4 #include "steadystate.h"
6 bool steadystate_enabled = false;
8 void steadystate_free(struct thread_data *td)
10 free(td->ss.iops_data);
12 td->ss.iops_data = NULL;
13 td->ss.bw_data = NULL;
16 static void steadystate_alloc(struct thread_data *td)
18 td->ss.bw_data = calloc(td->ss.dur, sizeof(uint64_t));
19 td->ss.iops_data = calloc(td->ss.dur, sizeof(uint64_t));
21 td->ss.state |= FIO_SS_DATA;
24 void steadystate_setup(void)
26 struct thread_data *td, *prev_td;
29 if (!steadystate_enabled)
33 * if group reporting is enabled, identify the last td
34 * for each group and use it for storing steady state
43 if (!td->o.group_reporting) {
44 steadystate_alloc(td);
48 if (prev_groupid != td->groupid) {
50 steadystate_alloc(prev_td);
51 prev_groupid = td->groupid;
56 if (prev_td && prev_td->o.group_reporting)
57 steadystate_alloc(prev_td);
60 static bool steadystate_slope(uint64_t iops, uint64_t bw,
61 struct thread_data *td)
65 struct steadystate_data *ss = &td->ss;
68 ss->bw_data[ss->tail] = bw;
69 ss->iops_data[ss->tail] = iops;
71 if (ss->state & FIO_SS_IOPS)
76 if (ss->state & FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
77 if (!(ss->state & FIO_SS_BUFFER_FULL)) {
78 /* first time through */
79 for(i = 0, ss->sum_y = 0; i < ss->dur; i++) {
80 if (ss->state & FIO_SS_IOPS)
81 ss->sum_y += ss->iops_data[i];
83 ss->sum_y += ss->bw_data[i];
84 j = (ss->head + i) % ss->dur;
85 if (ss->state & FIO_SS_IOPS)
86 ss->sum_xy += i * ss->iops_data[j];
88 ss->sum_xy += i * ss->bw_data[j];
90 ss->state |= FIO_SS_BUFFER_FULL;
91 } else { /* easy to update the sums */
92 ss->sum_y -= ss->oldest_y;
94 ss->sum_xy = ss->sum_xy - ss->sum_y + ss->dur * new_val;
97 if (ss->state & FIO_SS_IOPS)
98 ss->oldest_y = ss->iops_data[ss->head];
100 ss->oldest_y = ss->bw_data[ss->head];
103 * calculate slope as (sum_xy - sum_x * sum_y / n) / (sum_(x^2)
104 * - (sum_x)^2 / n) This code assumes that all x values are
105 * equally spaced when they are often off by a few milliseconds.
106 * This assumption greatly simplifies the calculations.
108 ss->slope = (ss->sum_xy - (double) ss->sum_x * ss->sum_y / ss->dur) /
109 (ss->sum_x_sq - (double) ss->sum_x * ss->sum_x / ss->dur);
110 if (ss->state & FIO_SS_PCT)
111 ss->criterion = 100.0 * ss->slope / (ss->sum_y / ss->dur);
113 ss->criterion = ss->slope;
115 dprint(FD_STEADYSTATE, "sum_y: %llu, sum_xy: %llu, slope: %f, "
116 "criterion: %f, limit: %f\n",
117 (unsigned long long) ss->sum_y,
118 (unsigned long long) ss->sum_xy,
119 ss->slope, ss->criterion, ss->limit);
121 result = ss->criterion * (ss->criterion < 0.0 ? -1.0 : 1.0);
122 if (result < ss->limit)
126 ss->tail = (ss->tail + 1) % ss->dur;
127 if (ss->tail <= ss->head)
128 ss->head = (ss->head + 1) % ss->dur;
133 static bool steadystate_deviation(uint64_t iops, uint64_t bw,
134 struct thread_data *td)
140 struct steadystate_data *ss = &td->ss;
142 ss->bw_data[ss->tail] = bw;
143 ss->iops_data[ss->tail] = iops;
145 if (ss->state & FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
146 if (!(ss->state & FIO_SS_BUFFER_FULL)) {
147 /* first time through */
148 for(i = 0, ss->sum_y = 0; i < ss->dur; i++)
149 if (ss->state & FIO_SS_IOPS)
150 ss->sum_y += ss->iops_data[i];
152 ss->sum_y += ss->bw_data[i];
153 ss->state |= FIO_SS_BUFFER_FULL;
154 } else { /* easy to update the sum */
155 ss->sum_y -= ss->oldest_y;
156 if (ss->state & FIO_SS_IOPS)
157 ss->sum_y += ss->iops_data[ss->tail];
159 ss->sum_y += ss->bw_data[ss->tail];
162 if (ss->state & FIO_SS_IOPS)
163 ss->oldest_y = ss->iops_data[ss->head];
165 ss->oldest_y = ss->bw_data[ss->head];
167 mean = (double) ss->sum_y / ss->dur;
170 for (i = 0; i < ss->dur; i++) {
171 if (ss->state & FIO_SS_IOPS)
172 diff = ss->iops_data[i] - mean;
174 diff = ss->bw_data[i] - mean;
175 ss->deviation = max(ss->deviation, diff * (diff < 0.0 ? -1.0 : 1.0));
178 if (ss->state & FIO_SS_PCT)
179 ss->criterion = 100.0 * ss->deviation / mean;
181 ss->criterion = ss->deviation;
183 dprint(FD_STEADYSTATE, "sum_y: %llu, mean: %f, max diff: %f, "
184 "objective: %f, limit: %f\n",
185 (unsigned long long) ss->sum_y, mean,
186 ss->deviation, ss->criterion, ss->limit);
188 if (ss->criterion < ss->limit)
192 ss->tail = (ss->tail + 1) % ss->dur;
193 if (ss->tail <= ss->head)
194 ss->head = (ss->head + 1) % ss->dur;
199 int steadystate_check(void)
201 int i, j, ddir, prev_groupid, group_ramp_time_over = 0;
202 unsigned long rate_time;
203 struct thread_data *td, *td2;
205 uint64_t group_bw = 0, group_iops = 0;
206 uint64_t td_iops, td_bytes;
211 const bool needs_lock = td_async_processing(td);
212 struct steadystate_data *ss = &td->ss;
214 if (!ss->dur || td->runstate <= TD_SETTING_UP ||
215 td->runstate >= TD_EXITED || !ss->state ||
216 ss->state & FIO_SS_ATTAINED)
221 if (!td->o.group_reporting ||
222 (td->o.group_reporting && td->groupid != prev_groupid)) {
225 group_ramp_time_over = 0;
227 prev_groupid = td->groupid;
229 fio_gettime(&now, NULL);
230 if (ss->ramp_time && !(ss->state & FIO_SS_RAMP_OVER)) {
232 * Begin recording data one second after ss->ramp_time
235 if (utime_since(&td->epoch, &now) >= (ss->ramp_time + 1000000L))
236 ss->state |= FIO_SS_RAMP_OVER;
242 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
243 td_iops += td->io_blocks[ddir];
244 td_bytes += td->io_bytes[ddir];
248 __td_io_u_unlock(td);
250 rate_time = mtime_since(&ss->prev_time, &now);
251 memcpy(&ss->prev_time, &now, sizeof(now));
253 if (ss->state & FIO_SS_RAMP_OVER) {
254 group_bw += 1000 * (td_bytes - ss->prev_bytes) / rate_time;
255 group_iops += 1000 * (td_iops - ss->prev_iops) / rate_time;
256 ++group_ramp_time_over;
258 ss->prev_iops = td_iops;
259 ss->prev_bytes = td_bytes;
261 if (td->o.group_reporting && !(ss->state & FIO_SS_DATA))
265 * Don't begin checking criterion until ss->ramp_time is over
266 * for at least one thread in group
268 if (!group_ramp_time_over)
271 dprint(FD_STEADYSTATE, "steadystate_check() thread: %d, "
272 "groupid: %u, rate_msec: %ld, "
273 "iops: %llu, bw: %llu, head: %d, tail: %d\n",
274 i, td->groupid, rate_time,
275 (unsigned long long) group_iops,
276 (unsigned long long) group_bw,
279 if (ss->state & FIO_SS_SLOPE)
280 ret = steadystate_slope(group_iops, group_bw, td);
282 ret = steadystate_deviation(group_iops, group_bw, td);
285 if (td->o.group_reporting) {
286 for_each_td(td2, j) {
287 if (td2->groupid == td->groupid) {
288 td2->ss.state |= FIO_SS_ATTAINED;
289 fio_mark_td_terminate(td2);
293 ss->state |= FIO_SS_ATTAINED;
294 fio_mark_td_terminate(td);
301 int td_steadystate_init(struct thread_data *td)
303 struct steadystate_data *ss = &td->ss;
304 struct thread_options *o = &td->o;
305 struct thread_data *td2;
308 memset(ss, 0, sizeof(*ss));
311 steadystate_enabled = true;
312 o->ss_dur /= 1000000L;
314 /* put all steady state info in one place */
316 ss->limit = o->ss_limit.u.f;
317 ss->ramp_time = o->ss_ramp_time;
319 ss->state = o->ss_state;
320 if (!td->ss.ramp_time)
321 ss->state |= FIO_SS_RAMP_OVER;
323 ss->sum_x = o->ss_dur * (o->ss_dur - 1) / 2;
324 ss->sum_x_sq = (o->ss_dur - 1) * (o->ss_dur) * (2*o->ss_dur - 1) / 6;
327 /* make sure that ss options are consistent within reporting group */
328 for_each_td(td2, j) {
329 if (td2->groupid == td->groupid) {
330 struct steadystate_data *ss2 = &td2->ss;
332 if (ss2->dur != ss->dur ||
333 ss2->limit != ss->limit ||
334 ss2->ramp_time != ss->ramp_time ||
335 ss2->state != ss->state ||
336 ss2->sum_x != ss->sum_x ||
337 ss2->sum_x_sq != ss->sum_x_sq) {
338 td_verror(td, EINVAL, "job rejected: steadystate options must be consistent within reporting groups");
347 uint64_t steadystate_bw_mean(struct thread_stat *ts)
355 for (i = 0, sum = 0; i < ts->ss_dur; i++)
356 sum += ts->ss_bw_data[i];
358 return sum / ts->ss_dur;
361 uint64_t steadystate_iops_mean(struct thread_stat *ts)
369 for (i = 0, sum = 0; i < ts->ss_dur; i++)
370 sum += ts->ss_iops_data[i];
372 return sum / ts->ss_dur;