projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
README: update for newer client job file argument format
[fio.git] / stat.c
... / ...
CommitLineData
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 "lib/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
117unsigned 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)
120{
121 unsigned long sum = 0;
122 unsigned int len, i, j = 0;
123 unsigned int oval_len = 0;
124 unsigned int *ovals = NULL;
125 int is_last;
126
127 *minv = -1U;
128 *maxv = 0;
129
130 len = 0;
131 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
132 len++;
133
134 if (!len)
135 return 0;
136
137 /*
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.
141 */
142 if (len > 1)
143 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
144
145 /*
146 * Calculate bucket values, note down max and min values
147 */
148 is_last = 0;
149 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
150 sum += io_u_plat[i];
151 while (sum >= (plist[j].u.f / 100.0 * nr)) {
152 assert(plist[j].u.f <= 100.0);
153
154 if (j == oval_len) {
155 oval_len += 100;
156 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
157 }
158
159 ovals[j] = plat_idx_to_val(i);
160 if (ovals[j] < *minv)
161 *minv = ovals[j];
162 if (ovals[j] > *maxv)
163 *maxv = ovals[j];
164
165 is_last = (j == len - 1);
166 if (is_last)
167 break;
168
169 j++;
170 }
171 }
172
173 *output = ovals;
174 return len;
175}
176
177/*
178 * Find and display the p-th percentile of clat
179 */
180static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
181 fio_fp64_t *plist)
182{
183 unsigned int len, j = 0, minv, maxv;
184 unsigned int *ovals;
185 int is_last, scale_down;
186
187 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
188 if (!len)
189 goto out;
190
191 /*
192 * We default to usecs, but if the value range is such that we
193 * should scale down to msecs, do that.
194 */
195 if (minv > 2000 && maxv > 99999) {
196 scale_down = 1;
197 log_info(" clat percentiles (msec):\n |");
198 } else {
199 scale_down = 0;
200 log_info(" clat percentiles (usec):\n |");
201 }
202
203 for (j = 0; j < len; j++) {
204 char fbuf[8];
205
206 /* for formatting */
207 if (j != 0 && (j % 4) == 0)
208 log_info(" |");
209
210 /* end of the list */
211 is_last = (j == len - 1);
212
213 if (plist[j].u.f < 10.0)
214 sprintf(fbuf, " %2.2f", plist[j].u.f);
215 else
216 sprintf(fbuf, "%2.2f", plist[j].u.f);
217
218 if (scale_down)
219 ovals[j] = (ovals[j] + 999) / 1000;
220
221 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
222
223 if (is_last)
224 break;
225
226 if (j % 4 == 3) /* for formatting */
227 log_info("\n");
228 }
229
230out:
231 if (ovals)
232 free(ovals);
233}
234
235int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
236 double *mean, double *dev)
237{
238 double n = is->samples;
239
240 if (is->samples == 0)
241 return 0;
242
243 *min = is->min_val;
244 *max = is->max_val;
245
246 n = (double) is->samples;
247 *mean = is->mean.u.f;
248
249 if (n > 1.0)
250 *dev = sqrt(is->S.u.f / (n - 1.0));
251 else
252 *dev = 0;
253
254 return 1;
255}
256
257void show_group_stats(struct group_run_stats *rs)
258{
259 char *p1, *p2, *p3, *p4;
260 const char *ddir_str[] = { " READ", " WRITE" };
261 int i;
262
263 log_info("\nRun status group %d (all jobs):\n", rs->groupid);
264
265 for (i = 0; i <= DDIR_WRITE; i++) {
266 const int i2p = is_power_of_2(rs->kb_base);
267
268 if (!rs->max_run[i])
269 continue;
270
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);
275
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],
279 rs->max_run[i]);
280
281 free(p1);
282 free(p2);
283 free(p3);
284 free(p4);
285 }
286}
287
288void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
289{
290 int i;
291
292 /*
293 * Do depth distribution calculations
294 */
295 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
296 if (total) {
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])
300 io_u_dist[i] = 0.1;
301 } else
302 io_u_dist[i] = 0.0;
303 }
304}
305
306static void stat_calc_lat(struct thread_stat *ts, double *dst,
307 unsigned int *src, int nr)
308{
309 unsigned long total = ts_total_io_u(ts);
310 int i;
311
312 /*
313 * Do latency distribution calculations
314 */
315 for (i = 0; i < nr; i++) {
316 if (total) {
317 dst[i] = (double) src[i] / (double) total;
318 dst[i] *= 100.0;
319 if (dst[i] < 0.01 && src[i])
320 dst[i] = 0.01;
321 } else
322 dst[i] = 0.0;
323 }
324}
325
326void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
327{
328 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
329}
330
331void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
332{
333 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
334}
335
336static void display_lat(const char *name, unsigned long min, unsigned long max,
337 double mean, double dev)
338{
339 const char *base = "(usec)";
340 char *minp, *maxp;
341
342 if (!usec_to_msec(&min, &max, &mean, &dev))
343 base = "(msec)";
344
345 minp = num2str(min, 6, 1, 0);
346 maxp = num2str(max, 6, 1, 0);
347
348 log_info(" %s %s: min=%s, max=%s, avg=%5.02f,"
349 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
350
351 free(minp);
352 free(maxp);
353}
354
355static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
356 int ddir)
357{
358 const char *ddir_str[] = { "read ", "write" };
359 unsigned long min, max, runt;
360 unsigned long long bw, iops;
361 double mean, dev;
362 char *io_p, *bw_p, *iops_p;
363 int i2p;
364
365 assert(ddir_rw(ddir));
366
367 if (!ts->runtime[ddir])
368 return;
369
370 i2p = is_power_of_2(rs->kb_base);
371 runt = ts->runtime[ddir];
372
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);
376
377 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
378 iops_p = num2str(iops, 6, 1, 0);
379
380 log_info(" %s: io=%sB, bw=%sB/s, iops=%s, runt=%6llumsec\n",
381 ddir_str[ddir], io_p, bw_p, iops_p,
382 ts->runtime[ddir]);
383
384 free(io_p);
385 free(bw_p);
386 free(iops_p);
387
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);
394
395 if (ts->clat_percentiles) {
396 show_clat_percentiles(ts->io_u_plat[ddir],
397 ts->clat_stat[ddir].samples,
398 ts->percentile_list);
399 }
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";
403
404 if (rs->agg[ddir]) {
405 p_of_agg = mean * 100 / (double) rs->agg[ddir];
406 if (p_of_agg > 100.0)
407 p_of_agg = 100.0;
408 }
409
410 if (mean > 999999.9) {
411 min /= 1000.0;
412 max /= 1000.0;
413 mean /= 1000.0;
414 dev /= 1000.0;
415 bw_str = "MB";
416 }
417
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);
421 }
422}
423
424static int show_lat(double *io_u_lat, int nr, const char **ranges,
425 const char *msg)
426{
427 int new_line = 1, i, line = 0, shown = 0;
428
429 for (i = 0; i < nr; i++) {
430 if (io_u_lat[i] <= 0.0)
431 continue;
432 shown = 1;
433 if (new_line) {
434 if (line)
435 log_info("\n");
436 log_info(" lat (%s) : ", msg);
437 new_line = 0;
438 line = 0;
439 }
440 if (line)
441 log_info(", ");
442 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
443 line++;
444 if (line == 5)
445 new_line = 1;
446 }
447
448 if (shown)
449 log_info("\n");
450
451 return shown;
452}
453
454static void show_lat_u(double *io_u_lat_u)
455{
456 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
457 "250=", "500=", "750=", "1000=", };
458
459 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
460}
461
462static void show_lat_m(double *io_u_lat_m)
463{
464 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
465 "250=", "500=", "750=", "1000=", "2000=",
466 ">=2000=", };
467
468 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
469}
470
471static void show_latencies(struct thread_stat *ts)
472{
473 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
474 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
475
476 stat_calc_lat_u(ts, io_u_lat_u);
477 stat_calc_lat_m(ts, io_u_lat_m);
478
479 show_lat_u(io_u_lat_u);
480 show_lat_m(io_u_lat_m);
481}
482
483void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
484{
485 double usr_cpu, sys_cpu;
486 unsigned long runtime;
487 double io_u_dist[FIO_IO_U_MAP_NR];
488 time_t time_p;
489 char time_buf[64];
490
491
492 if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
493 !(ts->total_io_u[0] + ts->total_io_u[1]))
494 return;
495
496 time(&time_p);
497 ctime_r((const time_t *) &time_p, time_buf);
498
499 if (!ts->error) {
500 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
501 ts->name, ts->groupid, ts->members,
502 ts->error, (int) ts->pid, time_buf);
503 } else {
504 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
505 ts->name, ts->groupid, ts->members,
506 ts->error, ts->verror, (int) ts->pid,
507 time_buf);
508 }
509
510 if (strlen(ts->description))
511 log_info(" Description : [%s]\n", ts->description);
512
513 if (ts->io_bytes[DDIR_READ])
514 show_ddir_status(rs, ts, DDIR_READ);
515 if (ts->io_bytes[DDIR_WRITE])
516 show_ddir_status(rs, ts, DDIR_WRITE);
517
518 show_latencies(ts);
519
520 runtime = ts->total_run_time;
521 if (runtime) {
522 double runt = (double) runtime;
523
524 usr_cpu = (double) ts->usr_time * 100 / runt;
525 sys_cpu = (double) ts->sys_time * 100 / runt;
526 } else {
527 usr_cpu = 0;
528 sys_cpu = 0;
529 }
530
531 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
532 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
533
534 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
535 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
536 " 16=%3.1f%%, 32=%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
541 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
542 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
543 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
544 io_u_dist[1], io_u_dist[2],
545 io_u_dist[3], io_u_dist[4],
546 io_u_dist[5], io_u_dist[6]);
547 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
548 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
549 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
550 io_u_dist[1], io_u_dist[2],
551 io_u_dist[3], io_u_dist[4],
552 io_u_dist[5], io_u_dist[6]);
553 log_info(" issued : total=r=%lu/w=%lu/d=%lu,"
554 " short=r=%lu/w=%lu/d=%lu\n",
555 ts->total_io_u[0], ts->total_io_u[1],
556 ts->total_io_u[2],
557 ts->short_io_u[0], ts->short_io_u[1],
558 ts->short_io_u[2]);
559 if (ts->continue_on_error) {
560 log_info(" errors : total=%lu, first_error=%d/<%s>\n",
561 ts->total_err_count,
562 ts->first_error,
563 strerror(ts->first_error));
564 }
565}
566
567static void show_ddir_status_terse(struct thread_stat *ts,
568 struct group_run_stats *rs, int ddir)
569{
570 unsigned long min, max;
571 unsigned long long bw, iops;
572 unsigned int *ovals = NULL;
573 double mean, dev;
574 unsigned int len, minv, maxv;
575 int i;
576
577 assert(ddir_rw(ddir));
578
579 iops = bw = 0;
580 if (ts->runtime[ddir]) {
581 uint64_t runt = ts->runtime[ddir];
582
583 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
584 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
585 }
586
587 log_info(";%llu;%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, iops,
588 ts->runtime[ddir]);
589
590 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
591 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
592 else
593 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
594
595 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
596 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
597 else
598 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
599
600 if (ts->clat_percentiles) {
601 len = calc_clat_percentiles(ts->io_u_plat[ddir],
602 ts->clat_stat[ddir].samples,
603 ts->percentile_list, &ovals, &maxv,
604 &minv);
605 } else
606 len = 0;
607
608 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
609 if (i >= len) {
610 log_info(";0%%=0");
611 continue;
612 }
613 log_info(";%2.2f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
614 }
615
616 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
617 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
618 else
619 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
620
621 if (ovals)
622 free(ovals);
623
624 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
625 double p_of_agg = 100.0;
626
627 if (rs->agg[ddir]) {
628 p_of_agg = mean * 100 / (double) rs->agg[ddir];
629 if (p_of_agg > 100.0)
630 p_of_agg = 100.0;
631 }
632
633 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
634 } else
635 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
636}
637
638static void show_thread_status_terse_v2(struct thread_stat *ts,
639 struct group_run_stats *rs)
640{
641 double io_u_dist[FIO_IO_U_MAP_NR];
642 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
643 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
644 double usr_cpu, sys_cpu;
645 int i;
646
647 /* General Info */
648 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
649 /* Log Read Status */
650 show_ddir_status_terse(ts, rs, 0);
651 /* Log Write Status */
652 show_ddir_status_terse(ts, rs, 1);
653
654 /* CPU Usage */
655 if (ts->total_run_time) {
656 double runt = (double) ts->total_run_time;
657
658 usr_cpu = (double) ts->usr_time * 100 / runt;
659 sys_cpu = (double) ts->sys_time * 100 / runt;
660 } else {
661 usr_cpu = 0;
662 sys_cpu = 0;
663 }
664
665 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
666 ts->minf);
667
668 /* Calc % distribution of IO depths, usecond, msecond latency */
669 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
670 stat_calc_lat_u(ts, io_u_lat_u);
671 stat_calc_lat_m(ts, io_u_lat_m);
672
673 /* Only show fixed 7 I/O depth levels*/
674 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
675 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
676 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
677
678 /* Microsecond latency */
679 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
680 log_info(";%3.2f%%", io_u_lat_u[i]);
681 /* Millisecond latency */
682 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
683 log_info(";%3.2f%%", io_u_lat_m[i]);
684 /* Additional output if continue_on_error set - default off*/
685 if (ts->continue_on_error)
686 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
687 log_info("\n");
688
689 /* Additional output if description is set */
690 if (ts->description)
691 log_info(";%s", ts->description);
692
693 log_info("\n");
694}
695
696#define FIO_TERSE_VERSION "3"
697
698static void show_thread_status_terse_v3(struct thread_stat *ts,
699 struct group_run_stats *rs)
700{
701 double io_u_dist[FIO_IO_U_MAP_NR];
702 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
703 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
704 double usr_cpu, sys_cpu;
705 int i;
706
707 /* General Info */
708 log_info("%s;%s;%s;%d;%d", FIO_TERSE_VERSION, fio_version_string,
709 ts->name, ts->groupid, ts->error);
710 /* Log Read Status */
711 show_ddir_status_terse(ts, rs, 0);
712 /* Log Write Status */
713 show_ddir_status_terse(ts, rs, 1);
714
715 /* CPU Usage */
716 if (ts->total_run_time) {
717 double runt = (double) ts->total_run_time;
718
719 usr_cpu = (double) ts->usr_time * 100 / runt;
720 sys_cpu = (double) ts->sys_time * 100 / runt;
721 } else {
722 usr_cpu = 0;
723 sys_cpu = 0;
724 }
725
726 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
727 ts->minf);
728
729 /* Calc % distribution of IO depths, usecond, msecond latency */
730 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
731 stat_calc_lat_u(ts, io_u_lat_u);
732 stat_calc_lat_m(ts, io_u_lat_m);
733
734 /* Only show fixed 7 I/O depth levels*/
735 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
736 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
737 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
738
739 /* Microsecond latency */
740 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
741 log_info(";%3.2f%%", io_u_lat_u[i]);
742 /* Millisecond latency */
743 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
744 log_info(";%3.2f%%", io_u_lat_m[i]);
745
746 /* disk util stats, if any */
747 show_disk_util(1);
748
749 /* Additional output if continue_on_error set - default off*/
750 if (ts->continue_on_error)
751 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
752
753 /* Additional output if description is set */
754 if (strlen(ts->description))
755 log_info(";%s", ts->description);
756
757 log_info("\n");
758}
759
760static void show_thread_status_terse(struct thread_stat *ts,
761 struct group_run_stats *rs)
762{
763 if (terse_version == 2)
764 show_thread_status_terse_v2(ts, rs);
765 else if (terse_version == 3)
766 show_thread_status_terse_v3(ts, rs);
767 else
768 log_err("fio: bad terse version!? %d\n", terse_version);
769}
770
771static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
772{
773 double mean, S;
774
775 if (src->samples == 0)
776 return;
777
778 dst->min_val = min(dst->min_val, src->min_val);
779 dst->max_val = max(dst->max_val, src->max_val);
780
781 /*
782 * Compute new mean and S after the merge
783 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
784 * #Parallel_algorithm>
785 */
786 if (nr == 1) {
787 mean = src->mean.u.f;
788 S = src->S.u.f;
789 } else {
790 double delta = src->mean.u.f - dst->mean.u.f;
791
792 mean = ((src->mean.u.f * src->samples) +
793 (dst->mean.u.f * dst->samples)) /
794 (dst->samples + src->samples);
795
796 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
797 (dst->samples * src->samples) /
798 (dst->samples + src->samples);
799 }
800
801 dst->samples += src->samples;
802 dst->mean.u.f = mean;
803 dst->S.u.f = S;
804}
805
806void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
807{
808 int i;
809
810 for (i = 0; i < 2; i++) {
811 if (dst->max_run[i] < src->max_run[i])
812 dst->max_run[i] = src->max_run[i];
813 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
814 dst->min_run[i] = src->min_run[i];
815 if (dst->max_bw[i] < src->max_bw[i])
816 dst->max_bw[i] = src->max_bw[i];
817 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
818 dst->min_bw[i] = src->min_bw[i];
819
820 dst->io_kb[i] += src->io_kb[i];
821 dst->agg[i] += src->agg[i];
822 }
823
824}
825
826void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
827{
828 int l, k;
829
830 for (l = 0; l <= DDIR_WRITE; l++) {
831 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
832 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
833 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
834 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
835
836 dst->io_bytes[l] += src->io_bytes[l];
837
838 if (dst->runtime[l] < src->runtime[l])
839 dst->runtime[l] = src->runtime[l];
840 }
841
842 dst->usr_time += src->usr_time;
843 dst->sys_time += src->sys_time;
844 dst->ctx += src->ctx;
845 dst->majf += src->majf;
846 dst->minf += src->minf;
847
848 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
849 dst->io_u_map[k] += src->io_u_map[k];
850 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
851 dst->io_u_submit[k] += src->io_u_submit[k];
852 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
853 dst->io_u_complete[k] += src->io_u_complete[k];
854 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
855 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
856 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
857 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
858
859 for (k = 0; k <= 2; k++) {
860 dst->total_io_u[k] += src->total_io_u[k];
861 dst->short_io_u[k] += src->short_io_u[k];
862 }
863
864 for (k = 0; k <= DDIR_WRITE; k++) {
865 int m;
866 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
867 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
868 }
869
870 dst->total_run_time += src->total_run_time;
871 dst->total_submit += src->total_submit;
872 dst->total_complete += src->total_complete;
873}
874
875void init_group_run_stat(struct group_run_stats *gs)
876{
877 memset(gs, 0, sizeof(*gs));
878 gs->min_bw[0] = gs->min_run[0] = ~0UL;
879 gs->min_bw[1] = gs->min_run[1] = ~0UL;
880}
881
882void init_thread_stat(struct thread_stat *ts)
883{
884 int j;
885
886 memset(ts, 0, sizeof(*ts));
887
888 for (j = 0; j <= DDIR_WRITE; j++) {
889 ts->lat_stat[j].min_val = -1UL;
890 ts->clat_stat[j].min_val = -1UL;
891 ts->slat_stat[j].min_val = -1UL;
892 ts->bw_stat[j].min_val = -1UL;
893 }
894 ts->groupid = -1;
895}
896
897void show_run_stats(void)
898{
899 struct group_run_stats *runstats, *rs;
900 struct thread_data *td;
901 struct thread_stat *threadstats, *ts;
902 int i, j, nr_ts, last_ts, idx;
903 int kb_base_warned = 0;
904
905 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
906
907 for (i = 0; i < groupid + 1; i++)
908 init_group_run_stat(&runstats[i]);
909
910 /*
911 * find out how many threads stats we need. if group reporting isn't
912 * enabled, it's one-per-td.
913 */
914 nr_ts = 0;
915 last_ts = -1;
916 for_each_td(td, i) {
917 if (!td->o.group_reporting) {
918 nr_ts++;
919 continue;
920 }
921 if (last_ts == td->groupid)
922 continue;
923
924 last_ts = td->groupid;
925 nr_ts++;
926 }
927
928 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
929
930 for (i = 0; i < nr_ts; i++)
931 init_thread_stat(&threadstats[i]);
932
933 j = 0;
934 last_ts = -1;
935 idx = 0;
936 for_each_td(td, i) {
937 if (idx && (!td->o.group_reporting ||
938 (td->o.group_reporting && last_ts != td->groupid))) {
939 idx = 0;
940 j++;
941 }
942
943 last_ts = td->groupid;
944
945 ts = &threadstats[j];
946
947 ts->clat_percentiles = td->o.clat_percentiles;
948 if (td->o.overwrite_plist)
949 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
950 else
951 memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list));
952
953 idx++;
954 ts->members++;
955
956 if (ts->groupid == -1) {
957 /*
958 * These are per-group shared already
959 */
960 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
961 if (td->o.description)
962 strncpy(ts->description, td->o.description,
963 FIO_JOBNAME_SIZE);
964 else
965 memset(ts->description, 0, FIO_JOBNAME_SIZE);
966
967 /*
968 * If multiple entries in this group, this is
969 * the first member.
970 */
971 ts->thread_number = td->thread_number;
972 ts->groupid = td->groupid;
973
974 /*
975 * first pid in group, not very useful...
976 */
977 ts->pid = td->pid;
978
979 ts->kb_base = td->o.kb_base;
980 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
981 log_info("fio: kb_base differs for jobs in group, using"
982 " %u as the base\n", ts->kb_base);
983 kb_base_warned = 1;
984 }
985
986 ts->continue_on_error = td->o.continue_on_error;
987 ts->total_err_count += td->total_err_count;
988 ts->first_error = td->first_error;
989 if (!ts->error) {
990 if (!td->error && td->o.continue_on_error &&
991 td->first_error) {
992 ts->error = td->first_error;
993 strcpy(ts->verror, td->verror);
994 } else if (td->error) {
995 ts->error = td->error;
996 strcpy(ts->verror, td->verror);
997 }
998 }
999
1000 sum_thread_stats(ts, &td->ts, idx);
1001 }
1002
1003 for (i = 0; i < nr_ts; i++) {
1004 unsigned long long bw;
1005
1006 ts = &threadstats[i];
1007 rs = &runstats[ts->groupid];
1008 rs->kb_base = ts->kb_base;
1009
1010 for (j = 0; j <= DDIR_WRITE; j++) {
1011 if (!ts->runtime[j])
1012 continue;
1013 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1014 rs->min_run[j] = ts->runtime[j];
1015 if (ts->runtime[j] > rs->max_run[j])
1016 rs->max_run[j] = ts->runtime[j];
1017
1018 bw = 0;
1019 if (ts->runtime[j]) {
1020 unsigned long runt = ts->runtime[j];
1021 unsigned long long kb;
1022
1023 kb = ts->io_bytes[j] / rs->kb_base;
1024 bw = kb * 1000 / runt;
1025 }
1026 if (bw < rs->min_bw[j])
1027 rs->min_bw[j] = bw;
1028 if (bw > rs->max_bw[j])
1029 rs->max_bw[j] = bw;
1030
1031 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1032 }
1033 }
1034
1035 for (i = 0; i < groupid + 1; i++) {
1036 rs = &runstats[i];
1037
1038 if (rs->max_run[0])
1039 rs->agg[0] = (rs->io_kb[0] * 1000) / rs->max_run[0];
1040 if (rs->max_run[1])
1041 rs->agg[1] = (rs->io_kb[1] * 1000) / rs->max_run[1];
1042 }
1043
1044 /*
1045 * don't overwrite last signal output
1046 */
1047 if (!terse_output)
1048 log_info("\n");
1049
1050 for (i = 0; i < nr_ts; i++) {
1051 ts = &threadstats[i];
1052 rs = &runstats[ts->groupid];
1053
1054 if (is_backend)
1055 fio_server_send_ts(ts, rs);
1056 else if (terse_output)
1057 show_thread_status_terse(ts, rs);
1058 else
1059 show_thread_status(ts, rs);
1060 }
1061
1062 for (i = 0; i < groupid + 1; i++) {
1063 rs = &runstats[i];
1064
1065 rs->groupid = i;
1066 if (is_backend)
1067 fio_server_send_gs(rs);
1068 else if (!terse_output)
1069 show_group_stats(rs);
1070 }
1071
1072 if (is_backend)
1073 fio_server_send_du();
1074 else if (!terse_output)
1075 show_disk_util(0);
1076
1077 free(runstats);
1078 free(threadstats);
1079}
1080
1081static void *__show_running_run_stats(void *arg)
1082{
1083 struct thread_data *td;
1084 unsigned long long *rt;
1085 struct timeval tv;
1086 int i;
1087
1088 rt = malloc(thread_number * sizeof(unsigned long long));
1089 fio_gettime(&tv, NULL);
1090
1091 for_each_td(td, i) {
1092 rt[i] = mtime_since(&td->start, &tv);
1093 if (td_read(td) && td->io_bytes[DDIR_READ])
1094 td->ts.runtime[DDIR_READ] += rt[i];
1095 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1096 td->ts.runtime[DDIR_WRITE] += rt[i];
1097
1098 update_rusage_stat(td);
1099 td->ts.io_bytes[0] = td->io_bytes[0];
1100 td->ts.io_bytes[1] = td->io_bytes[1];
1101 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1102 }
1103
1104 show_run_stats();
1105
1106 for_each_td(td, i) {
1107 if (td_read(td) && td->io_bytes[DDIR_READ])
1108 td->ts.runtime[DDIR_READ] -= rt[i];
1109 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1110 td->ts.runtime[DDIR_WRITE] -= rt[i];
1111 }
1112
1113 free(rt);
1114 return NULL;
1115}
1116
1117/*
1118 * Called from signal handler. It _should_ be safe to just run this inline
1119 * in the sig handler, but we should be disturbing the system less by just
1120 * creating a thread to do it.
1121 */
1122void show_running_run_stats(void)
1123{
1124 pthread_t thread;
1125
1126 pthread_create(&thread, NULL, __show_running_run_stats, NULL);
1127 pthread_detach(thread);
1128}
1129
1130static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1131{
1132 double val = data;
1133 double delta;
1134
1135 if (data > is->max_val)
1136 is->max_val = data;
1137 if (data < is->min_val)
1138 is->min_val = data;
1139
1140 delta = val - is->mean.u.f;
1141 if (delta) {
1142 is->mean.u.f += delta / (is->samples + 1.0);
1143 is->S.u.f += delta * (val - is->mean.u.f);
1144 }
1145
1146 is->samples++;
1147}
1148
1149static void __add_log_sample(struct io_log *iolog, unsigned long val,
1150 enum fio_ddir ddir, unsigned int bs,
1151 unsigned long t)
1152{
1153 const int nr_samples = iolog->nr_samples;
1154
1155 if (!iolog->nr_samples)
1156 iolog->avg_last = t;
1157
1158 if (iolog->nr_samples == iolog->max_samples) {
1159 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
1160
1161 iolog->log = realloc(iolog->log, new_size);
1162 iolog->max_samples <<= 1;
1163 }
1164
1165 iolog->log[nr_samples].val = val;
1166 iolog->log[nr_samples].time = t;
1167 iolog->log[nr_samples].ddir = ddir;
1168 iolog->log[nr_samples].bs = bs;
1169 iolog->nr_samples++;
1170}
1171
1172static inline void reset_io_stat(struct io_stat *ios)
1173{
1174 ios->max_val = ios->min_val = ios->samples = 0;
1175 ios->mean.u.f = ios->S.u.f = 0;
1176}
1177
1178static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1179 unsigned long val, enum fio_ddir ddir,
1180 unsigned int bs)
1181{
1182 unsigned long elapsed, this_window;
1183
1184 if (!ddir_rw(ddir))
1185 return;
1186
1187 elapsed = mtime_since_now(&td->epoch);
1188
1189 /*
1190 * If no time averaging, just add the log sample.
1191 */
1192 if (!iolog->avg_msec) {
1193 __add_log_sample(iolog, val, ddir, bs, elapsed);
1194 return;
1195 }
1196
1197 /*
1198 * Add the sample. If the time period has passed, then
1199 * add that entry to the log and clear.
1200 */
1201 add_stat_sample(&iolog->avg_window[ddir], val);
1202
1203 /*
1204 * If period hasn't passed, adding the above sample is all we
1205 * need to do.
1206 */
1207 this_window = elapsed - iolog->avg_last;
1208 if (this_window < iolog->avg_msec)
1209 return;
1210
1211 /*
1212 * Note an entry in the log. Use the mean from the logged samples,
1213 * making sure to properly round up. Only write a log entry if we
1214 * had actual samples done.
1215 */
1216 if (iolog->avg_window[DDIR_READ].samples) {
1217 unsigned long mr;
1218
1219 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1220 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed);
1221 }
1222 if (iolog->avg_window[DDIR_WRITE].samples) {
1223 unsigned long mw;
1224
1225 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1226 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed);
1227 }
1228
1229 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1230 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1231 iolog->avg_last = elapsed;
1232}
1233
1234void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1235{
1236 struct io_log *iolog;
1237
1238 if (!ddir_rw(ddir))
1239 return;
1240
1241 iolog = agg_io_log[ddir];
1242 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
1243}
1244
1245static void add_clat_percentile_sample(struct thread_stat *ts,
1246 unsigned long usec, enum fio_ddir ddir)
1247{
1248 unsigned int idx = plat_val_to_idx(usec);
1249 assert(idx < FIO_IO_U_PLAT_NR);
1250
1251 ts->io_u_plat[ddir][idx]++;
1252}
1253
1254void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1255 unsigned long usec, unsigned int bs)
1256{
1257 struct thread_stat *ts = &td->ts;
1258
1259 if (!ddir_rw(ddir))
1260 return;
1261
1262 add_stat_sample(&ts->clat_stat[ddir], usec);
1263
1264 if (td->clat_log)
1265 add_log_sample(td, td->clat_log, usec, ddir, bs);
1266
1267 if (ts->clat_percentiles)
1268 add_clat_percentile_sample(ts, usec, ddir);
1269}
1270
1271void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1272 unsigned long usec, unsigned int bs)
1273{
1274 struct thread_stat *ts = &td->ts;
1275
1276 if (!ddir_rw(ddir))
1277 return;
1278
1279 add_stat_sample(&ts->slat_stat[ddir], usec);
1280
1281 if (td->slat_log)
1282 add_log_sample(td, td->slat_log, usec, ddir, bs);
1283}
1284
1285void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1286 unsigned long usec, unsigned int bs)
1287{
1288 struct thread_stat *ts = &td->ts;
1289
1290 if (!ddir_rw(ddir))
1291 return;
1292
1293 add_stat_sample(&ts->lat_stat[ddir], usec);
1294
1295 if (td->lat_log)
1296 add_log_sample(td, td->lat_log, usec, ddir, bs);
1297}
1298
1299void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1300 struct timeval *t)
1301{
1302 struct thread_stat *ts = &td->ts;
1303 unsigned long spent, rate;
1304
1305 if (!ddir_rw(ddir))
1306 return;
1307
1308 spent = mtime_since(&td->bw_sample_time, t);
1309 if (spent < td->o.bw_avg_time)
1310 return;
1311
1312 /*
1313 * Compute both read and write rates for the interval.
1314 */
1315 for (ddir = DDIR_READ; ddir <= DDIR_WRITE; ddir++) {
1316 uint64_t delta;
1317
1318 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1319 if (!delta)
1320 continue; /* No entries for interval */
1321
1322 rate = delta * 1000 / spent / 1024;
1323 add_stat_sample(&ts->bw_stat[ddir], rate);
1324
1325 if (td->bw_log)
1326 add_log_sample(td, td->bw_log, rate, ddir, bs);
1327
1328 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1329 }
1330
1331 fio_gettime(&td->bw_sample_time, NULL);
1332}
1333
1334void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
1335 struct timeval *t)
1336{
1337 struct thread_stat *ts = &td->ts;
1338 unsigned long spent, iops;
1339
1340 if (!ddir_rw(ddir))
1341 return;
1342
1343 spent = mtime_since(&td->iops_sample_time, t);
1344 if (spent < td->o.iops_avg_time)
1345 return;
1346
1347 /*
1348 * Compute both read and write rates for the interval.
1349 */
1350 for (ddir = DDIR_READ; ddir <= DDIR_WRITE; ddir++) {
1351 uint64_t delta;
1352
1353 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1354 if (!delta)
1355 continue; /* No entries for interval */
1356
1357 iops = (delta * 1000) / spent;
1358 add_stat_sample(&ts->iops_stat[ddir], iops);
1359
1360 if (td->iops_log)
1361 add_log_sample(td, td->iops_log, iops, ddir, 0);
1362
1363 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1364 }
1365
1366 fio_gettime(&td->iops_sample_time, NULL);
1367}
fio - Flexible IO Tester