steadystate.c

   1 #include <stdlib.h>
   2
   3 #include "fio.h"
   4 #include "steadystate.h"
   5 #include "helper_thread.h"
   6
   7 bool steadystate_enabled = false;
   8
   9 static void steadystate_alloc(struct thread_data *td)
  10 {
  11         int i;
  12
  13         td->ss.bw_data = malloc(td->ss.dur * sizeof(uint64_t));
  14         td->ss.iops_data = malloc(td->ss.dur * sizeof(uint64_t));
  15         /* initialize so that it is obvious if the cache is not full in the output */
  16         for (i = 0; i < td->ss.dur; i++)
  17                 td->ss.iops_data[i] = td->ss.bw_data[i] = 0;
  18
  19         td->ss.state |= __FIO_SS_DATA;
  20 }
  21
  22 void steadystate_setup(void)
  23 {
  24         int i, prev_groupid;
  25         struct thread_data *td, *prev_td;
  26
  27         if (!steadystate_enabled)
  28                 return;
  29
  30         /*
  31          * if group reporting is enabled, identify the last td
  32          * for each group and use it for storing steady state
  33          * data
  34          */
  35         prev_groupid = -1;
  36         prev_td = NULL;
  37         for_each_td(td, i) {
  38                 if (!td->ss.dur)
  39                         continue;
  40
  41                 if (!td->o.group_reporting) {
  42                         steadystate_alloc(td);
  43                         continue;
  44                 }
  45
  46                 if (prev_groupid != td->groupid) {
  47                         if (prev_td != NULL) {
  48                                 steadystate_alloc(prev_td);
  49                         }
  50                         prev_groupid = td->groupid;
  51                 }
  52                 prev_td = td;
  53         }
  54
  55         if (prev_td != NULL && prev_td->o.group_reporting) {
  56                 steadystate_alloc(prev_td);
  57         }
  58 }
  59
  60 static bool steadystate_slope(unsigned long iops, unsigned long bw,
  61                               struct thread_data *td)
  62 {
  63         int i, j;
  64         double result;
  65         struct steadystate_data *ss = &td->ss;
  66         unsigned long new_val;
  67
  68         ss->bw_data[ss->tail] = bw;
  69         ss->iops_data[ss->tail] = iops;
  70
  71         if (ss->state & __FIO_SS_IOPS)
  72                 new_val = iops;
  73         else
  74                 new_val = bw;
  75
  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];
  82                                 else
  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];
  87                                 else
  88                                         ss->sum_xy += i * ss->bw_data[j];
  89                         }
  90                         ss->state |= __FIO_SS_BUFFER_FULL;
  91                 } else {                /* easy to update the sums */
  92                         ss->sum_y -= ss->oldest_y;
  93                         ss->sum_y += new_val;
  94                         ss->sum_xy = ss->sum_xy - ss->sum_y + ss->dur * new_val;
  95                 }
  96
  97                 if (ss->state & __FIO_SS_IOPS)
  98                         ss->oldest_y = ss->iops_data[ss->head];
  99                 else
 100                         ss->oldest_y = ss->bw_data[ss->head];
 101
 102                 /*
 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.
 107                  */
 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);
 112                 else
 113                         ss->criterion = ss->slope;
 114
 115                 dprint(FD_STEADYSTATE, "sum_y: %llu, sum_xy: %llu, slope: %f, "
 116                                         "criterion: %f, limit: %f\n",
 117                                         ss->sum_y, ss->sum_xy, ss->slope,
 118                                         ss->criterion, ss->limit);
 119
 120                 result = ss->criterion * (ss->criterion < 0.0 ? -1.0 : 1.0);
 121                 if (result < ss->limit)
 122                         return true;
 123         }
 124
 125         ss->tail = (ss->tail + 1) % ss->dur;
 126         if (ss->tail <= ss->head)
 127                 ss->head = (ss->head + 1) % ss->dur;
 128
 129         return false;
 130 }
 131
 132 static bool steadystate_deviation(unsigned long iops, unsigned long bw,
 133                                   struct thread_data *td)
 134 {
 135         int i;
 136         double diff;
 137         double mean;
 138
 139         struct steadystate_data *ss = &td->ss;
 140
 141         ss->bw_data[ss->tail] = bw;
 142         ss->iops_data[ss->tail] = iops;
 143
 144         if (ss->state & __FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
 145                 if (!(ss->state & __FIO_SS_BUFFER_FULL)) {
 146                         /* first time through */
 147                         for(i = 0, ss->sum_y = 0; i < ss->dur; i++)
 148                                 if (ss->state & __FIO_SS_IOPS)
 149                                         ss->sum_y += ss->iops_data[i];
 150                                 else
 151                                         ss->sum_y += ss->bw_data[i];
 152                         ss->state |= __FIO_SS_BUFFER_FULL;
 153                 } else {                /* easy to update the sum */
 154                         ss->sum_y -= ss->oldest_y;
 155                         if (ss->state & __FIO_SS_IOPS)
 156                                 ss->sum_y += ss->iops_data[ss->tail];
 157                         else
 158                                 ss->sum_y += ss->bw_data[ss->tail];
 159                 }
 160
 161                 if (ss->state & __FIO_SS_IOPS)
 162                         ss->oldest_y = ss->iops_data[ss->head];
 163                 else
 164                         ss->oldest_y = ss->bw_data[ss->head];
 165
 166                 mean = (double) ss->sum_y / ss->dur;
 167                 ss->deviation = 0.0;
 168
 169                 for (i = 0; i < ss->dur; i++) {
 170                         if (ss->state & __FIO_SS_IOPS)
 171                                 diff = ss->iops_data[i] - mean;
 172                         else
 173                                 diff = ss->bw_data[i] - mean;
 174                         ss->deviation = max(ss->deviation, diff * (diff < 0.0 ? -1.0 : 1.0));
 175                 }
 176
 177                 if (ss->state & __FIO_SS_PCT)
 178                         ss->criterion = 100.0 * ss->deviation / mean;
 179                 else
 180                         ss->criterion = ss->deviation;
 181
 182                 dprint(FD_STEADYSTATE, "sum_y: %llu, mean: %f, max diff: %f, "
 183                                         "objective: %f, limit: %f\n",
 184                                         ss->sum_y, mean, ss->deviation,
 185                                         ss->criterion, ss->limit);
 186
 187                 if (ss->criterion < ss->limit)
 188                         return true;
 189         }
 190
 191         ss->tail = (ss->tail + 1) % ss->dur;
 192         if (ss->tail <= ss->head)
 193                 ss->head = (ss->head + 1) % ss->dur;
 194
 195         return false;
 196 }
 197
 198 void steadystate_check(void)
 199 {
 200         int i, j, ddir, prev_groupid, group_ramp_time_over = 0;
 201         unsigned long rate_time;
 202         struct thread_data *td, *td2;
 203         struct timeval now;
 204         unsigned long group_bw = 0, group_iops = 0;
 205         unsigned long long td_iops;
 206         unsigned long long td_bytes;
 207         bool ret;
 208
 209         prev_groupid = -1;
 210         for_each_td(td, i) {
 211                 struct steadystate_data *ss = &td->ss;
 212
 213                 if (!ss->dur || td->runstate <= TD_SETTING_UP ||
 214                     td->runstate >= TD_EXITED || (ss->state & __FIO_SS_ATTAINED))
 215                         continue;
 216
 217                 td_iops = 0;
 218                 td_bytes = 0;
 219                 if (!td->o.group_reporting ||
 220                     (td->o.group_reporting && td->groupid != prev_groupid)) {
 221                         group_bw = 0;
 222                         group_iops = 0;
 223                         group_ramp_time_over = 0;
 224                 }
 225                 prev_groupid = td->groupid;
 226
 227                 fio_gettime(&now, NULL);
 228                 if (ss->ramp_time && !(ss->state & __FIO_SS_RAMP_OVER)) {
 229                         /*
 230                          * Begin recording data one second after ss->ramp_time
 231                          * has elapsed
 232                          */
 233                         if (utime_since(&td->epoch, &now) >= (ss->ramp_time + 1000000L))
 234                                 ss->state |= __FIO_SS_RAMP_OVER;
 235                 }
 236
 237                 td_io_u_lock(td);
 238                 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
 239                         td_iops += td->io_blocks[ddir];
 240                         td_bytes += td->io_bytes[ddir];
 241                 }
 242                 td_io_u_unlock(td);
 243
 244                 rate_time = mtime_since(&ss->prev_time, &now);
 245                 memcpy(&ss->prev_time, &now, sizeof(now));
 246
 247                 /*
 248                  * Begin monitoring when job starts but don't actually use
 249                  * data in checking stopping criterion until ss->ramp_time is
 250                  * over. This ensures that we will have a sane value in
 251                  * prev_iops/bw the first time through after ss->ramp_time
 252                  * is done.
 253                  */
 254                 if (ss->state & __FIO_SS_RAMP_OVER) {
 255                         group_bw += 1000 * (td_bytes - ss->prev_bytes) / rate_time;
 256                         group_iops += 1000 * (td_iops - ss->prev_iops) / rate_time;
 257                         ++group_ramp_time_over;
 258                 }
 259                 ss->prev_iops = td_iops;
 260                 ss->prev_bytes = td_bytes;
 261
 262                 if (td->o.group_reporting && !(ss->state & __FIO_SS_DATA))
 263                         continue;
 264
 265                 /*
 266                  * Don't begin checking criterion until ss->ramp_time is over
 267                  * for at least one thread in group
 268                  */
 269                 if (!group_ramp_time_over)
 270                         continue;
 271
 272                 dprint(FD_STEADYSTATE, "steadystate_check() thread: %d, "
 273                                         "groupid: %u, rate_msec: %ld, "
 274                                         "iops: %lu, bw: %lu, head: %d, tail: %d\n",
 275                                         i, td->groupid, rate_time, group_iops,
 276                                         group_bw, ss->head, ss->tail);
 277
 278                 if (td->o.ss & __FIO_SS_SLOPE)
 279                         ret = steadystate_slope(group_iops, group_bw, td);
 280                 else
 281                         ret = steadystate_deviation(group_iops, group_bw, td);
 282
 283                 if (ret) {
 284                         if (td->o.group_reporting) {
 285                                 for_each_td(td2, j) {
 286                                         if (td2->groupid == td->groupid) {
 287                                                 td2->ss.state |= __FIO_SS_ATTAINED;
 288                                                 fio_mark_td_terminate(td2);
 289                                         }
 290                                 }
 291                         } else {
 292                                 ss->state |= __FIO_SS_ATTAINED;
 293                                 fio_mark_td_terminate(td);
 294                         }
 295                 }
 296         }
 297 }
 298
 299 int td_steadystate_init(struct thread_data *td)
 300 {
 301         struct steadystate_data *ss = &td->ss;
 302         struct thread_options *o = &td->o;
 303         struct thread_data *td2;
 304         int j;
 305
 306         memset(ss, 0, sizeof(*ss));
 307
 308         if (o->ss_dur) {
 309                 steadystate_enabled = true;
 310                 o->ss_dur /= 1000000L;
 311
 312                 /* put all steady state info in one place */
 313                 ss->dur = o->ss_dur;
 314                 ss->limit = o->ss_limit.u.f;
 315                 ss->ramp_time = o->ss_ramp_time;
 316
 317                 ss->state = o->ss;
 318                 if (!td->ss.ramp_time)
 319                         ss->state |= __FIO_SS_RAMP_OVER;
 320
 321                 ss->sum_x = o->ss_dur * (o->ss_dur - 1) / 2;
 322                 ss->sum_x_sq = (o->ss_dur - 1) * (o->ss_dur) * (2*o->ss_dur - 1) / 6;
 323         }
 324
 325         /* make sure that ss options are consistent within reporting group */
 326         for_each_td(td2, j) {
 327                 if (td2->groupid == td->groupid) {
 328                         struct steadystate_data *ss2 = &td2->ss;
 329
 330                         if (ss2->dur != ss->dur ||
 331                             ss2->limit != ss->limit ||
 332                             ss2->ramp_time != ss->ramp_time ||
 333                             ss2->state != ss->state ||
 334                             ss2->sum_x != ss->sum_x ||
 335                             ss2->sum_x_sq != ss->sum_x_sq) {
 336                                 td_verror(td, EINVAL, "job rejected: steadystate options must be consistent within reporting groups");
 337                                 return 1;
 338                         }
 339                 }
 340         }
 341
 342         return 0;
 343 }
 344
 345 unsigned long long steadystate_bw_mean(struct thread_stat *ts)
 346 {
 347         int i;
 348         unsigned long long sum;
 349
 350         for (i = 0, sum = 0; i < ts->ss_dur; i++)
 351                 sum += ts->ss_bw_data[i];
 352
 353         return sum / ts->ss_dur;
 354 }
 355
 356 unsigned long long steadystate_iops_mean(struct thread_stat *ts)
 357 {
 358         int i;
 359         unsigned long long sum;
 360
 361         for (i = 0, sum = 0; i < ts->ss_dur; i++)
 362                 sum += ts->ss_iops_data[i];
 363
 364         return sum / ts->ss_dur;
 365 }