Add gtod_reduce option
This cuts down even more on gtod() calls. While the three disable_X
options halved the gettimeofday() call count, this can reduce it to
less than 1 percent of what it otherwise would have been.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
diff --git a/HOWTO b/HOWTO
index 1d1aa61..f4efd28 100644
--- a/HOWTO
+++ b/HOWTO
@@ -848,6 +848,13 @@
disable_bw=bool Disable measurements of throughput/bandwidth numbers. See
disable_clat.
+gtod_reduce=bool Enable all of the gettimeofday() reducing options
+ (disable_clat, disable_slat, disable_bw) plus reduce
+ precision of the timeout somewhat to really shrink
+ the gettimeofday() call count. With this option enabled,
+ we only do about 0.4% of the gtod() calls we would have
+ done if all time keeping was enabled.
+
6.0 Interpreting the output
---------------------------
diff --git a/fio.c b/fio.c
index 7fb9490..08011c9 100644
--- a/fio.c
+++ b/fio.c
@@ -156,6 +156,17 @@
sigaction(SIGILL, &act, NULL);
}
+static inline int should_check_rate(struct thread_data *td)
+{
+ /*
+ * No minimum rate set, always ok
+ */
+ if (!td->o.ratemin && !td->o.rate_iops_min)
+ return 0;
+
+ return 1;
+}
+
/*
* Check if we are above the minimum rate given.
*/
@@ -167,12 +178,6 @@
unsigned long rate;
/*
- * No minimum rate set, always ok
- */
- if (!td->o.ratemin && !td->o.rate_iops_min)
- return 0;
-
- /*
* allow a 2 second settle period in the beginning
*/
if (mtime_since(&td->start, now) < 2000)
@@ -340,6 +345,12 @@
return 0;
}
+static inline void update_tv_cache(struct thread_data *td)
+{
+ if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
+ fio_gettime(&td->tv_cache, NULL);
+}
+
/*
* The main verify engine. Runs over the writes we previously submitted,
* reads the blocks back in, and checks the crc/md5 of the data.
@@ -377,7 +388,9 @@
if (!io_u)
break;
- if (runtime_exceeded(td, &io_u->start_time)) {
+ update_tv_cache(td);
+
+ if (runtime_exceeded(td, &td->tv_cache)) {
put_io_u(td, io_u);
td->terminate = 1;
break;
@@ -490,7 +503,6 @@
*/
static void do_io(struct thread_data *td)
{
- struct timeval s;
unsigned long usec;
unsigned int i;
int ret = 0;
@@ -514,9 +526,9 @@
if (!io_u)
break;
- memcpy(&s, &io_u->start_time, sizeof(s));
+ update_tv_cache(td);
- if (runtime_exceeded(td, &s)) {
+ if (runtime_exceeded(td, &td->tv_cache)) {
put_io_u(td, io_u);
td->terminate = 1;
break;
@@ -564,7 +576,9 @@
requeue_io_u(td, &io_u);
} else {
sync_done:
- fio_gettime(&comp_time, NULL);
+ if (should_check_rate(td))
+ fio_gettime(&comp_time, NULL);
+
bytes_done = io_u_sync_complete(td, io_u);
if (bytes_done < 0)
ret = bytes_done;
@@ -603,7 +617,8 @@
if (full && !min_evts)
min_evts = 1;
- fio_gettime(&comp_time, NULL);
+ if (should_check_rate(td))
+ fio_gettime(&comp_time, NULL);
do {
ret = io_u_queued_complete(td, min_evts);
@@ -624,8 +639,8 @@
* of completions except the very first one which may look
* a little bursty
*/
- if (!in_ramp_time(td)) {
- usec = utime_since(&s, &comp_time);
+ if (!in_ramp_time(td) && should_check_rate(td)) {
+ usec = utime_since(&td->tv_cache, &comp_time);
rate_throttle(td, usec, bytes_done);
@@ -977,6 +992,7 @@
while (keep_running(td)) {
fio_gettime(&td->start, NULL);
memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
+ memcpy(&td->tv_cache, &td->start, sizeof(td->start));
if (td->o.ratemin)
memcpy(&td->lastrate, &td->ts.stat_sample_time,
diff --git a/fio.h b/fio.h
index ed1257c..3e39aea 100644
--- a/fio.h
+++ b/fio.h
@@ -499,6 +499,7 @@
unsigned int disable_clat;
unsigned int disable_slat;
unsigned int disable_bw;
+ unsigned int gtod_reduce;
char *read_iolog_file;
char *write_iolog_file;
@@ -612,6 +613,9 @@
struct timeval epoch; /* time job was started */
struct timeval rw_end[2];
struct timeval last_issue;
+ struct timeval tv_cache;
+ unsigned int tv_cache_nr;
+ unsigned int tv_cache_mask;
unsigned int rw_end_set[2];
unsigned int ramp_time_over;
diff --git a/io_u.c b/io_u.c
index 62a76b9..8863fa7 100644
--- a/io_u.c
+++ b/io_u.c
@@ -851,7 +851,8 @@
out:
if (!td_io_prep(td, io_u)) {
- fio_gettime(&io_u->start_time, NULL);
+ if (!td->o.disable_slat)
+ fio_gettime(&io_u->start_time, NULL);
return io_u;
}
err_put:
diff --git a/options.c b/options.c
index d4fc184..f068ebd 100644
--- a/options.c
+++ b/options.c
@@ -432,6 +432,20 @@
return 0;
}
+static int str_gtod_reduce_cb(void *data, int *il)
+{
+ struct thread_data *td = data;
+ int val = *il;
+
+ td->o.disable_clat = !!val;
+ td->o.disable_slat = !!val;
+ td->o.disable_bw = !!val;
+ if (val)
+ td->tv_cache_mask = 63;
+
+ return 0;
+}
+
#define __stringify_1(x) #x
#define __stringify(x) __stringify_1(x)
@@ -1291,10 +1305,18 @@
},
#endif
{
+ .name = "gtod_reduce",
+ .type = FIO_OPT_BOOL,
+ .help = "Greatly reduce number of gettimeofday() calls",
+ .cb = str_gtod_reduce_cb,
+ .def = "0",
+ },
+ {
.name = "disable_clat",
.type = FIO_OPT_BOOL,
.off1 = td_var_offset(disable_clat),
.help = "Disable completion latency numbers",
+ .parent = "gtod_reduce",
.def = "0",
},
{
@@ -1302,6 +1324,7 @@
.type = FIO_OPT_BOOL,
.off1 = td_var_offset(disable_slat),
.help = "Disable submissionn latency numbers",
+ .parent = "gtod_reduce",
.def = "0",
},
{
@@ -1309,6 +1332,7 @@
.type = FIO_OPT_BOOL,
.off1 = td_var_offset(disable_bw),
.help = "Disable bandwidth logging",
+ .parent = "gtod_reduce",
.def = "0",
},
{