* what an access pattern would look like.
*
* For instance, the following would generate a zipf distribution
- * with theta 1.2, using 262144 (1 GB / 4096) values and split the reporting into
- * 20 buckets:
+ * with theta 1.2, using 262144 (1 GiB / 4096) values and split the
+ * reporting into 20 buckets:
*
* ./t/fio-genzipf -t zipf -i 1.2 -g 1 -b 4096 -o 20
*
*/
#include <stdio.h>
#include <stdlib.h>
-#include <fcntl.h>
#include <string.h>
#include <unistd.h>
#include "../lib/zipf.h"
+#include "../lib/gauss.h"
#include "../flist.h"
#include "../hash.h"
TYPE_NONE = 0,
TYPE_ZIPF,
TYPE_PARETO,
+ TYPE_NORMAL,
+};
+static const char *dist_types[] = { "None", "Zipf", "Pareto", "Normal" };
+
+enum {
+ OUTPUT_NORMAL,
+ OUTPUT_CSV,
};
-static const char *dist_types[] = { "None", "Zipf", "Pareto" };
static int dist_type = TYPE_ZIPF;
-static unsigned long gb_size = 500;
+static unsigned long gib_size = 500;
static unsigned long block_size = 4096;
static unsigned long output_nranges = DEF_NR_OUTPUT;
static double percentage;
static double dist_val;
-static int output_csv = 0;
+static int output_type = OUTPUT_NORMAL;
#define DEF_ZIPF_VAL 1.2
#define DEF_PARETO_VAL 0.3
+static unsigned int hashv(unsigned long long val)
+{
+ return jhash(&val, sizeof(val), 0) & (hash_size - 1);
+}
+
static struct node *hash_lookup(unsigned long long val)
{
- struct flist_head *l = &hash[hash_long(val, hash_bits)];
+ struct flist_head *l = &hash[hashv(val)];
struct flist_head *entry;
struct node *n;
return NULL;
}
-static struct node *hash_insert(struct node *n, unsigned long long val)
+static void hash_insert(struct node *n, unsigned long long val)
{
- struct flist_head *l = &hash[hash_long(val, hash_bits)];
+ struct flist_head *l = &hash[hashv(val)];
n->val = val;
n->hits = 1;
flist_add_tail(&n->list, l);
- return n;
}
static void usage(void)
printf("genzipf: test zipf/pareto values for fio input\n");
printf("\t-h\tThis help screen\n");
printf("\t-p\tGenerate size of data set that are hit by this percentage\n");
- printf("\t-t\tDistribution type (zipf or pareto)\n");
- printf("\t-i\tDistribution algorithm input (zipf theta or pareto power)\n");
+ printf("\t-t\tDistribution type (zipf, pareto, or normal)\n");
+ printf("\t-i\tDistribution algorithm input (zipf theta, pareto power,\n"
+ "\t\tor normal %% deviation)\n");
printf("\t-b\tBlock size of a given range (in bytes)\n");
printf("\t-g\tSize of data set (in gigabytes)\n");
printf("\t-o\tNumber of output rows\n");
dist_type = TYPE_ZIPF;
else if (!strncmp(optarg, "pareto", 6))
dist_type = TYPE_PARETO;
+ else if (!strncmp(optarg, "normal", 6))
+ dist_type = TYPE_NORMAL;
else {
printf("wrong dist type: %s\n", optarg);
return 1;
}
break;
case 'g':
- gb_size = strtoul(optarg, NULL, 10);
+ gib_size = strtoul(optarg, NULL, 10);
break;
case 'i':
dist_val = atof(optarg);
output_nranges = strtoul(optarg, NULL, 10);
break;
case 'c':
- output_csv = 1;
+ output_type = OUTPUT_CSV;
break;
default:
printf("bad option %c\n", c);
return n2->hits - n1->hits;
}
+static void output_csv(struct node *nodes, unsigned long nnodes)
+{
+ unsigned long i;
+
+ printf("rank, count\n");
+ for (i = 0; i < nnodes; i++)
+ printf("%lu, %lu\n", i, nodes[i].hits);
+}
+
+static void output_normal(struct node *nodes, unsigned long nnodes,
+ unsigned long nranges)
+{
+ unsigned long i, j, cur_vals, interval_step, next_interval, total_vals;
+ unsigned long blocks = percentage * nnodes / 100;
+ double hit_percent_sum = 0;
+ unsigned long long hit_sum = 0;
+ double perc, perc_i;
+ struct output_sum *output_sums;
+
+ interval_step = (nnodes - 1) / output_nranges + 1;
+ next_interval = interval_step;
+ output_sums = malloc(output_nranges * sizeof(struct output_sum));
+
+ for (i = 0; i < output_nranges; i++) {
+ output_sums[i].output = 0.0;
+ output_sums[i].nranges = 0;
+ }
+
+ j = total_vals = cur_vals = 0;
+
+ for (i = 0; i < nnodes; i++) {
+ struct output_sum *os = &output_sums[j];
+ struct node *node = &nodes[i];
+ cur_vals += node->hits;
+ total_vals += node->hits;
+ os->nranges += node->hits;
+ if (i == (next_interval) -1 || i == nnodes - 1) {
+ os->output = (double) cur_vals / (double) nranges;
+ os->output *= 100.0;
+ cur_vals = 0;
+ next_interval += interval_step;
+ j++;
+ }
+
+ if (percentage) {
+ if (total_vals >= blocks) {
+ double cs = (double) i * block_size / (1024.0 * 1024.0);
+ char p = 'M';
+
+ if (cs > 1024.0) {
+ cs /= 1024.0;
+ p = 'G';
+ }
+ if (cs > 1024.0) {
+ cs /= 1024.0;
+ p = 'T';
+ }
+
+ printf("%.2f%% of hits satisfied in %.3f%cB of cache\n", percentage, cs, p);
+ percentage = 0.0;
+ }
+ }
+ }
+
+ perc_i = 100.0 / (double)output_nranges;
+ perc = 0.0;
+
+ printf("\n Rows Hits %% Sum %% # Hits Size\n");
+ printf("-----------------------------------------------------------------------\n");
+ for (i = 0; i < output_nranges; i++) {
+ struct output_sum *os = &output_sums[i];
+ double gb = (double)os->nranges * block_size / 1024.0;
+ char p = 'K';
+
+ if (gb > 1024.0) {
+ p = 'M';
+ gb /= 1024.0;
+ }
+ if (gb > 1024.0) {
+ p = 'G';
+ gb /= 1024.0;
+ }
+
+ perc += perc_i;
+ hit_percent_sum += os->output;
+ hit_sum += os->nranges;
+ printf("%s %6.2f%%\t%6.2f%%\t\t%6.2f%%\t\t%8u\t%6.2f%c\n",
+ i ? "|->" : "Top", perc, os->output, hit_percent_sum,
+ os->nranges, gb, p);
+ }
+
+ printf("-----------------------------------------------------------------------\n");
+ printf("Total\t\t\t\t\t\t%8llu\n", hit_sum);
+ free(output_sums);
+}
+
int main(int argc, char *argv[])
{
unsigned long offset;
- unsigned long i, j, k, nr_vals, cur_vals, interval_step, next_interval, total_vals, nnodes;
unsigned long long nranges;
- struct output_sum *output_sums;
+ unsigned long nnodes;
struct node *nodes;
- double perc, perc_i;
struct zipf_state zs;
+ struct gauss_state gs;
+ int i, j;
if (parse_options(argc, argv))
return 1;
- if( !output_csv )
- 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);
+ if (output_type != OUTPUT_CSV)
+ printf("Generating %s distribution with %f input and %lu GiB size and %lu block_size.\n",
+ dist_types[dist_type], dist_val, gib_size, block_size);
- nranges = gb_size * 1024 * 1024 * 1024ULL;
+ nranges = gib_size * 1024 * 1024 * 1024ULL;
nranges /= block_size;
if (dist_type == TYPE_ZIPF)
zipf_init(&zs, nranges, dist_val, 1);
- else
+ else if (dist_type == TYPE_PARETO)
pareto_init(&zs, nranges, dist_val, 1);
+ else
+ gauss_init(&gs, nranges, dist_val, 1);
hash_bits = 0;
hash_size = nranges;
hash_size = 1 << hash_bits;
- hash = malloc(hash_size * sizeof(struct flist_head));
+ hash = calloc(hash_size, sizeof(struct flist_head));
for (i = 0; i < hash_size; i++)
INIT_FLIST_HEAD(&hash[i]);
if (dist_type == TYPE_ZIPF)
offset = zipf_next(&zs);
- else
+ else if (dist_type == TYPE_PARETO)
offset = pareto_next(&zs);
+ else
+ offset = gauss_next(&gs);
n = hash_lookup(offset);
if (n)
hash_insert(&nodes[j], offset);
j++;
}
-
}
qsort(nodes, j, sizeof(struct node), node_cmp);
nnodes = j;
- nr_vals = nnodes;
-
- if (output_csv) {
- printf("rank, count\n");
- for (k = 0; k < nnodes; k++) {
- printf("%lu, %lu\n", k, nodes[k].hits);
- }
- } else {
- unsigned long blocks = percentage * nranges / 100;
- double hit_percent_sum = 0;
- unsigned long long hit_sum = 0;
- interval_step = (nr_vals - 1) / output_nranges + 1;
- next_interval = interval_step;
- output_sums = malloc(output_nranges * sizeof(struct output_sum));
- for (i = 0; i < output_nranges; i++) {
- output_sums[i].output = 0.0;
- output_sums[i].nranges = 0;
- }
-
- j = total_vals = cur_vals = 0;
-
- for (k = 0; k < nnodes; k++) {
- struct output_sum *os = &output_sums[j];
- struct node *node = &nodes[k];
- cur_vals += node->hits;
- total_vals += node->hits;
- os->nranges += node->hits;
- if (k == (next_interval) -1 || k == (nnodes - 1)) {
- os->output = (double)(cur_vals) / (double)nranges;
- os->output *= 100.0;
- cur_vals = 0;
- next_interval += interval_step;
- j++;
- }
-
- if (percentage) {
- if (total_vals >= blocks) {
- double cs = k * block_size / (1024 * 1024);
- char p = 'M';
-
- if (cs > 1024.0) {
- cs /= 1024.0;
- p = 'G';
- }
- if (cs > 1024.0) {
- cs /= 1024.0;
- p = 'T';
- }
-
- printf("%.2f%% of hits satisfied in %.3f%cB of cache\n", percentage, cs, p);
- percentage = 0.0;
- }
- }
- }
-
- perc_i = 100.0 / (double)output_nranges;
- perc = 0.0;
- printf("\n Rows Hits %% No Hits Size\n");
- printf("--------------------------------------------------------\n");
- for (i = 0; i < output_nranges; i++) {
- struct output_sum *os = &output_sums[i];
- double gb = (double)os->nranges * block_size / 1024.0;
- char p = 'K';
-
- if (gb > 1024.0) {
- p = 'M';
- gb /= 1024.0;
- }
- if (gb > 1024.0) {
- p = 'G';
- gb /= 1024.0;
- }
-
- perc += perc_i;
- hit_percent_sum += os->output;
- hit_sum += os->nranges;
- printf("%s %6.2f%%\t%6.2f%%\t\t%8u\t%6.2f%c\n",
- i ? "|->" : "Top", perc, os->output, os->nranges,
- gb, p);
- }
- printf("--------------------------------------------------------\n");
- printf("Total\t\t%6.2f%%\t\t%8llu\n", hit_percent_sum, hit_sum);
- free(output_sums);
- }
+ if (output_type == OUTPUT_CSV)
+ output_csv(nodes, nnodes);
+ else
+ output_normal(nodes, nnodes, nranges);
free(hash);
free(nodes);
projects / fio.git / blobdiff