13 * Cheesy number->string conversion, complete with carry rounding error.
15 static char *num2str(unsigned long num, int maxlen, int base, int pow2)
17 char postfix[] = { ' ', 'K', 'M', 'G', 'P', 'E' };
18 unsigned int thousand;
29 for (i = 0; base > 1; i++)
35 len = sprintf(buf, "%'lu", num);
38 buf[len] = postfix[i];
44 if ((num % thousand) >= (thousand / 2))
55 void update_rusage_stat(struct thread_data *td)
57 struct thread_stat *ts = &td->ts;
59 getrusage(RUSAGE_SELF, &ts->ru_end);
61 ts->usr_time += mtime_since(&ts->ru_start.ru_utime,
62 &ts->ru_end.ru_utime);
63 ts->sys_time += mtime_since(&ts->ru_start.ru_stime,
64 &ts->ru_end.ru_stime);
65 ts->ctx += ts->ru_end.ru_nvcsw + ts->ru_end.ru_nivcsw
66 - (ts->ru_start.ru_nvcsw + ts->ru_start.ru_nivcsw);
67 ts->minf += ts->ru_end.ru_minflt - ts->ru_start.ru_minflt;
68 ts->majf += ts->ru_end.ru_majflt - ts->ru_start.ru_majflt;
70 memcpy(&ts->ru_start, &ts->ru_end, sizeof(ts->ru_end));
73 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
74 double *mean, double *dev)
76 double n = is->samples;
84 n = (double) is->samples;
88 *dev = sqrt(is->S / (n - 1.0));
95 static void show_group_stats(struct group_run_stats *rs, int id)
97 char *p1, *p2, *p3, *p4;
98 const char *ddir_str[] = { " READ", " WRITE" };
101 log_info("\nRun status group %d (all jobs):\n", id);
103 for (i = 0; i <= DDIR_WRITE; i++) {
107 p1 = num2str(rs->io_kb[i], 6, 1024, 1);
108 p2 = num2str(rs->agg[i], 6, 1024, 1);
109 p3 = num2str(rs->min_bw[i], 6, 1024, 1);
110 p4 = num2str(rs->max_bw[i], 6, 1024, 1);
112 log_info("%s: io=%siB, aggrb=%siB/s, minb=%siB/s, maxb=%siB/s,"
113 " mint=%llumsec, maxt=%llumsec\n", ddir_str[i], p1, p2,
114 p3, p4, rs->min_run[i],
124 #define ts_total_io_u(ts) \
125 ((ts)->total_io_u[0] + (ts)->total_io_u[1])
127 static void stat_calc_dist(unsigned int *map, unsigned long total,
133 * Do depth distribution calculations
135 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
137 io_u_dist[i] = (double) map[i] / (double) total;
138 io_u_dist[i] *= 100.0;
139 if (io_u_dist[i] < 0.1 && map[i])
146 static void stat_calc_lat(struct thread_stat *ts, double *dst,
147 unsigned int *src, int nr)
149 unsigned long total = ts_total_io_u(ts);
153 * Do latency distribution calculations
155 for (i = 0; i < nr; i++) {
157 dst[i] = (double) src[i] / (double) total;
159 if (dst[i] < 0.01 && src[i])
166 static void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
168 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
171 static void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
173 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
176 static int usec_to_msec(unsigned long *min, unsigned long *max, double *mean,
179 if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) {
190 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
193 const char *ddir_str[] = { "read ", "write" };
194 unsigned long min, max, runt;
195 unsigned long long bw, iops;
197 char *io_p, *bw_p, *iops_p;
199 if (!ts->runtime[ddir])
202 runt = ts->runtime[ddir];
204 bw = (1000 * ts->io_bytes[ddir]) / runt;
205 io_p = num2str(ts->io_bytes[ddir] >> 10, 6, 1024, 1);
206 bw_p = num2str(bw >> 10, 6, 1024, 1);
208 iops = (1000 * ts->total_io_u[ddir]) / runt;
209 iops_p = num2str(iops, 6, 1, 0);
211 log_info(" %s: io=%siB, bw=%siB/s, iops=%s, runt=%6lumsec\n",
212 ddir_str[ddir], io_p, bw_p, iops_p,
219 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
220 const char *base = "(usec)";
223 if (!usec_to_msec(&min, &max, &mean, &dev))
226 minp = num2str(min, 6, 1, 0);
227 maxp = num2str(max, 6, 1, 0);
229 log_info(" slat %s: min=%s, max=%s, avg=%5.02f,"
230 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
235 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
236 const char *base = "(usec)";
239 if (!usec_to_msec(&min, &max, &mean, &dev))
242 minp = num2str(min, 6, 1, 0);
243 maxp = num2str(max, 6, 1, 0);
245 log_info(" clat %s: min=%s, max=%s, avg=%5.02f,"
246 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
251 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
254 p_of_agg = mean * 100 / (double) rs->agg[ddir];
255 log_info(" bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%,"
256 " avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg,
261 static void show_lat(double *io_u_lat, int nr, const char **ranges,
264 int new_line = 1, i, line = 0;
266 for (i = 0; i < nr; i++) {
267 if (io_u_lat[i] <= 0.0)
272 log_info(" lat (%s): ", msg);
278 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
285 static void show_lat_u(double *io_u_lat_u)
287 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
288 "250=", "500=", "750=", "1000=", };
290 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
293 static void show_lat_m(double *io_u_lat_m)
295 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
296 "250=", "500=", "750=", "1000=", "2000=",
299 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
302 static void show_latencies(double *io_u_lat_u, double *io_u_lat_m)
304 show_lat_u(io_u_lat_u);
306 show_lat_m(io_u_lat_m);
310 static void show_thread_status(struct thread_stat *ts,
311 struct group_run_stats *rs)
313 double usr_cpu, sys_cpu;
314 unsigned long runtime;
315 double io_u_dist[FIO_IO_U_MAP_NR];
316 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
317 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
319 if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
320 !(ts->total_io_u[0] + ts->total_io_u[1]))
324 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
325 ts->name, ts->groupid, ts->members,
326 ts->error, (int) ts->pid);
328 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
329 ts->name, ts->groupid, ts->members,
330 ts->error, ts->verror, (int) ts->pid);
334 log_info(" Description : [%s]\n", ts->description);
336 if (ts->io_bytes[DDIR_READ])
337 show_ddir_status(rs, ts, DDIR_READ);
338 if (ts->io_bytes[DDIR_WRITE])
339 show_ddir_status(rs, ts, DDIR_WRITE);
341 runtime = ts->total_run_time;
343 double runt = (double) runtime;
345 usr_cpu = (double) ts->usr_time * 100 / runt;
346 sys_cpu = (double) ts->sys_time * 100 / runt;
352 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
353 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
355 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
356 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
357 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
358 io_u_dist[1], io_u_dist[2],
359 io_u_dist[3], io_u_dist[4],
360 io_u_dist[5], io_u_dist[6]);
362 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
363 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
364 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
365 io_u_dist[1], io_u_dist[2],
366 io_u_dist[3], io_u_dist[4],
367 io_u_dist[5], io_u_dist[6]);
368 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
369 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
370 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
371 io_u_dist[1], io_u_dist[2],
372 io_u_dist[3], io_u_dist[4],
373 io_u_dist[5], io_u_dist[6]);
374 log_info(" issued r/w: total=%lu/%lu, short=%lu/%lu\n",
375 ts->total_io_u[0], ts->total_io_u[1],
376 ts->short_io_u[0], ts->short_io_u[1]);
377 stat_calc_lat_u(ts, io_u_lat_u);
378 stat_calc_lat_m(ts, io_u_lat_m);
379 show_latencies(io_u_lat_u, io_u_lat_m);
382 static void show_ddir_status_terse(struct thread_stat *ts,
383 struct group_run_stats *rs, int ddir)
385 unsigned long min, max;
386 unsigned long long bw;
390 if (ts->runtime[ddir])
391 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
393 log_info(";%llu;%llu;%lu", ts->io_bytes[ddir] >> 10, bw,
396 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
397 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
399 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
401 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
402 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
404 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
406 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
409 p_of_agg = mean * 100 / (double) rs->agg[ddir];
410 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
412 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
416 static void show_thread_status_terse(struct thread_stat *ts,
417 struct group_run_stats *rs)
419 double io_u_dist[FIO_IO_U_MAP_NR];
420 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
421 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
422 double usr_cpu, sys_cpu;
425 log_info("%s;%d;%d", ts->name, ts->groupid, ts->error);
427 show_ddir_status_terse(ts, rs, 0);
428 show_ddir_status_terse(ts, rs, 1);
430 if (ts->total_run_time) {
431 double runt = (double) ts->total_run_time;
433 usr_cpu = (double) ts->usr_time * 100 / runt;
434 sys_cpu = (double) ts->sys_time * 100 / runt;
440 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
443 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
444 stat_calc_lat_u(ts, io_u_lat_u);
445 stat_calc_lat_m(ts, io_u_lat_m);
447 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
448 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
449 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
451 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
452 log_info(";%3.2f%%", io_u_lat_u[i]);
453 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
454 log_info(";%3.2f%%", io_u_lat_m[i]);
458 log_info(";%s", ts->description);
463 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
467 dst->min_val = min(dst->min_val, src->min_val);
468 dst->max_val = max(dst->max_val, src->max_val);
469 dst->samples += src->samples;
472 * Needs a new method for calculating stddev, we cannot just
473 * average them we do below for nr > 1
479 mean = ((src->mean * (double) (nr - 1))
480 + dst->mean) / ((double) nr);
481 S = ((src->S * (double) (nr - 1)) + dst->S) / ((double) nr);
488 void show_run_stats(void)
490 struct group_run_stats *runstats, *rs;
491 struct thread_data *td;
492 struct thread_stat *threadstats, *ts;
493 int i, j, k, l, nr_ts, last_ts, idx;
495 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
497 for (i = 0; i < groupid + 1; i++) {
500 memset(rs, 0, sizeof(*rs));
501 rs->min_bw[0] = rs->min_run[0] = ~0UL;
502 rs->min_bw[1] = rs->min_run[1] = ~0UL;
506 * find out how many threads stats we need. if group reporting isn't
507 * enabled, it's one-per-td.
512 if (!td->o.group_reporting) {
516 if (last_ts == td->groupid)
519 last_ts = td->groupid;
523 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
525 for (i = 0; i < nr_ts; i++) {
526 ts = &threadstats[i];
528 memset(ts, 0, sizeof(*ts));
529 for (j = 0; j <= DDIR_WRITE; j++) {
530 ts->clat_stat[j].min_val = -1UL;
531 ts->slat_stat[j].min_val = -1UL;
532 ts->bw_stat[j].min_val = -1UL;
541 if (idx && (!td->o.group_reporting ||
542 (td->o.group_reporting && last_ts != td->groupid))) {
547 last_ts = td->groupid;
549 ts = &threadstats[j];
554 if (ts->groupid == -1) {
556 * These are per-group shared already
558 ts->name = td->o.name;
559 ts->description = td->o.description;
560 ts->groupid = td->groupid;
563 * first pid in group, not very useful...
568 if (td->error && !ts->error) {
569 ts->error = td->error;
570 ts->verror = td->verror;
573 for (l = 0; l <= DDIR_WRITE; l++) {
574 sum_stat(&ts->clat_stat[l], &td->ts.clat_stat[l], idx);
575 sum_stat(&ts->slat_stat[l], &td->ts.slat_stat[l], idx);
576 sum_stat(&ts->bw_stat[l], &td->ts.bw_stat[l], idx);
578 ts->stat_io_bytes[l] += td->ts.stat_io_bytes[l];
579 ts->io_bytes[l] += td->ts.io_bytes[l];
581 if (ts->runtime[l] < td->ts.runtime[l])
582 ts->runtime[l] = td->ts.runtime[l];
585 ts->usr_time += td->ts.usr_time;
586 ts->sys_time += td->ts.sys_time;
587 ts->ctx += td->ts.ctx;
588 ts->majf += td->ts.majf;
589 ts->minf += td->ts.minf;
591 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
592 ts->io_u_map[k] += td->ts.io_u_map[k];
593 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
594 ts->io_u_submit[k] += td->ts.io_u_submit[k];
595 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
596 ts->io_u_complete[k] += td->ts.io_u_complete[k];
597 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
598 ts->io_u_lat_u[k] += td->ts.io_u_lat_u[k];
599 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
600 ts->io_u_lat_m[k] += td->ts.io_u_lat_m[k];
603 for (k = 0; k <= DDIR_WRITE; k++) {
604 ts->total_io_u[k] += td->ts.total_io_u[k];
605 ts->short_io_u[k] += td->ts.short_io_u[k];
608 ts->total_run_time += td->ts.total_run_time;
609 ts->total_submit += td->ts.total_submit;
610 ts->total_complete += td->ts.total_complete;
613 for (i = 0; i < nr_ts; i++) {
614 unsigned long long bw;
616 ts = &threadstats[i];
617 rs = &runstats[ts->groupid];
619 for (j = 0; j <= DDIR_WRITE; j++) {
622 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
623 rs->min_run[j] = ts->runtime[j];
624 if (ts->runtime[j] > rs->max_run[j])
625 rs->max_run[j] = ts->runtime[j];
628 if (ts->runtime[j]) {
631 runt = ts->runtime[j] * 1024 / 1000;
632 bw = ts->io_bytes[j] / runt;
634 if (bw < rs->min_bw[j])
636 if (bw > rs->max_bw[j])
639 rs->io_kb[j] += ts->io_bytes[j] >> 10;
643 for (i = 0; i < groupid + 1; i++) {
644 unsigned long max_run[2];
647 max_run[0] = rs->max_run[0] * 1024 / 1000;
648 max_run[1] = rs->max_run[1] * 1024 / 1000;
651 rs->agg[0] = (rs->io_kb[0]*1024) / max_run[0];
653 rs->agg[1] = (rs->io_kb[1]*1024) / max_run[1];
657 * don't overwrite last signal output
662 for (i = 0; i < nr_ts; i++) {
663 ts = &threadstats[i];
664 rs = &runstats[ts->groupid];
667 show_thread_status_terse(ts, rs);
669 show_thread_status(ts, rs);
673 for (i = 0; i < groupid + 1; i++)
674 show_group_stats(&runstats[i], i);
683 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
688 if (data > is->max_val)
690 if (data < is->min_val)
693 delta = val - is->mean;
695 is->mean += delta / (is->samples + 1.0);
696 is->S += delta * (val - is->mean);
702 static void __add_log_sample(struct io_log *iolog, unsigned long val,
703 enum fio_ddir ddir, unsigned long time)
705 if (iolog->nr_samples == iolog->max_samples) {
706 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
708 iolog->log = realloc(iolog->log, new_size);
709 iolog->max_samples <<= 1;
712 iolog->log[iolog->nr_samples].val = val;
713 iolog->log[iolog->nr_samples].time = time;
714 iolog->log[iolog->nr_samples].ddir = ddir;
718 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
719 unsigned long val, enum fio_ddir ddir)
721 __add_log_sample(iolog, val, ddir, mtime_since_now(&td->epoch));
724 void add_agg_sample(unsigned long val, enum fio_ddir ddir)
726 struct io_log *iolog = agg_io_log[ddir];
728 __add_log_sample(iolog, val, ddir, mtime_since_genesis());
731 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
734 struct thread_stat *ts = &td->ts;
736 add_stat_sample(&ts->clat_stat[ddir], usec);
739 add_log_sample(td, ts->clat_log, usec, ddir);
742 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
745 struct thread_stat *ts = &td->ts;
747 add_stat_sample(&ts->slat_stat[ddir], usec);
750 add_log_sample(td, ts->slat_log, usec, ddir);
753 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir,
756 struct thread_stat *ts = &td->ts;
757 unsigned long spent = mtime_since(&ts->stat_sample_time[ddir], t);
760 if (spent < td->o.bw_avg_time)
763 rate = (td->this_io_bytes[ddir] - ts->stat_io_bytes[ddir]) * 1000 / spent / 1024;
764 add_stat_sample(&ts->bw_stat[ddir], rate);
767 add_log_sample(td, ts->bw_log, rate, ddir);
769 fio_gettime(&ts->stat_sample_time[ddir], NULL);
770 ts->stat_io_bytes[ddir] = td->this_io_bytes[ddir];