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 void steadystate_free(struct thread_data *td)
  10 {
  11         free(td->ss.iops_data);
  12         free(td->ss.bw_data);
  13         td->ss.iops_data = NULL;
  14         td->ss.bw_data = NULL;
  15 }
  16
  17 static void steadystate_alloc(struct thread_data *td)
  18 {
  19         td->ss.bw_data = calloc(td->ss.dur, sizeof(uint64_t));
  20         td->ss.iops_data = calloc(td->ss.dur, sizeof(uint64_t));
  21
  22         td->ss.state |= FIO_SS_DATA;
  23 }
  24
  25 void steadystate_setup(void)
  26 {
  27         int i, prev_groupid;
  28         struct thread_data *td, *prev_td;
  29
  30         if (!steadystate_enabled)
  31                 return;
  32
  33         /*
  34          * if group reporting is enabled, identify the last td
  35          * for each group and use it for storing steady state
  36          * data
  37          */
  38         prev_groupid = -1;
  39         prev_td = NULL;
  40         for_each_td(td, i) {
  41                 if (!td->ss.dur)
  42                         continue;
  43
  44                 if (!td->o.group_reporting) {
  45                         steadystate_alloc(td);
  46                         continue;
  47                 }
  48
  49                 if (prev_groupid != td->groupid) {
  50                         if (prev_td != NULL) {
  51                                 steadystate_alloc(prev_td);
  52                         }
  53                         prev_groupid = td->groupid;
  54                 }
  55                 prev_td = td;
  56         }
  57
  58         if (prev_td != NULL && prev_td->o.group_reporting) {
  59                 steadystate_alloc(prev_td);
  60         }
  61 }
  62
  63 static bool steadystate_slope(uint64_t iops, uint64_t bw,
  64                               struct thread_data *td)
  65 {
  66         int i, j;
  67         double result;
  68         struct steadystate_data *ss = &td->ss;
  69         uint64_t new_val;
  70
  71         ss->bw_data[ss->tail] = bw;
  72         ss->iops_data[ss->tail] = iops;
  73
  74         if (ss->state & FIO_SS_IOPS)
  75                 new_val = iops;
  76         else
  77                 new_val = bw;
  78
  79         if (ss->state & FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
  80                 if (!(ss->state & FIO_SS_BUFFER_FULL)) {
  81                         /* first time through */
  82                         for(i = 0, ss->sum_y = 0; i < ss->dur; i++) {
  83                                 if (ss->state & FIO_SS_IOPS)
  84                                         ss->sum_y += ss->iops_data[i];
  85                                 else
  86                                         ss->sum_y += ss->bw_data[i];
  87                                 j = (ss->head + i) % ss->dur;
  88                                 if (ss->state & FIO_SS_IOPS)
  89                                         ss->sum_xy += i * ss->iops_data[j];
  90                                 else
  91                                         ss->sum_xy += i * ss->bw_data[j];
  92                         }
  93                         ss->state |= FIO_SS_BUFFER_FULL;
  94                 } else {                /* easy to update the sums */
  95                         ss->sum_y -= ss->oldest_y;
  96                         ss->sum_y += new_val;
  97                         ss->sum_xy = ss->sum_xy - ss->sum_y + ss->dur * new_val;
  98                 }
  99
 100                 if (ss->state & FIO_SS_IOPS)
 101                         ss->oldest_y = ss->iops_data[ss->head];
 102                 else
 103                         ss->oldest_y = ss->bw_data[ss->head];
 104
 105                 /*
 106                  * calculate slope as (sum_xy - sum_x * sum_y / n) / (sum_(x^2)
 107                  * - (sum_x)^2 / n) This code assumes that all x values are
 108                  * equally spaced when they are often off by a few milliseconds.
 109                  * This assumption greatly simplifies the calculations.
 110                  */
 111                 ss->slope = (ss->sum_xy - (double) ss->sum_x * ss->sum_y / ss->dur) /
 112                                 (ss->sum_x_sq - (double) ss->sum_x * ss->sum_x / ss->dur);
 113                 if (ss->state & FIO_SS_PCT)
 114                         ss->criterion = 100.0 * ss->slope / (ss->sum_y / ss->dur);
 115                 else
 116                         ss->criterion = ss->slope;
 117
 118                 dprint(FD_STEADYSTATE, "sum_y: %llu, sum_xy: %llu, slope: %f, "
 119                                         "criterion: %f, limit: %f\n",
 120                                         (unsigned long long) ss->sum_y,
 121                                         (unsigned long long) ss->sum_xy,
 122                                         ss->slope, ss->criterion, ss->limit);
 123
 124                 result = ss->criterion * (ss->criterion < 0.0 ? -1.0 : 1.0);
 125                 if (result < ss->limit)
 126                         return true;
 127         }
 128
 129         ss->tail = (ss->tail + 1) % ss->dur;
 130         if (ss->tail <= ss->head)
 131                 ss->head = (ss->head + 1) % ss->dur;
 132
 133         return false;
 134 }
 135
 136 static bool steadystate_deviation(uint64_t iops, uint64_t bw,
 137                                   struct thread_data *td)
 138 {
 139         int i;
 140         double diff;
 141         double mean;
 142
 143         struct steadystate_data *ss = &td->ss;
 144
 145         ss->bw_data[ss->tail] = bw;
 146         ss->iops_data[ss->tail] = iops;
 147
 148         if (ss->state & FIO_SS_BUFFER_FULL || ss->tail - ss->head == ss->dur - 1) {
 149                 if (!(ss->state & FIO_SS_BUFFER_FULL)) {
 150                         /* first time through */
 151                         for(i = 0, ss->sum_y = 0; i < ss->dur; i++)
 152                                 if (ss->state & FIO_SS_IOPS)
 153                                         ss->sum_y += ss->iops_data[i];
 154                                 else
 155                                         ss->sum_y += ss->bw_data[i];
 156                         ss->state |= FIO_SS_BUFFER_FULL;
 157                 } else {                /* easy to update the sum */
 158                         ss->sum_y -= ss->oldest_y;
 159                         if (ss->state & FIO_SS_IOPS)
 160                                 ss->sum_y += ss->iops_data[ss->tail];
 161                         else
 162                                 ss->sum_y += ss->bw_data[ss->tail];
 163                 }
 164
 165                 if (ss->state & FIO_SS_IOPS)
 166                         ss->oldest_y = ss->iops_data[ss->head];
 167                 else
 168                         ss->oldest_y = ss->bw_data[ss->head];
 169
 170                 mean = (double) ss->sum_y / ss->dur;
 171                 ss->deviation = 0.0;
 172
 173                 for (i = 0; i < ss->dur; i++) {
 174                         if (ss->state & FIO_SS_IOPS)
 175                                 diff = ss->iops_data[i] - mean;
 176                         else
 177                                 diff = ss->bw_data[i] - mean;
 178                         ss->deviation = max(ss->deviation, diff * (diff < 0.0 ? -1.0 : 1.0));
 179                 }
 180
 181                 if (ss->state & FIO_SS_PCT)
 182                         ss->criterion = 100.0 * ss->deviation / mean;
 183                 else
 184                         ss->criterion = ss->deviation;
 185
 186                 dprint(FD_STEADYSTATE, "sum_y: %llu, mean: %f, max diff: %f, "
 187                                         "objective: %f, limit: %f\n",
 188                                         (unsigned long long) ss->sum_y, mean,
 189                                         ss->deviation, ss->criterion, ss->limit);
 190
 191                 if (ss->criterion < ss->limit)
 192                         return true;
 193         }
 194
 195         ss->tail = (ss->tail + 1) % ss->dur;
 196         if (ss->tail <= ss->head)
 197                 ss->head = (ss->head + 1) % ss->dur;
 198
 199         return false;
 200 }
 201
 202 void steadystate_check(void)
 203 {
 204         int i, j, ddir, prev_groupid, group_ramp_time_over = 0;
 205         unsigned long rate_time;
 206         struct thread_data *td, *td2;
 207         struct timespec now;
 208         uint64_t group_bw = 0, group_iops = 0;
 209         uint64_t td_iops, td_bytes;
 210         bool ret;
 211
 212         prev_groupid = -1;
 213         for_each_td(td, i) {
 214                 struct steadystate_data *ss = &td->ss;
 215
 216                 if (!ss->dur || td->runstate <= TD_SETTING_UP ||
 217                     td->runstate >= TD_EXITED || !ss->state ||
 218                     ss->state & FIO_SS_ATTAINED)
 219                         continue;
 220
 221                 td_iops = 0;
 222                 td_bytes = 0;
 223                 if (!td->o.group_reporting ||
 224                     (td->o.group_reporting && td->groupid != prev_groupid)) {
 225                         group_bw = 0;
 226                         group_iops = 0;
 227                         group_ramp_time_over = 0;
 228                 }
 229                 prev_groupid = td->groupid;
 230
 231                 fio_gettime(&now, NULL);
 232                 if (ss->ramp_time && !(ss->state & FIO_SS_RAMP_OVER)) {
 233                         /*
 234                          * Begin recording data one second after ss->ramp_time
 235                          * has elapsed
 236                          */
 237                         if (utime_since(&td->epoch, &now) >= (ss->ramp_time + 1000000L))
 238                                 ss->state |= FIO_SS_RAMP_OVER;
 239                 }
 240
 241                 td_io_u_lock(td);
 242                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
 243                         td_iops += td->io_blocks[ddir];
 244                         td_bytes += td->io_bytes[ddir];
 245                 }
 246                 td_io_u_unlock(td);
 247
 248                 rate_time = mtime_since(&ss->prev_time, &now);
 249                 memcpy(&ss->prev_time, &now, sizeof(now));
 250
 251                 /*
 252                  * Begin monitoring when job starts but don't actually use
 253                  * data in checking stopping criterion until ss->ramp_time is
 254                  * over. This ensures that we will have a sane value in
 255                  * prev_iops/bw the first time through after ss->ramp_time
 256                  * is done.
 257                  */
 258                 if (ss->state & FIO_SS_RAMP_OVER) {
 259                         group_bw += 1000 * (td_bytes - ss->prev_bytes) / rate_time;
 260                         group_iops += 1000 * (td_iops - ss->prev_iops) / rate_time;
 261                         ++group_ramp_time_over;
 262                 }
 263                 ss->prev_iops = td_iops;
 264                 ss->prev_bytes = td_bytes;
 265
 266                 if (td->o.group_reporting && !(ss->state & FIO_SS_DATA))
 267                         continue;
 268
 269                 /*
 270                  * Don't begin checking criterion until ss->ramp_time is over
 271                  * for at least one thread in group
 272                  */
 273                 if (!group_ramp_time_over)
 274                         continue;
 275
 276                 dprint(FD_STEADYSTATE, "steadystate_check() thread: %d, "
 277                                         "groupid: %u, rate_msec: %ld, "
 278                                         "iops: %llu, bw: %llu, head: %d, tail: %d\n",
 279                                         i, td->groupid, rate_time,
 280                                         (unsigned long long) group_iops,
 281                                         (unsigned long long) group_bw,
 282                                         ss->head, ss->tail);
 283
 284                 if (ss->state & FIO_SS_SLOPE)
 285                         ret = steadystate_slope(group_iops, group_bw, td);
 286                 else
 287                         ret = steadystate_deviation(group_iops, group_bw, td);
 288
 289                 if (ret) {
 290                         if (td->o.group_reporting) {
 291                                 for_each_td(td2, j) {
 292                                         if (td2->groupid == td->groupid) {
 293                                                 td2->ss.state |= FIO_SS_ATTAINED;
 294                                                 fio_mark_td_terminate(td2);
 295                                         }
 296                                 }
 297                         } else {
 298                                 ss->state |= FIO_SS_ATTAINED;
 299                                 fio_mark_td_terminate(td);
 300                         }
 301                 }
 302         }
 303 }
 304
 305 int td_steadystate_init(struct thread_data *td)
 306 {
 307         struct steadystate_data *ss = &td->ss;
 308         struct thread_options *o = &td->o;
 309         struct thread_data *td2;
 310         int j;
 311
 312         memset(ss, 0, sizeof(*ss));
 313
 314         if (o->ss_dur) {
 315                 steadystate_enabled = true;
 316                 o->ss_dur /= 1000000L;
 317
 318                 /* put all steady state info in one place */
 319                 ss->dur = o->ss_dur;
 320                 ss->limit = o->ss_limit.u.f;
 321                 ss->ramp_time = o->ss_ramp_time;
 322
 323                 ss->state = o->ss_state;
 324                 if (!td->ss.ramp_time)
 325                         ss->state |= FIO_SS_RAMP_OVER;
 326
 327                 ss->sum_x = o->ss_dur * (o->ss_dur - 1) / 2;
 328                 ss->sum_x_sq = (o->ss_dur - 1) * (o->ss_dur) * (2*o->ss_dur - 1) / 6;
 329         }
 330
 331         /* make sure that ss options are consistent within reporting group */
 332         for_each_td(td2, j) {
 333                 if (td2->groupid == td->groupid) {
 334                         struct steadystate_data *ss2 = &td2->ss;
 335
 336                         if (ss2->dur != ss->dur ||
 337                             ss2->limit != ss->limit ||
 338                             ss2->ramp_time != ss->ramp_time ||
 339                             ss2->state != ss->state ||
 340                             ss2->sum_x != ss->sum_x ||
 341                             ss2->sum_x_sq != ss->sum_x_sq) {
 342                                 td_verror(td, EINVAL, "job rejected: steadystate options must be consistent within reporting groups");
 343                                 return 1;
 344                         }
 345                 }
 346         }
 347
 348         return 0;
 349 }
 350
 351 uint64_t steadystate_bw_mean(struct thread_stat *ts)
 352 {
 353         int i;
 354         uint64_t sum;
 355
 356         for (i = 0, sum = 0; i < ts->ss_dur; i++)
 357                 sum += ts->ss_bw_data[i];
 358
 359         return sum / ts->ss_dur;
 360 }
 361
 362 uint64_t steadystate_iops_mean(struct thread_stat *ts)
 363 {
 364         int i;
 365         uint64_t sum;
 366
 367         for (i = 0, sum = 0; i < ts->ss_dur; i++)
 368                 sum += ts->ss_iops_data[i];
 369
 370         return sum / ts->ss_dur;
 371 }