;
 253                 }
 254         }
 255
 256         tsc_end = rte_rdtsc();
 257         tsc_duration = tsc_end - tsc_start;
 258
 259         if (type == RTE_COMP_COMPRESS)
 260                 test_data->comp_tsc_duration[level] =
 261                                 tsc_duration / num_iter;
 262         else
 263                 test_data->decomp_tsc_duration[level] =
 264                                 tsc_duration / num_iter;
 265
 266 end:
 267         rte_mempool_put_bulk(test_data->op_pool, (void **)ops, allocated);
 268         rte_compressdev_private_xform_free(dev_id, priv_xform);
 269         rte_free(ops);
 270         return res;
 271 }
 272
 273 int
 274 cperf_benchmark(struct comp_test_data *test_data, uint8_t level)
 275 {
 276         int i, ret = EXIT_SUCCESS;
 277
 278         /*
 279          * Run the tests twice, discarding the first performance
 280          * results, before the cache is warmed up
 281          */
 282         for (i = 0; i < 2; i++) {
 283                 if (main_loop(test_data, level, RTE_COMP_COMPRESS) < 0) {
 284                         ret = EXIT_FAILURE;
 285                         goto end;
 286                 }
 287         }
 288
 289         for (i = 0; i < 2; i++) {
 290                 if (main_loop(test_data, level, RTE_COMP_DECOMPRESS) < 0) {
 291                         ret = EXIT_FAILURE;
 292                         goto end;
 293                 }
 294         }
 295
 296         test_data->comp_tsc_byte =
 297                         (double)(test_data->comp_tsc_duration[level]) /
 298                                         test_data->input_data_sz;
 299
 300         test_data->decomp_tsc_byte =
 301                         (double)(test_data->decomp_tsc_duration[level]) /
 302                                         test_data->input_data_sz;
 303
 304         test_data->comp_gbps = rte_get_tsc_hz() / test_data->comp_tsc_byte * 8 /
 305                         1000000000;
 306
 307         test_data->decomp_gbps = rte_get_tsc_hz() / test_data->decomp_tsc_byte
 308                         * 8 / 1000000000;
 309 end:
 310         return ret;
 311 }