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, 1000, 1);
108 p2 = num2str(rs->agg[i], 6, 1000, 1);
109 p3 = num2str(rs->min_bw[i], 6, 1000, 1);
110 p4 = num2str(rs->max_bw[i], 6, 1000, 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;
195 unsigned long long bw, iops;
197 char *io_p, *bw_p, *iops_p;
199 if (!ts->runtime[ddir])
202 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
203 iops = (1000 * ts->total_io_u[ddir]) / ts->runtime[ddir];
204 io_p = num2str(ts->io_bytes[ddir] >> 10, 6, 1000, 1);
205 bw_p = num2str(bw, 6, 1000, 1);
206 iops_p = num2str(iops, 6, 1, 0);
208 log_info(" %s: io=%siB, bw=%siB/s, iops=%s, runt=%6lumsec\n",
209 ddir_str[ddir], io_p, bw_p, iops_p,
216 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
217 const char *base = "(usec)";
220 if (!usec_to_msec(&min, &max, &mean, &dev))
223 minp = num2str(min, 6, 1, 0);
224 maxp = num2str(max, 6, 1, 0);
226 log_info(" slat %s: min=%s, max=%s, avg=%5.02f,"
227 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
232 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
233 const char *base = "(usec)";
236 if (!usec_to_msec(&min, &max, &mean, &dev))
239 minp = num2str(min, 6, 1, 0);
240 maxp = num2str(max, 6, 1, 0);
242 log_info(" clat %s: min=%s, max=%s, avg=%5.02f,"
243 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
248 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
251 p_of_agg = mean * 100 / (double) rs->agg[ddir];
252 log_info(" bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%,"
253 " avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg,
258 static void show_lat(double *io_u_lat, int nr, const char **ranges,
261 int new_line = 1, i, line = 0;
263 for (i = 0; i < nr; i++) {
264 if (io_u_lat[i] <= 0.0)
269 log_info(" lat (%s): ", msg);
275 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
282 static void show_lat_u(double *io_u_lat_u)
284 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
285 "250=", "500=", "750=", "1000=", };
287 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
290 static void show_lat_m(double *io_u_lat_m)
292 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
293 "250=", "500=", "750=", "1000=", "2000=",
296 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
299 static void show_latencies(double *io_u_lat_u, double *io_u_lat_m)
301 show_lat_u(io_u_lat_u);
303 show_lat_m(io_u_lat_m);
307 static void show_thread_status(struct thread_stat *ts,
308 struct group_run_stats *rs)
310 double usr_cpu, sys_cpu;
311 unsigned long runtime;
312 double io_u_dist[FIO_IO_U_MAP_NR];
313 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
314 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
316 if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
317 !(ts->total_io_u[0] + ts->total_io_u[1]))
321 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
322 ts->name, ts->groupid, ts->members,
323 ts->error, (int) ts->pid);
325 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
326 ts->name, ts->groupid, ts->members,
327 ts->error, ts->verror, (int) ts->pid);
331 log_info(" Description : [%s]\n", ts->description);
333 if (ts->io_bytes[DDIR_READ])
334 show_ddir_status(rs, ts, DDIR_READ);
335 if (ts->io_bytes[DDIR_WRITE])
336 show_ddir_status(rs, ts, DDIR_WRITE);
338 runtime = ts->total_run_time;
340 double runt = (double) runtime;
342 usr_cpu = (double) ts->usr_time * 100 / runt;
343 sys_cpu = (double) ts->sys_time * 100 / runt;
349 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
350 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
352 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
353 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
354 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
355 io_u_dist[1], io_u_dist[2],
356 io_u_dist[3], io_u_dist[4],
357 io_u_dist[5], io_u_dist[6]);
359 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
360 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
361 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
362 io_u_dist[1], io_u_dist[2],
363 io_u_dist[3], io_u_dist[4],
364 io_u_dist[5], io_u_dist[6]);
365 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
366 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
367 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
368 io_u_dist[1], io_u_dist[2],
369 io_u_dist[3], io_u_dist[4],
370 io_u_dist[5], io_u_dist[6]);
371 log_info(" issued r/w: total=%lu/%lu, short=%lu/%lu\n",
372 ts->total_io_u[0], ts->total_io_u[1],
373 ts->short_io_u[0], ts->short_io_u[1]);
374 stat_calc_lat_u(ts, io_u_lat_u);
375 stat_calc_lat_m(ts, io_u_lat_m);
376 show_latencies(io_u_lat_u, io_u_lat_m);
379 static void show_ddir_status_terse(struct thread_stat *ts,
380 struct group_run_stats *rs, int ddir)
382 unsigned long min, max;
383 unsigned long long bw;
387 if (ts->runtime[ddir])
388 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
390 log_info(";%llu;%llu;%lu", ts->io_bytes[ddir] >> 10, bw,
393 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
394 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
396 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
398 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
399 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
401 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
403 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
406 p_of_agg = mean * 100 / (double) rs->agg[ddir];
407 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
409 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
413 static void show_thread_status_terse(struct thread_stat *ts,
414 struct group_run_stats *rs)
416 double io_u_dist[FIO_IO_U_MAP_NR];
417 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
418 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
419 double usr_cpu, sys_cpu;
422 log_info("%s;%d;%d", ts->name, ts->groupid, ts->error);
424 show_ddir_status_terse(ts, rs, 0);
425 show_ddir_status_terse(ts, rs, 1);
427 if (ts->total_run_time) {
428 double runt = (double) ts->total_run_time;
430 usr_cpu = (double) ts->usr_time * 100 / runt;
431 sys_cpu = (double) ts->sys_time * 100 / runt;
437 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
440 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
441 stat_calc_lat_u(ts, io_u_lat_u);
442 stat_calc_lat_m(ts, io_u_lat_m);
444 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
445 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
446 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
448 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
449 log_info(";%3.2f%%", io_u_lat_u[i]);
450 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
451 log_info(";%3.2f%%", io_u_lat_m[i]);
455 log_info(";%s", ts->description);
460 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
464 dst->min_val = min(dst->min_val, src->min_val);
465 dst->max_val = max(dst->max_val, src->max_val);
466 dst->samples += src->samples;
469 * Needs a new method for calculating stddev, we cannot just
470 * average them we do below for nr > 1
476 mean = ((src->mean * (double) (nr - 1))
477 + dst->mean) / ((double) nr);
478 S = ((src->S * (double) (nr - 1)) + dst->S) / ((double) nr);
485 void show_run_stats(void)
487 struct group_run_stats *runstats, *rs;
488 struct thread_data *td;
489 struct thread_stat *threadstats, *ts;
490 int i, j, k, l, nr_ts, last_ts, idx;
492 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
494 for (i = 0; i < groupid + 1; i++) {
497 memset(rs, 0, sizeof(*rs));
498 rs->min_bw[0] = rs->min_run[0] = ~0UL;
499 rs->min_bw[1] = rs->min_run[1] = ~0UL;
503 * find out how many threads stats we need. if group reporting isn't
504 * enabled, it's one-per-td.
509 if (!td->o.group_reporting) {
513 if (last_ts == td->groupid)
516 last_ts = td->groupid;
520 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
522 for (i = 0; i < nr_ts; i++) {
523 ts = &threadstats[i];
525 memset(ts, 0, sizeof(*ts));
526 for (j = 0; j <= DDIR_WRITE; j++) {
527 ts->clat_stat[j].min_val = -1UL;
528 ts->slat_stat[j].min_val = -1UL;
529 ts->bw_stat[j].min_val = -1UL;
538 if (idx && (!td->o.group_reporting ||
539 (td->o.group_reporting && last_ts != td->groupid))) {
544 last_ts = td->groupid;
546 ts = &threadstats[j];
551 if (ts->groupid == -1) {
553 * These are per-group shared already
555 ts->name = td->o.name;
556 ts->description = td->o.description;
557 ts->groupid = td->groupid;
560 * first pid in group, not very useful...
565 if (td->error && !ts->error) {
566 ts->error = td->error;
567 ts->verror = td->verror;
570 for (l = 0; l <= DDIR_WRITE; l++) {
571 sum_stat(&ts->clat_stat[l], &td->ts.clat_stat[l], idx);
572 sum_stat(&ts->slat_stat[l], &td->ts.slat_stat[l], idx);
573 sum_stat(&ts->bw_stat[l], &td->ts.bw_stat[l], idx);
575 ts->stat_io_bytes[l] += td->ts.stat_io_bytes[l];
576 ts->io_bytes[l] += td->ts.io_bytes[l];
578 if (ts->runtime[l] < td->ts.runtime[l])
579 ts->runtime[l] = td->ts.runtime[l];
582 ts->usr_time += td->ts.usr_time;
583 ts->sys_time += td->ts.sys_time;
584 ts->ctx += td->ts.ctx;
585 ts->majf += td->ts.majf;
586 ts->minf += td->ts.minf;
588 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
589 ts->io_u_map[k] += td->ts.io_u_map[k];
590 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
591 ts->io_u_submit[k] += td->ts.io_u_submit[k];
592 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
593 ts->io_u_complete[k] += td->ts.io_u_complete[k];
594 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
595 ts->io_u_lat_u[k] += td->ts.io_u_lat_u[k];
596 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
597 ts->io_u_lat_m[k] += td->ts.io_u_lat_m[k];
600 for (k = 0; k <= DDIR_WRITE; k++) {
601 ts->total_io_u[k] += td->ts.total_io_u[k];
602 ts->short_io_u[k] += td->ts.short_io_u[k];
605 ts->total_run_time += td->ts.total_run_time;
606 ts->total_submit += td->ts.total_submit;
607 ts->total_complete += td->ts.total_complete;
610 for (i = 0; i < nr_ts; i++) {
611 unsigned long long bw;
613 ts = &threadstats[i];
614 rs = &runstats[ts->groupid];
616 for (j = 0; j <= DDIR_WRITE; j++) {
619 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
620 rs->min_run[j] = ts->runtime[j];
621 if (ts->runtime[j] > rs->max_run[j])
622 rs->max_run[j] = ts->runtime[j];
627 / (unsigned long long) ts->runtime[j];
628 if (bw < rs->min_bw[j])
630 if (bw > rs->max_bw[j])
633 rs->io_kb[j] += ts->io_bytes[j] >> 10;
637 for (i = 0; i < groupid + 1; i++) {
641 rs->agg[0] = (rs->io_kb[0]*1024) / rs->max_run[0];
643 rs->agg[1] = (rs->io_kb[1]*1024) / rs->max_run[1];
647 * don't overwrite last signal output
652 for (i = 0; i < nr_ts; i++) {
653 ts = &threadstats[i];
654 rs = &runstats[ts->groupid];
657 show_thread_status_terse(ts, rs);
659 show_thread_status(ts, rs);
663 for (i = 0; i < groupid + 1; i++)
664 show_group_stats(&runstats[i], i);
673 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
678 if (data > is->max_val)
680 if (data < is->min_val)
683 delta = val - is->mean;
685 is->mean += delta / (is->samples + 1.0);
686 is->S += delta * (val - is->mean);
692 static void __add_log_sample(struct io_log *iolog, unsigned long val,
693 enum fio_ddir ddir, unsigned long time)
695 if (iolog->nr_samples == iolog->max_samples) {
696 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
698 iolog->log = realloc(iolog->log, new_size);
699 iolog->max_samples <<= 1;
702 iolog->log[iolog->nr_samples].val = val;
703 iolog->log[iolog->nr_samples].time = time;
704 iolog->log[iolog->nr_samples].ddir = ddir;
708 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
709 unsigned long val, enum fio_ddir ddir)
711 __add_log_sample(iolog, val, ddir, mtime_since_now(&td->epoch));
714 void add_agg_sample(unsigned long val, enum fio_ddir ddir)
716 struct io_log *iolog = agg_io_log[ddir];
718 __add_log_sample(iolog, val, ddir, mtime_since_genesis());
721 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
724 struct thread_stat *ts = &td->ts;
726 add_stat_sample(&ts->clat_stat[ddir], usec);
729 add_log_sample(td, ts->clat_log, usec, ddir);
732 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
735 struct thread_stat *ts = &td->ts;
737 add_stat_sample(&ts->slat_stat[ddir], usec);
740 add_log_sample(td, ts->slat_log, usec, ddir);
743 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir,
746 struct thread_stat *ts = &td->ts;
747 unsigned long spent = mtime_since(&ts->stat_sample_time[ddir], t);
750 if (spent < td->o.bw_avg_time)
753 rate = (td->this_io_bytes[ddir] - ts->stat_io_bytes[ddir]) / spent;
754 add_stat_sample(&ts->bw_stat[ddir], rate);
757 add_log_sample(td, ts->bw_log, rate, ddir);
759 fio_gettime(&ts->stat_sample_time[ddir], NULL);
760 ts->stat_io_bytes[ddir] = td->this_io_bytes[ddir];