stat.c revision 37f0c1ae23ad1716403d3d113c3dfdf41c47e329
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} 116 117/* 118 * Find and display the p-th percentile of clat 119 */ 120static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, 121 fio_fp64_t *plist) 122{ 123 unsigned long sum = 0; 124 unsigned int len, i, j = 0; 125 int is_last = 0; 126 127 len = 0; 128 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0) 129 len++; 130 131 /* 132 * Sort the percentile list. Note that it may already be sorted if 133 * we are using the default values, but since it's a short list this 134 * isn't a worry. Also note that this does not work for NaN values. 135 */ 136 if (len > 1) 137 qsort((void*)plist, len, sizeof(plist[0]), double_cmp); 138 139 log_info(" clat percentiles (usec) :"); 140 141 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) { 142 sum += io_u_plat[i]; 143 while (sum >= (plist[j].u.f / 100.0 * nr)) { 144 assert(plist[j].u.f <= 100.0); 145 146 /* for formatting */ 147 if (j != 0 && (j % 4) == 0) 148 log_info(" "); 149 150 /* end of the list */ 151 is_last = (j == len - 1); 152 153 log_info(" %2.2fth=%u%c", plist[j], plat_idx_to_val(i), 154 (is_last? '\n' : ',')); 155 156 if (is_last) 157 break; 158 159 if (j % 4 == 3) /* for formatting */ 160 log_info("\n"); 161 j++; 162 } 163 } 164} 165 166static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, 167 double *mean, double *dev) 168{ 169 double n = is->samples; 170 171 if (is->samples == 0) 172 return 0; 173 174 *min = is->min_val; 175 *max = is->max_val; 176 177 n = (double) is->samples; 178 *mean = is->mean.u.f; 179 180 if (n > 1.0) 181 *dev = sqrt(is->S.u.f / (n - 1.0)); 182 else 183 *dev = 0; 184 185 return 1; 186} 187 188void show_group_stats(struct group_run_stats *rs) 189{ 190 char *p1, *p2, *p3, *p4; 191 const char *ddir_str[] = { " READ", " WRITE" }; 192 int i; 193 194 log_info("\nRun status group %d (all jobs):\n", rs->groupid); 195 196 for (i = 0; i <= DDIR_WRITE; i++) { 197 const int i2p = is_power_of_2(rs->kb_base); 198 199 if (!rs->max_run[i]) 200 continue; 201 202 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p); 203 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p); 204 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p); 205 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p); 206 207 log_info("%s: io=%sB, aggrb=%sB/s, minb=%sB/s, maxb=%sB/s," 208 " mint=%llumsec, maxt=%llumsec\n", ddir_str[i], p1, p2, 209 p3, p4, rs->min_run[i], 210 rs->max_run[i]); 211 212 free(p1); 213 free(p2); 214 free(p3); 215 free(p4); 216 } 217} 218 219#define ts_total_io_u(ts) \ 220 ((ts)->total_io_u[0] + (ts)->total_io_u[1]) 221 222static void stat_calc_dist(unsigned int *map, unsigned long total, 223 double *io_u_dist) 224{ 225 int i; 226 227 /* 228 * Do depth distribution calculations 229 */ 230 for (i = 0; i < FIO_IO_U_MAP_NR; i++) { 231 if (total) { 232 io_u_dist[i] = (double) map[i] / (double) total; 233 io_u_dist[i] *= 100.0; 234 if (io_u_dist[i] < 0.1 && map[i]) 235 io_u_dist[i] = 0.1; 236 } else 237 io_u_dist[i] = 0.0; 238 } 239} 240 241static void stat_calc_lat(struct thread_stat *ts, double *dst, 242 unsigned int *src, int nr) 243{ 244 unsigned long total = ts_total_io_u(ts); 245 int i; 246 247 /* 248 * Do latency distribution calculations 249 */ 250 for (i = 0; i < nr; i++) { 251 if (total) { 252 dst[i] = (double) src[i] / (double) total; 253 dst[i] *= 100.0; 254 if (dst[i] < 0.01 && src[i]) 255 dst[i] = 0.01; 256 } else 257 dst[i] = 0.0; 258 } 259} 260 261static void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat) 262{ 263 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR); 264} 265 266static void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat) 267{ 268 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR); 269} 270 271static int usec_to_msec(unsigned long *min, unsigned long *max, double *mean, 272 double *dev) 273{ 274 if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) { 275 *min /= 1000; 276 *max /= 1000; 277 *mean /= 1000.0; 278 *dev /= 1000.0; 279 return 0; 280 } 281 282 return 1; 283} 284 285static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts, 286 int ddir) 287{ 288 const char *ddir_str[] = { "read ", "write" }; 289 unsigned long min, max, runt; 290 unsigned long long bw, iops; 291 double mean, dev; 292 char *io_p, *bw_p, *iops_p; 293 int i2p; 294 295 assert(ddir_rw(ddir)); 296 297 if (!ts->runtime[ddir]) 298 return; 299 300 i2p = is_power_of_2(rs->kb_base); 301 runt = ts->runtime[ddir]; 302 303 bw = (1000 * ts->io_bytes[ddir]) / runt; 304 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p); 305 bw_p = num2str(bw, 6, 1, i2p); 306 307 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt; 308 iops_p = num2str(iops, 6, 1, 0); 309 310 log_info(" %s: io=%sB, bw=%sB/s, iops=%s, runt=%6llumsec\n", 311 ddir_str[ddir], io_p, bw_p, iops_p, 312 ts->runtime[ddir]); 313 314 free(io_p); 315 free(bw_p); 316 free(iops_p); 317 318 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) { 319 const char *base = "(usec)"; 320 char *minp, *maxp; 321 322 if (!usec_to_msec(&min, &max, &mean, &dev)) 323 base = "(msec)"; 324 325 minp = num2str(min, 6, 1, 0); 326 maxp = num2str(max, 6, 1, 0); 327 328 log_info(" slat %s: min=%s, max=%s, avg=%5.02f," 329 " stdev=%5.02f\n", base, minp, maxp, mean, dev); 330 331 free(minp); 332 free(maxp); 333 } 334 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) { 335 const char *base = "(usec)"; 336 char *minp, *maxp; 337 338 if (!usec_to_msec(&min, &max, &mean, &dev)) 339 base = "(msec)"; 340 341 minp = num2str(min, 6, 1, 0); 342 maxp = num2str(max, 6, 1, 0); 343 344 log_info(" clat %s: min=%s, max=%s, avg=%5.02f," 345 " stdev=%5.02f\n", base, minp, maxp, mean, dev); 346 347 free(minp); 348 free(maxp); 349 } 350 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) { 351 const char *base = "(usec)"; 352 char *minp, *maxp; 353 354 if (!usec_to_msec(&min, &max, &mean, &dev)) 355 base = "(msec)"; 356 357 minp = num2str(min, 6, 1, 0); 358 maxp = num2str(max, 6, 1, 0); 359 360 log_info(" lat %s: min=%s, max=%s, avg=%5.02f," 361 " stdev=%5.02f\n", base, minp, maxp, mean, dev); 362 363 free(minp); 364 free(maxp); 365 } 366 if (ts->clat_percentiles) { 367 show_clat_percentiles(ts->io_u_plat[ddir], 368 ts->clat_stat[ddir].samples, 369 ts->percentile_list); 370 } 371 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 372 double p_of_agg; 373 374 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 375 log_info(" bw (KB/s) : min=%5lu, max=%5lu, per=%3.2f%%," 376 " avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg, 377 mean, dev); 378 } 379} 380 381static void show_lat(double *io_u_lat, int nr, const char **ranges, 382 const char *msg) 383{ 384 int new_line = 1, i, line = 0; 385 386 for (i = 0; i < nr; i++) { 387 if (io_u_lat[i] <= 0.0) 388 continue; 389 if (new_line) { 390 if (line) 391 log_info("\n"); 392 log_info(" lat (%s): ", msg); 393 new_line = 0; 394 line = 0; 395 } 396 if (line) 397 log_info(", "); 398 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]); 399 line++; 400 if (line == 5) 401 new_line = 1; 402 } 403} 404 405static void show_lat_u(double *io_u_lat_u) 406{ 407 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", 408 "250=", "500=", "750=", "1000=", }; 409 410 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec"); 411} 412 413static void show_lat_m(double *io_u_lat_m) 414{ 415 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", 416 "250=", "500=", "750=", "1000=", "2000=", 417 ">=2000=", }; 418 419 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec"); 420} 421 422static void show_latencies(double *io_u_lat_u, double *io_u_lat_m) 423{ 424 show_lat_u(io_u_lat_u); 425 log_info("\n"); 426 show_lat_m(io_u_lat_m); 427 log_info("\n"); 428} 429 430void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs) 431{ 432 double usr_cpu, sys_cpu; 433 unsigned long runtime; 434 double io_u_dist[FIO_IO_U_MAP_NR]; 435 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 436 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 437 438 if (!(ts->io_bytes[0] + ts->io_bytes[1]) && 439 !(ts->total_io_u[0] + ts->total_io_u[1])) 440 return; 441 442 if (!ts->error) { 443 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n", 444 ts->name, ts->groupid, ts->members, 445 ts->error, (int) ts->pid); 446 } else { 447 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n", 448 ts->name, ts->groupid, ts->members, 449 ts->error, ts->verror, (int) ts->pid); 450 } 451 452 if (ts->description) 453 log_info(" Description : [%s]\n", ts->description); 454 455 if (ts->io_bytes[DDIR_READ]) 456 show_ddir_status(rs, ts, DDIR_READ); 457 if (ts->io_bytes[DDIR_WRITE]) 458 show_ddir_status(rs, ts, DDIR_WRITE); 459 460 runtime = ts->total_run_time; 461 if (runtime) { 462 double runt = (double) runtime; 463 464 usr_cpu = (double) ts->usr_time * 100 / runt; 465 sys_cpu = (double) ts->sys_time * 100 / runt; 466 } else { 467 usr_cpu = 0; 468 sys_cpu = 0; 469 } 470 471 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu," 472 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf); 473 474 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 475 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%," 476 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 477 io_u_dist[1], io_u_dist[2], 478 io_u_dist[3], io_u_dist[4], 479 io_u_dist[5], io_u_dist[6]); 480 481 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist); 482 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 483 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 484 io_u_dist[1], io_u_dist[2], 485 io_u_dist[3], io_u_dist[4], 486 io_u_dist[5], io_u_dist[6]); 487 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist); 488 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 489 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 490 io_u_dist[1], io_u_dist[2], 491 io_u_dist[3], io_u_dist[4], 492 io_u_dist[5], io_u_dist[6]); 493 log_info(" issued r/w/d: total=%lu/%lu/%lu, short=%lu/%lu/%lu\n", 494 ts->total_io_u[0], ts->total_io_u[1], 495 ts->total_io_u[2], 496 ts->short_io_u[0], ts->short_io_u[1], 497 ts->short_io_u[2]); 498 stat_calc_lat_u(ts, io_u_lat_u); 499 stat_calc_lat_m(ts, io_u_lat_m); 500 show_latencies(io_u_lat_u, io_u_lat_m); 501 if (ts->continue_on_error) { 502 log_info(" errors : total=%lu, first_error=%d/<%s>\n", 503 ts->total_err_count, 504 ts->first_error, 505 strerror(ts->first_error)); 506 } 507} 508 509static void show_ddir_status_terse(struct thread_stat *ts, 510 struct group_run_stats *rs, int ddir) 511{ 512 unsigned long min, max; 513 unsigned long long bw; 514 double mean, dev; 515 516 assert(ddir_rw(ddir)); 517 518 bw = 0; 519 if (ts->runtime[ddir]) 520 bw = ts->io_bytes[ddir] / ts->runtime[ddir]; 521 522 log_info(";%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, 523 ts->runtime[ddir]); 524 525 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) 526 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 527 else 528 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 529 530 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) 531 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 532 else 533 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 534 535 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) 536 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 537 else 538 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 539 540 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 541 double p_of_agg; 542 543 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 544 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev); 545 } else 546 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0); 547} 548 549#define FIO_TERSE_VERSION "2" 550 551static void show_thread_status_terse(struct thread_stat *ts, 552 struct group_run_stats *rs) 553{ 554 double io_u_dist[FIO_IO_U_MAP_NR]; 555 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 556 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 557 double usr_cpu, sys_cpu; 558 int i; 559 560 /* General Info */ 561 log_info("%s;%s;%d;%d", FIO_TERSE_VERSION, ts->name, ts->groupid, 562 ts->error); 563 /* Log Read Status */ 564 show_ddir_status_terse(ts, rs, 0); 565 /* Log Write Status */ 566 show_ddir_status_terse(ts, rs, 1); 567 568 /* CPU Usage */ 569 if (ts->total_run_time) { 570 double runt = (double) ts->total_run_time; 571 572 usr_cpu = (double) ts->usr_time * 100 / runt; 573 sys_cpu = (double) ts->sys_time * 100 / runt; 574 } else { 575 usr_cpu = 0; 576 sys_cpu = 0; 577 } 578 579 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf, 580 ts->minf); 581 582 /* Calc % distribution of IO depths, usecond, msecond latency */ 583 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 584 stat_calc_lat_u(ts, io_u_lat_u); 585 stat_calc_lat_m(ts, io_u_lat_m); 586 587 /* Only show fixed 7 I/O depth levels*/ 588 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 589 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], 590 io_u_dist[4], io_u_dist[5], io_u_dist[6]); 591 592 /* Microsecond latency */ 593 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) 594 log_info(";%3.2f%%", io_u_lat_u[i]); 595 /* Millisecond latency */ 596 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) 597 log_info(";%3.2f%%", io_u_lat_m[i]); 598 /* Additional output if continue_on_error set - default off*/ 599 if (ts->continue_on_error) 600 log_info(";%lu;%d", ts->total_err_count, ts->first_error); 601 log_info("\n"); 602 603 /* Additional output if description is set */ 604 if (ts->description) 605 log_info(";%s", ts->description); 606 607 log_info("\n"); 608} 609 610static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr) 611{ 612 double mean, S; 613 614 if (src->samples == 0) 615 return; 616 617 dst->min_val = min(dst->min_val, src->min_val); 618 dst->max_val = max(dst->max_val, src->max_val); 619 620 /* 621 * Compute new mean and S after the merge 622 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 623 * #Parallel_algorithm> 624 */ 625 if (nr == 1) { 626 mean = src->mean.u.f; 627 S = src->S.u.f; 628 } else { 629 double delta = src->mean.u.f - dst->mean.u.f; 630 631 mean = ((src->mean.u.f * src->samples) + 632 (dst->mean.u.f * dst->samples)) / 633 (dst->samples + src->samples); 634 635 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) * 636 (dst->samples * src->samples) / 637 (dst->samples + src->samples); 638 } 639 640 dst->samples += src->samples; 641 dst->mean.u.f = mean; 642 dst->S.u.f = S; 643} 644 645void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src) 646{ 647 int i; 648 649 for (i = 0; i < 2; i++) { 650 if (dst->max_run[i] < src->max_run[i]) 651 dst->max_run[i] = src->max_run[i]; 652 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i]) 653 dst->min_run[i] = src->min_run[i]; 654 if (dst->max_bw[i] < src->max_bw[i]) 655 dst->max_bw[i] = src->max_bw[i]; 656 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i]) 657 dst->min_bw[i] = src->min_bw[i]; 658 659 dst->io_kb[i] += src->io_kb[i]; 660 dst->agg[i] += src->agg[i]; 661 } 662 663} 664 665void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr) 666{ 667 int l, k; 668 669 for (l = 0; l <= DDIR_WRITE; l++) { 670 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr); 671 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr); 672 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr); 673 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr); 674 675 dst->io_bytes[l] += src->io_bytes[l]; 676 677 if (dst->runtime[l] < src->runtime[l]) 678 dst->runtime[l] = src->runtime[l]; 679 } 680 681 dst->usr_time += src->usr_time; 682 dst->sys_time += src->sys_time; 683 dst->ctx += src->ctx; 684 dst->majf += src->majf; 685 dst->minf += src->minf; 686 687 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 688 dst->io_u_map[k] += src->io_u_map[k]; 689 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 690 dst->io_u_submit[k] += src->io_u_submit[k]; 691 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 692 dst->io_u_complete[k] += src->io_u_complete[k]; 693 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++) 694 dst->io_u_lat_u[k] += src->io_u_lat_u[k]; 695 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++) 696 dst->io_u_lat_m[k] += src->io_u_lat_m[k]; 697 698 for (k = 0; k <= 2; k++) { 699 dst->total_io_u[k] += src->total_io_u[k]; 700 dst->short_io_u[k] += src->short_io_u[k]; 701 } 702 703 for (k = 0; k <= DDIR_WRITE; k++) { 704 int m; 705 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) 706 dst->io_u_plat[k][m] += src->io_u_plat[k][m]; 707 } 708 709 dst->total_run_time += src->total_run_time; 710 dst->total_submit += src->total_submit; 711 dst->total_complete += src->total_complete; 712} 713 714void init_group_run_stat(struct group_run_stats *gs) 715{ 716 memset(gs, 0, sizeof(*gs)); 717 gs->min_bw[0] = gs->min_run[0] = ~0UL; 718 gs->min_bw[1] = gs->min_run[1] = ~0UL; 719} 720 721void init_thread_stat(struct thread_stat *ts) 722{ 723 int j; 724 725 memset(ts, 0, sizeof(*ts)); 726 727 for (j = 0; j <= DDIR_WRITE; j++) { 728 ts->lat_stat[j].min_val = -1UL; 729 ts->clat_stat[j].min_val = -1UL; 730 ts->slat_stat[j].min_val = -1UL; 731 ts->bw_stat[j].min_val = -1UL; 732 } 733 ts->groupid = -1; 734} 735 736void show_run_stats(void) 737{ 738 struct group_run_stats *runstats, *rs; 739 struct thread_data *td; 740 struct thread_stat *threadstats, *ts; 741 int i, j, nr_ts, last_ts, idx; 742 int kb_base_warned = 0; 743 744 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1)); 745 746 for (i = 0; i < groupid + 1; i++) 747 init_group_run_stat(&runstats[i]); 748 749 /* 750 * find out how many threads stats we need. if group reporting isn't 751 * enabled, it's one-per-td. 752 */ 753 nr_ts = 0; 754 last_ts = -1; 755 for_each_td(td, i) { 756 if (!td->o.group_reporting) { 757 nr_ts++; 758 continue; 759 } 760 if (last_ts == td->groupid) 761 continue; 762 763 last_ts = td->groupid; 764 nr_ts++; 765 } 766 767 threadstats = malloc(nr_ts * sizeof(struct thread_stat)); 768 769 for (i = 0; i < nr_ts; i++) 770 init_thread_stat(&threadstats[i]); 771 772 j = 0; 773 last_ts = -1; 774 idx = 0; 775 for_each_td(td, i) { 776 if (idx && (!td->o.group_reporting || 777 (td->o.group_reporting && last_ts != td->groupid))) { 778 idx = 0; 779 j++; 780 } 781 782 last_ts = td->groupid; 783 784 ts = &threadstats[j]; 785 786 ts->clat_percentiles = td->o.clat_percentiles; 787 if (td->o.overwrite_plist) 788 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list)); 789 else 790 memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list)); 791 792 idx++; 793 ts->members++; 794 795 if (ts->groupid == -1) { 796 /* 797 * These are per-group shared already 798 */ 799 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE); 800 if (td->o.description) 801 strncpy(ts->description, td->o.description, 802 FIO_JOBNAME_SIZE); 803 else 804 memset(ts->description, 0, FIO_JOBNAME_SIZE); 805 806 ts->groupid = td->groupid; 807 808 /* 809 * first pid in group, not very useful... 810 */ 811 ts->pid = td->pid; 812 813 ts->kb_base = td->o.kb_base; 814 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) { 815 log_info("fio: kb_base differs for jobs in group, using" 816 " %u as the base\n", ts->kb_base); 817 kb_base_warned = 1; 818 } 819 820 ts->continue_on_error = td->o.continue_on_error; 821 ts->total_err_count += td->total_err_count; 822 ts->first_error = td->first_error; 823 if (!ts->error) { 824 if (!td->error && td->o.continue_on_error && 825 td->first_error) { 826 ts->error = td->first_error; 827 strcpy(ts->verror, td->verror); 828 } else if (td->error) { 829 ts->error = td->error; 830 strcpy(ts->verror, td->verror); 831 } 832 } 833 834 sum_thread_stats(ts, &td->ts, idx); 835 } 836 837 for (i = 0; i < nr_ts; i++) { 838 unsigned long long bw; 839 840 ts = &threadstats[i]; 841 rs = &runstats[ts->groupid]; 842 rs->kb_base = ts->kb_base; 843 844 for (j = 0; j <= DDIR_WRITE; j++) { 845 if (!ts->runtime[j]) 846 continue; 847 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j]) 848 rs->min_run[j] = ts->runtime[j]; 849 if (ts->runtime[j] > rs->max_run[j]) 850 rs->max_run[j] = ts->runtime[j]; 851 852 bw = 0; 853 if (ts->runtime[j]) { 854 unsigned long runt; 855 856 runt = ts->runtime[j]; 857 bw = ts->io_bytes[j] / runt; 858 } 859 if (bw < rs->min_bw[j]) 860 rs->min_bw[j] = bw; 861 if (bw > rs->max_bw[j]) 862 rs->max_bw[j] = bw; 863 864 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base; 865 } 866 } 867 868 for (i = 0; i < groupid + 1; i++) { 869 unsigned long max_run[2]; 870 871 rs = &runstats[i]; 872 max_run[0] = rs->max_run[0]; 873 max_run[1] = rs->max_run[1]; 874 875 if (rs->max_run[0]) 876 rs->agg[0] = (rs->io_kb[0] * 1000) / max_run[0]; 877 if (rs->max_run[1]) 878 rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1]; 879 } 880 881 /* 882 * don't overwrite last signal output 883 */ 884 if (!terse_output) 885 log_info("\n"); 886 887 for (i = 0; i < nr_ts; i++) { 888 ts = &threadstats[i]; 889 rs = &runstats[ts->groupid]; 890 891 if (is_backend) 892 fio_server_send_ts(ts, rs); 893 else if (terse_output) 894 show_thread_status_terse(ts, rs); 895 else 896 show_thread_status(ts, rs); 897 } 898 899 if (!terse_output) { 900 for (i = 0; i < groupid + 1; i++) { 901 rs = &runstats[i]; 902 903 rs->groupid = i; 904 if (is_backend) 905 fio_server_send_gs(rs); 906 else 907 show_group_stats(rs); 908 } 909 910 show_disk_util(); 911 } 912 913 free(runstats); 914 free(threadstats); 915} 916 917static inline void add_stat_sample(struct io_stat *is, unsigned long data) 918{ 919 double val = data; 920 double delta; 921 922 if (data > is->max_val) 923 is->max_val = data; 924 if (data < is->min_val) 925 is->min_val = data; 926 927 delta = val - is->mean.u.f; 928 if (delta) { 929 is->mean.u.f += delta / (is->samples + 1.0); 930 is->S.u.f += delta * (val - is->mean.u.f); 931 } 932 933 is->samples++; 934} 935 936static void __add_log_sample(struct io_log *iolog, unsigned long val, 937 enum fio_ddir ddir, unsigned int bs, 938 unsigned long t) 939{ 940 const int nr_samples = iolog->nr_samples; 941 942 if (iolog->nr_samples == iolog->max_samples) { 943 int new_size = sizeof(struct io_sample) * iolog->max_samples*2; 944 945 iolog->log = realloc(iolog->log, new_size); 946 iolog->max_samples <<= 1; 947 } 948 949 iolog->log[nr_samples].val = val; 950 iolog->log[nr_samples].time = t; 951 iolog->log[nr_samples].ddir = ddir; 952 iolog->log[nr_samples].bs = bs; 953 iolog->nr_samples++; 954} 955 956static void add_log_sample(struct thread_data *td, struct io_log *iolog, 957 unsigned long val, enum fio_ddir ddir, 958 unsigned int bs) 959{ 960 if (!ddir_rw(ddir)) 961 return; 962 963 __add_log_sample(iolog, val, ddir, bs, mtime_since_now(&td->epoch)); 964} 965 966void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs) 967{ 968 struct io_log *iolog; 969 970 if (!ddir_rw(ddir)) 971 return; 972 973 iolog = agg_io_log[ddir]; 974 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis()); 975} 976 977static void add_clat_percentile_sample(struct thread_stat *ts, 978 unsigned long usec, enum fio_ddir ddir) 979{ 980 unsigned int idx = plat_val_to_idx(usec); 981 assert(idx < FIO_IO_U_PLAT_NR); 982 983 ts->io_u_plat[ddir][idx]++; 984} 985 986void add_clat_sample(struct thread_data *td, enum fio_ddir ddir, 987 unsigned long usec, unsigned int bs) 988{ 989 struct thread_stat *ts = &td->ts; 990 991 if (!ddir_rw(ddir)) 992 return; 993 994 add_stat_sample(&ts->clat_stat[ddir], usec); 995 996 if (td->clat_log) 997 add_log_sample(td, td->clat_log, usec, ddir, bs); 998 999 if (ts->clat_percentiles) 1000 add_clat_percentile_sample(ts, usec, ddir); 1001} 1002 1003void add_slat_sample(struct thread_data *td, enum fio_ddir ddir, 1004 unsigned long usec, unsigned int bs) 1005{ 1006 struct thread_stat *ts = &td->ts; 1007 1008 if (!ddir_rw(ddir)) 1009 return; 1010 1011 add_stat_sample(&ts->slat_stat[ddir], usec); 1012 1013 if (td->slat_log) 1014 add_log_sample(td, td->slat_log, usec, ddir, bs); 1015} 1016 1017void add_lat_sample(struct thread_data *td, enum fio_ddir ddir, 1018 unsigned long usec, unsigned int bs) 1019{ 1020 struct thread_stat *ts = &td->ts; 1021 1022 if (!ddir_rw(ddir)) 1023 return; 1024 1025 add_stat_sample(&ts->lat_stat[ddir], usec); 1026 1027 if (td->lat_log) 1028 add_log_sample(td, td->lat_log, usec, ddir, bs); 1029} 1030 1031void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs, 1032 struct timeval *t) 1033{ 1034 struct thread_stat *ts = &td->ts; 1035 unsigned long spent, rate; 1036 1037 if (!ddir_rw(ddir)) 1038 return; 1039 1040 spent = mtime_since(&td->bw_sample_time, t); 1041 if (spent < td->o.bw_avg_time) 1042 return; 1043 1044 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) * 1045 1000 / spent / 1024; 1046 add_stat_sample(&ts->bw_stat[ddir], rate); 1047 1048 if (td->bw_log) 1049 add_log_sample(td, td->bw_log, rate, ddir, bs); 1050 1051 fio_gettime(&td->bw_sample_time, NULL); 1052 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir]; 1053} 1054 1055void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, 1056 struct timeval *t) 1057{ 1058 struct thread_stat *ts = &td->ts; 1059 unsigned long spent, iops; 1060 1061 if (!ddir_rw(ddir)) 1062 return; 1063 1064 spent = mtime_since(&td->iops_sample_time, t); 1065 if (spent < td->o.iops_avg_time) 1066 return; 1067 1068 iops = ((td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]) * 1000) / spent; 1069 1070 add_stat_sample(&ts->iops_stat[ddir], iops); 1071 1072 if (td->iops_log) { 1073 assert(iops); 1074 add_log_sample(td, td->iops_log, iops, ddir, 0); 1075 } 1076 1077 fio_gettime(&td->iops_sample_time, NULL); 1078 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir]; 1079} 1080