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