12 #include "lib/ieee754.h"
14 void update_rusage_stat(struct thread_data *td)
16 struct thread_stat *ts = &td->ts;
18 getrusage(RUSAGE_SELF, &td->ru_end);
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;
29 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
33 * Given a latency, return the index of the corresponding bucket in
34 * the structure tracking percentiles.
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.
41 static unsigned int plat_val_to_idx(unsigned int val)
43 unsigned int msb, error_bits, base, offset, idx;
45 /* Find MSB starting from bit 0 */
49 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
52 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
53 * all bits of the sample as index
55 if (msb <= FIO_IO_U_PLAT_BITS)
58 /* Compute the number of error bits to discard*/
59 error_bits = msb - FIO_IO_U_PLAT_BITS;
61 /* Compute the number of buckets before the group */
62 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
65 * Discard the error bits and apply the mask to find the
66 * index for the buckets in the group
68 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
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);
78 * Convert the given index of the bucket array to the value
79 * represented by the bucket
81 static unsigned int plat_idx_to_val(unsigned int idx)
83 unsigned int error_bits, k, base;
85 assert(idx < FIO_IO_U_PLAT_NR);
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))
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);
96 /* Find its bucket number of the group */
97 k = idx % FIO_IO_U_PLAT_VAL;
99 /* Return the mean of the range of the bucket */
100 return base + ((k + 0.5) * (1 << error_bits));
103 static int double_cmp(const void *a, const void *b)
105 const fio_fp64_t fa = *(const fio_fp64_t *) a;
106 const fio_fp64_t fb = *(const fio_fp64_t *) b;
111 else if (fa.u.f < fb.u.f)
117 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
118 fio_fp64_t *plist, unsigned int **output,
119 unsigned int *maxv, unsigned int *minv)
121 unsigned long sum = 0;
122 unsigned int len, i, j = 0;
123 unsigned int oval_len = 0;
124 unsigned int *ovals = NULL;
131 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
138 * Sort the percentile list. Note that it may already be sorted if
139 * we are using the default values, but since it's a short list this
140 * isn't a worry. Also note that this does not work for NaN values.
143 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
146 * Calculate bucket values, note down max and min values
149 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
151 while (sum >= (plist[j].u.f / 100.0 * nr)) {
152 assert(plist[j].u.f <= 100.0);
156 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
159 ovals[j] = plat_idx_to_val(i);
160 if (ovals[j] < *minv)
162 if (ovals[j] > *maxv)
165 is_last = (j == len - 1);
178 * Find and display the p-th percentile of clat
180 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
183 unsigned int len, j = 0, minv, maxv;
185 int is_last, scale_down;
187 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
192 * We default to usecs, but if the value range is such that we
193 * should scale down to msecs, do that.
195 if (minv > 2000 && maxv > 99999) {
197 log_info(" clat percentiles (msec):\n |");
200 log_info(" clat percentiles (usec):\n |");
203 for (j = 0; j < len; j++) {
207 if (j != 0 && (j % 4) == 0)
210 /* end of the list */
211 is_last = (j == len - 1);
213 if (plist[j].u.f < 10.0)
214 sprintf(fbuf, " %2.2f", plist[j].u.f);
216 sprintf(fbuf, "%2.2f", plist[j].u.f);
219 ovals[j] = (ovals[j] + 999) / 1000;
221 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
226 if (j % 4 == 3) /* for formatting */
235 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
236 double *mean, double *dev)
238 double n = is->samples;
240 if (is->samples == 0)
246 n = (double) is->samples;
247 *mean = is->mean.u.f;
250 *dev = sqrt(is->S.u.f / (n - 1.0));
257 void show_group_stats(struct group_run_stats *rs)
259 char *p1, *p2, *p3, *p4;
260 const char *ddir_str[] = { " READ", " WRITE" };
263 log_info("\nRun status group %d (all jobs):\n", rs->groupid);
265 for (i = 0; i <= DDIR_WRITE; i++) {
266 const int i2p = is_power_of_2(rs->kb_base);
271 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p);
272 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p);
273 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p);
274 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p);
276 log_info("%s: io=%sB, aggrb=%sB/s, minb=%sB/s, maxb=%sB/s,"
277 " mint=%llumsec, maxt=%llumsec\n", ddir_str[i], p1, p2,
278 p3, p4, rs->min_run[i],
288 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
293 * Do depth distribution calculations
295 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
297 io_u_dist[i] = (double) map[i] / (double) total;
298 io_u_dist[i] *= 100.0;
299 if (io_u_dist[i] < 0.1 && map[i])
306 static void stat_calc_lat(struct thread_stat *ts, double *dst,
307 unsigned int *src, int nr)
309 unsigned long total = ts_total_io_u(ts);
313 * Do latency distribution calculations
315 for (i = 0; i < nr; i++) {
317 dst[i] = (double) src[i] / (double) total;
319 if (dst[i] < 0.01 && src[i])
326 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
328 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
331 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
333 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
336 static void display_lat(const char *name, unsigned long min, unsigned long max,
337 double mean, double dev)
339 const char *base = "(usec)";
342 if (!usec_to_msec(&min, &max, &mean, &dev))
345 minp = num2str(min, 6, 1, 0);
346 maxp = num2str(max, 6, 1, 0);
348 log_info(" %s %s: min=%s, max=%s, avg=%5.02f,"
349 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
355 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
358 const char *ddir_str[] = { "read ", "write" };
359 unsigned long min, max, runt;
360 unsigned long long bw, iops;
362 char *io_p, *bw_p, *iops_p;
365 assert(ddir_rw(ddir));
367 if (!ts->runtime[ddir])
370 i2p = is_power_of_2(rs->kb_base);
371 runt = ts->runtime[ddir];
373 bw = (1000 * ts->io_bytes[ddir]) / runt;
374 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p);
375 bw_p = num2str(bw, 6, 1, i2p);
377 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
378 iops_p = num2str(iops, 6, 1, 0);
380 log_info(" %s: io=%sB, bw=%sB/s, iops=%s, runt=%6llumsec\n",
381 ddir_str[ddir], io_p, bw_p, iops_p,
388 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
389 display_lat("slat", min, max, mean, dev);
390 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
391 display_lat("clat", min, max, mean, dev);
392 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
393 display_lat(" lat", min, max, mean, dev);
395 if (ts->clat_percentiles) {
396 show_clat_percentiles(ts->io_u_plat[ddir],
397 ts->clat_stat[ddir].samples,
398 ts->percentile_list);
400 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
401 double p_of_agg = 100.0;
402 const char *bw_str = "KB";
405 p_of_agg = mean * 100 / (double) rs->agg[ddir];
406 if (p_of_agg > 100.0)
410 if (mean > 999999.9) {
418 log_info(" bw (%s/s) : min=%5lu, max=%5lu, per=%3.2f%%,"
419 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
420 p_of_agg, mean, dev);
424 static int show_lat(double *io_u_lat, int nr, const char **ranges,
427 int new_line = 1, i, line = 0, shown = 0;
429 for (i = 0; i < nr; i++) {
430 if (io_u_lat[i] <= 0.0)
436 log_info(" lat (%s) : ", msg);
442 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
454 static void show_lat_u(double *io_u_lat_u)
456 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
457 "250=", "500=", "750=", "1000=", };
459 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
462 static void show_lat_m(double *io_u_lat_m)
464 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
465 "250=", "500=", "750=", "1000=", "2000=",
468 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
471 static void show_latencies(struct thread_stat *ts)
473 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
474 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
476 stat_calc_lat_u(ts, io_u_lat_u);
477 stat_calc_lat_m(ts, io_u_lat_m);
479 show_lat_u(io_u_lat_u);
480 show_lat_m(io_u_lat_m);
483 void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
485 double usr_cpu, sys_cpu;
486 unsigned long runtime;
487 double io_u_dist[FIO_IO_U_MAP_NR];
489 if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
490 !(ts->total_io_u[0] + ts->total_io_u[1]))
494 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
495 ts->name, ts->groupid, ts->members,
496 ts->error, (int) ts->pid);
498 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
499 ts->name, ts->groupid, ts->members,
500 ts->error, ts->verror, (int) ts->pid);
503 if (strlen(ts->description))
504 log_info(" Description : [%s]\n", ts->description);
506 if (ts->io_bytes[DDIR_READ])
507 show_ddir_status(rs, ts, DDIR_READ);
508 if (ts->io_bytes[DDIR_WRITE])
509 show_ddir_status(rs, ts, DDIR_WRITE);
513 runtime = ts->total_run_time;
515 double runt = (double) runtime;
517 usr_cpu = (double) ts->usr_time * 100 / runt;
518 sys_cpu = (double) ts->sys_time * 100 / runt;
524 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
525 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
527 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
528 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
529 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
530 io_u_dist[1], io_u_dist[2],
531 io_u_dist[3], io_u_dist[4],
532 io_u_dist[5], io_u_dist[6]);
534 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
535 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
536 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
537 io_u_dist[1], io_u_dist[2],
538 io_u_dist[3], io_u_dist[4],
539 io_u_dist[5], io_u_dist[6]);
540 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
541 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
542 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
543 io_u_dist[1], io_u_dist[2],
544 io_u_dist[3], io_u_dist[4],
545 io_u_dist[5], io_u_dist[6]);
546 log_info(" issued : total=r=%lu/w=%lu/d=%lu,"
547 " short=r=%lu/w=%lu/d=%lu\n",
548 ts->total_io_u[0], ts->total_io_u[1],
550 ts->short_io_u[0], ts->short_io_u[1],
552 if (ts->continue_on_error) {
553 log_info(" errors : total=%lu, first_error=%d/<%s>\n",
556 strerror(ts->first_error));
560 static void show_ddir_status_terse(struct thread_stat *ts,
561 struct group_run_stats *rs, int ddir)
563 unsigned long min, max;
564 unsigned long long bw, iops;
565 unsigned int *ovals = NULL;
567 unsigned int len, minv, maxv;
570 assert(ddir_rw(ddir));
573 if (ts->runtime[ddir]) {
574 uint64_t runt = ts->runtime[ddir];
576 bw = ts->io_bytes[ddir] / runt;
577 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
580 log_info(";%llu;%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, iops,
583 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
584 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
586 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
588 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
589 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
591 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
593 if (ts->clat_percentiles) {
594 len = calc_clat_percentiles(ts->io_u_plat[ddir],
595 ts->clat_stat[ddir].samples,
596 ts->percentile_list, &ovals, &maxv,
601 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
606 log_info(";%2.2f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
609 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
610 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
612 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
617 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
618 double p_of_agg = 100.0;
621 p_of_agg = mean * 100 / (double) rs->agg[ddir];
622 if (p_of_agg > 100.0)
626 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
628 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
631 static void show_thread_status_terse_v2(struct thread_stat *ts,
632 struct group_run_stats *rs)
634 double io_u_dist[FIO_IO_U_MAP_NR];
635 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
636 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
637 double usr_cpu, sys_cpu;
641 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
642 /* Log Read Status */
643 show_ddir_status_terse(ts, rs, 0);
644 /* Log Write Status */
645 show_ddir_status_terse(ts, rs, 1);
648 if (ts->total_run_time) {
649 double runt = (double) ts->total_run_time;
651 usr_cpu = (double) ts->usr_time * 100 / runt;
652 sys_cpu = (double) ts->sys_time * 100 / runt;
658 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
661 /* Calc % distribution of IO depths, usecond, msecond latency */
662 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
663 stat_calc_lat_u(ts, io_u_lat_u);
664 stat_calc_lat_m(ts, io_u_lat_m);
666 /* Only show fixed 7 I/O depth levels*/
667 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
668 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
669 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
671 /* Microsecond latency */
672 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
673 log_info(";%3.2f%%", io_u_lat_u[i]);
674 /* Millisecond latency */
675 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
676 log_info(";%3.2f%%", io_u_lat_m[i]);
677 /* Additional output if continue_on_error set - default off*/
678 if (ts->continue_on_error)
679 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
682 /* Additional output if description is set */
684 log_info(";%s", ts->description);
689 #define FIO_TERSE_VERSION "3"
691 static void show_thread_status_terse_v3(struct thread_stat *ts,
692 struct group_run_stats *rs)
694 double io_u_dist[FIO_IO_U_MAP_NR];
695 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
696 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
697 double usr_cpu, sys_cpu;
701 log_info("%s;%s;%s;%d;%d", FIO_TERSE_VERSION, fio_version_string,
702 ts->name, ts->groupid, ts->error);
703 /* Log Read Status */
704 show_ddir_status_terse(ts, rs, 0);
705 /* Log Write Status */
706 show_ddir_status_terse(ts, rs, 1);
709 if (ts->total_run_time) {
710 double runt = (double) ts->total_run_time;
712 usr_cpu = (double) ts->usr_time * 100 / runt;
713 sys_cpu = (double) ts->sys_time * 100 / runt;
719 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
722 /* Calc % distribution of IO depths, usecond, msecond latency */
723 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
724 stat_calc_lat_u(ts, io_u_lat_u);
725 stat_calc_lat_m(ts, io_u_lat_m);
727 /* Only show fixed 7 I/O depth levels*/
728 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
729 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
730 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
732 /* Microsecond latency */
733 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
734 log_info(";%3.2f%%", io_u_lat_u[i]);
735 /* Millisecond latency */
736 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
737 log_info(";%3.2f%%", io_u_lat_m[i]);
739 /* disk util stats, if any */
742 /* Additional output if continue_on_error set - default off*/
743 if (ts->continue_on_error)
744 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
746 /* Additional output if description is set */
747 if (strlen(ts->description))
748 log_info(";%s", ts->description);
753 static void show_thread_status_terse(struct thread_stat *ts,
754 struct group_run_stats *rs)
756 if (terse_version == 2)
757 show_thread_status_terse_v2(ts, rs);
758 else if (terse_version == 3)
759 show_thread_status_terse_v3(ts, rs);
761 log_err("fio: bad terse version!? %d\n", terse_version);
764 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
768 if (src->samples == 0)
771 dst->min_val = min(dst->min_val, src->min_val);
772 dst->max_val = max(dst->max_val, src->max_val);
775 * Compute new mean and S after the merge
776 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
777 * #Parallel_algorithm>
780 mean = src->mean.u.f;
783 double delta = src->mean.u.f - dst->mean.u.f;
785 mean = ((src->mean.u.f * src->samples) +
786 (dst->mean.u.f * dst->samples)) /
787 (dst->samples + src->samples);
789 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
790 (dst->samples * src->samples) /
791 (dst->samples + src->samples);
794 dst->samples += src->samples;
795 dst->mean.u.f = mean;
799 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
803 for (i = 0; i < 2; i++) {
804 if (dst->max_run[i] < src->max_run[i])
805 dst->max_run[i] = src->max_run[i];
806 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
807 dst->min_run[i] = src->min_run[i];
808 if (dst->max_bw[i] < src->max_bw[i])
809 dst->max_bw[i] = src->max_bw[i];
810 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
811 dst->min_bw[i] = src->min_bw[i];
813 dst->io_kb[i] += src->io_kb[i];
814 dst->agg[i] += src->agg[i];
819 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
823 for (l = 0; l <= DDIR_WRITE; l++) {
824 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
825 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
826 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
827 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
829 dst->io_bytes[l] += src->io_bytes[l];
831 if (dst->runtime[l] < src->runtime[l])
832 dst->runtime[l] = src->runtime[l];
835 dst->usr_time += src->usr_time;
836 dst->sys_time += src->sys_time;
837 dst->ctx += src->ctx;
838 dst->majf += src->majf;
839 dst->minf += src->minf;
841 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
842 dst->io_u_map[k] += src->io_u_map[k];
843 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
844 dst->io_u_submit[k] += src->io_u_submit[k];
845 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
846 dst->io_u_complete[k] += src->io_u_complete[k];
847 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
848 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
849 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
850 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
852 for (k = 0; k <= 2; k++) {
853 dst->total_io_u[k] += src->total_io_u[k];
854 dst->short_io_u[k] += src->short_io_u[k];
857 for (k = 0; k <= DDIR_WRITE; k++) {
859 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
860 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
863 dst->total_run_time += src->total_run_time;
864 dst->total_submit += src->total_submit;
865 dst->total_complete += src->total_complete;
868 void init_group_run_stat(struct group_run_stats *gs)
870 memset(gs, 0, sizeof(*gs));
871 gs->min_bw[0] = gs->min_run[0] = ~0UL;
872 gs->min_bw[1] = gs->min_run[1] = ~0UL;
875 void init_thread_stat(struct thread_stat *ts)
879 memset(ts, 0, sizeof(*ts));
881 for (j = 0; j <= DDIR_WRITE; j++) {
882 ts->lat_stat[j].min_val = -1UL;
883 ts->clat_stat[j].min_val = -1UL;
884 ts->slat_stat[j].min_val = -1UL;
885 ts->bw_stat[j].min_val = -1UL;
890 void show_run_stats(void)
892 struct group_run_stats *runstats, *rs;
893 struct thread_data *td;
894 struct thread_stat *threadstats, *ts;
895 int i, j, nr_ts, last_ts, idx;
896 int kb_base_warned = 0;
898 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
900 for (i = 0; i < groupid + 1; i++)
901 init_group_run_stat(&runstats[i]);
904 * find out how many threads stats we need. if group reporting isn't
905 * enabled, it's one-per-td.
910 if (!td->o.group_reporting) {
914 if (last_ts == td->groupid)
917 last_ts = td->groupid;
921 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
923 for (i = 0; i < nr_ts; i++)
924 init_thread_stat(&threadstats[i]);
930 if (idx && (!td->o.group_reporting ||
931 (td->o.group_reporting && last_ts != td->groupid))) {
936 last_ts = td->groupid;
938 ts = &threadstats[j];
940 ts->clat_percentiles = td->o.clat_percentiles;
941 if (td->o.overwrite_plist)
942 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
944 memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list));
949 if (ts->groupid == -1) {
951 * These are per-group shared already
953 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
954 if (td->o.description)
955 strncpy(ts->description, td->o.description,
958 memset(ts->description, 0, FIO_JOBNAME_SIZE);
961 * If multiple entries in this group, this is
964 ts->thread_number = td->thread_number;
965 ts->groupid = td->groupid;
968 * first pid in group, not very useful...
972 ts->kb_base = td->o.kb_base;
973 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
974 log_info("fio: kb_base differs for jobs in group, using"
975 " %u as the base\n", ts->kb_base);
979 ts->continue_on_error = td->o.continue_on_error;
980 ts->total_err_count += td->total_err_count;
981 ts->first_error = td->first_error;
983 if (!td->error && td->o.continue_on_error &&
985 ts->error = td->first_error;
986 strcpy(ts->verror, td->verror);
987 } else if (td->error) {
988 ts->error = td->error;
989 strcpy(ts->verror, td->verror);
993 sum_thread_stats(ts, &td->ts, idx);
996 for (i = 0; i < nr_ts; i++) {
997 unsigned long long bw;
999 ts = &threadstats[i];
1000 rs = &runstats[ts->groupid];
1001 rs->kb_base = ts->kb_base;
1003 for (j = 0; j <= DDIR_WRITE; j++) {
1004 if (!ts->runtime[j])
1006 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1007 rs->min_run[j] = ts->runtime[j];
1008 if (ts->runtime[j] > rs->max_run[j])
1009 rs->max_run[j] = ts->runtime[j];
1012 if (ts->runtime[j]) {
1013 unsigned long runt = ts->runtime[j];
1014 unsigned long long kb;
1016 kb = ts->io_bytes[j] / rs->kb_base;
1017 bw = kb * 1000 / runt;
1019 if (bw < rs->min_bw[j])
1021 if (bw > rs->max_bw[j])
1024 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1028 for (i = 0; i < groupid + 1; i++) {
1032 rs->agg[0] = (rs->io_kb[0] * 1000) / rs->max_run[0];
1034 rs->agg[1] = (rs->io_kb[1] * 1000) / rs->max_run[1];
1038 * don't overwrite last signal output
1043 for (i = 0; i < nr_ts; i++) {
1044 ts = &threadstats[i];
1045 rs = &runstats[ts->groupid];
1048 fio_server_send_ts(ts, rs);
1049 else if (terse_output)
1050 show_thread_status_terse(ts, rs);
1052 show_thread_status(ts, rs);
1055 for (i = 0; i < groupid + 1; i++) {
1060 fio_server_send_gs(rs);
1061 else if (!terse_output)
1062 show_group_stats(rs);
1066 fio_server_send_du();
1067 else if (!terse_output)
1074 static void *__show_running_run_stats(void *arg)
1076 struct thread_data *td;
1077 unsigned long long *rt;
1081 rt = malloc(thread_number * sizeof(unsigned long long));
1082 fio_gettime(&tv, NULL);
1084 for_each_td(td, i) {
1085 rt[i] = mtime_since(&td->start, &tv);
1086 if (td_read(td) && td->io_bytes[DDIR_READ])
1087 td->ts.runtime[DDIR_READ] += rt[i];
1088 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1089 td->ts.runtime[DDIR_WRITE] += rt[i];
1091 update_rusage_stat(td);
1092 td->ts.io_bytes[0] = td->io_bytes[0];
1093 td->ts.io_bytes[1] = td->io_bytes[1];
1094 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1099 for_each_td(td, i) {
1100 if (td_read(td) && td->io_bytes[DDIR_READ])
1101 td->ts.runtime[DDIR_READ] -= rt[i];
1102 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1103 td->ts.runtime[DDIR_WRITE] -= rt[i];
1111 * Called from signal handler. It _should_ be safe to just run this inline
1112 * in the sig handler, but we should be disturbing the system less by just
1113 * creating a thread to do it.
1115 void show_running_run_stats(void)
1119 pthread_create(&thread, NULL, __show_running_run_stats, NULL);
1120 pthread_detach(thread);
1123 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1128 if (data > is->max_val)
1130 if (data < is->min_val)
1133 delta = val - is->mean.u.f;
1135 is->mean.u.f += delta / (is->samples + 1.0);
1136 is->S.u.f += delta * (val - is->mean.u.f);
1142 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1143 enum fio_ddir ddir, unsigned int bs,
1146 const int nr_samples = iolog->nr_samples;
1148 if (!iolog->nr_samples)
1149 iolog->avg_last = t;
1151 if (iolog->nr_samples == iolog->max_samples) {
1152 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
1154 iolog->log = realloc(iolog->log, new_size);
1155 iolog->max_samples <<= 1;
1158 iolog->log[nr_samples].val = val;
1159 iolog->log[nr_samples].time = t;
1160 iolog->log[nr_samples].ddir = ddir;
1161 iolog->log[nr_samples].bs = bs;
1162 iolog->nr_samples++;
1165 static inline void reset_io_stat(struct io_stat *ios)
1167 ios->max_val = ios->min_val = ios->samples = 0;
1168 ios->mean.u.f = ios->S.u.f = 0;
1171 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1172 unsigned long val, enum fio_ddir ddir,
1175 unsigned long elapsed, this_window;
1180 elapsed = mtime_since_now(&td->epoch);
1183 * If no time averaging, just add the log sample.
1185 if (!iolog->avg_msec) {
1186 __add_log_sample(iolog, val, ddir, bs, elapsed);
1191 * Add the sample. If the time period has passed, then
1192 * add that entry to the log and clear.
1194 add_stat_sample(&iolog->avg_window[ddir], val);
1197 * If period hasn't passed, adding the above sample is all we
1200 this_window = elapsed - iolog->avg_last;
1201 if (this_window < iolog->avg_msec)
1205 * Note an entry in the log. Use the mean from the logged samples,
1206 * making sure to properly round up. Only write a log entry if we
1207 * had actual samples done.
1209 if (iolog->avg_window[DDIR_READ].samples) {
1212 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1213 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed);
1215 if (iolog->avg_window[DDIR_WRITE].samples) {
1218 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1219 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed);
1222 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1223 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1224 iolog->avg_last = elapsed;
1227 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1229 struct io_log *iolog;
1234 iolog = agg_io_log[ddir];
1235 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
1238 static void add_clat_percentile_sample(struct thread_stat *ts,
1239 unsigned long usec, enum fio_ddir ddir)
1241 unsigned int idx = plat_val_to_idx(usec);
1242 assert(idx < FIO_IO_U_PLAT_NR);
1244 ts->io_u_plat[ddir][idx]++;
1247 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1248 unsigned long usec, unsigned int bs)
1250 struct thread_stat *ts = &td->ts;
1255 add_stat_sample(&ts->clat_stat[ddir], usec);
1258 add_log_sample(td, td->clat_log, usec, ddir, bs);
1260 if (ts->clat_percentiles)
1261 add_clat_percentile_sample(ts, usec, ddir);
1264 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1265 unsigned long usec, unsigned int bs)
1267 struct thread_stat *ts = &td->ts;
1272 add_stat_sample(&ts->slat_stat[ddir], usec);
1275 add_log_sample(td, td->slat_log, usec, ddir, bs);
1278 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1279 unsigned long usec, unsigned int bs)
1281 struct thread_stat *ts = &td->ts;
1286 add_stat_sample(&ts->lat_stat[ddir], usec);
1289 add_log_sample(td, td->lat_log, usec, ddir, bs);
1292 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1295 struct thread_stat *ts = &td->ts;
1296 unsigned long spent, rate;
1301 spent = mtime_since(&td->bw_sample_time, t);
1302 if (spent < td->o.bw_avg_time)
1306 * Compute both read and write rates for the interval.
1308 for (ddir = DDIR_READ; ddir <= DDIR_WRITE; ddir++) {
1311 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1313 continue; /* No entries for interval */
1315 rate = delta * 1000 / spent / 1024;
1316 add_stat_sample(&ts->bw_stat[ddir], rate);
1319 add_log_sample(td, td->bw_log, rate, ddir, bs);
1321 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1324 fio_gettime(&td->bw_sample_time, NULL);
1327 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
1330 struct thread_stat *ts = &td->ts;
1331 unsigned long spent, iops;
1336 spent = mtime_since(&td->iops_sample_time, t);
1337 if (spent < td->o.iops_avg_time)
1341 * Compute both read and write rates for the interval.
1343 for (ddir = DDIR_READ; ddir <= DDIR_WRITE; ddir++) {
1346 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1348 continue; /* No entries for interval */
1350 iops = (delta * 1000) / spent;
1351 add_stat_sample(&ts->iops_stat[ddir], iops);
1354 add_log_sample(td, td->iops_log, iops, ddir, 0);
1356 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1359 fio_gettime(&td->iops_sample_time, NULL);