The following changes since commit 16ada7542f32b38c1189390a97b595948079b100:
sg: don't return success if non bdev/char is used (2015-03-31 13:41:18 -0600)
are available in the git repository at:
git://git.kernel.dk/fio.git master
for you to fetch changes up to 87f5fce38c1b8fc979cea10c926973c2f2a76217:
genzip: switch to jenkins hash (2015-04-08 16:25:49 -0600)
----------------------------------------------------------------
Fabrice Bacchella (1):
fio-genzipf results output are slightly wrong, it doesn't help to understand how to use it: * The last row was not displayed * some useless variables were used * the calculation for the output table missed a few events, so the sum of 'No Hits' column was wrong * Wrong sample in genzipf.c header, used old style arguments * the help message for -o options was misleading. * useless define of DEF_NR. * default of default row number to a more common value of 20, instead of 23.
Jens Axboe (8):
Add support for normal/gaussian random distributions
Merge branch 'master' of ssh://git.kernel.dk/data/git/fio
Make normal distribution takes deviations as a percentage
genzipf: add support for normal distribution
genzip: add summed hit percentages
genzip: cleanups
Unify gauss and zipf/pareto input values
genzip: switch to jenkins hash
Makefile | 4 +-
cconv.c | 2 +
file.h | 6 +-
filesetup.c | 4 +-
fio.h | 1 +
io_u.c | 11 +++
lib/gauss.c | 56 ++++++++++++
lib/gauss.h | 17 ++++
options.c | 14 ++-
t/genzipf.c | 251 ++++++++++++++++++++++++++++++------------------------
thread_options.h | 3 +
11 files changed, 252 insertions(+), 117 deletions(-)
create mode 100644 lib/gauss.c
create mode 100644 lib/gauss.h
---
Diff of recent changes:
diff --git a/Makefile b/Makefile
index 52e515b..50f7468 100644
--- a/Makefile
+++ b/Makefile
@@ -36,7 +36,7 @@ SOURCE := gettime.c ioengines.c init.c stat.c log.c time.c filesetup.c \
lib/lfsr.c gettime-thread.c helpers.c lib/flist_sort.c \
lib/hweight.c lib/getrusage.c idletime.c td_error.c \
profiles/tiobench.c profiles/act.c io_u_queue.c filelock.c \
- lib/tp.c lib/bloom.c
+ lib/tp.c lib/bloom.c lib/gauss.c
ifdef CONFIG_LIBHDFS
HDFSFLAGS= -I $(JAVA_HOME)/include -I $(JAVA_HOME)/include/linux -I $(FIO_LIBHDFS_INCLUDE)
@@ -182,7 +182,7 @@ T_IEEE_OBJS += lib/ieee754.o
T_IEEE_PROGS = t/ieee754
T_ZIPF_OBS = t/genzipf.o
-T_ZIPF_OBJS += t/log.o lib/ieee754.o lib/rand.o lib/zipf.o t/genzipf.o
+T_ZIPF_OBJS += t/log.o lib/ieee754.o lib/rand.o lib/zipf.o lib/gauss.o t/genzipf.o
T_ZIPF_PROGS = t/fio-genzipf
T_AXMAP_OBJS = t/axmap.o
diff --git a/cconv.c b/cconv.c
index 0fca764..68f119f 100644
--- a/cconv.c
+++ b/cconv.c
@@ -165,6 +165,7 @@ void convert_thread_options_to_cpu(struct thread_options *o,
o->random_distribution = le32_to_cpu(top->random_distribution);
o->zipf_theta.u.f = fio_uint64_to_double(le64_to_cpu(top->zipf_theta.u.i));
o->pareto_h.u.f = fio_uint64_to_double(le64_to_cpu(top->pareto_h.u.i));
+ o->gauss_dev.u.f = fio_uint64_to_double(le64_to_cpu(top->gauss_dev.u.i));
o->random_generator = le32_to_cpu(top->random_generator);
o->hugepage_size = le32_to_cpu(top->hugepage_size);
o->rw_min_bs = le32_to_cpu(top->rw_min_bs);
@@ -339,6 +340,7 @@ void convert_thread_options_to_net(struct thread_options_pack *top,
top->random_distribution = cpu_to_le32(o->random_distribution);
top->zipf_theta.u.i = __cpu_to_le64(fio_double_to_uint64(o->zipf_theta.u.f));
top->pareto_h.u.i = __cpu_to_le64(fio_double_to_uint64(o->pareto_h.u.f));
+ top->gauss_dev.u.i = __cpu_to_le64(fio_double_to_uint64(o->gauss_dev.u.f));
top->random_generator = cpu_to_le32(o->random_generator);
top->hugepage_size = cpu_to_le32(o->hugepage_size);
top->rw_min_bs = cpu_to_le32(o->rw_min_bs);
diff --git a/file.h b/file.h
index fc47310..22ec742 100644
--- a/file.h
+++ b/file.h
@@ -8,6 +8,7 @@
#include "lib/zipf.h"
#include "lib/axmap.h"
#include "lib/lfsr.h"
+#include "lib/gauss.h"
/*
* The type of object we are working on
@@ -119,7 +120,10 @@ struct fio_file {
/*
* Used for zipf random distribution
*/
- struct zipf_state zipf;
+ union {
+ struct zipf_state zipf;
+ struct gauss_state gauss;
+ };
int references;
enum fio_file_flags flags;
diff --git a/filesetup.c b/filesetup.c
index 0fb5589..09e877f 100644
--- a/filesetup.c
+++ b/filesetup.c
@@ -998,8 +998,10 @@ static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
- else
+ else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
+ else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
+ gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
return 1;
}
diff --git a/fio.h b/fio.h
index f688084..0fb86ea 100644
--- a/fio.h
+++ b/fio.h
@@ -649,6 +649,7 @@ enum {
FIO_RAND_DIST_RANDOM = 0,
FIO_RAND_DIST_ZIPF,
FIO_RAND_DIST_PARETO,
+ FIO_RAND_DIST_GAUSS,
};
#define FIO_DEF_ZIPF 1.1
diff --git a/io_u.c b/io_u.c
index 975d242..1606512 100644
--- a/io_u.c
+++ b/io_u.c
@@ -149,6 +149,15 @@ static int __get_next_rand_offset_pareto(struct thread_data *td,
return 0;
}
+static int __get_next_rand_offset_gauss(struct thread_data *td,
+ struct fio_file *f, enum fio_ddir ddir,
+ uint64_t *b)
+{
+ *b = gauss_next(&f->gauss);
+ return 0;
+}
+
+
static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
{
struct rand_off *r1 = flist_entry(a, struct rand_off, list);
@@ -166,6 +175,8 @@ static int get_off_from_method(struct thread_data *td, struct fio_file *f,
return __get_next_rand_offset_zipf(td, f, ddir, b);
else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
return __get_next_rand_offset_pareto(td, f, ddir, b);
+ else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
+ return __get_next_rand_offset_gauss(td, f, ddir, b);
log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
return 1;
diff --git a/lib/gauss.c b/lib/gauss.c
new file mode 100644
index 0000000..1bb6c41
--- /dev/null
+++ b/lib/gauss.c
@@ -0,0 +1,56 @@
+#include <math.h>
+#include <string.h>
+#include <stdio.h>
+#include "../hash.h"
+#include "gauss.h"
+
+#define GAUSS_ITERS 12
+
+static int gauss_dev(struct gauss_state *gs)
+{
+ unsigned int r;
+ int vr;
+
+ if (!gs->stddev)
+ return 0;
+
+ r = __rand(&gs->r);
+ vr = gs->stddev * (r / (FRAND_MAX + 1.0));
+
+ return vr - gs->stddev / 2;
+}
+
+unsigned long long gauss_next(struct gauss_state *gs)
+{
+ unsigned long long sum = 0;
+ int i;
+
+ for (i = 0; i < GAUSS_ITERS; i++)
+ sum += __rand(&gs->r) % (gs->nranges + 1);
+
+ sum = (sum + GAUSS_ITERS - 1) / GAUSS_ITERS;
+
+ if (gs->stddev) {
+ int dev = gauss_dev(gs);
+
+ while (dev + sum >= gs->nranges)
+ dev /= 2;
+ sum += dev;
+ }
+
+ return __hash_u64(sum) % gs->nranges;
+}
+
+void gauss_init(struct gauss_state *gs, unsigned long nranges, double dev,
+ unsigned int seed)
+{
+ memset(gs, 0, sizeof(*gs));
+ init_rand_seed(&gs->r, seed);
+ gs->nranges = nranges;
+
+ if (dev != 0.0) {
+ gs->stddev = ceil((double) (nranges * 100.0) / dev);
+ if (gs->stddev > nranges / 2)
+ gs->stddev = nranges / 2;
+ }
+}
diff --git a/lib/gauss.h b/lib/gauss.h
new file mode 100644
index 0000000..a76df3f
--- /dev/null
+++ b/lib/gauss.h
@@ -0,0 +1,17 @@
+#ifndef FIO_GAUSS_H
+#define FIO_GAUSS_H
+
+#include <inttypes.h>
+#include "rand.h"
+
+struct gauss_state {
+ struct frand_state r;
+ uint64_t nranges;
+ unsigned int stddev;
+};
+
+void gauss_init(struct gauss_state *gs, unsigned long nranges, double dev,
+ unsigned int seed);
+unsigned long long gauss_next(struct gauss_state *gs);
+
+#endif
diff --git a/options.c b/options.c
index 337fecd..95e0e0c 100644
--- a/options.c
+++ b/options.c
@@ -718,6 +718,8 @@ static int str_random_distribution_cb(void *data, const char *str)
val = FIO_DEF_ZIPF;
else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
val = FIO_DEF_PARETO;
+ else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
+ val = 0.0;
else
return 0;
@@ -736,12 +738,18 @@ static int str_random_distribution_cb(void *data, const char *str)
return 1;
}
td->o.zipf_theta.u.f = val;
- } else {
+ } else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) {
if (val <= 0.00 || val >= 1.00) {
log_err("fio: pareto input out of range (0 < input < 1.0)\n");
return 1;
}
td->o.pareto_h.u.f = val;
+ } else {
+ if (val <= 0.00 || val >= 100.0) {
+ log_err("fio: normal deviation out of range (0 < input < 100.0)\n");
+ return 1;
+ }
+ td->o.gauss_dev.u.f = val;
}
return 0;
@@ -1875,6 +1883,10 @@ struct fio_option fio_options[FIO_MAX_OPTS] = {
.oval = FIO_RAND_DIST_PARETO,
.help = "Pareto distribution",
},
+ { .ival = "normal",
+ .oval = FIO_RAND_DIST_GAUSS,
+ .help = "Normal (gaussian) distribution",
+ },
},
.category = FIO_OPT_C_IO,
.group = FIO_OPT_G_RANDOM,
diff --git a/t/genzipf.c b/t/genzipf.c
index c5f098c..ff0729e 100644
--- a/t/genzipf.c
+++ b/t/genzipf.c
@@ -3,10 +3,10 @@
* 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
+ * with theta 1.2, using 262144 (1 GB / 4096) values and split the reporting into
* 20 buckets:
*
- * t/genzipf zipf 1.2 100000 20
+ * ./t/fio-genzipf -t zipf -i 1.2 -g 1 -b 4096 -o 20
*
* Only the distribution type (zipf or pareto) and spread input need
* to be given, if not given defaults are used.
@@ -19,11 +19,11 @@
#include <unistd.h>
#include "../lib/zipf.h"
+#include "../lib/gauss.h"
#include "../flist.h"
#include "../hash.h"
-#define DEF_NR 1000000
-#define DEF_NR_OUTPUT 23
+#define DEF_NR_OUTPUT 20
struct node {
struct flist_head list;
@@ -39,8 +39,14 @@ enum {
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;
@@ -48,14 +54,19 @@ 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;
@@ -68,14 +79,13 @@ static struct node *hash_lookup(unsigned long long val)
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)
@@ -83,11 +93,12 @@ 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 columns\n");
+ printf("\t-o\tNumber of output rows\n");
printf("\t-c\tOutput ranges in CSV format\n");
}
@@ -112,6 +123,8 @@ static int parse_options(int argc, char *argv[])
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;
@@ -128,7 +141,7 @@ static int parse_options(int argc, char *argv[])
output_nranges = strtoul(optarg, NULL, 10);
break;
case 'c':
- output_csv = 1;
+ output_type = OUTPUT_CSV;
break;
default:
printf("bad option %c\n", c);
@@ -168,20 +181,116 @@ static int node_cmp(const void *p1, const void *p2)
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 = 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 %% 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, 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 )
+ if (output_type != 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;
@@ -189,8 +298,10 @@ int main(int argc, char *argv[])
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;
@@ -199,19 +310,21 @@ int main(int argc, char *argv[])
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]);
nodes = malloc(nranges * sizeof(struct node));
- for (nr_vals = i = j = 0; i < nranges; i++) {
+ for (i = j = 0; i < nranges; i++) {
struct node *n;
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)
@@ -220,101 +333,15 @@ int main(int argc, char *argv[])
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);
- }
+ if (output_type == OUTPUT_CSV)
+ output_csv(nodes, nnodes);
+ else
+ output_normal(nodes, nnodes, nranges);
free(hash);
free(nodes);
diff --git a/thread_options.h b/thread_options.h
index 5a2428a..ee1114d 100644
--- a/thread_options.h
+++ b/thread_options.h
@@ -129,6 +129,7 @@ struct thread_options {
fio_fp64_t zipf_theta;
fio_fp64_t pareto_h;
+ fio_fp64_t gauss_dev;
unsigned int random_generator;
@@ -355,8 +356,10 @@ struct thread_options_pack {
uint32_t random_distribution;
uint32_t pad;
+
fio_fp64_t zipf_theta;
fio_fp64_t pareto_h;
+ fio_fp64_t gauss_dev;
uint32_t random_generator;
--
To unsubscribe from this list: send the line "unsubscribe fio" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
