2 * Generate/analyze pareto/zipf distributions to better understand
3 * what an access pattern would look like.
5 * For instance, the following would generate a zipf distribution
6 * with theta 1.2, using 262144 (1 GB / 4096) values and split the reporting into
9 * ./t/fio-genzipf -t zipf -i 1.2 -g 1 -b 4096 -o 20
11 * Only the distribution type (zipf or pareto) and spread input need
12 * to be given, if not given defaults are used.
21 #include "../lib/zipf.h"
22 #include "../lib/gauss.h"
26 #define DEF_NR_OUTPUT 20
29 struct flist_head list;
30 unsigned long long val;
34 static struct flist_head *hash;
35 static unsigned long hash_bits = 24;
36 static unsigned long hash_size = 1 << 24;
44 static const char *dist_types[] = { "None", "Zipf", "Pareto", "Normal" };
51 static int dist_type = TYPE_ZIPF;
52 static unsigned long gb_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;
59 #define DEF_ZIPF_VAL 1.2
60 #define DEF_PARETO_VAL 0.3
62 static unsigned int hashv(unsigned long long val)
64 return jhash(&val, sizeof(val), 0) & (hash_size - 1);
67 static struct node *hash_lookup(unsigned long long val)
69 struct flist_head *l = &hash[hashv(val)];
70 struct flist_head *entry;
73 flist_for_each(entry, l) {
74 n = flist_entry(entry, struct node, list);
82 static void hash_insert(struct node *n, unsigned long long val)
84 struct flist_head *l = &hash[hashv(val)];
88 flist_add_tail(&n->list, l);
91 static void usage(void)
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");
105 static int parse_options(int argc, char *argv[])
107 const char *optstring = "t:g:i:o:b:p:ch";
108 int c, dist_val_set = 0;
110 while ((c = getopt(argc, argv, optstring)) != -1) {
116 percentage = atof(optarg);
119 block_size = strtoul(optarg, NULL, 10);
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;
129 printf("wrong dist type: %s\n", optarg);
134 gb_size = strtoul(optarg, NULL, 10);
137 dist_val = atof(optarg);
141 output_nranges = strtoul(optarg, NULL, 10);
144 output_type = OUTPUT_CSV;
147 printf("bad option %c\n", c);
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");
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");
165 dist_val = DEF_ZIPF_VAL;
173 unsigned int nranges;
176 static int node_cmp(const void *p1, const void *p2)
178 const struct node *n1 = p1;
179 const struct node *n2 = p2;
181 return n2->hits - n1->hits;
184 static void output_csv(struct node *nodes, unsigned long nnodes)
188 printf("rank, count\n");
189 for (i = 0; i < nnodes; i++)
190 printf("%lu, %lu\n", i, nodes[i].hits);
193 static void output_normal(struct node *nodes, unsigned long nnodes,
194 unsigned long nranges)
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;
201 struct output_sum *output_sums;
203 interval_step = (nnodes - 1) / output_nranges + 1;
204 next_interval = interval_step;
205 output_sums = malloc(output_nranges * sizeof(struct output_sum));
207 for (i = 0; i < output_nranges; i++) {
208 output_sums[i].output = 0.0;
209 output_sums[i].nranges = 0;
212 j = total_vals = cur_vals = 0;
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;
224 next_interval += interval_step;
229 if (total_vals >= blocks) {
230 double cs = i * block_size / (1024 * 1024);
242 printf("%.2f%% of hits satisfied in %.3f%cB of cache\n", percentage, cs, p);
248 perc_i = 100.0 / (double)output_nranges;
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;
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,
275 printf("-----------------------------------------------------------------------\n");
276 printf("Total\t\t\t\t\t\t%8llu\n", hit_sum);
280 int main(int argc, char *argv[])
282 unsigned long offset;
283 unsigned long long nranges;
284 unsigned long nnodes;
286 struct zipf_state zs;
287 struct gauss_state gs;
290 if (parse_options(argc, argv))
293 if (output_type != OUTPUT_CSV)
294 printf("Generating %s distribution with %f input and %lu GB size and %lu block_size.\n", dist_types[dist_type], dist_val, gb_size, block_size);
296 nranges = gb_size * 1024 * 1024 * 1024ULL;
297 nranges /= block_size;
299 if (dist_type == TYPE_ZIPF)
300 zipf_init(&zs, nranges, dist_val, 1);
301 else if (dist_type == TYPE_PARETO)
302 pareto_init(&zs, nranges, dist_val, 1);
304 gauss_init(&gs, nranges, dist_val, 1);
308 while ((hash_size >>= 1) != 0)
311 hash_size = 1 << hash_bits;
313 hash = calloc(hash_size, sizeof(struct flist_head));
314 for (i = 0; i < hash_size; i++)
315 INIT_FLIST_HEAD(&hash[i]);
317 nodes = malloc(nranges * sizeof(struct node));
319 for (i = j = 0; i < nranges; i++) {
322 if (dist_type == TYPE_ZIPF)
323 offset = zipf_next(&zs);
324 else if (dist_type == TYPE_PARETO)
325 offset = pareto_next(&zs);
327 offset = gauss_next(&gs);
329 n = hash_lookup(offset);
333 hash_insert(&nodes[j], offset);
338 qsort(nodes, j, sizeof(struct node), node_cmp);
341 if (output_type == OUTPUT_CSV)
342 output_csv(nodes, nnodes);
344 output_normal(nodes, nnodes, nranges);