steadystate: Use calloc(3)
[fio.git] / steadystate.c
index 1f53ae46a88b504c225144a13563635422c72657..43c715c96422dddfcbca6f8fed9871c80ee4ffd2 100644 (file)
@@ -8,13 +8,10 @@ bool steadystate_enabled = false;
 
 static void steadystate_alloc(struct thread_data *td)
 {
-       int i;
+       td->ss.bw_data = calloc(td->ss.dur, sizeof(uint64_t));
+       td->ss.iops_data = calloc(td->ss.dur, sizeof(uint64_t));
 
-       td->ss.bw_data = malloc(td->ss.dur * sizeof(unsigned long));
-       td->ss.iops_data = malloc(td->ss.dur * sizeof(unsigned long));
-       /* initialize so that it is obvious if the cache is not full in the output */
-       for (i = 0; i < td->ss.dur; i++)
-               td->ss.iops_data[i] = td->ss.bw_data[i] = 0;
+       td->ss.state |= __FIO_SS_DATA;
 }
 
 void steadystate_setup(void)
@@ -33,7 +30,7 @@ void steadystate_setup(void)
        prev_groupid = -1;
        prev_td = NULL;
        for_each_td(td, i) {
-               if (td->ts.ss == NULL)
+               if (!td->ss.dur)
                        continue;
 
                if (!td->o.group_reporting) {
@@ -43,7 +40,6 @@ void steadystate_setup(void)
 
                if (prev_groupid != td->groupid) {
                        if (prev_td != NULL) {
-                               prev_td->ss.state |= __FIO_SS_LAST;
                                steadystate_alloc(prev_td);
                        }
                        prev_groupid = td->groupid;
@@ -52,18 +48,17 @@ void steadystate_setup(void)
        }
 
        if (prev_td != NULL && prev_td->o.group_reporting) {
-               prev_td->ss.state |= __FIO_SS_LAST;
                steadystate_alloc(prev_td);
        }
 }
 
-static bool steadystate_slope(unsigned long iops, unsigned long bw,
+static bool steadystate_slope(uint64_t iops, uint64_t bw,
                              struct thread_data *td)
 {
        int i, j;
        double result;
        struct steadystate_data *ss = &td->ss;
-       unsigned long new_val;
+       uint64_t new_val;
 
        ss->bw_data[ss->tail] = bw;
        ss->iops_data[ss->tail] = iops;
@@ -73,21 +68,21 @@ static bool steadystate_slope(unsigned long iops, unsigned long bw,
        else
                new_val = bw;
 
-       if (ss->tail < ss->head || (ss->tail - ss->head == ss->dur - 1)) {
-               if (ss->sum_y == 0) {   /* first time through */
-                       for(i = 0; i < ss->dur; i++) {
+       if (ss->state & __FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
+               if (!(ss->state & __FIO_SS_BUFFER_FULL)) {
+                       /* first time through */
+                       for(i = 0, ss->sum_y = 0; i < ss->dur; i++) {
                                if (ss->state & __FIO_SS_IOPS)
                                        ss->sum_y += ss->iops_data[i];
                                else
                                        ss->sum_y += ss->bw_data[i];
-                               j = ss->head + i;
-                               if (j >= ss->dur)
-                                       j -= ss->dur;
+                               j = (ss->head + i) % ss->dur;
                                if (ss->state & __FIO_SS_IOPS)
                                        ss->sum_xy += i * ss->iops_data[j];
                                else
                                        ss->sum_xy += i * ss->bw_data[j];
                        }
+                       ss->state |= __FIO_SS_BUFFER_FULL;
                } else {                /* easy to update the sums */
                        ss->sum_y -= ss->oldest_y;
                        ss->sum_y += new_val;
@@ -107,15 +102,16 @@ static bool steadystate_slope(unsigned long iops, unsigned long bw,
                 */
                ss->slope = (ss->sum_xy - (double) ss->sum_x * ss->sum_y / ss->dur) /
                                (ss->sum_x_sq - (double) ss->sum_x * ss->sum_x / ss->dur);
-               if (ss->pct)
+               if (ss->state & __FIO_SS_PCT)
                        ss->criterion = 100.0 * ss->slope / (ss->sum_y / ss->dur);
                else
                        ss->criterion = ss->slope;
 
                dprint(FD_STEADYSTATE, "sum_y: %llu, sum_xy: %llu, slope: %f, "
                                        "criterion: %f, limit: %f\n",
-                                       ss->sum_y, ss->sum_xy, ss->slope,
-                                       ss->criterion, ss->limit);
+                                       (unsigned long long) ss->sum_y,
+                                       (unsigned long long) ss->sum_xy,
+                                       ss->slope, ss->criterion, ss->limit);
 
                result = ss->criterion * (ss->criterion < 0.0 ? -1.0 : 1.0);
                if (result < ss->limit)
@@ -129,7 +125,7 @@ static bool steadystate_slope(unsigned long iops, unsigned long bw,
        return false;
 }
 
-static bool steadystate_deviation(unsigned long iops, unsigned long bw,
+static bool steadystate_deviation(uint64_t iops, uint64_t bw,
                                  struct thread_data *td)
 {
        int i;
@@ -141,13 +137,15 @@ static bool steadystate_deviation(unsigned long iops, unsigned long bw,
        ss->bw_data[ss->tail] = bw;
        ss->iops_data[ss->tail] = iops;
 
-       if (ss->tail < ss->head || (ss->tail - ss->head == ss->dur - 1)) {
-               if (ss->sum_y == 0) {   /* first time through */
-                       for(i = 0; i < ss->dur; i++)
+       if (ss->state & __FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
+               if (!(ss->state & __FIO_SS_BUFFER_FULL)) {
+                       /* first time through */
+                       for(i = 0, ss->sum_y = 0; i < ss->dur; i++)
                                if (ss->state & __FIO_SS_IOPS)
                                        ss->sum_y += ss->iops_data[i];
                                else
                                        ss->sum_y += ss->bw_data[i];
+                       ss->state |= __FIO_SS_BUFFER_FULL;
                } else {                /* easy to update the sum */
                        ss->sum_y -= ss->oldest_y;
                        if (ss->state & __FIO_SS_IOPS)
@@ -172,15 +170,15 @@ static bool steadystate_deviation(unsigned long iops, unsigned long bw,
                        ss->deviation = max(ss->deviation, diff * (diff < 0.0 ? -1.0 : 1.0));
                }
 
-               if (ss->pct)
+               if (ss->state & __FIO_SS_PCT)
                        ss->criterion = 100.0 * ss->deviation / mean;
                else
                        ss->criterion = ss->deviation;
 
                dprint(FD_STEADYSTATE, "sum_y: %llu, mean: %f, max diff: %f, "
                                        "objective: %f, limit: %f\n",
-                                       ss->sum_y, mean, ss->deviation,
-                                       ss->criterion, ss->limit);
+                                       (unsigned long long) ss->sum_y, mean,
+                                       ss->deviation, ss->criterion, ss->limit);
 
                if (ss->criterion < ss->limit)
                        return true;
@@ -199,9 +197,8 @@ void steadystate_check(void)
        unsigned long rate_time;
        struct thread_data *td, *td2;
        struct timeval now;
-       unsigned long group_bw = 0, group_iops = 0;
-       unsigned long long td_iops;
-       unsigned long long td_bytes;
+       uint64_t group_bw = 0, group_iops = 0;
+       uint64_t td_iops, td_bytes;
        bool ret;
 
        prev_groupid = -1;
@@ -209,7 +206,8 @@ void steadystate_check(void)
                struct steadystate_data *ss = &td->ss;
 
                if (!ss->dur || td->runstate <= TD_SETTING_UP ||
-                   td->runstate >= TD_EXITED || (ss->state & __FIO_SS_ATTAINED))
+                   td->runstate >= TD_EXITED || !ss->state ||
+                   ss->state & __FIO_SS_ATTAINED)
                        continue;
 
                td_iops = 0;
@@ -257,7 +255,7 @@ void steadystate_check(void)
                ss->prev_iops = td_iops;
                ss->prev_bytes = td_bytes;
 
-               if (td->o.group_reporting && !(ss->state & __FIO_SS_LAST))
+               if (td->o.group_reporting && !(ss->state & __FIO_SS_DATA))
                        continue;
 
                /*
@@ -269,11 +267,13 @@ void steadystate_check(void)
 
                dprint(FD_STEADYSTATE, "steadystate_check() thread: %d, "
                                        "groupid: %u, rate_msec: %ld, "
-                                       "iops: %lu, bw: %lu, head: %d, tail: %d\n",
-                                       i, td->groupid, rate_time, group_iops,
-                                       group_bw, ss->head, ss->tail);
+                                       "iops: %llu, bw: %llu, head: %d, tail: %d\n",
+                                       i, td->groupid, rate_time,
+                                       (unsigned long long) group_iops,
+                                       (unsigned long long) group_bw,
+                                       ss->head, ss->tail);
 
-               if (td->o.ss & __FIO_SS_SLOPE)
+               if (ss->state & __FIO_SS_SLOPE)
                        ret = steadystate_slope(group_iops, group_bw, td);
                else
                        ret = steadystate_deviation(group_iops, group_bw, td);
@@ -294,31 +294,70 @@ void steadystate_check(void)
        }
 }
 
-void td_steadystate_init(struct thread_data *td)
+int td_steadystate_init(struct thread_data *td)
 {
        struct steadystate_data *ss = &td->ss;
        struct thread_options *o = &td->o;
+       struct thread_data *td2;
+       int j;
 
        memset(ss, 0, sizeof(*ss));
 
-       if (!o->ss_dur)
-               return;
+       if (o->ss_dur) {
+               steadystate_enabled = true;
+               o->ss_dur /= 1000000L;
+
+               /* put all steady state info in one place */
+               ss->dur = o->ss_dur;
+               ss->limit = o->ss_limit.u.f;
+               ss->ramp_time = o->ss_ramp_time;
+
+               ss->state = o->ss_state;
+               if (!td->ss.ramp_time)
+                       ss->state |= __FIO_SS_RAMP_OVER;
+
+               ss->sum_x = o->ss_dur * (o->ss_dur - 1) / 2;
+               ss->sum_x_sq = (o->ss_dur - 1) * (o->ss_dur) * (2*o->ss_dur - 1) / 6;
+       }
+
+       /* make sure that ss options are consistent within reporting group */
+       for_each_td(td2, j) {
+               if (td2->groupid == td->groupid) {
+                       struct steadystate_data *ss2 = &td2->ss;
+
+                       if (ss2->dur != ss->dur ||
+                           ss2->limit != ss->limit ||
+                           ss2->ramp_time != ss->ramp_time ||
+                           ss2->state != ss->state ||
+                           ss2->sum_x != ss->sum_x ||
+                           ss2->sum_x_sq != ss->sum_x_sq) {
+                               td_verror(td, EINVAL, "job rejected: steadystate options must be consistent within reporting groups");
+                               return 1;
+                       }
+               }
+       }
+
+       return 0;
+}
 
-       steadystate_enabled = true;
-       o->ss_dur /= 1000000L;
+uint64_t steadystate_bw_mean(struct thread_stat *ts)
+{
+       int i;
+       uint64_t sum;
 
-       /* put all steady state info in one place */
-       ss->dur = o->ss_dur;
-       ss->limit = o->ss_limit.u.f;
-       ss->ramp_time = o->ss_ramp_time;
-       ss->pct = o->ss_pct;
+       for (i = 0, sum = 0; i < ts->ss_dur; i++)
+               sum += ts->ss_bw_data[i];
 
-       ss->state = o->ss;
-       if (!td->ss.ramp_time)
-               ss->state |= __FIO_SS_RAMP_OVER;
+       return sum / ts->ss_dur;
+}
+
+uint64_t steadystate_iops_mean(struct thread_stat *ts)
+{
+       int i;
+       uint64_t sum;
 
-       ss->sum_x = o->ss_dur * (o->ss_dur - 1) / 2;
-       ss->sum_x_sq = (o->ss_dur - 1) * (o->ss_dur) * (2*o->ss_dur - 1) / 6;
+       for (i = 0, sum = 0; i < ts->ss_dur; i++)
+               sum += ts->ss_iops_data[i];
 
-       td->ts.ss = ss;
+       return sum / ts->ss_dur;
 }