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 static unsigned int calc_clat_percentiles(unsigned int *io_u_plat,
118 unsigned long nr, fio_fp64_t *plist,
119 unsigned int **output,
123 unsigned long sum = 0;
124 unsigned int len, i, j = 0;
125 unsigned int oval_len = 0;
126 unsigned int *ovals = NULL;
133 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
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.
145 qsort((void*)plist, len, sizeof(plist[0]), double_cmp);
148 * Calculate bucket values, note down max and min values
151 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
153 while (sum >= (plist[j].u.f / 100.0 * nr)) {
154 assert(plist[j].u.f <= 100.0);
158 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
161 ovals[j] = plat_idx_to_val(i);
162 if (ovals[j] < *minv)
164 if (ovals[j] > *maxv)
167 is_last = (j == len - 1);
180 * Find and display the p-th percentile of clat
182 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
185 unsigned int len, j = 0, minv, maxv;
187 int is_last, scale_down;
189 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
194 * We default to usecs, but if the value range is such that we
195 * should scale down to msecs, do that.
197 if (minv > 2000 && maxv > 99999) {
199 log_info(" clat percentiles (msec):\n |");
202 log_info(" clat percentiles (usec):\n |");
205 for (j = 0; j < len; j++) {
209 if (j != 0 && (j % 4) == 0)
212 /* end of the list */
213 is_last = (j == len - 1);
215 if (plist[j].u.f < 10.0)
216 sprintf(fbuf, " %2.2f", plist[j].u.f);
218 sprintf(fbuf, "%2.2f", plist[j].u.f);
221 ovals[j] = (ovals[j] + 999) / 1000;
223 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
228 if (j % 4 == 3) /* for formatting */
237 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
238 double *mean, double *dev)
240 double n = is->samples;
242 if (is->samples == 0)
248 n = (double) is->samples;
249 *mean = is->mean.u.f;
252 *dev = sqrt(is->S.u.f / (n - 1.0));
259 void show_group_stats(struct group_run_stats *rs)
261 char *p1, *p2, *p3, *p4;
262 const char *ddir_str[] = { " READ", " WRITE" };
265 log_info("\nRun status group %d (all jobs):\n", rs->groupid);
267 for (i = 0; i <= DDIR_WRITE; i++) {
268 const int i2p = is_power_of_2(rs->kb_base);
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);
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],
290 #define ts_total_io_u(ts) \
291 ((ts)->total_io_u[0] + (ts)->total_io_u[1])
293 static void stat_calc_dist(unsigned int *map, unsigned long total,
299 * Do depth distribution calculations
301 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
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])
312 static void stat_calc_lat(struct thread_stat *ts, double *dst,
313 unsigned int *src, int nr)
315 unsigned long total = ts_total_io_u(ts);
319 * Do latency distribution calculations
321 for (i = 0; i < nr; i++) {
323 dst[i] = (double) src[i] / (double) total;
325 if (dst[i] < 0.01 && src[i])
332 static void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
334 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
337 static void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
339 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
342 static void display_lat(const char *name, unsigned long min, unsigned long max,
343 double mean, double dev)
345 const char *base = "(usec)";
348 if (!usec_to_msec(&min, &max, &mean, &dev))
351 minp = num2str(min, 6, 1, 0);
352 maxp = num2str(max, 6, 1, 0);
354 log_info(" %s %s: min=%s, max=%s, avg=%5.02f,"
355 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
361 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
364 const char *ddir_str[] = { "read ", "write" };
365 unsigned long min, max, runt;
366 unsigned long long bw, iops;
368 char *io_p, *bw_p, *iops_p;
371 assert(ddir_rw(ddir));
373 if (!ts->runtime[ddir])
376 i2p = is_power_of_2(rs->kb_base);
377 runt = ts->runtime[ddir];
379 bw = (1000 * ts->io_bytes[ddir]) / runt;
380 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p);
381 bw_p = num2str(bw, 6, 1, i2p);
383 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
384 iops_p = num2str(iops, 6, 1, 0);
386 log_info(" %s: io=%sB, bw=%sB/s, iops=%s, runt=%6llumsec\n",
387 ddir_str[ddir], io_p, bw_p, iops_p,
394 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
395 display_lat("slat", min, max, mean, dev);
396 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
397 display_lat("clat", min, max, mean, dev);
398 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
399 display_lat(" lat", min, max, mean, dev);
401 if (ts->clat_percentiles) {
402 show_clat_percentiles(ts->io_u_plat[ddir],
403 ts->clat_stat[ddir].samples,
404 ts->percentile_list);
406 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
407 double p_of_agg = 100.0;
408 const char *bw_str = "KB";
411 p_of_agg = mean * 100 / (double) rs->agg[ddir];
412 if (p_of_agg > 100.0)
416 if (mean > 999999.9) {
424 log_info(" bw (%s/s) : min=%5lu, max=%5lu, per=%3.2f%%,"
425 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
426 p_of_agg, mean, dev);
430 static int show_lat(double *io_u_lat, int nr, const char **ranges,
433 int new_line = 1, i, line = 0, shown = 0;
435 for (i = 0; i < nr; i++) {
436 if (io_u_lat[i] <= 0.0)
442 log_info(" lat (%s) : ", msg);
448 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
460 static void show_lat_u(double *io_u_lat_u)
462 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
463 "250=", "500=", "750=", "1000=", };
465 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
468 static void show_lat_m(double *io_u_lat_m)
470 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
471 "250=", "500=", "750=", "1000=", "2000=",
474 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
477 static void show_latencies(struct thread_stat *ts, double *io_u_lat_u,
480 stat_calc_lat_u(ts, io_u_lat_u);
481 stat_calc_lat_m(ts, io_u_lat_m);
483 show_lat_u(io_u_lat_u);
484 show_lat_m(io_u_lat_m);
487 void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
489 double usr_cpu, sys_cpu;
490 unsigned long runtime;
491 double io_u_dist[FIO_IO_U_MAP_NR];
492 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
493 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
495 if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
496 !(ts->total_io_u[0] + ts->total_io_u[1]))
500 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
501 ts->name, ts->groupid, ts->members,
502 ts->error, (int) ts->pid);
504 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
505 ts->name, ts->groupid, ts->members,
506 ts->error, ts->verror, (int) ts->pid);
509 if (strlen(ts->description))
510 log_info(" Description : [%s]\n", ts->description);
512 if (ts->io_bytes[DDIR_READ])
513 show_ddir_status(rs, ts, DDIR_READ);
514 if (ts->io_bytes[DDIR_WRITE])
515 show_ddir_status(rs, ts, DDIR_WRITE);
517 show_latencies(ts, io_u_lat_u, io_u_lat_m);
519 runtime = ts->total_run_time;
521 double runt = (double) runtime;
523 usr_cpu = (double) ts->usr_time * 100 / runt;
524 sys_cpu = (double) ts->sys_time * 100 / runt;
530 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
531 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
533 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
534 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
535 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
536 io_u_dist[1], io_u_dist[2],
537 io_u_dist[3], io_u_dist[4],
538 io_u_dist[5], io_u_dist[6]);
540 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
541 log_info(" submit : 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 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
547 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
548 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
549 io_u_dist[1], io_u_dist[2],
550 io_u_dist[3], io_u_dist[4],
551 io_u_dist[5], io_u_dist[6]);
552 log_info(" issued : total=r=%lu/w=%lu/d=%lu,"
553 " short=r=%lu/w=%lu/d=%lu\n",
554 ts->total_io_u[0], ts->total_io_u[1],
556 ts->short_io_u[0], ts->short_io_u[1],
558 if (ts->continue_on_error) {
559 log_info(" errors : total=%lu, first_error=%d/<%s>\n",
562 strerror(ts->first_error));
566 static void show_ddir_status_terse(struct thread_stat *ts,
567 struct group_run_stats *rs, int ddir)
569 unsigned long min, max;
570 unsigned long long bw, iops;
571 unsigned int *ovals = NULL;
573 unsigned int len, minv, maxv;
576 assert(ddir_rw(ddir));
579 if (ts->runtime[ddir]) {
580 uint64_t runt = ts->runtime[ddir];
582 bw = ts->io_bytes[ddir] / runt;
583 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
586 log_info(";%llu;%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, iops,
589 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
590 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
592 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
594 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
595 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
597 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
599 if (ts->clat_percentiles) {
600 len = calc_clat_percentiles(ts->io_u_plat[ddir],
601 ts->clat_stat[ddir].samples,
602 ts->percentile_list, &ovals, &maxv,
607 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
612 log_info(";%2.2f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
615 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
616 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
618 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
623 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
624 double p_of_agg = 100.0;
627 p_of_agg = mean * 100 / (double) rs->agg[ddir];
628 if (p_of_agg > 100.0)
632 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
634 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
637 static void show_thread_status_terse_v2(struct thread_stat *ts,
638 struct group_run_stats *rs)
640 double io_u_dist[FIO_IO_U_MAP_NR];
641 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
642 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
643 double usr_cpu, sys_cpu;
647 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
648 /* Log Read Status */
649 show_ddir_status_terse(ts, rs, 0);
650 /* Log Write Status */
651 show_ddir_status_terse(ts, rs, 1);
654 if (ts->total_run_time) {
655 double runt = (double) ts->total_run_time;
657 usr_cpu = (double) ts->usr_time * 100 / runt;
658 sys_cpu = (double) ts->sys_time * 100 / runt;
664 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
667 /* Calc % distribution of IO depths, usecond, msecond latency */
668 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
669 stat_calc_lat_u(ts, io_u_lat_u);
670 stat_calc_lat_m(ts, io_u_lat_m);
672 /* Only show fixed 7 I/O depth levels*/
673 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
674 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
675 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
677 /* Microsecond latency */
678 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
679 log_info(";%3.2f%%", io_u_lat_u[i]);
680 /* Millisecond latency */
681 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
682 log_info(";%3.2f%%", io_u_lat_m[i]);
683 /* Additional output if continue_on_error set - default off*/
684 if (ts->continue_on_error)
685 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
688 /* Additional output if description is set */
690 log_info(";%s", ts->description);
695 #define FIO_TERSE_VERSION "3"
697 static void show_thread_status_terse_v3(struct thread_stat *ts,
698 struct group_run_stats *rs)
700 double io_u_dist[FIO_IO_U_MAP_NR];
701 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
702 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
703 double usr_cpu, sys_cpu;
707 log_info("%s;%s;%s;%d;%d", FIO_TERSE_VERSION, fio_version_string,
708 ts->name, ts->groupid, ts->error);
709 /* Log Read Status */
710 show_ddir_status_terse(ts, rs, 0);
711 /* Log Write Status */
712 show_ddir_status_terse(ts, rs, 1);
715 if (ts->total_run_time) {
716 double runt = (double) ts->total_run_time;
718 usr_cpu = (double) ts->usr_time * 100 / runt;
719 sys_cpu = (double) ts->sys_time * 100 / runt;
725 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
728 /* Calc % distribution of IO depths, usecond, msecond latency */
729 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
730 stat_calc_lat_u(ts, io_u_lat_u);
731 stat_calc_lat_m(ts, io_u_lat_m);
733 /* Only show fixed 7 I/O depth levels*/
734 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
735 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
736 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
738 /* Microsecond latency */
739 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
740 log_info(";%3.2f%%", io_u_lat_u[i]);
741 /* Millisecond latency */
742 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
743 log_info(";%3.2f%%", io_u_lat_m[i]);
745 /* disk util stats, if any */
748 /* Additional output if continue_on_error set - default off*/
749 if (ts->continue_on_error)
750 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
753 /* Additional output if description is set */
754 if (strlen(ts->description))
755 log_info(";%s", ts->description);
758 static void show_thread_status_terse(struct thread_stat *ts,
759 struct group_run_stats *rs)
761 if (terse_version == 2)
762 show_thread_status_terse_v2(ts, rs);
763 else if (terse_version == 3)
764 show_thread_status_terse_v3(ts, rs);
766 log_err("fio: bad terse version!? %d\n", terse_version);
769 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
773 if (src->samples == 0)
776 dst->min_val = min(dst->min_val, src->min_val);
777 dst->max_val = max(dst->max_val, src->max_val);
780 * Compute new mean and S after the merge
781 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
782 * #Parallel_algorithm>
785 mean = src->mean.u.f;
788 double delta = src->mean.u.f - dst->mean.u.f;
790 mean = ((src->mean.u.f * src->samples) +
791 (dst->mean.u.f * dst->samples)) /
792 (dst->samples + src->samples);
794 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
795 (dst->samples * src->samples) /
796 (dst->samples + src->samples);
799 dst->samples += src->samples;
800 dst->mean.u.f = mean;
804 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
808 for (i = 0; i < 2; i++) {
809 if (dst->max_run[i] < src->max_run[i])
810 dst->max_run[i] = src->max_run[i];
811 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
812 dst->min_run[i] = src->min_run[i];
813 if (dst->max_bw[i] < src->max_bw[i])
814 dst->max_bw[i] = src->max_bw[i];
815 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
816 dst->min_bw[i] = src->min_bw[i];
818 dst->io_kb[i] += src->io_kb[i];
819 dst->agg[i] += src->agg[i];
824 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
828 for (l = 0; l <= DDIR_WRITE; l++) {
829 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
830 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
831 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
832 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
834 dst->io_bytes[l] += src->io_bytes[l];
836 if (dst->runtime[l] < src->runtime[l])
837 dst->runtime[l] = src->runtime[l];
840 dst->usr_time += src->usr_time;
841 dst->sys_time += src->sys_time;
842 dst->ctx += src->ctx;
843 dst->majf += src->majf;
844 dst->minf += src->minf;
846 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
847 dst->io_u_map[k] += src->io_u_map[k];
848 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
849 dst->io_u_submit[k] += src->io_u_submit[k];
850 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
851 dst->io_u_complete[k] += src->io_u_complete[k];
852 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
853 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
854 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
855 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
857 for (k = 0; k <= 2; k++) {
858 dst->total_io_u[k] += src->total_io_u[k];
859 dst->short_io_u[k] += src->short_io_u[k];
862 for (k = 0; k <= DDIR_WRITE; k++) {
864 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
865 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
868 dst->total_run_time += src->total_run_time;
869 dst->total_submit += src->total_submit;
870 dst->total_complete += src->total_complete;
873 void init_group_run_stat(struct group_run_stats *gs)
875 memset(gs, 0, sizeof(*gs));
876 gs->min_bw[0] = gs->min_run[0] = ~0UL;
877 gs->min_bw[1] = gs->min_run[1] = ~0UL;
880 void init_thread_stat(struct thread_stat *ts)
884 memset(ts, 0, sizeof(*ts));
886 for (j = 0; j <= DDIR_WRITE; j++) {
887 ts->lat_stat[j].min_val = -1UL;
888 ts->clat_stat[j].min_val = -1UL;
889 ts->slat_stat[j].min_val = -1UL;
890 ts->bw_stat[j].min_val = -1UL;
895 void show_run_stats(void)
897 struct group_run_stats *runstats, *rs;
898 struct thread_data *td;
899 struct thread_stat *threadstats, *ts;
900 int i, j, nr_ts, last_ts, idx;
901 int kb_base_warned = 0;
903 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
905 for (i = 0; i < groupid + 1; i++)
906 init_group_run_stat(&runstats[i]);
909 * find out how many threads stats we need. if group reporting isn't
910 * enabled, it's one-per-td.
915 if (!td->o.group_reporting) {
919 if (last_ts == td->groupid)
922 last_ts = td->groupid;
926 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
928 for (i = 0; i < nr_ts; i++)
929 init_thread_stat(&threadstats[i]);
935 if (idx && (!td->o.group_reporting ||
936 (td->o.group_reporting && last_ts != td->groupid))) {
941 last_ts = td->groupid;
943 ts = &threadstats[j];
945 ts->clat_percentiles = td->o.clat_percentiles;
946 if (td->o.overwrite_plist)
947 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
949 memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list));
954 if (ts->groupid == -1) {
956 * These are per-group shared already
958 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
959 if (td->o.description)
960 strncpy(ts->description, td->o.description,
963 memset(ts->description, 0, FIO_JOBNAME_SIZE);
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]) {
1015 runt = ts->runtime[j];
1016 bw = ts->io_bytes[j] / runt;
1018 if (bw < rs->min_bw[j])
1020 if (bw > rs->max_bw[j])
1023 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1027 for (i = 0; i < groupid + 1; i++) {
1028 unsigned long max_run[2];
1031 max_run[0] = rs->max_run[0];
1032 max_run[1] = rs->max_run[1];
1035 rs->agg[0] = (rs->io_kb[0] * 1000) / max_run[0];
1037 rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1];
1041 * don't overwrite last signal output
1046 for (i = 0; i < nr_ts; i++) {
1047 ts = &threadstats[i];
1048 rs = &runstats[ts->groupid];
1051 fio_server_send_ts(ts, rs);
1052 else if (terse_output)
1053 show_thread_status_terse(ts, rs);
1055 show_thread_status(ts, rs);
1058 for (i = 0; i < groupid + 1; i++) {
1063 fio_server_send_gs(rs);
1064 else if (!terse_output)
1065 show_group_stats(rs);
1069 fio_server_send_du();
1070 else if (!terse_output)
1079 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1084 if (data > is->max_val)
1086 if (data < is->min_val)
1089 delta = val - is->mean.u.f;
1091 is->mean.u.f += delta / (is->samples + 1.0);
1092 is->S.u.f += delta * (val - is->mean.u.f);
1098 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1099 enum fio_ddir ddir, unsigned int bs,
1102 const int nr_samples = iolog->nr_samples;
1104 if (!iolog->nr_samples)
1105 iolog->avg_last = t;
1107 if (iolog->nr_samples == iolog->max_samples) {
1108 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
1110 iolog->log = realloc(iolog->log, new_size);
1111 iolog->max_samples <<= 1;
1114 iolog->log[nr_samples].val = val;
1115 iolog->log[nr_samples].time = t;
1116 iolog->log[nr_samples].ddir = ddir;
1117 iolog->log[nr_samples].bs = bs;
1118 iolog->nr_samples++;
1121 static inline void reset_io_stat(struct io_stat *ios)
1123 ios->max_val = ios->min_val = ios->samples = 0;
1124 ios->mean.u.f = ios->S.u.f = 0;
1127 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1128 unsigned long val, enum fio_ddir ddir,
1131 unsigned long elapsed, this_window;
1136 elapsed = mtime_since_now(&td->epoch);
1139 * If no time averaging, just add the log sample.
1141 if (!iolog->avg_msec) {
1142 __add_log_sample(iolog, val, ddir, bs, elapsed);
1147 * Add the sample. If the time period has passed, then
1148 * add that entry to the log and clear.
1150 add_stat_sample(&iolog->avg_window[ddir], val);
1153 * If period hasn't passed, adding the above sample is all we
1156 this_window = elapsed - iolog->avg_last;
1157 if (this_window < iolog->avg_msec)
1161 * Note an entry in the log. Use the mean from the logged samples,
1162 * making sure to properly round up. Only write a log entry if we
1163 * had actual samples done.
1165 if (iolog->avg_window[DDIR_READ].samples) {
1168 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1169 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed);
1171 if (iolog->avg_window[DDIR_WRITE].samples) {
1174 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1175 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed);
1178 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1179 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1180 iolog->avg_last = elapsed;
1183 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1185 struct io_log *iolog;
1190 iolog = agg_io_log[ddir];
1191 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
1194 static void add_clat_percentile_sample(struct thread_stat *ts,
1195 unsigned long usec, enum fio_ddir ddir)
1197 unsigned int idx = plat_val_to_idx(usec);
1198 assert(idx < FIO_IO_U_PLAT_NR);
1200 ts->io_u_plat[ddir][idx]++;
1203 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1204 unsigned long usec, unsigned int bs)
1206 struct thread_stat *ts = &td->ts;
1211 add_stat_sample(&ts->clat_stat[ddir], usec);
1214 add_log_sample(td, td->clat_log, usec, ddir, bs);
1216 if (ts->clat_percentiles)
1217 add_clat_percentile_sample(ts, usec, ddir);
1220 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1221 unsigned long usec, unsigned int bs)
1223 struct thread_stat *ts = &td->ts;
1228 add_stat_sample(&ts->slat_stat[ddir], usec);
1231 add_log_sample(td, td->slat_log, usec, ddir, bs);
1234 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1235 unsigned long usec, unsigned int bs)
1237 struct thread_stat *ts = &td->ts;
1242 add_stat_sample(&ts->lat_stat[ddir], usec);
1245 add_log_sample(td, td->lat_log, usec, ddir, bs);
1248 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1251 struct thread_stat *ts = &td->ts;
1252 unsigned long spent, rate;
1257 spent = mtime_since(&td->bw_sample_time, t);
1258 if (spent < td->o.bw_avg_time)
1262 * Compute both read and write rates for the interval.
1264 for (ddir = DDIR_READ; ddir <= DDIR_WRITE; ddir++) {
1267 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1269 continue; /* No entries for interval */
1271 rate = delta * 1000 / spent / 1024;
1272 add_stat_sample(&ts->bw_stat[ddir], rate);
1275 add_log_sample(td, td->bw_log, rate, ddir, bs);
1277 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1280 fio_gettime(&td->bw_sample_time, NULL);
1283 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
1286 struct thread_stat *ts = &td->ts;
1287 unsigned long spent, iops;
1292 spent = mtime_since(&td->iops_sample_time, t);
1293 if (spent < td->o.iops_avg_time)
1297 * Compute both read and write rates for the interval.
1299 for (ddir = DDIR_READ; ddir <= DDIR_WRITE; ddir++) {
1302 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1304 continue; /* No entries for interval */
1306 iops = (delta * 1000) / spent;
1307 add_stat_sample(&ts->iops_stat[ddir], iops);
1310 add_log_sample(td, td->iops_log, iops, ddir, 0);
1312 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1315 fio_gettime(&td->iops_sample_time, NULL);