4 #include "steadystate.h"
5 #include "helper_thread.h"
7 bool steadystate_enabled = false;
9 void steadystate_free(struct thread_data *td)
11 free(td->ss.iops_data);
13 td->ss.iops_data = NULL;
14 td->ss.bw_data = NULL;
17 static void steadystate_alloc(struct thread_data *td)
19 td->ss.bw_data = calloc(td->ss.dur, sizeof(uint64_t));
20 td->ss.iops_data = calloc(td->ss.dur, sizeof(uint64_t));
22 td->ss.state |= FIO_SS_DATA;
25 void steadystate_setup(void)
27 struct thread_data *td, *prev_td;
30 if (!steadystate_enabled)
34 * if group reporting is enabled, identify the last td
35 * for each group and use it for storing steady state
44 if (!td->o.group_reporting) {
45 steadystate_alloc(td);
49 if (prev_groupid != td->groupid) {
51 steadystate_alloc(prev_td);
52 prev_groupid = td->groupid;
57 if (prev_td && prev_td->o.group_reporting)
58 steadystate_alloc(prev_td);
61 static bool steadystate_slope(uint64_t iops, uint64_t bw,
62 struct thread_data *td)
66 struct steadystate_data *ss = &td->ss;
69 ss->bw_data[ss->tail] = bw;
70 ss->iops_data[ss->tail] = iops;
72 if (ss->state & FIO_SS_IOPS)
77 if (ss->state & FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
78 if (!(ss->state & FIO_SS_BUFFER_FULL)) {
79 /* first time through */
80 for(i = 0, ss->sum_y = 0; i < ss->dur; i++) {
81 if (ss->state & FIO_SS_IOPS)
82 ss->sum_y += ss->iops_data[i];
84 ss->sum_y += ss->bw_data[i];
85 j = (ss->head + i) % ss->dur;
86 if (ss->state & FIO_SS_IOPS)
87 ss->sum_xy += i * ss->iops_data[j];
89 ss->sum_xy += i * ss->bw_data[j];
91 ss->state |= FIO_SS_BUFFER_FULL;
92 } else { /* easy to update the sums */
93 ss->sum_y -= ss->oldest_y;
95 ss->sum_xy = ss->sum_xy - ss->sum_y + ss->dur * new_val;
98 if (ss->state & FIO_SS_IOPS)
99 ss->oldest_y = ss->iops_data[ss->head];
101 ss->oldest_y = ss->bw_data[ss->head];
104 * calculate slope as (sum_xy - sum_x * sum_y / n) / (sum_(x^2)
105 * - (sum_x)^2 / n) This code assumes that all x values are
106 * equally spaced when they are often off by a few milliseconds.
107 * This assumption greatly simplifies the calculations.
109 ss->slope = (ss->sum_xy - (double) ss->sum_x * ss->sum_y / ss->dur) /
110 (ss->sum_x_sq - (double) ss->sum_x * ss->sum_x / ss->dur);
111 if (ss->state & FIO_SS_PCT)
112 ss->criterion = 100.0 * ss->slope / (ss->sum_y / ss->dur);
114 ss->criterion = ss->slope;
116 dprint(FD_STEADYSTATE, "sum_y: %llu, sum_xy: %llu, slope: %f, "
117 "criterion: %f, limit: %f\n",
118 (unsigned long long) ss->sum_y,
119 (unsigned long long) ss->sum_xy,
120 ss->slope, ss->criterion, ss->limit);
122 result = ss->criterion * (ss->criterion < 0.0 ? -1.0 : 1.0);
123 if (result < ss->limit)
127 ss->tail = (ss->tail + 1) % ss->dur;
128 if (ss->tail <= ss->head)
129 ss->head = (ss->head + 1) % ss->dur;
134 static bool steadystate_deviation(uint64_t iops, uint64_t bw,
135 struct thread_data *td)
141 struct steadystate_data *ss = &td->ss;
143 ss->bw_data[ss->tail] = bw;
144 ss->iops_data[ss->tail] = iops;
146 if (ss->state & FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
147 if (!(ss->state & FIO_SS_BUFFER_FULL)) {
148 /* first time through */
149 for(i = 0, ss->sum_y = 0; i < ss->dur; i++)
150 if (ss->state & FIO_SS_IOPS)
151 ss->sum_y += ss->iops_data[i];
153 ss->sum_y += ss->bw_data[i];
154 ss->state |= FIO_SS_BUFFER_FULL;
155 } else { /* easy to update the sum */
156 ss->sum_y -= ss->oldest_y;
157 if (ss->state & FIO_SS_IOPS)
158 ss->sum_y += ss->iops_data[ss->tail];
160 ss->sum_y += ss->bw_data[ss->tail];
163 if (ss->state & FIO_SS_IOPS)
164 ss->oldest_y = ss->iops_data[ss->head];
166 ss->oldest_y = ss->bw_data[ss->head];
168 mean = (double) ss->sum_y / ss->dur;
171 for (i = 0; i < ss->dur; i++) {
172 if (ss->state & FIO_SS_IOPS)
173 diff = ss->iops_data[i] - mean;
175 diff = ss->bw_data[i] - mean;
176 ss->deviation = max(ss->deviation, diff * (diff < 0.0 ? -1.0 : 1.0));
179 if (ss->state & FIO_SS_PCT)
180 ss->criterion = 100.0 * ss->deviation / mean;
182 ss->criterion = ss->deviation;
184 dprint(FD_STEADYSTATE, "sum_y: %llu, mean: %f, max diff: %f, "
185 "objective: %f, limit: %f\n",
186 (unsigned long long) ss->sum_y, mean,
187 ss->deviation, ss->criterion, ss->limit);
189 if (ss->criterion < ss->limit)
193 ss->tail = (ss->tail + 1) % ss->dur;
194 if (ss->tail <= ss->head)
195 ss->head = (ss->head + 1) % ss->dur;
200 void steadystate_check(void)
202 int i, j, ddir, prev_groupid, group_ramp_time_over = 0;
203 unsigned long rate_time;
204 struct thread_data *td, *td2;
206 uint64_t group_bw = 0, group_iops = 0;
207 uint64_t td_iops, td_bytes;
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;
240 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
241 td_iops += td->io_blocks[ddir];
242 td_bytes += td->io_bytes[ddir];
246 rate_time = mtime_since(&ss->prev_time, &now);
247 memcpy(&ss->prev_time, &now, sizeof(now));
250 * Begin monitoring when job starts but don't actually use
251 * data in checking stopping criterion until ss->ramp_time is
252 * over. This ensures that we will have a sane value in
253 * prev_iops/bw the first time through after ss->ramp_time
256 if (ss->state & FIO_SS_RAMP_OVER) {
257 group_bw += 1000 * (td_bytes - ss->prev_bytes) / rate_time;
258 group_iops += 1000 * (td_iops - ss->prev_iops) / rate_time;
259 ++group_ramp_time_over;
261 ss->prev_iops = td_iops;
262 ss->prev_bytes = td_bytes;
264 if (td->o.group_reporting && !(ss->state & FIO_SS_DATA))
268 * Don't begin checking criterion until ss->ramp_time is over
269 * for at least one thread in group
271 if (!group_ramp_time_over)
274 dprint(FD_STEADYSTATE, "steadystate_check() thread: %d, "
275 "groupid: %u, rate_msec: %ld, "
276 "iops: %llu, bw: %llu, head: %d, tail: %d\n",
277 i, td->groupid, rate_time,
278 (unsigned long long) group_iops,
279 (unsigned long long) group_bw,
282 if (ss->state & FIO_SS_SLOPE)
283 ret = steadystate_slope(group_iops, group_bw, td);
285 ret = steadystate_deviation(group_iops, group_bw, td);
288 if (td->o.group_reporting) {
289 for_each_td(td2, j) {
290 if (td2->groupid == td->groupid) {
291 td2->ss.state |= FIO_SS_ATTAINED;
292 fio_mark_td_terminate(td2);
296 ss->state |= FIO_SS_ATTAINED;
297 fio_mark_td_terminate(td);
303 int td_steadystate_init(struct thread_data *td)
305 struct steadystate_data *ss = &td->ss;
306 struct thread_options *o = &td->o;
307 struct thread_data *td2;
310 memset(ss, 0, sizeof(*ss));
313 steadystate_enabled = true;
314 o->ss_dur /= 1000000L;
316 /* put all steady state info in one place */
318 ss->limit = o->ss_limit.u.f;
319 ss->ramp_time = o->ss_ramp_time;
321 ss->state = o->ss_state;
322 if (!td->ss.ramp_time)
323 ss->state |= FIO_SS_RAMP_OVER;
325 ss->sum_x = o->ss_dur * (o->ss_dur - 1) / 2;
326 ss->sum_x_sq = (o->ss_dur - 1) * (o->ss_dur) * (2*o->ss_dur - 1) / 6;
329 /* make sure that ss options are consistent within reporting group */
330 for_each_td(td2, j) {
331 if (td2->groupid == td->groupid) {
332 struct steadystate_data *ss2 = &td2->ss;
334 if (ss2->dur != ss->dur ||
335 ss2->limit != ss->limit ||
336 ss2->ramp_time != ss->ramp_time ||
337 ss2->state != ss->state ||
338 ss2->sum_x != ss->sum_x ||
339 ss2->sum_x_sq != ss->sum_x_sq) {
340 td_verror(td, EINVAL, "job rejected: steadystate options must be consistent within reporting groups");
349 uint64_t steadystate_bw_mean(struct thread_stat *ts)
354 for (i = 0, sum = 0; i < ts->ss_dur; i++)
355 sum += ts->ss_bw_data[i];
357 return sum / ts->ss_dur;
360 uint64_t steadystate_iops_mean(struct thread_stat *ts)
365 for (i = 0, sum = 0; i < ts->ss_dur; i++)
366 sum += ts->ss_iops_data[i];
368 return sum / ts->ss_dur;