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