defaults to 100.0, meaning that all I/Os must be equal or below to the value
set by :option:`latency_target`.
+.. option:: latency_run=bool
+
+ Used with :option:`latency_target`. If false (default), fio will find
+ the highest queue depth that meets :option:`latency_target` and exit. If
+ true, fio will continue running and try to meet :option:`latency_target`
+ by adjusting queue depth.
+
.. option:: max_latency=time
If set, fio will exit the job with an ETIMEDOUT error if it exceeds this
o->latency_window = le64_to_cpu(top->latency_window);
o->max_latency = le64_to_cpu(top->max_latency);
o->latency_percentile.u.f = fio_uint64_to_double(le64_to_cpu(top->latency_percentile.u.i));
+ o->latency_run = le32_to_cpu(top->latency_run);
o->compress_percentage = le32_to_cpu(top->compress_percentage);
o->compress_chunk = le32_to_cpu(top->compress_chunk);
o->dedupe_percentage = le32_to_cpu(top->dedupe_percentage);
top->latency_window = __cpu_to_le64(o->latency_window);
top->max_latency = __cpu_to_le64(o->max_latency);
top->latency_percentile.u.i = __cpu_to_le64(fio_double_to_uint64(o->latency_percentile.u.f));
+ top->latency_run = __cpu_to_le32(o->latency_run);
top->compress_percentage = cpu_to_le32(o->compress_percentage);
top->compress_chunk = cpu_to_le32(o->compress_chunk);
top->dedupe_percentage = cpu_to_le32(o->dedupe_percentage);
defaults to 100.0, meaning that all I/Os must be equal or below to the value
set by \fBlatency_target\fR.
.TP
+.BI latency_run \fR=\fPbool
+Used with \fBlatency_target\fR. If false (default), fio will find the highest
+queue depth that meets \fBlatency_target\fR and exit. If true, fio will continue
+running and try to meet \fBlatency_target\fR by adjusting queue depth.
+.TP
.BI max_latency \fR=\fPtime
If set, fio will exit the job with an ETIMEDOUT error if it exceeds this
maximum latency. When the unit is omitted, the value is interpreted in
unsigned int latency_qd_high;
unsigned int latency_qd_low;
unsigned int latency_failed;
+ unsigned int latency_stable_count;
uint64_t latency_ios;
int latency_end_run;
td->latency_qd_low--;
td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
+ td->latency_stable_count = 0;
dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
td->latency_qd_low = td->latency_qd;
+ if (td->latency_qd + 1 == td->latency_qd_high) {
+ /*
+ * latency_qd will not incease on lat_target_success(), so
+ * called stable. If we stick with this queue depth, the
+ * final latency is likely lower than latency_target. Fix
+ * this by increasing latency_qd_high slowly. Use a naive
+ * heuristic here. If we get lat_target_success() 3 times
+ * in a row, increase latency_qd_high by 1.
+ */
+ if (++td->latency_stable_count >= 3) {
+ td->latency_qd_high++;
+ td->latency_stable_count = 0;
+ }
+ }
+
/*
* If we haven't failed yet, we double up to a failing value instead
* of bisecting from highest possible queue depth. If we have set
* Same as last one, we are done. Let it run a latency cycle, so
* we get only the results from the targeted depth.
*/
- if (td->latency_qd == qd) {
+ if (!o->latency_run && td->latency_qd == qd) {
if (td->latency_end_run) {
dprint(FD_RATE, "We are done\n");
td->done = 1;
.category = FIO_OPT_C_IO,
.group = FIO_OPT_G_LATPROF,
},
+ {
+ .name = "latency_run",
+ .lname = "Latency Run",
+ .type = FIO_OPT_BOOL,
+ .off1 = offsetof(struct thread_options, latency_run),
+ .help = "Keep adjusting queue depth to match latency_target",
+ .def = "0",
+ .category = FIO_OPT_C_IO,
+ .group = FIO_OPT_G_LATPROF,
+ },
{
.name = "invalidate",
.lname = "Cache invalidate",
};
enum {
- FIO_SERVER_VER = 82,
+ FIO_SERVER_VER = 83,
FIO_SERVER_MAX_FRAGMENT_PDU = 1024,
FIO_SERVER_MAX_CMD_MB = 2048,
unsigned long long latency_target;
unsigned long long latency_window;
fio_fp64_t latency_percentile;
+ uint32_t latency_run;
unsigned int sig_figs;
uint64_t latency_window;
uint64_t max_latency;
fio_fp64_t latency_percentile;
+ uint32_t latency_run;
uint32_t sig_figs;