| 1 | /* |
| 2 | * Generate/analyze pareto/zipf distributions to better understand |
| 3 | * what an access pattern would look like. |
| 4 | * |
| 5 | * For instance, the following would generate a zipf distribution |
| 6 | * with theta 1.2, using 262144 (1 GiB / 4096) values and split the |
| 7 | * reporting into 20 buckets: |
| 8 | * |
| 9 | * ./t/fio-genzipf -t zipf -i 1.2 -g 1 -b 4096 -o 20 |
| 10 | * |
| 11 | * Only the distribution type (zipf or pareto) and spread input need |
| 12 | * to be given, if not given defaults are used. |
| 13 | * |
| 14 | */ |
| 15 | #include <stdio.h> |
| 16 | #include <stdlib.h> |
| 17 | #include <fcntl.h> |
| 18 | #include <string.h> |
| 19 | #include <unistd.h> |
| 20 | |
| 21 | #include "../lib/zipf.h" |
| 22 | #include "../lib/gauss.h" |
| 23 | #include "../flist.h" |
| 24 | #include "../hash.h" |
| 25 | |
| 26 | #define DEF_NR_OUTPUT 20 |
| 27 | |
| 28 | struct node { |
| 29 | struct flist_head list; |
| 30 | unsigned long long val; |
| 31 | unsigned long hits; |
| 32 | }; |
| 33 | |
| 34 | static struct flist_head *hash; |
| 35 | static unsigned long hash_bits = 24; |
| 36 | static unsigned long hash_size = 1 << 24; |
| 37 | |
| 38 | enum { |
| 39 | TYPE_NONE = 0, |
| 40 | TYPE_ZIPF, |
| 41 | TYPE_PARETO, |
| 42 | TYPE_NORMAL, |
| 43 | }; |
| 44 | static const char *dist_types[] = { "None", "Zipf", "Pareto", "Normal" }; |
| 45 | |
| 46 | enum { |
| 47 | OUTPUT_NORMAL, |
| 48 | OUTPUT_CSV, |
| 49 | }; |
| 50 | |
| 51 | static int dist_type = TYPE_ZIPF; |
| 52 | static unsigned long gib_size = 500; |
| 53 | static unsigned long block_size = 4096; |
| 54 | static unsigned long output_nranges = DEF_NR_OUTPUT; |
| 55 | static double percentage; |
| 56 | static double dist_val; |
| 57 | static int output_type = OUTPUT_NORMAL; |
| 58 | |
| 59 | #define DEF_ZIPF_VAL 1.2 |
| 60 | #define DEF_PARETO_VAL 0.3 |
| 61 | |
| 62 | static unsigned int hashv(unsigned long long val) |
| 63 | { |
| 64 | return jhash(&val, sizeof(val), 0) & (hash_size - 1); |
| 65 | } |
| 66 | |
| 67 | static struct node *hash_lookup(unsigned long long val) |
| 68 | { |
| 69 | struct flist_head *l = &hash[hashv(val)]; |
| 70 | struct flist_head *entry; |
| 71 | struct node *n; |
| 72 | |
| 73 | flist_for_each(entry, l) { |
| 74 | n = flist_entry(entry, struct node, list); |
| 75 | if (n->val == val) |
| 76 | return n; |
| 77 | } |
| 78 | |
| 79 | return NULL; |
| 80 | } |
| 81 | |
| 82 | static void hash_insert(struct node *n, unsigned long long val) |
| 83 | { |
| 84 | struct flist_head *l = &hash[hashv(val)]; |
| 85 | |
| 86 | n->val = val; |
| 87 | n->hits = 1; |
| 88 | flist_add_tail(&n->list, l); |
| 89 | } |
| 90 | |
| 91 | static void usage(void) |
| 92 | { |
| 93 | printf("genzipf: test zipf/pareto values for fio input\n"); |
| 94 | printf("\t-h\tThis help screen\n"); |
| 95 | printf("\t-p\tGenerate size of data set that are hit by this percentage\n"); |
| 96 | printf("\t-t\tDistribution type (zipf, pareto, or normal)\n"); |
| 97 | printf("\t-i\tDistribution algorithm input (zipf theta, pareto power,\n" |
| 98 | "\t\tor normal %% deviation)\n"); |
| 99 | printf("\t-b\tBlock size of a given range (in bytes)\n"); |
| 100 | printf("\t-g\tSize of data set (in gigabytes)\n"); |
| 101 | printf("\t-o\tNumber of output rows\n"); |
| 102 | printf("\t-c\tOutput ranges in CSV format\n"); |
| 103 | } |
| 104 | |
| 105 | static int parse_options(int argc, char *argv[]) |
| 106 | { |
| 107 | const char *optstring = "t:g:i:o:b:p:ch"; |
| 108 | int c, dist_val_set = 0; |
| 109 | |
| 110 | while ((c = getopt(argc, argv, optstring)) != -1) { |
| 111 | switch (c) { |
| 112 | case 'h': |
| 113 | usage(); |
| 114 | return 1; |
| 115 | case 'p': |
| 116 | percentage = atof(optarg); |
| 117 | break; |
| 118 | case 'b': |
| 119 | block_size = strtoul(optarg, NULL, 10); |
| 120 | break; |
| 121 | case 't': |
| 122 | if (!strncmp(optarg, "zipf", 4)) |
| 123 | dist_type = TYPE_ZIPF; |
| 124 | else if (!strncmp(optarg, "pareto", 6)) |
| 125 | dist_type = TYPE_PARETO; |
| 126 | else if (!strncmp(optarg, "normal", 6)) |
| 127 | dist_type = TYPE_NORMAL; |
| 128 | else { |
| 129 | printf("wrong dist type: %s\n", optarg); |
| 130 | return 1; |
| 131 | } |
| 132 | break; |
| 133 | case 'g': |
| 134 | gib_size = strtoul(optarg, NULL, 10); |
| 135 | break; |
| 136 | case 'i': |
| 137 | dist_val = atof(optarg); |
| 138 | dist_val_set = 1; |
| 139 | break; |
| 140 | case 'o': |
| 141 | output_nranges = strtoul(optarg, NULL, 10); |
| 142 | break; |
| 143 | case 'c': |
| 144 | output_type = OUTPUT_CSV; |
| 145 | break; |
| 146 | default: |
| 147 | printf("bad option %c\n", c); |
| 148 | return 1; |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | if (dist_type == TYPE_PARETO) { |
| 153 | if ((dist_val >= 1.00 || dist_val < 0.00)) { |
| 154 | printf("pareto input must be > 0.00 and < 1.00\n"); |
| 155 | return 1; |
| 156 | } |
| 157 | if (!dist_val_set) |
| 158 | dist_val = DEF_PARETO_VAL; |
| 159 | } else if (dist_type == TYPE_ZIPF) { |
| 160 | if (dist_val == 1.0) { |
| 161 | printf("zipf input must be different than 1.0\n"); |
| 162 | return 1; |
| 163 | } |
| 164 | if (!dist_val_set) |
| 165 | dist_val = DEF_ZIPF_VAL; |
| 166 | } |
| 167 | |
| 168 | return 0; |
| 169 | } |
| 170 | |
| 171 | struct output_sum { |
| 172 | double output; |
| 173 | unsigned int nranges; |
| 174 | }; |
| 175 | |
| 176 | static int node_cmp(const void *p1, const void *p2) |
| 177 | { |
| 178 | const struct node *n1 = p1; |
| 179 | const struct node *n2 = p2; |
| 180 | |
| 181 | return n2->hits - n1->hits; |
| 182 | } |
| 183 | |
| 184 | static void output_csv(struct node *nodes, unsigned long nnodes) |
| 185 | { |
| 186 | unsigned long i; |
| 187 | |
| 188 | printf("rank, count\n"); |
| 189 | for (i = 0; i < nnodes; i++) |
| 190 | printf("%lu, %lu\n", i, nodes[i].hits); |
| 191 | } |
| 192 | |
| 193 | static void output_normal(struct node *nodes, unsigned long nnodes, |
| 194 | unsigned long nranges) |
| 195 | { |
| 196 | unsigned long i, j, cur_vals, interval_step, next_interval, total_vals; |
| 197 | unsigned long blocks = percentage * nnodes / 100; |
| 198 | double hit_percent_sum = 0; |
| 199 | unsigned long long hit_sum = 0; |
| 200 | double perc, perc_i; |
| 201 | struct output_sum *output_sums; |
| 202 | |
| 203 | interval_step = (nnodes - 1) / output_nranges + 1; |
| 204 | next_interval = interval_step; |
| 205 | output_sums = malloc(output_nranges * sizeof(struct output_sum)); |
| 206 | |
| 207 | for (i = 0; i < output_nranges; i++) { |
| 208 | output_sums[i].output = 0.0; |
| 209 | output_sums[i].nranges = 0; |
| 210 | } |
| 211 | |
| 212 | j = total_vals = cur_vals = 0; |
| 213 | |
| 214 | for (i = 0; i < nnodes; i++) { |
| 215 | struct output_sum *os = &output_sums[j]; |
| 216 | struct node *node = &nodes[i]; |
| 217 | cur_vals += node->hits; |
| 218 | total_vals += node->hits; |
| 219 | os->nranges += node->hits; |
| 220 | if (i == (next_interval) -1 || i == nnodes - 1) { |
| 221 | os->output = (double) cur_vals / (double) nranges; |
| 222 | os->output *= 100.0; |
| 223 | cur_vals = 0; |
| 224 | next_interval += interval_step; |
| 225 | j++; |
| 226 | } |
| 227 | |
| 228 | if (percentage) { |
| 229 | if (total_vals >= blocks) { |
| 230 | double cs = (double) i * block_size / (1024.0 * 1024.0); |
| 231 | char p = 'M'; |
| 232 | |
| 233 | if (cs > 1024.0) { |
| 234 | cs /= 1024.0; |
| 235 | p = 'G'; |
| 236 | } |
| 237 | if (cs > 1024.0) { |
| 238 | cs /= 1024.0; |
| 239 | p = 'T'; |
| 240 | } |
| 241 | |
| 242 | printf("%.2f%% of hits satisfied in %.3f%cB of cache\n", percentage, cs, p); |
| 243 | percentage = 0.0; |
| 244 | } |
| 245 | } |
| 246 | } |
| 247 | |
| 248 | perc_i = 100.0 / (double)output_nranges; |
| 249 | perc = 0.0; |
| 250 | |
| 251 | printf("\n Rows Hits %% Sum %% # Hits Size\n"); |
| 252 | printf("-----------------------------------------------------------------------\n"); |
| 253 | for (i = 0; i < output_nranges; i++) { |
| 254 | struct output_sum *os = &output_sums[i]; |
| 255 | double gb = (double)os->nranges * block_size / 1024.0; |
| 256 | char p = 'K'; |
| 257 | |
| 258 | if (gb > 1024.0) { |
| 259 | p = 'M'; |
| 260 | gb /= 1024.0; |
| 261 | } |
| 262 | if (gb > 1024.0) { |
| 263 | p = 'G'; |
| 264 | gb /= 1024.0; |
| 265 | } |
| 266 | |
| 267 | perc += perc_i; |
| 268 | hit_percent_sum += os->output; |
| 269 | hit_sum += os->nranges; |
| 270 | printf("%s %6.2f%%\t%6.2f%%\t\t%6.2f%%\t\t%8u\t%6.2f%c\n", |
| 271 | i ? "|->" : "Top", perc, os->output, hit_percent_sum, |
| 272 | os->nranges, gb, p); |
| 273 | } |
| 274 | |
| 275 | printf("-----------------------------------------------------------------------\n"); |
| 276 | printf("Total\t\t\t\t\t\t%8llu\n", hit_sum); |
| 277 | free(output_sums); |
| 278 | } |
| 279 | |
| 280 | int main(int argc, char *argv[]) |
| 281 | { |
| 282 | unsigned long offset; |
| 283 | unsigned long long nranges; |
| 284 | unsigned long nnodes; |
| 285 | struct node *nodes; |
| 286 | struct zipf_state zs; |
| 287 | struct gauss_state gs; |
| 288 | int i, j; |
| 289 | |
| 290 | if (parse_options(argc, argv)) |
| 291 | return 1; |
| 292 | |
| 293 | if (output_type != OUTPUT_CSV) |
| 294 | printf("Generating %s distribution with %f input and %lu GiB size and %lu block_size.\n", |
| 295 | dist_types[dist_type], dist_val, gib_size, block_size); |
| 296 | |
| 297 | nranges = gib_size * 1024 * 1024 * 1024ULL; |
| 298 | nranges /= block_size; |
| 299 | |
| 300 | if (dist_type == TYPE_ZIPF) |
| 301 | zipf_init(&zs, nranges, dist_val, 1); |
| 302 | else if (dist_type == TYPE_PARETO) |
| 303 | pareto_init(&zs, nranges, dist_val, 1); |
| 304 | else |
| 305 | gauss_init(&gs, nranges, dist_val, 1); |
| 306 | |
| 307 | hash_bits = 0; |
| 308 | hash_size = nranges; |
| 309 | while ((hash_size >>= 1) != 0) |
| 310 | hash_bits++; |
| 311 | |
| 312 | hash_size = 1 << hash_bits; |
| 313 | |
| 314 | hash = calloc(hash_size, sizeof(struct flist_head)); |
| 315 | for (i = 0; i < hash_size; i++) |
| 316 | INIT_FLIST_HEAD(&hash[i]); |
| 317 | |
| 318 | nodes = malloc(nranges * sizeof(struct node)); |
| 319 | |
| 320 | for (i = j = 0; i < nranges; i++) { |
| 321 | struct node *n; |
| 322 | |
| 323 | if (dist_type == TYPE_ZIPF) |
| 324 | offset = zipf_next(&zs); |
| 325 | else if (dist_type == TYPE_PARETO) |
| 326 | offset = pareto_next(&zs); |
| 327 | else |
| 328 | offset = gauss_next(&gs); |
| 329 | |
| 330 | n = hash_lookup(offset); |
| 331 | if (n) |
| 332 | n->hits++; |
| 333 | else { |
| 334 | hash_insert(&nodes[j], offset); |
| 335 | j++; |
| 336 | } |
| 337 | } |
| 338 | |
| 339 | qsort(nodes, j, sizeof(struct node), node_cmp); |
| 340 | nnodes = j; |
| 341 | |
| 342 | if (output_type == OUTPUT_CSV) |
| 343 | output_csv(nodes, nnodes); |
| 344 | else |
| 345 | output_normal(nodes, nnodes, nranges); |
| 346 | |
| 347 | free(hash); |
| 348 | free(nodes); |
| 349 | return 0; |
| 350 | } |