stat.c revision f2f788dd732d97c2c3a5f5dd93223a7bfafcc410
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 "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("Run 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; 444 445 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 446 log_info(" bw (KB/s) : min=%5lu, max=%5lu, per=%3.2f%%," 447 " avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg, 448 mean, dev); 449 } 450} 451 452static void show_lat(double *io_u_lat, int nr, const char **ranges, 453 const char *msg) 454{ 455 int new_line = 1, i, line = 0; 456 457 for (i = 0; i < nr; i++) { 458 if (io_u_lat[i] <= 0.0) 459 continue; 460 if (new_line) { 461 if (line) 462 log_info("\n"); 463 log_info(" lat (%s): ", msg); 464 new_line = 0; 465 line = 0; 466 } 467 if (line) 468 log_info(", "); 469 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]); 470 line++; 471 if (line == 5) 472 new_line = 1; 473 } 474} 475 476static void show_lat_u(double *io_u_lat_u) 477{ 478 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", 479 "250=", "500=", "750=", "1000=", }; 480 481 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec"); 482} 483 484static void show_lat_m(double *io_u_lat_m) 485{ 486 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", 487 "250=", "500=", "750=", "1000=", "2000=", 488 ">=2000=", }; 489 490 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec"); 491} 492 493static void show_latencies(double *io_u_lat_u, double *io_u_lat_m) 494{ 495 show_lat_u(io_u_lat_u); 496 log_info("\n"); 497 show_lat_m(io_u_lat_m); 498 log_info("\n"); 499} 500 501void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs) 502{ 503 double usr_cpu, sys_cpu; 504 unsigned long runtime; 505 double io_u_dist[FIO_IO_U_MAP_NR]; 506 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 507 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 508 509 if (!(ts->io_bytes[0] + ts->io_bytes[1]) && 510 !(ts->total_io_u[0] + ts->total_io_u[1])) 511 return; 512 513 if (!ts->error) { 514 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n", 515 ts->name, ts->groupid, ts->members, 516 ts->error, (int) ts->pid); 517 } else { 518 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n", 519 ts->name, ts->groupid, ts->members, 520 ts->error, ts->verror, (int) ts->pid); 521 } 522 523 if (ts->description) 524 log_info(" Description : [%s]\n", ts->description); 525 526 if (ts->io_bytes[DDIR_READ]) 527 show_ddir_status(rs, ts, DDIR_READ); 528 if (ts->io_bytes[DDIR_WRITE]) 529 show_ddir_status(rs, ts, DDIR_WRITE); 530 531 runtime = ts->total_run_time; 532 if (runtime) { 533 double runt = (double) runtime; 534 535 usr_cpu = (double) ts->usr_time * 100 / runt; 536 sys_cpu = (double) ts->sys_time * 100 / runt; 537 } else { 538 usr_cpu = 0; 539 sys_cpu = 0; 540 } 541 542 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu," 543 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf); 544 545 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 546 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%," 547 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 548 io_u_dist[1], io_u_dist[2], 549 io_u_dist[3], io_u_dist[4], 550 io_u_dist[5], io_u_dist[6]); 551 552 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist); 553 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 554 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 555 io_u_dist[1], io_u_dist[2], 556 io_u_dist[3], io_u_dist[4], 557 io_u_dist[5], io_u_dist[6]); 558 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist); 559 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 560 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 561 io_u_dist[1], io_u_dist[2], 562 io_u_dist[3], io_u_dist[4], 563 io_u_dist[5], io_u_dist[6]); 564 log_info(" issued r/w/d: total=%lu/%lu/%lu, short=%lu/%lu/%lu\n", 565 ts->total_io_u[0], ts->total_io_u[1], 566 ts->total_io_u[2], 567 ts->short_io_u[0], ts->short_io_u[1], 568 ts->short_io_u[2]); 569 stat_calc_lat_u(ts, io_u_lat_u); 570 stat_calc_lat_m(ts, io_u_lat_m); 571 show_latencies(io_u_lat_u, io_u_lat_m); 572 if (ts->continue_on_error) { 573 log_info(" errors : total=%lu, first_error=%d/<%s>\n", 574 ts->total_err_count, 575 ts->first_error, 576 strerror(ts->first_error)); 577 } 578} 579 580static void show_ddir_status_terse(struct thread_stat *ts, 581 struct group_run_stats *rs, int ddir) 582{ 583 unsigned long min, max; 584 unsigned long long bw, iops; 585 unsigned int *ovals = NULL; 586 double mean, dev; 587 unsigned int len, minv, maxv; 588 int i; 589 590 assert(ddir_rw(ddir)); 591 592 iops = bw = 0; 593 if (ts->runtime[ddir]) { 594 uint64_t runt = ts->runtime[ddir]; 595 596 bw = ts->io_bytes[ddir] / runt; 597 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt; 598 } 599 600 log_info(";%llu;%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, iops, 601 ts->runtime[ddir]); 602 603 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) 604 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 605 else 606 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 607 608 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) 609 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 610 else 611 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 612 613 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) 614 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 615 else 616 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 617 618 if (ts->clat_percentiles) { 619 len = calc_clat_percentiles(ts->io_u_plat[ddir], 620 ts->clat_stat[ddir].samples, 621 ts->percentile_list, &ovals, &maxv, 622 &minv); 623 } else 624 len = 0; 625 626 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) { 627 if (i >= len) { 628 log_info(";0%%=0"); 629 continue; 630 } 631 log_info(";%2.2f%%=%u", ts->percentile_list[i].u.f, ovals[i]); 632 } 633 if (ovals) 634 free(ovals); 635 636 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 637 double p_of_agg; 638 639 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 640 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev); 641 } else 642 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0); 643} 644 645#define FIO_TERSE_VERSION "3" 646 647static void show_thread_status_terse(struct thread_stat *ts, 648 struct group_run_stats *rs) 649{ 650 double io_u_dist[FIO_IO_U_MAP_NR]; 651 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 652 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 653 double usr_cpu, sys_cpu; 654 int i; 655 656 /* General Info */ 657 log_info("%s;%s;%d;%d", FIO_TERSE_VERSION, ts->name, ts->groupid, 658 ts->error); 659 /* Log Read Status */ 660 show_ddir_status_terse(ts, rs, 0); 661 /* Log Write Status */ 662 show_ddir_status_terse(ts, rs, 1); 663 664 /* CPU Usage */ 665 if (ts->total_run_time) { 666 double runt = (double) ts->total_run_time; 667 668 usr_cpu = (double) ts->usr_time * 100 / runt; 669 sys_cpu = (double) ts->sys_time * 100 / runt; 670 } else { 671 usr_cpu = 0; 672 sys_cpu = 0; 673 } 674 675 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf, 676 ts->minf); 677 678 /* Calc % distribution of IO depths, usecond, msecond latency */ 679 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 680 stat_calc_lat_u(ts, io_u_lat_u); 681 stat_calc_lat_m(ts, io_u_lat_m); 682 683 /* Only show fixed 7 I/O depth levels*/ 684 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 685 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], 686 io_u_dist[4], io_u_dist[5], io_u_dist[6]); 687 688 /* Microsecond latency */ 689 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) 690 log_info(";%3.2f%%", io_u_lat_u[i]); 691 /* Millisecond latency */ 692 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) 693 log_info(";%3.2f%%", io_u_lat_m[i]); 694 695 /* disk util stats, if any */ 696 show_disk_util(1); 697 698 /* Additional output if continue_on_error set - default off*/ 699 if (ts->continue_on_error) 700 log_info(";%lu;%d", ts->total_err_count, ts->first_error); 701 log_info("\n"); 702 703 /* Additional output if description is set */ 704 if (ts->description) 705 log_info(";%s", ts->description); 706 707 log_info("\n"); 708} 709 710static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr) 711{ 712 double mean, S; 713 714 if (src->samples == 0) 715 return; 716 717 dst->min_val = min(dst->min_val, src->min_val); 718 dst->max_val = max(dst->max_val, src->max_val); 719 720 /* 721 * Compute new mean and S after the merge 722 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 723 * #Parallel_algorithm> 724 */ 725 if (nr == 1) { 726 mean = src->mean.u.f; 727 S = src->S.u.f; 728 } else { 729 double delta = src->mean.u.f - dst->mean.u.f; 730 731 mean = ((src->mean.u.f * src->samples) + 732 (dst->mean.u.f * dst->samples)) / 733 (dst->samples + src->samples); 734 735 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) * 736 (dst->samples * src->samples) / 737 (dst->samples + src->samples); 738 } 739 740 dst->samples += src->samples; 741 dst->mean.u.f = mean; 742 dst->S.u.f = S; 743} 744 745void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src) 746{ 747 int i; 748 749 for (i = 0; i < 2; i++) { 750 if (dst->max_run[i] < src->max_run[i]) 751 dst->max_run[i] = src->max_run[i]; 752 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i]) 753 dst->min_run[i] = src->min_run[i]; 754 if (dst->max_bw[i] < src->max_bw[i]) 755 dst->max_bw[i] = src->max_bw[i]; 756 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i]) 757 dst->min_bw[i] = src->min_bw[i]; 758 759 dst->io_kb[i] += src->io_kb[i]; 760 dst->agg[i] += src->agg[i]; 761 } 762 763} 764 765void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr) 766{ 767 int l, k; 768 769 for (l = 0; l <= DDIR_WRITE; l++) { 770 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr); 771 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr); 772 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr); 773 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr); 774 775 dst->io_bytes[l] += src->io_bytes[l]; 776 777 if (dst->runtime[l] < src->runtime[l]) 778 dst->runtime[l] = src->runtime[l]; 779 } 780 781 dst->usr_time += src->usr_time; 782 dst->sys_time += src->sys_time; 783 dst->ctx += src->ctx; 784 dst->majf += src->majf; 785 dst->minf += src->minf; 786 787 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 788 dst->io_u_map[k] += src->io_u_map[k]; 789 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 790 dst->io_u_submit[k] += src->io_u_submit[k]; 791 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 792 dst->io_u_complete[k] += src->io_u_complete[k]; 793 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++) 794 dst->io_u_lat_u[k] += src->io_u_lat_u[k]; 795 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++) 796 dst->io_u_lat_m[k] += src->io_u_lat_m[k]; 797 798 for (k = 0; k <= 2; k++) { 799 dst->total_io_u[k] += src->total_io_u[k]; 800 dst->short_io_u[k] += src->short_io_u[k]; 801 } 802 803 for (k = 0; k <= DDIR_WRITE; k++) { 804 int m; 805 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) 806 dst->io_u_plat[k][m] += src->io_u_plat[k][m]; 807 } 808 809 dst->total_run_time += src->total_run_time; 810 dst->total_submit += src->total_submit; 811 dst->total_complete += src->total_complete; 812} 813 814void init_group_run_stat(struct group_run_stats *gs) 815{ 816 memset(gs, 0, sizeof(*gs)); 817 gs->min_bw[0] = gs->min_run[0] = ~0UL; 818 gs->min_bw[1] = gs->min_run[1] = ~0UL; 819} 820 821void init_thread_stat(struct thread_stat *ts) 822{ 823 int j; 824 825 memset(ts, 0, sizeof(*ts)); 826 827 for (j = 0; j <= DDIR_WRITE; j++) { 828 ts->lat_stat[j].min_val = -1UL; 829 ts->clat_stat[j].min_val = -1UL; 830 ts->slat_stat[j].min_val = -1UL; 831 ts->bw_stat[j].min_val = -1UL; 832 } 833 ts->groupid = -1; 834} 835 836void show_run_stats(void) 837{ 838 struct group_run_stats *runstats, *rs; 839 struct thread_data *td; 840 struct thread_stat *threadstats, *ts; 841 int i, j, nr_ts, last_ts, idx; 842 int kb_base_warned = 0; 843 844 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1)); 845 846 for (i = 0; i < groupid + 1; i++) 847 init_group_run_stat(&runstats[i]); 848 849 /* 850 * find out how many threads stats we need. if group reporting isn't 851 * enabled, it's one-per-td. 852 */ 853 nr_ts = 0; 854 last_ts = -1; 855 for_each_td(td, i) { 856 if (!td->o.group_reporting) { 857 nr_ts++; 858 continue; 859 } 860 if (last_ts == td->groupid) 861 continue; 862 863 last_ts = td->groupid; 864 nr_ts++; 865 } 866 867 threadstats = malloc(nr_ts * sizeof(struct thread_stat)); 868 869 for (i = 0; i < nr_ts; i++) 870 init_thread_stat(&threadstats[i]); 871 872 j = 0; 873 last_ts = -1; 874 idx = 0; 875 for_each_td(td, i) { 876 if (idx && (!td->o.group_reporting || 877 (td->o.group_reporting && last_ts != td->groupid))) { 878 idx = 0; 879 j++; 880 } 881 882 last_ts = td->groupid; 883 884 ts = &threadstats[j]; 885 886 ts->clat_percentiles = td->o.clat_percentiles; 887 if (td->o.overwrite_plist) 888 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list)); 889 else 890 memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list)); 891 892 idx++; 893 ts->members++; 894 895 if (ts->groupid == -1) { 896 /* 897 * These are per-group shared already 898 */ 899 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE); 900 if (td->o.description) 901 strncpy(ts->description, td->o.description, 902 FIO_JOBNAME_SIZE); 903 else 904 memset(ts->description, 0, FIO_JOBNAME_SIZE); 905 906 ts->groupid = td->groupid; 907 908 /* 909 * first pid in group, not very useful... 910 */ 911 ts->pid = td->pid; 912 913 ts->kb_base = td->o.kb_base; 914 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) { 915 log_info("fio: kb_base differs for jobs in group, using" 916 " %u as the base\n", ts->kb_base); 917 kb_base_warned = 1; 918 } 919 920 ts->continue_on_error = td->o.continue_on_error; 921 ts->total_err_count += td->total_err_count; 922 ts->first_error = td->first_error; 923 if (!ts->error) { 924 if (!td->error && td->o.continue_on_error && 925 td->first_error) { 926 ts->error = td->first_error; 927 strcpy(ts->verror, td->verror); 928 } else if (td->error) { 929 ts->error = td->error; 930 strcpy(ts->verror, td->verror); 931 } 932 } 933 934 sum_thread_stats(ts, &td->ts, idx); 935 } 936 937 for (i = 0; i < nr_ts; i++) { 938 unsigned long long bw; 939 940 ts = &threadstats[i]; 941 rs = &runstats[ts->groupid]; 942 rs->kb_base = ts->kb_base; 943 944 for (j = 0; j <= DDIR_WRITE; j++) { 945 if (!ts->runtime[j]) 946 continue; 947 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j]) 948 rs->min_run[j] = ts->runtime[j]; 949 if (ts->runtime[j] > rs->max_run[j]) 950 rs->max_run[j] = ts->runtime[j]; 951 952 bw = 0; 953 if (ts->runtime[j]) { 954 unsigned long runt; 955 956 runt = ts->runtime[j]; 957 bw = ts->io_bytes[j] / runt; 958 } 959 if (bw < rs->min_bw[j]) 960 rs->min_bw[j] = bw; 961 if (bw > rs->max_bw[j]) 962 rs->max_bw[j] = bw; 963 964 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base; 965 } 966 } 967 968 for (i = 0; i < groupid + 1; i++) { 969 unsigned long max_run[2]; 970 971 rs = &runstats[i]; 972 max_run[0] = rs->max_run[0]; 973 max_run[1] = rs->max_run[1]; 974 975 if (rs->max_run[0]) 976 rs->agg[0] = (rs->io_kb[0] * 1000) / max_run[0]; 977 if (rs->max_run[1]) 978 rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1]; 979 } 980 981 /* 982 * don't overwrite last signal output 983 */ 984 if (!terse_output) 985 log_info("\n"); 986 987 for (i = 0; i < nr_ts; i++) { 988 ts = &threadstats[i]; 989 rs = &runstats[ts->groupid]; 990 991 if (is_backend) 992 fio_server_send_ts(ts, rs); 993 else if (terse_output) 994 show_thread_status_terse(ts, rs); 995 else 996 show_thread_status(ts, rs); 997 } 998 999 if (!terse_output) { 1000 for (i = 0; i < groupid + 1; i++) { 1001 rs = &runstats[i]; 1002 1003 rs->groupid = i; 1004 if (is_backend) 1005 fio_server_send_gs(rs); 1006 else 1007 show_group_stats(rs); 1008 } 1009 1010 if (is_backend) 1011 fio_server_send_du(); 1012 else 1013 show_disk_util(0); 1014 1015 free_disk_util(); 1016 } 1017 1018 free(runstats); 1019 free(threadstats); 1020} 1021 1022static inline void add_stat_sample(struct io_stat *is, unsigned long data) 1023{ 1024 double val = data; 1025 double delta; 1026 1027 if (data > is->max_val) 1028 is->max_val = data; 1029 if (data < is->min_val) 1030 is->min_val = data; 1031 1032 delta = val - is->mean.u.f; 1033 if (delta) { 1034 is->mean.u.f += delta / (is->samples + 1.0); 1035 is->S.u.f += delta * (val - is->mean.u.f); 1036 } 1037 1038 is->samples++; 1039} 1040 1041static void __add_log_sample(struct io_log *iolog, unsigned long val, 1042 enum fio_ddir ddir, unsigned int bs, 1043 unsigned long t) 1044{ 1045 const int nr_samples = iolog->nr_samples; 1046 1047 if (iolog->nr_samples == iolog->max_samples) { 1048 int new_size = sizeof(struct io_sample) * iolog->max_samples*2; 1049 1050 iolog->log = realloc(iolog->log, new_size); 1051 iolog->max_samples <<= 1; 1052 } 1053 1054 iolog->log[nr_samples].val = val; 1055 iolog->log[nr_samples].time = t; 1056 iolog->log[nr_samples].ddir = ddir; 1057 iolog->log[nr_samples].bs = bs; 1058 iolog->nr_samples++; 1059} 1060 1061static void add_log_sample(struct thread_data *td, struct io_log *iolog, 1062 unsigned long val, enum fio_ddir ddir, 1063 unsigned int bs) 1064{ 1065 if (!ddir_rw(ddir)) 1066 return; 1067 1068 __add_log_sample(iolog, val, ddir, bs, mtime_since_now(&td->epoch)); 1069} 1070 1071void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs) 1072{ 1073 struct io_log *iolog; 1074 1075 if (!ddir_rw(ddir)) 1076 return; 1077 1078 iolog = agg_io_log[ddir]; 1079 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis()); 1080} 1081 1082static void add_clat_percentile_sample(struct thread_stat *ts, 1083 unsigned long usec, enum fio_ddir ddir) 1084{ 1085 unsigned int idx = plat_val_to_idx(usec); 1086 assert(idx < FIO_IO_U_PLAT_NR); 1087 1088 ts->io_u_plat[ddir][idx]++; 1089} 1090 1091void add_clat_sample(struct thread_data *td, enum fio_ddir ddir, 1092 unsigned long usec, unsigned int bs) 1093{ 1094 struct thread_stat *ts = &td->ts; 1095 1096 if (!ddir_rw(ddir)) 1097 return; 1098 1099 add_stat_sample(&ts->clat_stat[ddir], usec); 1100 1101 if (td->clat_log) 1102 add_log_sample(td, td->clat_log, usec, ddir, bs); 1103 1104 if (ts->clat_percentiles) 1105 add_clat_percentile_sample(ts, usec, ddir); 1106} 1107 1108void add_slat_sample(struct thread_data *td, enum fio_ddir ddir, 1109 unsigned long usec, unsigned int bs) 1110{ 1111 struct thread_stat *ts = &td->ts; 1112 1113 if (!ddir_rw(ddir)) 1114 return; 1115 1116 add_stat_sample(&ts->slat_stat[ddir], usec); 1117 1118 if (td->slat_log) 1119 add_log_sample(td, td->slat_log, usec, ddir, bs); 1120} 1121 1122void add_lat_sample(struct thread_data *td, enum fio_ddir ddir, 1123 unsigned long usec, unsigned int bs) 1124{ 1125 struct thread_stat *ts = &td->ts; 1126 1127 if (!ddir_rw(ddir)) 1128 return; 1129 1130 add_stat_sample(&ts->lat_stat[ddir], usec); 1131 1132 if (td->lat_log) 1133 add_log_sample(td, td->lat_log, usec, ddir, bs); 1134} 1135 1136void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs, 1137 struct timeval *t) 1138{ 1139 struct thread_stat *ts = &td->ts; 1140 unsigned long spent, rate; 1141 1142 if (!ddir_rw(ddir)) 1143 return; 1144 1145 spent = mtime_since(&td->bw_sample_time, t); 1146 if (spent < td->o.bw_avg_time) 1147 return; 1148 1149 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) * 1150 1000 / spent / 1024; 1151 add_stat_sample(&ts->bw_stat[ddir], rate); 1152 1153 if (td->bw_log) 1154 add_log_sample(td, td->bw_log, rate, ddir, bs); 1155 1156 fio_gettime(&td->bw_sample_time, NULL); 1157 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir]; 1158} 1159 1160void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, 1161 struct timeval *t) 1162{ 1163 struct thread_stat *ts = &td->ts; 1164 unsigned long spent, iops; 1165 1166 if (!ddir_rw(ddir)) 1167 return; 1168 1169 spent = mtime_since(&td->iops_sample_time, t); 1170 if (spent < td->o.iops_avg_time) 1171 return; 1172 1173 iops = ((td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]) * 1000) / spent; 1174 1175 add_stat_sample(&ts->iops_stat[ddir], iops); 1176 1177 if (td->iops_log) { 1178 assert(iops); 1179 add_log_sample(td, td->iops_log, iops, ddir, 0); 1180 } 1181 1182 fio_gettime(&td->iops_sample_time, NULL); 1183 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir]; 1184} 1185