eta.c

   1 /*
   2  * Status and ETA code
   3  */
   4 #include <unistd.h>
   5 #include <fcntl.h>
   6 #include <string.h>
   7
   8 #include "fio.h"
   9
  10 static char __run_str[REAL_MAX_JOBS + 1];
  11 static char run_str[__THREAD_RUNSTR_SZ(REAL_MAX_JOBS)];
  12
  13 static void update_condensed_str(char *run_str, char *run_str_condensed)
  14 {
  15         int i, ci, last, nr;
  16         size_t len;
  17
  18         len = strlen(run_str);
  19         if (!len)
  20                 return;
  21
  22         last = 0;
  23         nr = 0;
  24         ci = 0;
  25         for (i = 0; i < len; i++) {
  26                 if (!last) {
  27 new:
  28                         run_str_condensed[ci] = run_str[i];
  29                         last = run_str[i];
  30                         nr = 1;
  31                         ci++;
  32                 } else if (last == run_str[i]) {
  33                         nr++;
  34                 } else {
  35                         ci += sprintf(&run_str_condensed[ci], "(%u),", nr);
  36                         goto new;
  37                 }
  38         }
  39
  40         if (nr)
  41                 ci += sprintf(&run_str_condensed[ci], "(%u)", nr);
  42
  43         run_str_condensed[ci + 1] = '\0';
  44 }
  45
  46 /*
  47  * Sets the status of the 'td' in the printed status map.
  48  */
  49 static void check_str_update(struct thread_data *td)
  50 {
  51         char c = __run_str[td->thread_number - 1];
  52
  53         switch (td->runstate) {
  54         case TD_REAPED:
  55                 if (td->error)
  56                         c = 'X';
  57                 else if (td->sig)
  58                         c = 'K';
  59                 else
  60                         c = '_';
  61                 break;
  62         case TD_EXITED:
  63                 c = 'E';
  64                 break;
  65         case TD_RAMP:
  66                 c = '/';
  67                 break;
  68         case TD_RUNNING:
  69                 if (td_rw(td)) {
  70                         if (td_random(td)) {
  71                                 if (td->o.rwmix[DDIR_READ] == 100)
  72                                         c = 'r';
  73                                 else if (td->o.rwmix[DDIR_WRITE] == 100)
  74                                         c = 'w';
  75                                 else
  76                                         c = 'm';
  77                         } else {
  78                                 if (td->o.rwmix[DDIR_READ] == 100)
  79                                         c = 'R';
  80                                 else if (td->o.rwmix[DDIR_WRITE] == 100)
  81                                         c = 'W';
  82                                 else
  83                                         c = 'M';
  84                         }
  85                 } else if (td_read(td)) {
  86                         if (td_random(td))
  87                                 c = 'r';
  88                         else
  89                                 c = 'R';
  90                 } else if (td_write(td)) {
  91                         if (td_random(td))
  92                                 c = 'w';
  93                         else
  94                                 c = 'W';
  95                 } else {
  96                         if (td_random(td))
  97                                 c = 'd';
  98                         else
  99                                 c = 'D';
 100                 }
 101                 break;
 102         case TD_PRE_READING:
 103                 c = 'p';
 104                 break;
 105         case TD_VERIFYING:
 106                 c = 'V';
 107                 break;
 108         case TD_FSYNCING:
 109                 c = 'F';
 110                 break;
 111         case TD_FINISHING:
 112                 c = 'f';
 113                 break;
 114         case TD_CREATED:
 115                 c = 'C';
 116                 break;
 117         case TD_INITIALIZED:
 118         case TD_SETTING_UP:
 119                 c = 'I';
 120                 break;
 121         case TD_NOT_CREATED:
 122                 c = 'P';
 123                 break;
 124         default:
 125                 log_err("state %d\n", td->runstate);
 126         }
 127
 128         __run_str[td->thread_number - 1] = c;
 129         update_condensed_str(__run_str, run_str);
 130 }
 131
 132 /*
 133  * Convert seconds to a printable string.
 134  */
 135 void eta_to_str(char *str, unsigned long eta_sec)
 136 {
 137         unsigned int d, h, m, s;
 138         int disp_hour = 0;
 139
 140         s = eta_sec % 60;
 141         eta_sec /= 60;
 142         m = eta_sec % 60;
 143         eta_sec /= 60;
 144         h = eta_sec % 24;
 145         eta_sec /= 24;
 146         d = eta_sec;
 147
 148         if (d) {
 149                 disp_hour = 1;
 150                 str += sprintf(str, "%02ud:", d);
 151         }
 152
 153         if (h || disp_hour)
 154                 str += sprintf(str, "%02uh:", h);
 155
 156         str += sprintf(str, "%02um:", m);
 157         str += sprintf(str, "%02us", s);
 158 }
 159
 160 /*
 161  * Best effort calculation of the estimated pending runtime of a job.
 162  */
 163 static int thread_eta(struct thread_data *td)
 164 {
 165         unsigned long long bytes_total, bytes_done;
 166         unsigned long eta_sec = 0;
 167         unsigned long elapsed;
 168         uint64_t timeout;
 169
 170         elapsed = (mtime_since_now(&td->epoch) + 999) / 1000;
 171         timeout = td->o.timeout / 1000000UL;
 172
 173         bytes_total = td->total_io_size;
 174
 175         if (td->o.fill_device && td->o.size  == -1ULL) {
 176                 if (!td->fill_device_size || td->fill_device_size == -1ULL)
 177                         return 0;
 178
 179                 bytes_total = td->fill_device_size;
 180         }
 181
 182         if (td->o.zone_size && td->o.zone_skip && bytes_total) {
 183                 unsigned int nr_zones;
 184                 uint64_t zone_bytes;
 185
 186                 zone_bytes = bytes_total + td->o.zone_size + td->o.zone_skip;
 187                 nr_zones = (zone_bytes - 1) / (td->o.zone_size + td->o.zone_skip);
 188                 bytes_total -= nr_zones * td->o.zone_skip;
 189         }
 190
 191         /*
 192          * if writing and verifying afterwards, bytes_total will be twice the
 193          * size. In a mixed workload, verify phase will be the size of the
 194          * first stage writes.
 195          */
 196         if (td->o.do_verify && td->o.verify && td_write(td)) {
 197                 if (td_rw(td)) {
 198                         unsigned int perc = 50;
 199
 200                         if (td->o.rwmix[DDIR_WRITE])
 201                                 perc = td->o.rwmix[DDIR_WRITE];
 202
 203                         bytes_total += (bytes_total * perc) / 100;
 204                 } else
 205                         bytes_total <<= 1;
 206         }
 207
 208         if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
 209                 double perc, perc_t;
 210
 211                 bytes_done = ddir_rw_sum(td->io_bytes);
 212
 213                 if (bytes_total) {
 214                         perc = (double) bytes_done / (double) bytes_total;
 215                         if (perc > 1.0)
 216                                 perc = 1.0;
 217                 } else
 218                         perc = 0.0;
 219
 220                 if (td->o.time_based) {
 221                         if (timeout) {
 222                                 perc_t = (double) elapsed / (double) timeout;
 223                                 if (perc_t < perc)
 224                                         perc = perc_t;
 225                         } else {
 226                                 /*
 227                                  * Will never hit, we can't have time_based
 228                                  * without a timeout set.
 229                                  */
 230                                 perc = 0.0;
 231                         }
 232                 }
 233
 234                 eta_sec = (unsigned long) (elapsed * (1.0 / perc)) - elapsed;
 235
 236                 if (td->o.timeout &&
 237                     eta_sec > (timeout + done_secs - elapsed))
 238                         eta_sec = timeout + done_secs - elapsed;
 239         } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
 240                         || td->runstate == TD_INITIALIZED
 241                         || td->runstate == TD_SETTING_UP
 242                         || td->runstate == TD_RAMP
 243                         || td->runstate == TD_PRE_READING) {
 244                 int t_eta = 0, r_eta = 0;
 245                 unsigned long long rate_bytes;
 246
 247                 /*
 248                  * We can only guess - assume it'll run the full timeout
 249                  * if given, otherwise assume it'll run at the specified rate.
 250                  */
 251                 if (td->o.timeout) {
 252                         uint64_t timeout = td->o.timeout;
 253                         uint64_t start_delay = td->o.start_delay;
 254                         uint64_t ramp_time = td->o.ramp_time;
 255
 256                         t_eta = timeout + start_delay + ramp_time;
 257                         t_eta /= 1000000ULL;
 258
 259                         if (in_ramp_time(td)) {
 260                                 unsigned long ramp_left;
 261
 262                                 ramp_left = mtime_since_now(&td->epoch);
 263                                 ramp_left = (ramp_left + 999) / 1000;
 264                                 if (ramp_left <= t_eta)
 265                                         t_eta -= ramp_left;
 266                         }
 267                 }
 268                 rate_bytes = ddir_rw_sum(td->o.rate);
 269                 if (rate_bytes) {
 270                         r_eta = (bytes_total / 1024) / rate_bytes;
 271                         r_eta += (td->o.start_delay / 1000000ULL);
 272                 }
 273
 274                 if (r_eta && t_eta)
 275                         eta_sec = min(r_eta, t_eta);
 276                 else if (r_eta)
 277                         eta_sec = r_eta;
 278                 else if (t_eta)
 279                         eta_sec = t_eta;
 280                 else
 281                         eta_sec = 0;
 282         } else {
 283                 /*
 284                  * thread is already done or waiting for fsync
 285                  */
 286                 eta_sec = 0;
 287         }
 288
 289         return eta_sec;
 290 }
 291
 292 static void calc_rate(int unified_rw_rep, unsigned long mtime,
 293                       unsigned long long *io_bytes,
 294                       unsigned long long *prev_io_bytes, unsigned int *rate)
 295 {
 296         int i;
 297
 298         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 299                 unsigned long long diff;
 300
 301                 diff = io_bytes[i] - prev_io_bytes[i];
 302                 if (unified_rw_rep) {
 303                         rate[i] = 0;
 304                         rate[0] += ((1000 * diff) / mtime) / 1024;
 305                 } else
 306                         rate[i] = ((1000 * diff) / mtime) / 1024;
 307
 308                 prev_io_bytes[i] = io_bytes[i];
 309         }
 310 }
 311
 312 static void calc_iops(int unified_rw_rep, unsigned long mtime,
 313                       unsigned long long *io_iops,
 314                       unsigned long long *prev_io_iops, unsigned int *iops)
 315 {
 316         int i;
 317
 318         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 319                 unsigned long long diff;
 320
 321                 diff = io_iops[i] - prev_io_iops[i];
 322                 if (unified_rw_rep) {
 323                         iops[i] = 0;
 324                         iops[0] += (diff * 1000) / mtime;
 325                 } else
 326                         iops[i] = (diff * 1000) / mtime;
 327
 328                 prev_io_iops[i] = io_iops[i];
 329         }
 330 }
 331
 332 /*
 333  * Print status of the jobs we know about. This includes rate estimates,
 334  * ETA, thread state, etc.
 335  */
 336 int calc_thread_status(struct jobs_eta *je, int force)
 337 {
 338         struct thread_data *td;
 339         int i, unified_rw_rep;
 340         unsigned long rate_time, disp_time, bw_avg_time, *eta_secs;
 341         unsigned long long io_bytes[DDIR_RWDIR_CNT];
 342         unsigned long long io_iops[DDIR_RWDIR_CNT];
 343         struct timeval now;
 344
 345         static unsigned long long rate_io_bytes[DDIR_RWDIR_CNT];
 346         static unsigned long long disp_io_bytes[DDIR_RWDIR_CNT];
 347         static unsigned long long disp_io_iops[DDIR_RWDIR_CNT];
 348         static struct timeval rate_prev_time, disp_prev_time;
 349
 350         if (!force) {
 351                 if (output_format != FIO_OUTPUT_NORMAL &&
 352                     f_out == stdout)
 353                         return 0;
 354                 if (temp_stall_ts || eta_print == FIO_ETA_NEVER)
 355                         return 0;
 356
 357                 if (!isatty(STDOUT_FILENO) && (eta_print != FIO_ETA_ALWAYS))
 358                         return 0;
 359         }
 360
 361         if (!ddir_rw_sum(rate_io_bytes))
 362                 fill_start_time(&rate_prev_time);
 363         if (!ddir_rw_sum(disp_io_bytes))
 364                 fill_start_time(&disp_prev_time);
 365
 366         eta_secs = malloc(thread_number * sizeof(unsigned long));
 367         memset(eta_secs, 0, thread_number * sizeof(unsigned long));
 368
 369         je->elapsed_sec = (mtime_since_genesis() + 999) / 1000;
 370
 371         io_bytes[DDIR_READ] = io_bytes[DDIR_WRITE] = io_bytes[DDIR_TRIM] = 0;
 372         io_iops[DDIR_READ] = io_iops[DDIR_WRITE] = io_iops[DDIR_TRIM] = 0;
 373         bw_avg_time = ULONG_MAX;
 374         unified_rw_rep = 0;
 375         for_each_td(td, i) {
 376                 unified_rw_rep += td->o.unified_rw_rep;
 377                 if (is_power_of_2(td->o.kb_base))
 378                         je->is_pow2 = 1;
 379                 je->unit_base = td->o.unit_base;
 380                 if (td->o.bw_avg_time < bw_avg_time)
 381                         bw_avg_time = td->o.bw_avg_time;
 382                 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING
 383                     || td->runstate == TD_FSYNCING
 384                     || td->runstate == TD_PRE_READING
 385                     || td->runstate == TD_FINISHING) {
 386                         je->nr_running++;
 387                         if (td_read(td)) {
 388                                 je->t_rate[0] += td->o.rate[DDIR_READ];
 389                                 je->t_iops[0] += td->o.rate_iops[DDIR_READ];
 390                                 je->m_rate[0] += td->o.ratemin[DDIR_READ];
 391                                 je->m_iops[0] += td->o.rate_iops_min[DDIR_READ];
 392                         }
 393                         if (td_write(td)) {
 394                                 je->t_rate[1] += td->o.rate[DDIR_WRITE];
 395                                 je->t_iops[1] += td->o.rate_iops[DDIR_WRITE];
 396                                 je->m_rate[1] += td->o.ratemin[DDIR_WRITE];
 397                                 je->m_iops[1] += td->o.rate_iops_min[DDIR_WRITE];
 398                         }
 399                         if (td_trim(td)) {
 400                                 je->t_rate[2] += td->o.rate[DDIR_TRIM];
 401                                 je->t_iops[2] += td->o.rate_iops[DDIR_TRIM];
 402                                 je->m_rate[2] += td->o.ratemin[DDIR_TRIM];
 403                                 je->m_iops[2] += td->o.rate_iops_min[DDIR_TRIM];
 404                         }
 405
 406                         je->files_open += td->nr_open_files;
 407                 } else if (td->runstate == TD_RAMP) {
 408                         je->nr_running++;
 409                         je->nr_ramp++;
 410                 } else if (td->runstate == TD_SETTING_UP) {
 411                         je->nr_running++;
 412                         je->nr_setting_up++;
 413                 } else if (td->runstate < TD_RUNNING)
 414                         je->nr_pending++;
 415
 416                 if (je->elapsed_sec >= 3)
 417                         eta_secs[i] = thread_eta(td);
 418                 else
 419                         eta_secs[i] = INT_MAX;
 420
 421                 check_str_update(td);
 422
 423                 if (td->runstate > TD_SETTING_UP) {
 424                         int ddir;
 425
 426                         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
 427                                 if (unified_rw_rep) {
 428                                         io_bytes[0] += td->io_bytes[ddir];
 429                                         io_iops[0] += td->io_blocks[ddir];
 430                                 } else {
 431                                         io_bytes[ddir] += td->io_bytes[ddir];
 432                                         io_iops[ddir] += td->io_blocks[ddir];
 433                                 }
 434                         }
 435                 }
 436         }
 437
 438         if (exitall_on_terminate)
 439                 je->eta_sec = INT_MAX;
 440         else
 441                 je->eta_sec = 0;
 442
 443         for_each_td(td, i) {
 444                 if (exitall_on_terminate) {
 445                         if (eta_secs[i] < je->eta_sec)
 446                                 je->eta_sec = eta_secs[i];
 447                 } else {
 448                         if (eta_secs[i] > je->eta_sec)
 449                                 je->eta_sec = eta_secs[i];
 450                 }
 451         }
 452
 453         free(eta_secs);
 454
 455         fio_gettime(&now, NULL);
 456         rate_time = mtime_since(&rate_prev_time, &now);
 457
 458         if (write_bw_log && rate_time > bw_avg_time && !in_ramp_time(td)) {
 459                 calc_rate(unified_rw_rep, rate_time, io_bytes, rate_io_bytes,
 460                                 je->rate);
 461                 memcpy(&rate_prev_time, &now, sizeof(now));
 462                 add_agg_sample(je->rate[DDIR_READ], DDIR_READ, 0);
 463                 add_agg_sample(je->rate[DDIR_WRITE], DDIR_WRITE, 0);
 464                 add_agg_sample(je->rate[DDIR_TRIM], DDIR_TRIM, 0);
 465         }
 466
 467         disp_time = mtime_since(&disp_prev_time, &now);
 468
 469         /*
 470          * Allow a little slack, the target is to print it every 1000 msecs
 471          */
 472         if (!force && disp_time < 900)
 473                 return 0;
 474
 475         calc_rate(unified_rw_rep, disp_time, io_bytes, disp_io_bytes, je->rate);
 476         calc_iops(unified_rw_rep, disp_time, io_iops, disp_io_iops, je->iops);
 477
 478         memcpy(&disp_prev_time, &now, sizeof(now));
 479
 480         if (!force && !je->nr_running && !je->nr_pending)
 481                 return 0;
 482
 483         je->nr_threads = thread_number;
 484         update_condensed_str(__run_str, run_str);
 485         memcpy(je->run_str, run_str, strlen(run_str));
 486         return 1;
 487 }
 488
 489 void display_thread_status(struct jobs_eta *je)
 490 {
 491         static struct timeval disp_eta_new_line;
 492         static int eta_new_line_init, eta_new_line_pending;
 493         static int linelen_last;
 494         static int eta_good;
 495         char output[REAL_MAX_JOBS + 512], *p = output;
 496         char eta_str[128];
 497         double perc = 0.0;
 498
 499         if (je->eta_sec != INT_MAX && je->elapsed_sec) {
 500                 perc = (double) je->elapsed_sec / (double) (je->elapsed_sec + je->eta_sec);
 501                 eta_to_str(eta_str, je->eta_sec);
 502         }
 503
 504         if (eta_new_line_pending) {
 505                 eta_new_line_pending = 0;
 506                 p += sprintf(p, "\n");
 507         }
 508
 509         p += sprintf(p, "Jobs: %d (f=%d)", je->nr_running, je->files_open);
 510         if (je->m_rate[0] || je->m_rate[1] || je->t_rate[0] || je->t_rate[1]) {
 511                 char *tr, *mr;
 512
 513                 mr = num2str(je->m_rate[0] + je->m_rate[1], 4, 0, je->is_pow2, 8);
 514                 tr = num2str(je->t_rate[0] + je->t_rate[1], 4, 0, je->is_pow2, 8);
 515                 p += sprintf(p, ", CR=%s/%s KB/s", tr, mr);
 516                 free(tr);
 517                 free(mr);
 518         } else if (je->m_iops[0] || je->m_iops[1] || je->t_iops[0] || je->t_iops[1]) {
 519                 p += sprintf(p, ", CR=%d/%d IOPS",
 520                                         je->t_iops[0] + je->t_iops[1],
 521                                         je->m_iops[0] + je->m_iops[1]);
 522         }
 523         if (je->eta_sec != INT_MAX && je->nr_running) {
 524                 char perc_str[32];
 525                 char *iops_str[DDIR_RWDIR_CNT];
 526                 char *rate_str[DDIR_RWDIR_CNT];
 527                 size_t left;
 528                 int l;
 529                 int ddir;
 530
 531                 if ((!je->eta_sec && !eta_good) || je->nr_ramp == je->nr_running)
 532                         strcpy(perc_str, "-.-% done");
 533                 else {
 534                         double mult = 100.0;
 535
 536                         if (je->nr_setting_up && je->nr_running)
 537                                 mult *= (1.0 - (double) je->nr_setting_up / (double) je->nr_running);
 538
 539                         eta_good = 1;
 540                         perc *= mult;
 541                         sprintf(perc_str, "%3.1f%% done", perc);
 542                 }
 543
 544                 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
 545                         rate_str[ddir] = num2str(je->rate[ddir], 5,
 546                                                 1024, je->is_pow2, je->unit_base);
 547                         iops_str[ddir] = num2str(je->iops[ddir], 4, 1, 0, 0);
 548                 }
 549
 550                 left = sizeof(output) - (p - output) - 1;
 551
 552                 l = snprintf(p, left, ": [%s] [%s] [%s/%s/%s /s] [%s/%s/%s iops] [eta %s]",
 553                                 je->run_str, perc_str, rate_str[DDIR_READ],
 554                                 rate_str[DDIR_WRITE], rate_str[DDIR_TRIM],
 555                                 iops_str[DDIR_READ], iops_str[DDIR_WRITE],
 556                                 iops_str[DDIR_TRIM], eta_str);
 557                 p += l;
 558                 if (l >= 0 && l < linelen_last)
 559                         p += sprintf(p, "%*s", linelen_last - l, "");
 560                 linelen_last = l;
 561
 562                 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
 563                         free(rate_str[ddir]);
 564                         free(iops_str[ddir]);
 565                 }
 566         }
 567         p += sprintf(p, "\r");
 568
 569         printf("%s", output);
 570
 571         if (!eta_new_line_init) {
 572                 fio_gettime(&disp_eta_new_line, NULL);
 573                 eta_new_line_init = 1;
 574         } else if (eta_new_line &&
 575                    mtime_since_now(&disp_eta_new_line) > eta_new_line * 1000) {
 576                 fio_gettime(&disp_eta_new_line, NULL);
 577                 eta_new_line_pending = 1;
 578         }
 579
 580         fflush(stdout);
 581 }
 582
 583 struct jobs_eta *get_jobs_eta(int force, size_t *size)
 584 {
 585         struct jobs_eta *je;
 586
 587         if (!thread_number)
 588                 return NULL;
 589
 590         *size = sizeof(*je) + THREAD_RUNSTR_SZ;
 591         je = malloc(*size);
 592         memset(je, 0, *size);
 593
 594         if (!calc_thread_status(je, 0)) {
 595                 free(je);
 596                 return NULL;
 597         }
 598
 599         return je;
 600 }
 601
 602 void print_thread_status(void)
 603 {
 604         struct jobs_eta *je;
 605         size_t size;
 606
 607         je = get_jobs_eta(0, &size);
 608         if (je)
 609                 display_thread_status(je);
 610
 611         free(je);
 612 }
 613
 614 void print_status_init(int thr_number)
 615 {
 616         __run_str[thr_number] = 'P';
 617         update_condensed_str(__run_str, run_str);
 618 }