[fio.git] / t / genzipf.c

/*
 * Generate/analyze pareto/zipf distributions to better understand
 * what an access pattern would look like.
 *
 * For instance, the following would generate a zipf distribution
 * with theta 1.2, using 100,000 values and split the reporting into
 * 20 buckets:
 *
 *	t/genzipf zipf 1.2 100000 20
 *
 * Only the distribution type (zipf or pareto) and spread input need
 * to be given, if not given defaults are used.
 *
 */
#include <stdio.h>
#include <stdlib.h>
#include <fcntl.h>
#include <string.h>
#include <unistd.h>

#include "../lib/zipf.h"
#include "../flist.h"
#include "../hash.h"

#define DEF_NR		1000000
#define DEF_NR_OUTPUT	23

struct node {
	struct flist_head list;
	unsigned long long val;
	unsigned long hits;
};

static struct flist_head *hash;
static unsigned long hash_bits = 24;
static unsigned long hash_size = 1 << 24;

enum {
	TYPE_NONE = 0,
	TYPE_ZIPF,
	TYPE_PARETO,
};
static const char *dist_types[] = { "None", "Zipf", "Pareto" };

static int dist_type = TYPE_ZIPF;
static unsigned long gb_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;

#define DEF_ZIPF_VAL	1.2
#define DEF_PARETO_VAL	0.3

static struct node *hash_lookup(unsigned long long val)
{
	struct flist_head *l = &hash[hash_long(val, hash_bits)];
	struct flist_head *entry;
	struct node *n;

	flist_for_each(entry, l) {
		n = flist_entry(entry, struct node, list);
		if (n->val == val)
			return n;
	}

	return NULL;
}

static struct node *hash_insert(struct node *n, unsigned long long val)
{
	struct flist_head *l = &hash[hash_long(val, hash_bits)];

	n->val = val;
	n->hits = 1;
	flist_add_tail(&n->list, l);
	return n;
}

static int parse_options(int argc, char *argv[])
{
	const char *optstring = "t:g:i:o:b:p:c";
	int c, dist_val_set = 0;

	while ((c = getopt(argc, argv, optstring)) != -1) {
		switch (c) {
		case 'p':
			percentage = atof(optarg);
			break;
		case 'b':
			block_size = strtoul(optarg, NULL, 10);
			break;
		case 't':
			if (!strncmp(optarg, "zipf", 4))
				dist_type = TYPE_ZIPF;
			else if (!strncmp(optarg, "pareto", 6))
				dist_type = TYPE_PARETO;
			else {
				printf("wrong dist type: %s\n", optarg);
				return 1;
			}
			break;
		case 'g':
			gb_size = strtoul(optarg, NULL, 10);
			break;
		case 'i':
			dist_val = atof(optarg);
			dist_val_set = 1;
			break;
		case 'o':
			output_nranges = strtoul(optarg, NULL, 10);
			break;
		case 'c':
			output_csv = 1;
			break;
		default:
			printf("bad option %c\n", c);
			return 1;
		}
	}

	if (dist_type == TYPE_PARETO) {
		if ((dist_val >= 1.00 || dist_val < 0.00)) {
			printf("pareto input must be > 0.00 and < 1.00\n");
			return 1;
		}
		if (!dist_val_set)
			dist_val = DEF_PARETO_VAL;
	} else if (dist_type == TYPE_ZIPF) {
		if (dist_val == 1.0) {
			printf("zipf input must be different than 1.0\n");
			return 1;
		}
		if (!dist_val_set)
			dist_val = DEF_ZIPF_VAL;
	}

	return 0;
}

struct output_sum {
	double output;
	unsigned int nranges;
};

static int node_cmp(const void *p1, const void *p2)
{
	const struct node *n1 = p1;
	const struct node *n2 = p2;

	return n2->hits - n1->hits;
}

int main(int argc, char *argv[])
{
	unsigned long offset;
	unsigned long i, j, k, nr_vals, cur_vals, interval, total_vals, nnodes;
	unsigned long long nranges;
	struct output_sum *output_sums;
	struct node *nodes;
	double perc, perc_i;
	struct zipf_state zs;

	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);

	nranges = gb_size * 1024 * 1024 * 1024ULL;
	nranges /= block_size;

	if (dist_type == TYPE_ZIPF)
		zipf_init(&zs, nranges, dist_val, 1);
	else
		pareto_init(&zs, nranges, dist_val, 1);

	hash_bits = 0;
	hash_size = nranges;
	while ((hash_size >>= 1) != 0)
		hash_bits++;

	hash_size = 1 << hash_bits;

	hash = malloc(hash_size * sizeof(struct flist_head));
	for (i = 0; i < hash_size; i++)
		INIT_FLIST_HEAD(&hash[i]);

	nodes = malloc(nranges * sizeof(struct node));

	for (nr_vals = i = j = 0; i < nranges; i++) {
		struct node *n;

		if (dist_type == TYPE_ZIPF)
			offset = zipf_next(&zs);
		else
			offset = pareto_next(&zs);

		n = hash_lookup(offset);
		if (n)
			n->hits++;
		else {
			hash_insert(&nodes[j], offset);
			j++;
		}

		nr_vals++;
	}

	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 {
		interval = (nr_vals + output_nranges - 1) / output_nranges;

		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 = 1;
		}

		total_vals = i = j = cur_vals = 0;

		for (k = 0; k < nnodes; k++) {
			struct output_sum *os = &output_sums[j];
			struct node *node = &nodes[k];

			if (i >= interval) {
				os->output =
				    (double)(cur_vals + 1) / (double)nranges;
				os->output *= 100.0;
				j++;
				cur_vals = node->hits;
				interval +=
				    (nr_vals + output_nranges -
				     1) / output_nranges;
			} else {
				cur_vals += node->hits;
				os->nranges += node->hits;
			}

			i++;
			total_vals += node->hits;

			if (percentage) {
				unsigned long blocks =
				    percentage * nranges / 100;

				if (total_vals >= blocks) {
					double cs =
					    i * 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 < j; 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;
			printf("%s %6.2f%%\t%6.2f%%\t\t%8u\t%6.2f%c\n",
			       i ? "|->" : "Top", perc, os->output, os->nranges,
			       gb, p);
		}

		free(output_sums);
	}

	free(hash);
	free(nodes);
	return 0;
}
Commit	Line	Data
	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 100,000 values and split the reporting into
	7	* 20 buckets:
	8	*
	9	* t/genzipf zipf 1.2 100000 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 "../flist.h"
	23	#include "../hash.h"
	24
	25	#define DEF_NR 1000000
	26	#define DEF_NR_OUTPUT 23
	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	};
	43	static const char *dist_types[] = { "None", "Zipf", "Pareto" };
	44
	45	static int dist_type = TYPE_ZIPF;
	46	static unsigned long gb_size = 500;
	47	static unsigned long block_size = 4096;
	48	static unsigned long output_nranges = DEF_NR_OUTPUT;
	49	static double percentage;
	50	static double dist_val;
	51	static int output_csv = 0;
	52
	53	#define DEF_ZIPF_VAL 1.2
	54	#define DEF_PARETO_VAL 0.3
	55
	56	static struct node *hash_lookup(unsigned long long val)
	57	{
	58	struct flist_head *l = &hash[hash_long(val, hash_bits)];
	59	struct flist_head *entry;
	60	struct node *n;
	61
	62	flist_for_each(entry, l) {
	63	n = flist_entry(entry, struct node, list);
	64	if (n->val == val)
	65	return n;
	66	}
	67
	68	return NULL;
	69	}
	70
	71	static struct node hash_insert(struct node n, unsigned long long val)
	72	{
	73	struct flist_head *l = &hash[hash_long(val, hash_bits)];
	74
	75	n->val = val;
	76	n->hits = 1;
	77	flist_add_tail(&n->list, l);
	78	return n;
	79	}
	80
	81	static int parse_options(int argc, char *argv[])
	82	{
	83	const char *optstring = "t:g:i:o:b:p:c";
	84	int c, dist_val_set = 0;
	85
	86	while ((c = getopt(argc, argv, optstring)) != -1) {
	87	switch (c) {
	88	case 'p':
	89	percentage = atof(optarg);
	90	break;
	91	case 'b':
	92	block_size = strtoul(optarg, NULL, 10);
	93	break;
	94	case 't':
	95	if (!strncmp(optarg, "zipf", 4))
	96	dist_type = TYPE_ZIPF;
	97	else if (!strncmp(optarg, "pareto", 6))
	98	dist_type = TYPE_PARETO;
	99	else {
	100	printf("wrong dist type: %s\n", optarg);
	101	return 1;
	102	}
	103	break;
	104	case 'g':
	105	gb_size = strtoul(optarg, NULL, 10);
	106	break;
	107	case 'i':
	108	dist_val = atof(optarg);
	109	dist_val_set = 1;
	110	break;
	111	case 'o':
	112	output_nranges = strtoul(optarg, NULL, 10);
	113	break;
	114	case 'c':
	115	output_csv = 1;
	116	break;
	117	default:
	118	printf("bad option %c\n", c);
	119	return 1;
	120	}
	121	}
	122
	123	if (dist_type == TYPE_PARETO) {
	124	if ((dist_val >= 1.00 \|\| dist_val < 0.00)) {
	125	printf("pareto input must be > 0.00 and < 1.00\n");
	126	return 1;
	127	}
	128	if (!dist_val_set)
	129	dist_val = DEF_PARETO_VAL;
	130	} else if (dist_type == TYPE_ZIPF) {
	131	if (dist_val == 1.0) {
	132	printf("zipf input must be different than 1.0\n");
	133	return 1;
	134	}
	135	if (!dist_val_set)
	136	dist_val = DEF_ZIPF_VAL;
	137	}
	138
	139	return 0;
	140	}
	141
	142	struct output_sum {
	143	double output;
	144	unsigned int nranges;
	145	};
	146
	147	static int node_cmp(const void p1, const void p2)
	148	{
	149	const struct node *n1 = p1;
	150	const struct node *n2 = p2;
	151
	152	return n2->hits - n1->hits;
	153	}
	154
	155	int main(int argc, char *argv[])
	156	{
	157	unsigned long offset;
	158	unsigned long i, j, k, nr_vals, cur_vals, interval, total_vals, nnodes;
	159	unsigned long long nranges;
	160	struct output_sum *output_sums;
	161	struct node *nodes;
	162	double perc, perc_i;
	163	struct zipf_state zs;
	164
	165	if (parse_options(argc, argv))
	166	return 1;
	167
	168	if( !output_csv )
	169	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);
	170
	171	nranges = gb_size * 1024 * 1024 * 1024ULL;
	172	nranges /= block_size;
	173
	174	if (dist_type == TYPE_ZIPF)
	175	zipf_init(&zs, nranges, dist_val, 1);
	176	else
	177	pareto_init(&zs, nranges, dist_val, 1);
	178
	179	hash_bits = 0;
	180	hash_size = nranges;
	181	while ((hash_size >>= 1) != 0)
	182	hash_bits++;
	183
	184	hash_size = 1 << hash_bits;
	185
	186	hash = malloc(hash_size * sizeof(struct flist_head));
	187	for (i = 0; i < hash_size; i++)
	188	INIT_FLIST_HEAD(&hash[i]);
	189
	190	nodes = malloc(nranges * sizeof(struct node));
	191
	192	for (nr_vals = i = j = 0; i < nranges; i++) {
	193	struct node *n;
	194
	195	if (dist_type == TYPE_ZIPF)
	196	offset = zipf_next(&zs);
	197	else
	198	offset = pareto_next(&zs);
	199
	200	n = hash_lookup(offset);
	201	if (n)
	202	n->hits++;
	203	else {
	204	hash_insert(&nodes[j], offset);
	205	j++;
	206	}
	207
	208	nr_vals++;
	209	}
	210
	211	qsort(nodes, j, sizeof(struct node), node_cmp);
	212	nnodes = j;
	213	nr_vals = nnodes;
	214
	215	if (output_csv) {
	216	printf("rank, count\n");
	217	for (k = 0; k < nnodes; k++)
	218	printf("%lu, %lu\n", k, nodes[k].hits);
	219	} else {
	220	interval = (nr_vals + output_nranges - 1) / output_nranges;
	221
	222	output_sums = malloc(output_nranges * sizeof(struct output_sum));
	223	for (i = 0; i < output_nranges; i++) {
	224	output_sums[i].output = 0.0;
	225	output_sums[i].nranges = 1;
	226	}
	227
	228	total_vals = i = j = cur_vals = 0;
	229
	230	for (k = 0; k < nnodes; k++) {
	231	struct output_sum *os = &output_sums[j];
	232	struct node *node = &nodes[k];
	233
	234	if (i >= interval) {
	235	os->output =
	236	(double)(cur_vals + 1) / (double)nranges;
	237	os->output *= 100.0;
	238	j++;
	239	cur_vals = node->hits;
	240	interval +=
	241	(nr_vals + output_nranges -
	242	1) / output_nranges;
	243	} else {
	244	cur_vals += node->hits;
	245	os->nranges += node->hits;
	246	}
	247
	248	i++;
	249	total_vals += node->hits;
	250
	251	if (percentage) {
	252	unsigned long blocks =
	253	percentage * nranges / 100;
	254
	255	if (total_vals >= blocks) {
	256	double cs =
	257	i * block_size / (1024 * 1024);
	258	char p = 'M';
	259
	260	if (cs > 1024.0) {
	261	cs /= 1024.0;
	262	p = 'G';
	263	}
	264	if (cs > 1024.0) {
	265	cs /= 1024.0;
	266	p = 'T';
	267	}
	268
	269	printf("%.2f%% of hits satisfied in %.3f%cB of cache\n", percentage, cs, p);
	270	percentage = 0.0;
	271	}
	272	}
	273	}
	274
	275	perc_i = 100.0 / (double)output_nranges;
	276	perc = 0.0;
	277
	278	printf("\n Rows Hits No Hits Size\n");
	279	printf("--------------------------------------------------------\n");
	280	for (i = 0; i < j; i++) {
	281	struct output_sum *os = &output_sums[i];
	282	double gb = (double)os->nranges * block_size / 1024.0;
	283	char p = 'K';
	284
	285	if (gb > 1024.0) {
	286	p = 'M';
	287	gb /= 1024.0;
	288	}
	289	if (gb > 1024.0) {
	290	p = 'G';
	291	gb /= 1024.0;
	292	}
	293
	294	perc += perc_i;
	295	printf("%s %6.2f%%\t%6.2f%%\t\t%8u\t%6.2f%c\n",
	296	i ? "\|->" : "Top", perc, os->output, os->nranges,
	297	gb, p);
	298	}
	299
	300	free(output_sums);
	301	}
	302
	303	free(hash);
	304	free(nodes);
	305	return 0;
	306	}