stat.c revision 7fb28d3661a5833d8be24a014a04ee4548ec1c16
1#include <stdio.h> 2#include <string.h> 3#include <sys/time.h> 4#include <sys/types.h> 5#include <sys/stat.h> 6#include <dirent.h> 7#include <libgen.h> 8#include <math.h> 9 10#include "fio.h" 11#include "diskutil.h" 12#include "lib/ieee754.h" 13 14void update_rusage_stat(struct thread_data *td) 15{ 16 struct thread_stat *ts = &td->ts; 17 18 getrusage(RUSAGE_SELF, &td->ru_end); 19 20 ts->usr_time += mtime_since(&td->ru_start.ru_utime, 21 &td->ru_end.ru_utime); 22 ts->sys_time += mtime_since(&td->ru_start.ru_stime, 23 &td->ru_end.ru_stime); 24 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw 25 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw); 26 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt; 27 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt; 28 29 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end)); 30} 31 32/* 33 * Given a latency, return the index of the corresponding bucket in 34 * the structure tracking percentiles. 35 * 36 * (1) find the group (and error bits) that the value (latency) 37 * belongs to by looking at its MSB. (2) find the bucket number in the 38 * group by looking at the index bits. 39 * 40 */ 41static unsigned int plat_val_to_idx(unsigned int val) 42{ 43 unsigned int msb, error_bits, base, offset, idx; 44 45 /* Find MSB starting from bit 0 */ 46 if (val == 0) 47 msb = 0; 48 else 49 msb = (sizeof(val)*8) - __builtin_clz(val) - 1; 50 51 /* 52 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use 53 * all bits of the sample as index 54 */ 55 if (msb <= FIO_IO_U_PLAT_BITS) 56 return val; 57 58 /* Compute the number of error bits to discard*/ 59 error_bits = msb - FIO_IO_U_PLAT_BITS; 60 61 /* Compute the number of buckets before the group */ 62 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS; 63 64 /* 65 * Discard the error bits and apply the mask to find the 66 * index for the buckets in the group 67 */ 68 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits); 69 70 /* Make sure the index does not exceed (array size - 1) */ 71 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1)? 72 (base + offset) : (FIO_IO_U_PLAT_NR - 1); 73 74 return idx; 75} 76 77/* 78 * Convert the given index of the bucket array to the value 79 * represented by the bucket 80 */ 81static unsigned int plat_idx_to_val(unsigned int idx) 82{ 83 unsigned int error_bits, k, base; 84 85 assert(idx < FIO_IO_U_PLAT_NR); 86 87 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use 88 * all bits of the sample as index */ 89 if (idx < (FIO_IO_U_PLAT_VAL << 1) ) 90 return idx; 91 92 /* Find the group and compute the minimum value of that group */ 93 error_bits = (idx >> FIO_IO_U_PLAT_BITS) -1; 94 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS); 95 96 /* Find its bucket number of the group */ 97 k = idx % FIO_IO_U_PLAT_VAL; 98 99 /* Return the mean of the range of the bucket */ 100 return base + ((k + 0.5) * (1 << error_bits)); 101} 102 103static int double_cmp(const void *a, const void *b) 104{ 105 const fio_fp64_t fa = *(const fio_fp64_t *) a; 106 const fio_fp64_t fb = *(const fio_fp64_t *) b; 107 int cmp = 0; 108 109 if (fa.u.f > fb.u.f) 110 cmp = 1; 111 else if (fa.u.f < fb.u.f) 112 cmp = -1; 113 114 return cmp; 115} 116 117static unsigned int calc_clat_percentiles(unsigned int *io_u_plat, 118 unsigned long nr, fio_fp64_t *plist, 119 unsigned int **output, 120 unsigned int *maxv, 121 unsigned int *minv) 122{ 123 unsigned long sum = 0; 124 unsigned int len, i, j = 0; 125 unsigned int oval_len = 0; 126 unsigned int *ovals = NULL; 127 int is_last; 128 129 *minv = -1U; 130 *maxv = 0; 131 132 len = 0; 133 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0) 134 len++; 135 136 if (!len) 137 return 0; 138 139 /* 140 * Sort the percentile list. Note that it may already be sorted if 141 * we are using the default values, but since it's a short list this 142 * isn't a worry. Also note that this does not work for NaN values. 143 */ 144 if (len > 1) 145 qsort((void*)plist, len, sizeof(plist[0]), double_cmp); 146 147 /* 148 * Calculate bucket values, note down max and min values 149 */ 150 is_last = 0; 151 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) { 152 sum += io_u_plat[i]; 153 while (sum >= (plist[j].u.f / 100.0 * nr)) { 154 assert(plist[j].u.f <= 100.0); 155 156 if (j == oval_len) { 157 oval_len += 100; 158 ovals = realloc(ovals, oval_len * sizeof(unsigned int)); 159 } 160 161 ovals[j] = plat_idx_to_val(i); 162 if (ovals[j] < *minv) 163 *minv = ovals[j]; 164 if (ovals[j] > *maxv) 165 *maxv = ovals[j]; 166 167 is_last = (j == len - 1); 168 if (is_last) 169 break; 170 171 j++; 172 } 173 } 174 175 *output = ovals; 176 return len; 177} 178 179/* 180 * Find and display the p-th percentile of clat 181 */ 182static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, 183 fio_fp64_t *plist) 184{ 185 unsigned int len, j = 0, minv, maxv; 186 unsigned int *ovals; 187 int is_last, scale_down; 188 189 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv); 190 if (!len) 191 goto out; 192 193 /* 194 * We default to usecs, but if the value range is such that we 195 * should scale down to msecs, do that. 196 */ 197 if (minv > 2000 && maxv > 99999) { 198 scale_down = 1; 199 log_info(" clat percentiles (msec):\n |"); 200 } else { 201 scale_down = 0; 202 log_info(" clat percentiles (usec):\n |"); 203 } 204 205 for (j = 0; j < len; j++) { 206 char fbuf[8]; 207 208 /* for formatting */ 209 if (j != 0 && (j % 4) == 0) 210 log_info(" |"); 211 212 /* end of the list */ 213 is_last = (j == len - 1); 214 215 if (plist[j].u.f < 10.0) 216 sprintf(fbuf, " %2.2f", plist[j].u.f); 217 else 218 sprintf(fbuf, "%2.2f", plist[j].u.f); 219 220 if (scale_down) 221 ovals[j] = (ovals[j] + 999) / 1000; 222 223 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ','); 224 225 if (is_last) 226 break; 227 228 if (j % 4 == 3) /* for formatting */ 229 log_info("\n"); 230 } 231 232out: 233 if (ovals) 234 free(ovals); 235} 236 237static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, 238 double *mean, double *dev) 239{ 240 double n = is->samples; 241 242 if (is->samples == 0) 243 return 0; 244 245 *min = is->min_val; 246 *max = is->max_val; 247 248 n = (double) is->samples; 249 *mean = is->mean.u.f; 250 251 if (n > 1.0) 252 *dev = sqrt(is->S.u.f / (n - 1.0)); 253 else 254 *dev = 0; 255 256 return 1; 257} 258 259void show_group_stats(struct group_run_stats *rs) 260{ 261 char *p1, *p2, *p3, *p4; 262 const char *ddir_str[] = { " READ", " WRITE" }; 263 int i; 264 265 log_info("\nRun status group %d (all jobs):\n", rs->groupid); 266 267 for (i = 0; i <= DDIR_WRITE; i++) { 268 const int i2p = is_power_of_2(rs->kb_base); 269 270 if (!rs->max_run[i]) 271 continue; 272 273 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p); 274 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p); 275 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p); 276 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p); 277 278 log_info("%s: io=%sB, aggrb=%sB/s, minb=%sB/s, maxb=%sB/s," 279 " mint=%llumsec, maxt=%llumsec\n", ddir_str[i], p1, p2, 280 p3, p4, rs->min_run[i], 281 rs->max_run[i]); 282 283 free(p1); 284 free(p2); 285 free(p3); 286 free(p4); 287 } 288} 289 290#define ts_total_io_u(ts) \ 291 ((ts)->total_io_u[0] + (ts)->total_io_u[1]) 292 293static void stat_calc_dist(unsigned int *map, unsigned long total, 294 double *io_u_dist) 295{ 296 int i; 297 298 /* 299 * Do depth distribution calculations 300 */ 301 for (i = 0; i < FIO_IO_U_MAP_NR; i++) { 302 if (total) { 303 io_u_dist[i] = (double) map[i] / (double) total; 304 io_u_dist[i] *= 100.0; 305 if (io_u_dist[i] < 0.1 && map[i]) 306 io_u_dist[i] = 0.1; 307 } else 308 io_u_dist[i] = 0.0; 309 } 310} 311 312static void stat_calc_lat(struct thread_stat *ts, double *dst, 313 unsigned int *src, int nr) 314{ 315 unsigned long total = ts_total_io_u(ts); 316 int i; 317 318 /* 319 * Do latency distribution calculations 320 */ 321 for (i = 0; i < nr; i++) { 322 if (total) { 323 dst[i] = (double) src[i] / (double) total; 324 dst[i] *= 100.0; 325 if (dst[i] < 0.01 && src[i]) 326 dst[i] = 0.01; 327 } else 328 dst[i] = 0.0; 329 } 330} 331 332static void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat) 333{ 334 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR); 335} 336 337static void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat) 338{ 339 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR); 340} 341 342static int usec_to_msec(unsigned long *min, unsigned long *max, double *mean, 343 double *dev) 344{ 345 if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) { 346 *min /= 1000; 347 *max /= 1000; 348 *mean /= 1000.0; 349 *dev /= 1000.0; 350 return 0; 351 } 352 353 return 1; 354} 355 356static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts, 357 int ddir) 358{ 359 const char *ddir_str[] = { "read ", "write" }; 360 unsigned long min, max, runt; 361 unsigned long long bw, iops; 362 double mean, dev; 363 char *io_p, *bw_p, *iops_p; 364 int i2p; 365 366 assert(ddir_rw(ddir)); 367 368 if (!ts->runtime[ddir]) 369 return; 370 371 i2p = is_power_of_2(rs->kb_base); 372 runt = ts->runtime[ddir]; 373 374 bw = (1000 * ts->io_bytes[ddir]) / runt; 375 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p); 376 bw_p = num2str(bw, 6, 1, i2p); 377 378 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt; 379 iops_p = num2str(iops, 6, 1, 0); 380 381 log_info(" %s: io=%sB, bw=%sB/s, iops=%s, runt=%6llumsec\n", 382 ddir_str[ddir], io_p, bw_p, iops_p, 383 ts->runtime[ddir]); 384 385 free(io_p); 386 free(bw_p); 387 free(iops_p); 388 389 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) { 390 const char *base = "(usec)"; 391 char *minp, *maxp; 392 393 if (!usec_to_msec(&min, &max, &mean, &dev)) 394 base = "(msec)"; 395 396 minp = num2str(min, 6, 1, 0); 397 maxp = num2str(max, 6, 1, 0); 398 399 log_info(" slat %s: min=%s, max=%s, avg=%5.02f," 400 " stdev=%5.02f\n", base, minp, maxp, mean, dev); 401 402 free(minp); 403 free(maxp); 404 } 405 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) { 406 const char *base = "(usec)"; 407 char *minp, *maxp; 408 409 if (!usec_to_msec(&min, &max, &mean, &dev)) 410 base = "(msec)"; 411 412 minp = num2str(min, 6, 1, 0); 413 maxp = num2str(max, 6, 1, 0); 414 415 log_info(" clat %s: min=%s, max=%s, avg=%5.02f," 416 " stdev=%5.02f\n", base, minp, maxp, mean, dev); 417 418 free(minp); 419 free(maxp); 420 } 421 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) { 422 const char *base = "(usec)"; 423 char *minp, *maxp; 424 425 if (!usec_to_msec(&min, &max, &mean, &dev)) 426 base = "(msec)"; 427 428 minp = num2str(min, 6, 1, 0); 429 maxp = num2str(max, 6, 1, 0); 430 431 log_info(" lat %s: min=%s, max=%s, avg=%5.02f," 432 " stdev=%5.02f\n", base, minp, maxp, mean, dev); 433 434 free(minp); 435 free(maxp); 436 } 437 if (ts->clat_percentiles) { 438 show_clat_percentiles(ts->io_u_plat[ddir], 439 ts->clat_stat[ddir].samples, 440 ts->percentile_list); 441 } 442 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 443 double p_of_agg; 444 const char *bw_str = "KB"; 445 446 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 447 if (p_of_agg > 100.0) 448 p_of_agg = 100.0; 449 450 if (mean > 999999.9) { 451 min /= 1000.0; 452 max /= 1000.0; 453 mean /= 1000.0; 454 dev /= 1000.0; 455 bw_str = "MB"; 456 } 457 458 log_info(" bw (%s/s) : min=%5lu, max=%5lu, per=%3.2f%%," 459 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max, 460 p_of_agg, mean, dev); 461 } 462} 463 464static int show_lat(double *io_u_lat, int nr, const char **ranges, 465 const char *msg) 466{ 467 int new_line = 1, i, line = 0, shown = 0; 468 469 for (i = 0; i < nr; i++) { 470 if (io_u_lat[i] <= 0.0) 471 continue; 472 shown = 1; 473 if (new_line) { 474 if (line) 475 log_info("\n"); 476 log_info(" lat (%s) : ", msg); 477 new_line = 0; 478 line = 0; 479 } 480 if (line) 481 log_info(", "); 482 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]); 483 line++; 484 if (line == 5) 485 new_line = 1; 486 } 487 488 if (shown) 489 log_info("\n"); 490 491 return shown; 492} 493 494static void show_lat_u(double *io_u_lat_u) 495{ 496 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", 497 "250=", "500=", "750=", "1000=", }; 498 499 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec"); 500} 501 502static void show_lat_m(double *io_u_lat_m) 503{ 504 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", 505 "250=", "500=", "750=", "1000=", "2000=", 506 ">=2000=", }; 507 508 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec"); 509} 510 511static void show_latencies(double *io_u_lat_u, double *io_u_lat_m) 512{ 513 show_lat_u(io_u_lat_u); 514 show_lat_m(io_u_lat_m); 515} 516 517void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs) 518{ 519 double usr_cpu, sys_cpu; 520 unsigned long runtime; 521 double io_u_dist[FIO_IO_U_MAP_NR]; 522 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 523 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 524 525 if (!(ts->io_bytes[0] + ts->io_bytes[1]) && 526 !(ts->total_io_u[0] + ts->total_io_u[1])) 527 return; 528 529 if (!ts->error) { 530 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n", 531 ts->name, ts->groupid, ts->members, 532 ts->error, (int) ts->pid); 533 } else { 534 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n", 535 ts->name, ts->groupid, ts->members, 536 ts->error, ts->verror, (int) ts->pid); 537 } 538 539 if (strlen(ts->description)) 540 log_info(" Description : [%s]\n", ts->description); 541 542 if (ts->io_bytes[DDIR_READ]) 543 show_ddir_status(rs, ts, DDIR_READ); 544 if (ts->io_bytes[DDIR_WRITE]) 545 show_ddir_status(rs, ts, DDIR_WRITE); 546 547 stat_calc_lat_u(ts, io_u_lat_u); 548 stat_calc_lat_m(ts, io_u_lat_m); 549 show_latencies(io_u_lat_u, io_u_lat_m); 550 551 runtime = ts->total_run_time; 552 if (runtime) { 553 double runt = (double) runtime; 554 555 usr_cpu = (double) ts->usr_time * 100 / runt; 556 sys_cpu = (double) ts->sys_time * 100 / runt; 557 } else { 558 usr_cpu = 0; 559 sys_cpu = 0; 560 } 561 562 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu," 563 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf); 564 565 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 566 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%," 567 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 568 io_u_dist[1], io_u_dist[2], 569 io_u_dist[3], io_u_dist[4], 570 io_u_dist[5], io_u_dist[6]); 571 572 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist); 573 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 574 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 575 io_u_dist[1], io_u_dist[2], 576 io_u_dist[3], io_u_dist[4], 577 io_u_dist[5], io_u_dist[6]); 578 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist); 579 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 580 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 581 io_u_dist[1], io_u_dist[2], 582 io_u_dist[3], io_u_dist[4], 583 io_u_dist[5], io_u_dist[6]); 584 log_info(" issued : total=r=%lu/w=%lu/d=%lu," 585 " short=r=%lu/w=%lu/d=%lu\n", 586 ts->total_io_u[0], ts->total_io_u[1], 587 ts->total_io_u[2], 588 ts->short_io_u[0], ts->short_io_u[1], 589 ts->short_io_u[2]); 590 if (ts->continue_on_error) { 591 log_info(" errors : total=%lu, first_error=%d/<%s>\n", 592 ts->total_err_count, 593 ts->first_error, 594 strerror(ts->first_error)); 595 } 596} 597 598static void show_ddir_status_terse(struct thread_stat *ts, 599 struct group_run_stats *rs, int ddir) 600{ 601 unsigned long min, max; 602 unsigned long long bw, iops; 603 unsigned int *ovals = NULL; 604 double mean, dev; 605 unsigned int len, minv, maxv; 606 int i; 607 608 assert(ddir_rw(ddir)); 609 610 iops = bw = 0; 611 if (ts->runtime[ddir]) { 612 uint64_t runt = ts->runtime[ddir]; 613 614 bw = ts->io_bytes[ddir] / runt; 615 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt; 616 } 617 618 log_info(";%llu;%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, iops, 619 ts->runtime[ddir]); 620 621 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) 622 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 623 else 624 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 625 626 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) 627 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 628 else 629 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 630 631 if (ts->clat_percentiles) { 632 len = calc_clat_percentiles(ts->io_u_plat[ddir], 633 ts->clat_stat[ddir].samples, 634 ts->percentile_list, &ovals, &maxv, 635 &minv); 636 } else 637 len = 0; 638 639 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) { 640 if (i >= len) { 641 log_info(";0%%=0"); 642 continue; 643 } 644 log_info(";%2.2f%%=%u", ts->percentile_list[i].u.f, ovals[i]); 645 } 646 647 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) 648 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 649 else 650 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 651 652 if (ovals) 653 free(ovals); 654 655 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 656 double p_of_agg; 657 658 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 659 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev); 660 } else 661 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0); 662} 663 664static void show_thread_status_terse_v2(struct thread_stat *ts, 665 struct group_run_stats *rs) 666{ 667 double io_u_dist[FIO_IO_U_MAP_NR]; 668 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 669 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 670 double usr_cpu, sys_cpu; 671 int i; 672 673 /* General Info */ 674 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error); 675 /* Log Read Status */ 676 show_ddir_status_terse(ts, rs, 0); 677 /* Log Write Status */ 678 show_ddir_status_terse(ts, rs, 1); 679 680 /* CPU Usage */ 681 if (ts->total_run_time) { 682 double runt = (double) ts->total_run_time; 683 684 usr_cpu = (double) ts->usr_time * 100 / runt; 685 sys_cpu = (double) ts->sys_time * 100 / runt; 686 } else { 687 usr_cpu = 0; 688 sys_cpu = 0; 689 } 690 691 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf, 692 ts->minf); 693 694 /* Calc % distribution of IO depths, usecond, msecond latency */ 695 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 696 stat_calc_lat_u(ts, io_u_lat_u); 697 stat_calc_lat_m(ts, io_u_lat_m); 698 699 /* Only show fixed 7 I/O depth levels*/ 700 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 701 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], 702 io_u_dist[4], io_u_dist[5], io_u_dist[6]); 703 704 /* Microsecond latency */ 705 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) 706 log_info(";%3.2f%%", io_u_lat_u[i]); 707 /* Millisecond latency */ 708 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) 709 log_info(";%3.2f%%", io_u_lat_m[i]); 710 /* Additional output if continue_on_error set - default off*/ 711 if (ts->continue_on_error) 712 log_info(";%lu;%d", ts->total_err_count, ts->first_error); 713 log_info("\n"); 714 715 /* Additional output if description is set */ 716 if (ts->description) 717 log_info(";%s", ts->description); 718 719 log_info("\n"); 720} 721 722#define FIO_TERSE_VERSION "3" 723 724static void show_thread_status_terse_v3(struct thread_stat *ts, 725 struct group_run_stats *rs) 726{ 727 double io_u_dist[FIO_IO_U_MAP_NR]; 728 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 729 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 730 double usr_cpu, sys_cpu; 731 int i; 732 733 /* General Info */ 734 log_info("%s;%s;%s;%d;%d", FIO_TERSE_VERSION, fio_version_string, 735 ts->name, ts->groupid, ts->error); 736 /* Log Read Status */ 737 show_ddir_status_terse(ts, rs, 0); 738 /* Log Write Status */ 739 show_ddir_status_terse(ts, rs, 1); 740 741 /* CPU Usage */ 742 if (ts->total_run_time) { 743 double runt = (double) ts->total_run_time; 744 745 usr_cpu = (double) ts->usr_time * 100 / runt; 746 sys_cpu = (double) ts->sys_time * 100 / runt; 747 } else { 748 usr_cpu = 0; 749 sys_cpu = 0; 750 } 751 752 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf, 753 ts->minf); 754 755 /* Calc % distribution of IO depths, usecond, msecond latency */ 756 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 757 stat_calc_lat_u(ts, io_u_lat_u); 758 stat_calc_lat_m(ts, io_u_lat_m); 759 760 /* Only show fixed 7 I/O depth levels*/ 761 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 762 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], 763 io_u_dist[4], io_u_dist[5], io_u_dist[6]); 764 765 /* Microsecond latency */ 766 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) 767 log_info(";%3.2f%%", io_u_lat_u[i]); 768 /* Millisecond latency */ 769 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) 770 log_info(";%3.2f%%", io_u_lat_m[i]); 771 772 /* disk util stats, if any */ 773 show_disk_util(1); 774 775 /* Additional output if continue_on_error set - default off*/ 776 if (ts->continue_on_error) 777 log_info(";%lu;%d", ts->total_err_count, ts->first_error); 778 log_info("\n"); 779 780 /* Additional output if description is set */ 781 if (strlen(ts->description)) 782 log_info(";%s", ts->description); 783} 784 785static void show_thread_status_terse(struct thread_stat *ts, 786 struct group_run_stats *rs) 787{ 788 if (terse_version == 2) 789 show_thread_status_terse_v2(ts, rs); 790 else if (terse_version == 3) 791 show_thread_status_terse_v3(ts, rs); 792 else 793 log_err("fio: bad terse version!? %d\n", terse_version); 794} 795 796static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr) 797{ 798 double mean, S; 799 800 if (src->samples == 0) 801 return; 802 803 dst->min_val = min(dst->min_val, src->min_val); 804 dst->max_val = max(dst->max_val, src->max_val); 805 806 /* 807 * Compute new mean and S after the merge 808 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 809 * #Parallel_algorithm> 810 */ 811 if (nr == 1) { 812 mean = src->mean.u.f; 813 S = src->S.u.f; 814 } else { 815 double delta = src->mean.u.f - dst->mean.u.f; 816 817 mean = ((src->mean.u.f * src->samples) + 818 (dst->mean.u.f * dst->samples)) / 819 (dst->samples + src->samples); 820 821 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) * 822 (dst->samples * src->samples) / 823 (dst->samples + src->samples); 824 } 825 826 dst->samples += src->samples; 827 dst->mean.u.f = mean; 828 dst->S.u.f = S; 829} 830 831void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src) 832{ 833 int i; 834 835 for (i = 0; i < 2; i++) { 836 if (dst->max_run[i] < src->max_run[i]) 837 dst->max_run[i] = src->max_run[i]; 838 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i]) 839 dst->min_run[i] = src->min_run[i]; 840 if (dst->max_bw[i] < src->max_bw[i]) 841 dst->max_bw[i] = src->max_bw[i]; 842 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i]) 843 dst->min_bw[i] = src->min_bw[i]; 844 845 dst->io_kb[i] += src->io_kb[i]; 846 dst->agg[i] += src->agg[i]; 847 } 848 849} 850 851void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr) 852{ 853 int l, k; 854 855 for (l = 0; l <= DDIR_WRITE; l++) { 856 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr); 857 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr); 858 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr); 859 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr); 860 861 dst->io_bytes[l] += src->io_bytes[l]; 862 863 if (dst->runtime[l] < src->runtime[l]) 864 dst->runtime[l] = src->runtime[l]; 865 } 866 867 dst->usr_time += src->usr_time; 868 dst->sys_time += src->sys_time; 869 dst->ctx += src->ctx; 870 dst->majf += src->majf; 871 dst->minf += src->minf; 872 873 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 874 dst->io_u_map[k] += src->io_u_map[k]; 875 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 876 dst->io_u_submit[k] += src->io_u_submit[k]; 877 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 878 dst->io_u_complete[k] += src->io_u_complete[k]; 879 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++) 880 dst->io_u_lat_u[k] += src->io_u_lat_u[k]; 881 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++) 882 dst->io_u_lat_m[k] += src->io_u_lat_m[k]; 883 884 for (k = 0; k <= 2; k++) { 885 dst->total_io_u[k] += src->total_io_u[k]; 886 dst->short_io_u[k] += src->short_io_u[k]; 887 } 888 889 for (k = 0; k <= DDIR_WRITE; k++) { 890 int m; 891 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) 892 dst->io_u_plat[k][m] += src->io_u_plat[k][m]; 893 } 894 895 dst->total_run_time += src->total_run_time; 896 dst->total_submit += src->total_submit; 897 dst->total_complete += src->total_complete; 898} 899 900void init_group_run_stat(struct group_run_stats *gs) 901{ 902 memset(gs, 0, sizeof(*gs)); 903 gs->min_bw[0] = gs->min_run[0] = ~0UL; 904 gs->min_bw[1] = gs->min_run[1] = ~0UL; 905} 906 907void init_thread_stat(struct thread_stat *ts) 908{ 909 int j; 910 911 memset(ts, 0, sizeof(*ts)); 912 913 for (j = 0; j <= DDIR_WRITE; j++) { 914 ts->lat_stat[j].min_val = -1UL; 915 ts->clat_stat[j].min_val = -1UL; 916 ts->slat_stat[j].min_val = -1UL; 917 ts->bw_stat[j].min_val = -1UL; 918 } 919 ts->groupid = -1; 920} 921 922void show_run_stats(void) 923{ 924 struct group_run_stats *runstats, *rs; 925 struct thread_data *td; 926 struct thread_stat *threadstats, *ts; 927 int i, j, nr_ts, last_ts, idx; 928 int kb_base_warned = 0; 929 930 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1)); 931 932 for (i = 0; i < groupid + 1; i++) 933 init_group_run_stat(&runstats[i]); 934 935 /* 936 * find out how many threads stats we need. if group reporting isn't 937 * enabled, it's one-per-td. 938 */ 939 nr_ts = 0; 940 last_ts = -1; 941 for_each_td(td, i) { 942 if (!td->o.group_reporting) { 943 nr_ts++; 944 continue; 945 } 946 if (last_ts == td->groupid) 947 continue; 948 949 last_ts = td->groupid; 950 nr_ts++; 951 } 952 953 threadstats = malloc(nr_ts * sizeof(struct thread_stat)); 954 955 for (i = 0; i < nr_ts; i++) 956 init_thread_stat(&threadstats[i]); 957 958 j = 0; 959 last_ts = -1; 960 idx = 0; 961 for_each_td(td, i) { 962 if (idx && (!td->o.group_reporting || 963 (td->o.group_reporting && last_ts != td->groupid))) { 964 idx = 0; 965 j++; 966 } 967 968 last_ts = td->groupid; 969 970 ts = &threadstats[j]; 971 972 ts->clat_percentiles = td->o.clat_percentiles; 973 if (td->o.overwrite_plist) 974 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list)); 975 else 976 memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list)); 977 978 idx++; 979 ts->members++; 980 981 if (ts->groupid == -1) { 982 /* 983 * These are per-group shared already 984 */ 985 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE); 986 if (td->o.description) 987 strncpy(ts->description, td->o.description, 988 FIO_JOBNAME_SIZE); 989 else 990 memset(ts->description, 0, FIO_JOBNAME_SIZE); 991 992 ts->groupid = td->groupid; 993 994 /* 995 * first pid in group, not very useful... 996 */ 997 ts->pid = td->pid; 998 999 ts->kb_base = td->o.kb_base; 1000 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) { 1001 log_info("fio: kb_base differs for jobs in group, using" 1002 " %u as the base\n", ts->kb_base); 1003 kb_base_warned = 1; 1004 } 1005 1006 ts->continue_on_error = td->o.continue_on_error; 1007 ts->total_err_count += td->total_err_count; 1008 ts->first_error = td->first_error; 1009 if (!ts->error) { 1010 if (!td->error && td->o.continue_on_error && 1011 td->first_error) { 1012 ts->error = td->first_error; 1013 strcpy(ts->verror, td->verror); 1014 } else if (td->error) { 1015 ts->error = td->error; 1016 strcpy(ts->verror, td->verror); 1017 } 1018 } 1019 1020 sum_thread_stats(ts, &td->ts, idx); 1021 } 1022 1023 for (i = 0; i < nr_ts; i++) { 1024 unsigned long long bw; 1025 1026 ts = &threadstats[i]; 1027 rs = &runstats[ts->groupid]; 1028 rs->kb_base = ts->kb_base; 1029 1030 for (j = 0; j <= DDIR_WRITE; j++) { 1031 if (!ts->runtime[j]) 1032 continue; 1033 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j]) 1034 rs->min_run[j] = ts->runtime[j]; 1035 if (ts->runtime[j] > rs->max_run[j]) 1036 rs->max_run[j] = ts->runtime[j]; 1037 1038 bw = 0; 1039 if (ts->runtime[j]) { 1040 unsigned long runt; 1041 1042 runt = ts->runtime[j]; 1043 bw = ts->io_bytes[j] / runt; 1044 } 1045 if (bw < rs->min_bw[j]) 1046 rs->min_bw[j] = bw; 1047 if (bw > rs->max_bw[j]) 1048 rs->max_bw[j] = bw; 1049 1050 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base; 1051 } 1052 } 1053 1054 for (i = 0; i < groupid + 1; i++) { 1055 unsigned long max_run[2]; 1056 1057 rs = &runstats[i]; 1058 max_run[0] = rs->max_run[0]; 1059 max_run[1] = rs->max_run[1]; 1060 1061 if (rs->max_run[0]) 1062 rs->agg[0] = (rs->io_kb[0] * 1000) / max_run[0]; 1063 if (rs->max_run[1]) 1064 rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1]; 1065 } 1066 1067 /* 1068 * don't overwrite last signal output 1069 */ 1070 if (!terse_output) 1071 log_info("\n"); 1072 1073 for (i = 0; i < nr_ts; i++) { 1074 ts = &threadstats[i]; 1075 rs = &runstats[ts->groupid]; 1076 1077 if (is_backend) 1078 fio_server_send_ts(ts, rs); 1079 else if (terse_output) 1080 show_thread_status_terse(ts, rs); 1081 else 1082 show_thread_status(ts, rs); 1083 } 1084 1085 for (i = 0; i < groupid + 1; i++) { 1086 rs = &runstats[i]; 1087 1088 rs->groupid = i; 1089 if (is_backend) 1090 fio_server_send_gs(rs); 1091 else if (!terse_output) 1092 show_group_stats(rs); 1093 } 1094 1095 if (is_backend) 1096 fio_server_send_du(); 1097 else if (!terse_output) 1098 show_disk_util(0); 1099 1100 free_disk_util(); 1101 1102 free(runstats); 1103 free(threadstats); 1104} 1105 1106static inline void add_stat_sample(struct io_stat *is, unsigned long data) 1107{ 1108 double val = data; 1109 double delta; 1110 1111 if (data > is->max_val) 1112 is->max_val = data; 1113 if (data < is->min_val) 1114 is->min_val = data; 1115 1116 delta = val - is->mean.u.f; 1117 if (delta) { 1118 is->mean.u.f += delta / (is->samples + 1.0); 1119 is->S.u.f += delta * (val - is->mean.u.f); 1120 } 1121 1122 is->samples++; 1123} 1124 1125static void __add_log_sample(struct io_log *iolog, unsigned long val, 1126 enum fio_ddir ddir, unsigned int bs, 1127 unsigned long t) 1128{ 1129 const int nr_samples = iolog->nr_samples; 1130 1131 if (!iolog->nr_samples) 1132 iolog->avg_last = t; 1133 1134 if (iolog->nr_samples == iolog->max_samples) { 1135 int new_size = sizeof(struct io_sample) * iolog->max_samples*2; 1136 1137 iolog->log = realloc(iolog->log, new_size); 1138 iolog->max_samples <<= 1; 1139 } 1140 1141 iolog->log[nr_samples].val = val; 1142 iolog->log[nr_samples].time = t; 1143 iolog->log[nr_samples].ddir = ddir; 1144 iolog->log[nr_samples].bs = bs; 1145 iolog->nr_samples++; 1146} 1147 1148static inline void reset_io_stat(struct io_stat *ios) 1149{ 1150 ios->max_val = ios->min_val = ios->samples = 0; 1151 ios->mean.u.f = ios->S.u.f = 0; 1152} 1153 1154static void add_log_sample(struct thread_data *td, struct io_log *iolog, 1155 unsigned long val, enum fio_ddir ddir, 1156 unsigned int bs) 1157{ 1158 unsigned long elapsed, this_window; 1159 1160 if (!ddir_rw(ddir)) 1161 return; 1162 1163 elapsed = mtime_since_now(&td->epoch); 1164 1165 /* 1166 * If no time averaging, just add the log sample. 1167 */ 1168 if (!iolog->avg_msec) { 1169 __add_log_sample(iolog, val, ddir, bs, elapsed); 1170 return; 1171 } 1172 1173 /* 1174 * Add the sample. If the time period has passed, then 1175 * add that entry to the log and clear. 1176 */ 1177 add_stat_sample(&iolog->avg_window[ddir], val); 1178 1179 /* 1180 * If period hasn't passed, adding the above sample is all we 1181 * need to do. 1182 */ 1183 this_window = elapsed - iolog->avg_last; 1184 if (this_window < iolog->avg_msec) 1185 return; 1186 1187 /* 1188 * Note an entry in the log. Use the mean from the logged samples, 1189 * making sure to properly round up. Only write a log entry if we 1190 * had actual samples done. 1191 */ 1192 if (iolog->avg_window[DDIR_READ].samples) { 1193 unsigned long mr; 1194 1195 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50; 1196 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed); 1197 } 1198 if (iolog->avg_window[DDIR_WRITE].samples) { 1199 unsigned long mw; 1200 1201 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50; 1202 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed); 1203 } 1204 1205 reset_io_stat(&iolog->avg_window[DDIR_READ]); 1206 reset_io_stat(&iolog->avg_window[DDIR_WRITE]); 1207 iolog->avg_last = elapsed; 1208} 1209 1210void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs) 1211{ 1212 struct io_log *iolog; 1213 1214 if (!ddir_rw(ddir)) 1215 return; 1216 1217 iolog = agg_io_log[ddir]; 1218 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis()); 1219} 1220 1221static void add_clat_percentile_sample(struct thread_stat *ts, 1222 unsigned long usec, enum fio_ddir ddir) 1223{ 1224 unsigned int idx = plat_val_to_idx(usec); 1225 assert(idx < FIO_IO_U_PLAT_NR); 1226 1227 ts->io_u_plat[ddir][idx]++; 1228} 1229 1230void add_clat_sample(struct thread_data *td, enum fio_ddir ddir, 1231 unsigned long usec, unsigned int bs) 1232{ 1233 struct thread_stat *ts = &td->ts; 1234 1235 if (!ddir_rw(ddir)) 1236 return; 1237 1238 add_stat_sample(&ts->clat_stat[ddir], usec); 1239 1240 if (td->clat_log) 1241 add_log_sample(td, td->clat_log, usec, ddir, bs); 1242 1243 if (ts->clat_percentiles) 1244 add_clat_percentile_sample(ts, usec, ddir); 1245} 1246 1247void add_slat_sample(struct thread_data *td, enum fio_ddir ddir, 1248 unsigned long usec, unsigned int bs) 1249{ 1250 struct thread_stat *ts = &td->ts; 1251 1252 if (!ddir_rw(ddir)) 1253 return; 1254 1255 add_stat_sample(&ts->slat_stat[ddir], usec); 1256 1257 if (td->slat_log) 1258 add_log_sample(td, td->slat_log, usec, ddir, bs); 1259} 1260 1261void add_lat_sample(struct thread_data *td, enum fio_ddir ddir, 1262 unsigned long usec, unsigned int bs) 1263{ 1264 struct thread_stat *ts = &td->ts; 1265 1266 if (!ddir_rw(ddir)) 1267 return; 1268 1269 add_stat_sample(&ts->lat_stat[ddir], usec); 1270 1271 if (td->lat_log) 1272 add_log_sample(td, td->lat_log, usec, ddir, bs); 1273} 1274 1275void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs, 1276 struct timeval *t) 1277{ 1278 struct thread_stat *ts = &td->ts; 1279 unsigned long spent, rate; 1280 1281 if (!ddir_rw(ddir)) 1282 return; 1283 1284 spent = mtime_since(&td->bw_sample_time, t); 1285 if (spent < td->o.bw_avg_time) 1286 return; 1287 1288 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) * 1289 1000 / spent / 1024; 1290 add_stat_sample(&ts->bw_stat[ddir], rate); 1291 1292 if (td->bw_log) 1293 add_log_sample(td, td->bw_log, rate, ddir, bs); 1294 1295 fio_gettime(&td->bw_sample_time, NULL); 1296 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir]; 1297} 1298 1299void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, 1300 struct timeval *t) 1301{ 1302 struct thread_stat *ts = &td->ts; 1303 unsigned long spent, iops; 1304 1305 if (!ddir_rw(ddir)) 1306 return; 1307 1308 spent = mtime_since(&td->iops_sample_time, t); 1309 if (spent < td->o.iops_avg_time) 1310 return; 1311 1312 iops = ((td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]) * 1000) / spent; 1313 1314 add_stat_sample(&ts->iops_stat[ddir], iops); 1315 1316 if (td->iops_log) { 1317 assert(iops); 1318 add_log_sample(td, td->iops_log, iops, ddir, 0); 1319 } 1320 1321 fio_gettime(&td->iops_sample_time, NULL); 1322 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir]; 1323} 1324