projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Unify cli latency display
[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
117static unsigned int calc_clat_percentiles(unsigned int *io_u_plat,
118 unsigned long nr, fio_fp64_t *plist,
119 unsigned int **output,
120 unsigned int *maxv,
121 unsigned int *minv)
122{
123 unsigned long sum = 0;
124 unsigned int len, i, j = 0;
125 unsigned int oval_len = 0;
126 unsigned int *ovals = NULL;
127 int is_last;
128
129 *minv = -1U;
130 *maxv = 0;
131
132 len = 0;
133 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
134 len++;
135
136 if (!len)
137 return 0;
138
139 /*
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.
143 */
144 if (len > 1)
145 qsort((void*)plist, len, sizeof(plist[0]), double_cmp);
146
147 /*
148 * Calculate bucket values, note down max and min values
149 */
150 is_last = 0;
151 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
152 sum += io_u_plat[i];
153 while (sum >= (plist[j].u.f / 100.0 * nr)) {
154 assert(plist[j].u.f <= 100.0);
155
156 if (j == oval_len) {
157 oval_len += 100;
158 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
159 }
160
161 ovals[j] = plat_idx_to_val(i);
162 if (ovals[j] < *minv)
163 *minv = ovals[j];
164 if (ovals[j] > *maxv)
165 *maxv = ovals[j];
166
167 is_last = (j == len - 1);
168 if (is_last)
169 break;
170
171 j++;
172 }
173 }
174
175 *output = ovals;
176 return len;
177}
178
179/*
180 * Find and display the p-th percentile of clat
181 */
182static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
183 fio_fp64_t *plist)
184{
185 unsigned int len, j = 0, minv, maxv;
186 unsigned int *ovals;
187 int is_last, scale_down;
188
189 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
190 if (!len)
191 goto out;
192
193 /*
194 * We default to usecs, but if the value range is such that we
195 * should scale down to msecs, do that.
196 */
197 if (minv > 2000 && maxv > 99999) {
198 scale_down = 1;
199 log_info(" clat percentiles (msec):\n |");
200 } else {
201 scale_down = 0;
202 log_info(" clat percentiles (usec):\n |");
203 }
204
205 for (j = 0; j < len; j++) {
206 char fbuf[8];
207
208 /* for formatting */
209 if (j != 0 && (j % 4) == 0)
210 log_info(" |");
211
212 /* end of the list */
213 is_last = (j == len - 1);
214
215 if (plist[j].u.f < 10.0)
216 sprintf(fbuf, " %2.2f", plist[j].u.f);
217 else
218 sprintf(fbuf, "%2.2f", plist[j].u.f);
219
220 if (scale_down)
221 ovals[j] = (ovals[j] + 999) / 1000;
222
223 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
224
225 if (is_last)
226 break;
227
228 if (j % 4 == 3) /* for formatting */
229 log_info("\n");
230 }
231
232out:
233 if (ovals)
234 free(ovals);
235}
236
237int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
238 double *mean, double *dev)
239{
240 double n = is->samples;
241
242 if (is->samples == 0)
243 return 0;
244
245 *min = is->min_val;
246 *max = is->max_val;
247
248 n = (double) is->samples;
249 *mean = is->mean.u.f;
250
251 if (n > 1.0)
252 *dev = sqrt(is->S.u.f / (n - 1.0));
253 else
254 *dev = 0;
255
256 return 1;
257}
258
259void show_group_stats(struct group_run_stats *rs)
260{
261 char *p1, *p2, *p3, *p4;
262 const char *ddir_str[] = { " READ", " WRITE" };
263 int i;
264
265 log_info("\nRun status group %d (all jobs):\n", rs->groupid);
266
267 for (i = 0; i <= DDIR_WRITE; i++) {
268 const int i2p = is_power_of_2(rs->kb_base);
269
270 if (!rs->max_run[i])
271 continue;
272
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);
277
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],
281 rs->max_run[i]);
282
283 free(p1);
284 free(p2);
285 free(p3);
286 free(p4);
287 }
288}
289
290#define ts_total_io_u(ts) \
291 ((ts)->total_io_u[0] + (ts)->total_io_u[1])
292
293static void stat_calc_dist(unsigned int *map, unsigned long total,
294 double *io_u_dist)
295{
296 int i;
297
298 /*
299 * Do depth distribution calculations
300 */
301 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
302 if (total) {
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])
306 io_u_dist[i] = 0.1;
307 } else
308 io_u_dist[i] = 0.0;
309 }
310}
311
312static void stat_calc_lat(struct thread_stat *ts, double *dst,
313 unsigned int *src, int nr)
314{
315 unsigned long total = ts_total_io_u(ts);
316 int i;
317
318 /*
319 * Do latency distribution calculations
320 */
321 for (i = 0; i < nr; i++) {
322 if (total) {
323 dst[i] = (double) src[i] / (double) total;
324 dst[i] *= 100.0;
325 if (dst[i] < 0.01 && src[i])
326 dst[i] = 0.01;
327 } else
328 dst[i] = 0.0;
329 }
330}
331
332static void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
333{
334 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
335}
336
337static void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
338{
339 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
340}
341
342static void display_lat(const char *name, unsigned long min, unsigned long max,
343 double mean, double dev)
344{
345 const char *base = "(usec)";
346 char *minp, *maxp;
347
348 if (!usec_to_msec(&min, &max, &mean, &dev))
349 base = "(msec)";
350
351 minp = num2str(min, 6, 1, 0);
352 maxp = num2str(max, 6, 1, 0);
353
354 log_info(" %s %s: min=%s, max=%s, avg=%5.02f,"
355 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
356
357 free(minp);
358 free(maxp);
359}
360
361static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
362 int ddir)
363{
364 const char *ddir_str[] = { "read ", "write" };
365 unsigned long min, max, runt;
366 unsigned long long bw, iops;
367 double mean, dev;
368 char *io_p, *bw_p, *iops_p;
369 int i2p;
370
371 assert(ddir_rw(ddir));
372
373 if (!ts->runtime[ddir])
374 return;
375
376 i2p = is_power_of_2(rs->kb_base);
377 runt = ts->runtime[ddir];
378
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);
382
383 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
384 iops_p = num2str(iops, 6, 1, 0);
385
386 log_info(" %s: io=%sB, bw=%sB/s, iops=%s, runt=%6llumsec\n",
387 ddir_str[ddir], io_p, bw_p, iops_p,
388 ts->runtime[ddir]);
389
390 free(io_p);
391 free(bw_p);
392 free(iops_p);
393
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);
400
401 if (ts->clat_percentiles) {
402 show_clat_percentiles(ts->io_u_plat[ddir],
403 ts->clat_stat[ddir].samples,
404 ts->percentile_list);
405 }
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";
409
410 if (rs->agg[ddir]) {
411 p_of_agg = mean * 100 / (double) rs->agg[ddir];
412 if (p_of_agg > 100.0)
413 p_of_agg = 100.0;
414 }
415
416 if (mean > 999999.9) {
417 min /= 1000.0;
418 max /= 1000.0;
419 mean /= 1000.0;
420 dev /= 1000.0;
421 bw_str = "MB";
422 }
423
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);
427 }
428}
429
430static int show_lat(double *io_u_lat, int nr, const char **ranges,
431 const char *msg)
432{
433 int new_line = 1, i, line = 0, shown = 0;
434
435 for (i = 0; i < nr; i++) {
436 if (io_u_lat[i] <= 0.0)
437 continue;
438 shown = 1;
439 if (new_line) {
440 if (line)
441 log_info("\n");
442 log_info(" lat (%s) : ", msg);
443 new_line = 0;
444 line = 0;
445 }
446 if (line)
447 log_info(", ");
448 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
449 line++;
450 if (line == 5)
451 new_line = 1;
452 }
453
454 if (shown)
455 log_info("\n");
456
457 return shown;
458}
459
460static void show_lat_u(double *io_u_lat_u)
461{
462 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
463 "250=", "500=", "750=", "1000=", };
464
465 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
466}
467
468static void show_lat_m(double *io_u_lat_m)
469{
470 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
471 "250=", "500=", "750=", "1000=", "2000=",
472 ">=2000=", };
473
474 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
475}
476
477static void show_latencies(double *io_u_lat_u, double *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 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
489 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
490
491 if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
492 !(ts->total_io_u[0] + ts->total_io_u[1]))
493 return;
494
495 if (!ts->error) {
496 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
497 ts->name, ts->groupid, ts->members,
498 ts->error, (int) ts->pid);
499 } else {
500 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
501 ts->name, ts->groupid, ts->members,
502 ts->error, ts->verror, (int) ts->pid);
503 }
504
505 if (strlen(ts->description))
506 log_info(" Description : [%s]\n", ts->description);
507
508 if (ts->io_bytes[DDIR_READ])
509 show_ddir_status(rs, ts, DDIR_READ);
510 if (ts->io_bytes[DDIR_WRITE])
511 show_ddir_status(rs, ts, DDIR_WRITE);
512
513 stat_calc_lat_u(ts, io_u_lat_u);
514 stat_calc_lat_m(ts, io_u_lat_m);
515 show_latencies(io_u_lat_u, io_u_lat_m);
516
517 runtime = ts->total_run_time;
518 if (runtime) {
519 double runt = (double) runtime;
520
521 usr_cpu = (double) ts->usr_time * 100 / runt;
522 sys_cpu = (double) ts->sys_time * 100 / runt;
523 } else {
524 usr_cpu = 0;
525 sys_cpu = 0;
526 }
527
528 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
529 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
530
531 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
532 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
533 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
534 io_u_dist[1], io_u_dist[2],
535 io_u_dist[3], io_u_dist[4],
536 io_u_dist[5], io_u_dist[6]);
537
538 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
539 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
540 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
541 io_u_dist[1], io_u_dist[2],
542 io_u_dist[3], io_u_dist[4],
543 io_u_dist[5], io_u_dist[6]);
544 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
545 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
546 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
547 io_u_dist[1], io_u_dist[2],
548 io_u_dist[3], io_u_dist[4],
549 io_u_dist[5], io_u_dist[6]);
550 log_info(" issued : total=r=%lu/w=%lu/d=%lu,"
551 " short=r=%lu/w=%lu/d=%lu\n",
552 ts->total_io_u[0], ts->total_io_u[1],
553 ts->total_io_u[2],
554 ts->short_io_u[0], ts->short_io_u[1],
555 ts->short_io_u[2]);
556 if (ts->continue_on_error) {
557 log_info(" errors : total=%lu, first_error=%d/<%s>\n",
558 ts->total_err_count,
559 ts->first_error,
560 strerror(ts->first_error));
561 }
562}
563
564static void show_ddir_status_terse(struct thread_stat *ts,
565 struct group_run_stats *rs, int ddir)
566{
567 unsigned long min, max;
568 unsigned long long bw, iops;
569 unsigned int *ovals = NULL;
570 double mean, dev;
571 unsigned int len, minv, maxv;
572 int i;
573
574 assert(ddir_rw(ddir));
575
576 iops = bw = 0;
577 if (ts->runtime[ddir]) {
578 uint64_t runt = ts->runtime[ddir];
579
580 bw = ts->io_bytes[ddir] / runt;
581 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
582 }
583
584 log_info(";%llu;%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, iops,
585 ts->runtime[ddir]);
586
587 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
588 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
589 else
590 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
591
592 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
593 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
594 else
595 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
596
597 if (ts->clat_percentiles) {
598 len = calc_clat_percentiles(ts->io_u_plat[ddir],
599 ts->clat_stat[ddir].samples,
600 ts->percentile_list, &ovals, &maxv,
601 &minv);
602 } else
603 len = 0;
604
605 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
606 if (i >= len) {
607 log_info(";0%%=0");
608 continue;
609 }
610 log_info(";%2.2f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
611 }
612
613 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
614 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
615 else
616 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
617
618 if (ovals)
619 free(ovals);
620
621 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
622 double p_of_agg = 100.0;
623
624 if (rs->agg[ddir]) {
625 p_of_agg = mean * 100 / (double) rs->agg[ddir];
626 if (p_of_agg > 100.0)
627 p_of_agg = 100.0;
628 }
629
630 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
631 } else
632 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
633}
634
635static void show_thread_status_terse_v2(struct thread_stat *ts,
636 struct group_run_stats *rs)
637{
638 double io_u_dist[FIO_IO_U_MAP_NR];
639 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
640 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
641 double usr_cpu, sys_cpu;
642 int i;
643
644 /* General Info */
645 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
646 /* Log Read Status */
647 show_ddir_status_terse(ts, rs, 0);
648 /* Log Write Status */
649 show_ddir_status_terse(ts, rs, 1);
650
651 /* CPU Usage */
652 if (ts->total_run_time) {
653 double runt = (double) ts->total_run_time;
654
655 usr_cpu = (double) ts->usr_time * 100 / runt;
656 sys_cpu = (double) ts->sys_time * 100 / runt;
657 } else {
658 usr_cpu = 0;
659 sys_cpu = 0;
660 }
661
662 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
663 ts->minf);
664
665 /* Calc % distribution of IO depths, usecond, msecond latency */
666 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
667 stat_calc_lat_u(ts, io_u_lat_u);
668 stat_calc_lat_m(ts, io_u_lat_m);
669
670 /* Only show fixed 7 I/O depth levels*/
671 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
672 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
673 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
674
675 /* Microsecond latency */
676 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
677 log_info(";%3.2f%%", io_u_lat_u[i]);
678 /* Millisecond latency */
679 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
680 log_info(";%3.2f%%", io_u_lat_m[i]);
681 /* Additional output if continue_on_error set - default off*/
682 if (ts->continue_on_error)
683 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
684 log_info("\n");
685
686 /* Additional output if description is set */
687 if (ts->description)
688 log_info(";%s", ts->description);
689
690 log_info("\n");
691}
692
693#define FIO_TERSE_VERSION "3"
694
695static void show_thread_status_terse_v3(struct thread_stat *ts,
696 struct group_run_stats *rs)
697{
698 double io_u_dist[FIO_IO_U_MAP_NR];
699 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
700 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
701 double usr_cpu, sys_cpu;
702 int i;
703
704 /* General Info */
705 log_info("%s;%s;%s;%d;%d", FIO_TERSE_VERSION, fio_version_string,
706 ts->name, ts->groupid, ts->error);
707 /* Log Read Status */
708 show_ddir_status_terse(ts, rs, 0);
709 /* Log Write Status */
710 show_ddir_status_terse(ts, rs, 1);
711
712 /* CPU Usage */
713 if (ts->total_run_time) {
714 double runt = (double) ts->total_run_time;
715
716 usr_cpu = (double) ts->usr_time * 100 / runt;
717 sys_cpu = (double) ts->sys_time * 100 / runt;
718 } else {
719 usr_cpu = 0;
720 sys_cpu = 0;
721 }
722
723 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
724 ts->minf);
725
726 /* Calc % distribution of IO depths, usecond, msecond latency */
727 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
728 stat_calc_lat_u(ts, io_u_lat_u);
729 stat_calc_lat_m(ts, io_u_lat_m);
730
731 /* Only show fixed 7 I/O depth levels*/
732 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
733 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
734 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
735
736 /* Microsecond latency */
737 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
738 log_info(";%3.2f%%", io_u_lat_u[i]);
739 /* Millisecond latency */
740 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
741 log_info(";%3.2f%%", io_u_lat_m[i]);
742
743 /* disk util stats, if any */
744 show_disk_util(1);
745
746 /* Additional output if continue_on_error set - default off*/
747 if (ts->continue_on_error)
748 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
749 log_info("\n");
750
751 /* Additional output if description is set */
752 if (strlen(ts->description))
753 log_info(";%s", ts->description);
754}
755
756static void show_thread_status_terse(struct thread_stat *ts,
757 struct group_run_stats *rs)
758{
759 if (terse_version == 2)
760 show_thread_status_terse_v2(ts, rs);
761 else if (terse_version == 3)
762 show_thread_status_terse_v3(ts, rs);
763 else
764 log_err("fio: bad terse version!? %d\n", terse_version);
765}
766
767static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
768{
769 double mean, S;
770
771 if (src->samples == 0)
772 return;
773
774 dst->min_val = min(dst->min_val, src->min_val);
775 dst->max_val = max(dst->max_val, src->max_val);
776
777 /*
778 * Compute new mean and S after the merge
779 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
780 * #Parallel_algorithm>
781 */
782 if (nr == 1) {
783 mean = src->mean.u.f;
784 S = src->S.u.f;
785 } else {
786 double delta = src->mean.u.f - dst->mean.u.f;
787
788 mean = ((src->mean.u.f * src->samples) +
789 (dst->mean.u.f * dst->samples)) /
790 (dst->samples + src->samples);
791
792 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
793 (dst->samples * src->samples) /
794 (dst->samples + src->samples);
795 }
796
797 dst->samples += src->samples;
798 dst->mean.u.f = mean;
799 dst->S.u.f = S;
800}
801
802void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
803{
804 int i;
805
806 for (i = 0; i < 2; i++) {
807 if (dst->max_run[i] < src->max_run[i])
808 dst->max_run[i] = src->max_run[i];
809 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
810 dst->min_run[i] = src->min_run[i];
811 if (dst->max_bw[i] < src->max_bw[i])
812 dst->max_bw[i] = src->max_bw[i];
813 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
814 dst->min_bw[i] = src->min_bw[i];
815
816 dst->io_kb[i] += src->io_kb[i];
817 dst->agg[i] += src->agg[i];
818 }
819
820}
821
822void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
823{
824 int l, k;
825
826 for (l = 0; l <= DDIR_WRITE; l++) {
827 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
828 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
829 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
830 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
831
832 dst->io_bytes[l] += src->io_bytes[l];
833
834 if (dst->runtime[l] < src->runtime[l])
835 dst->runtime[l] = src->runtime[l];
836 }
837
838 dst->usr_time += src->usr_time;
839 dst->sys_time += src->sys_time;
840 dst->ctx += src->ctx;
841 dst->majf += src->majf;
842 dst->minf += src->minf;
843
844 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
845 dst->io_u_map[k] += src->io_u_map[k];
846 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
847 dst->io_u_submit[k] += src->io_u_submit[k];
848 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
849 dst->io_u_complete[k] += src->io_u_complete[k];
850 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
851 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
852 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
853 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
854
855 for (k = 0; k <= 2; k++) {
856 dst->total_io_u[k] += src->total_io_u[k];
857 dst->short_io_u[k] += src->short_io_u[k];
858 }
859
860 for (k = 0; k <= DDIR_WRITE; k++) {
861 int m;
862 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
863 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
864 }
865
866 dst->total_run_time += src->total_run_time;
867 dst->total_submit += src->total_submit;
868 dst->total_complete += src->total_complete;
869}
870
871void init_group_run_stat(struct group_run_stats *gs)
872{
873 memset(gs, 0, sizeof(*gs));
874 gs->min_bw[0] = gs->min_run[0] = ~0UL;
875 gs->min_bw[1] = gs->min_run[1] = ~0UL;
876}
877
878void init_thread_stat(struct thread_stat *ts)
879{
880 int j;
881
882 memset(ts, 0, sizeof(*ts));
883
884 for (j = 0; j <= DDIR_WRITE; j++) {
885 ts->lat_stat[j].min_val = -1UL;
886 ts->clat_stat[j].min_val = -1UL;
887 ts->slat_stat[j].min_val = -1UL;
888 ts->bw_stat[j].min_val = -1UL;
889 }
890 ts->groupid = -1;
891}
892
893void show_run_stats(void)
894{
895 struct group_run_stats *runstats, *rs;
896 struct thread_data *td;
897 struct thread_stat *threadstats, *ts;
898 int i, j, nr_ts, last_ts, idx;
899 int kb_base_warned = 0;
900
901 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
902
903 for (i = 0; i < groupid + 1; i++)
904 init_group_run_stat(&runstats[i]);
905
906 /*
907 * find out how many threads stats we need. if group reporting isn't
908 * enabled, it's one-per-td.
909 */
910 nr_ts = 0;
911 last_ts = -1;
912 for_each_td(td, i) {
913 if (!td->o.group_reporting) {
914 nr_ts++;
915 continue;
916 }
917 if (last_ts == td->groupid)
918 continue;
919
920 last_ts = td->groupid;
921 nr_ts++;
922 }
923
924 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
925
926 for (i = 0; i < nr_ts; i++)
927 init_thread_stat(&threadstats[i]);
928
929 j = 0;
930 last_ts = -1;
931 idx = 0;
932 for_each_td(td, i) {
933 if (idx && (!td->o.group_reporting ||
934 (td->o.group_reporting && last_ts != td->groupid))) {
935 idx = 0;
936 j++;
937 }
938
939 last_ts = td->groupid;
940
941 ts = &threadstats[j];
942
943 ts->clat_percentiles = td->o.clat_percentiles;
944 if (td->o.overwrite_plist)
945 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
946 else
947 memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list));
948
949 idx++;
950 ts->members++;
951
952 if (ts->groupid == -1) {
953 /*
954 * These are per-group shared already
955 */
956 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
957 if (td->o.description)
958 strncpy(ts->description, td->o.description,
959 FIO_JOBNAME_SIZE);
960 else
961 memset(ts->description, 0, FIO_JOBNAME_SIZE);
962
963 ts->groupid = td->groupid;
964
965 /*
966 * first pid in group, not very useful...
967 */
968 ts->pid = td->pid;
969
970 ts->kb_base = td->o.kb_base;
971 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
972 log_info("fio: kb_base differs for jobs in group, using"
973 " %u as the base\n", ts->kb_base);
974 kb_base_warned = 1;
975 }
976
977 ts->continue_on_error = td->o.continue_on_error;
978 ts->total_err_count += td->total_err_count;
979 ts->first_error = td->first_error;
980 if (!ts->error) {
981 if (!td->error && td->o.continue_on_error &&
982 td->first_error) {
983 ts->error = td->first_error;
984 strcpy(ts->verror, td->verror);
985 } else if (td->error) {
986 ts->error = td->error;
987 strcpy(ts->verror, td->verror);
988 }
989 }
990
991 sum_thread_stats(ts, &td->ts, idx);
992 }
993
994 for (i = 0; i < nr_ts; i++) {
995 unsigned long long bw;
996
997 ts = &threadstats[i];
998 rs = &runstats[ts->groupid];
999 rs->kb_base = ts->kb_base;
1000
1001 for (j = 0; j <= DDIR_WRITE; j++) {
1002 if (!ts->runtime[j])
1003 continue;
1004 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1005 rs->min_run[j] = ts->runtime[j];
1006 if (ts->runtime[j] > rs->max_run[j])
1007 rs->max_run[j] = ts->runtime[j];
1008
1009 bw = 0;
1010 if (ts->runtime[j]) {
1011 unsigned long runt;
1012
1013 runt = ts->runtime[j];
1014 bw = ts->io_bytes[j] / runt;
1015 }
1016 if (bw < rs->min_bw[j])
1017 rs->min_bw[j] = bw;
1018 if (bw > rs->max_bw[j])
1019 rs->max_bw[j] = bw;
1020
1021 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1022 }
1023 }
1024
1025 for (i = 0; i < groupid + 1; i++) {
1026 unsigned long max_run[2];
1027
1028 rs = &runstats[i];
1029 max_run[0] = rs->max_run[0];
1030 max_run[1] = rs->max_run[1];
1031
1032 if (rs->max_run[0])
1033 rs->agg[0] = (rs->io_kb[0] * 1000) / max_run[0];
1034 if (rs->max_run[1])
1035 rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1];
1036 }
1037
1038 /*
1039 * don't overwrite last signal output
1040 */
1041 if (!terse_output)
1042 log_info("\n");
1043
1044 for (i = 0; i < nr_ts; i++) {
1045 ts = &threadstats[i];
1046 rs = &runstats[ts->groupid];
1047
1048 if (is_backend)
1049 fio_server_send_ts(ts, rs);
1050 else if (terse_output)
1051 show_thread_status_terse(ts, rs);
1052 else
1053 show_thread_status(ts, rs);
1054 }
1055
1056 for (i = 0; i < groupid + 1; i++) {
1057 rs = &runstats[i];
1058
1059 rs->groupid = i;
1060 if (is_backend)
1061 fio_server_send_gs(rs);
1062 else if (!terse_output)
1063 show_group_stats(rs);
1064 }
1065
1066 if (is_backend)
1067 fio_server_send_du();
1068 else if (!terse_output)
1069 show_disk_util(0);
1070
1071 free_disk_util();
1072
1073 free(runstats);
1074 free(threadstats);
1075}
1076
1077static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1078{
1079 double val = data;
1080 double delta;
1081
1082 if (data > is->max_val)
1083 is->max_val = data;
1084 if (data < is->min_val)
1085 is->min_val = data;
1086
1087 delta = val - is->mean.u.f;
1088 if (delta) {
1089 is->mean.u.f += delta / (is->samples + 1.0);
1090 is->S.u.f += delta * (val - is->mean.u.f);
1091 }
1092
1093 is->samples++;
1094}
1095
1096static void __add_log_sample(struct io_log *iolog, unsigned long val,
1097 enum fio_ddir ddir, unsigned int bs,
1098 unsigned long t)
1099{
1100 const int nr_samples = iolog->nr_samples;
1101
1102 if (!iolog->nr_samples)
1103 iolog->avg_last = t;
1104
1105 if (iolog->nr_samples == iolog->max_samples) {
1106 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
1107
1108 iolog->log = realloc(iolog->log, new_size);
1109 iolog->max_samples <<= 1;
1110 }
1111
1112 iolog->log[nr_samples].val = val;
1113 iolog->log[nr_samples].time = t;
1114 iolog->log[nr_samples].ddir = ddir;
1115 iolog->log[nr_samples].bs = bs;
1116 iolog->nr_samples++;
1117}
1118
1119static inline void reset_io_stat(struct io_stat *ios)
1120{
1121 ios->max_val = ios->min_val = ios->samples = 0;
1122 ios->mean.u.f = ios->S.u.f = 0;
1123}
1124
1125static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1126 unsigned long val, enum fio_ddir ddir,
1127 unsigned int bs)
1128{
1129 unsigned long elapsed, this_window;
1130
1131 if (!ddir_rw(ddir))
1132 return;
1133
1134 elapsed = mtime_since_now(&td->epoch);
1135
1136 /*
1137 * If no time averaging, just add the log sample.
1138 */
1139 if (!iolog->avg_msec) {
1140 __add_log_sample(iolog, val, ddir, bs, elapsed);
1141 return;
1142 }
1143
1144 /*
1145 * Add the sample. If the time period has passed, then
1146 * add that entry to the log and clear.
1147 */
1148 add_stat_sample(&iolog->avg_window[ddir], val);
1149
1150 /*
1151 * If period hasn't passed, adding the above sample is all we
1152 * need to do.
1153 */
1154 this_window = elapsed - iolog->avg_last;
1155 if (this_window < iolog->avg_msec)
1156 return;
1157
1158 /*
1159 * Note an entry in the log. Use the mean from the logged samples,
1160 * making sure to properly round up. Only write a log entry if we
1161 * had actual samples done.
1162 */
1163 if (iolog->avg_window[DDIR_READ].samples) {
1164 unsigned long mr;
1165
1166 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1167 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed);
1168 }
1169 if (iolog->avg_window[DDIR_WRITE].samples) {
1170 unsigned long mw;
1171
1172 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1173 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed);
1174 }
1175
1176 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1177 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1178 iolog->avg_last = elapsed;
1179}
1180
1181void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1182{
1183 struct io_log *iolog;
1184
1185 if (!ddir_rw(ddir))
1186 return;
1187
1188 iolog = agg_io_log[ddir];
1189 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
1190}
1191
1192static void add_clat_percentile_sample(struct thread_stat *ts,
1193 unsigned long usec, enum fio_ddir ddir)
1194{
1195 unsigned int idx = plat_val_to_idx(usec);
1196 assert(idx < FIO_IO_U_PLAT_NR);
1197
1198 ts->io_u_plat[ddir][idx]++;
1199}
1200
1201void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1202 unsigned long usec, unsigned int bs)
1203{
1204 struct thread_stat *ts = &td->ts;
1205
1206 if (!ddir_rw(ddir))
1207 return;
1208
1209 add_stat_sample(&ts->clat_stat[ddir], usec);
1210
1211 if (td->clat_log)
1212 add_log_sample(td, td->clat_log, usec, ddir, bs);
1213
1214 if (ts->clat_percentiles)
1215 add_clat_percentile_sample(ts, usec, ddir);
1216}
1217
1218void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1219 unsigned long usec, unsigned int bs)
1220{
1221 struct thread_stat *ts = &td->ts;
1222
1223 if (!ddir_rw(ddir))
1224 return;
1225
1226 add_stat_sample(&ts->slat_stat[ddir], usec);
1227
1228 if (td->slat_log)
1229 add_log_sample(td, td->slat_log, usec, ddir, bs);
1230}
1231
1232void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1233 unsigned long usec, unsigned int bs)
1234{
1235 struct thread_stat *ts = &td->ts;
1236
1237 if (!ddir_rw(ddir))
1238 return;
1239
1240 add_stat_sample(&ts->lat_stat[ddir], usec);
1241
1242 if (td->lat_log)
1243 add_log_sample(td, td->lat_log, usec, ddir, bs);
1244}
1245
1246void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1247 struct timeval *t)
1248{
1249 struct thread_stat *ts = &td->ts;
1250 unsigned long spent, rate;
1251
1252 if (!ddir_rw(ddir))
1253 return;
1254
1255 spent = mtime_since(&td->bw_sample_time, t);
1256 if (spent < td->o.bw_avg_time)
1257 return;
1258
1259 /*
1260 * Compute both read and write rates for the interval.
1261 */
1262 for (ddir = DDIR_READ; ddir <= DDIR_WRITE; ddir++) {
1263 uint64_t delta;
1264
1265 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1266 if (!delta)
1267 continue; /* No entries for interval */
1268
1269 rate = delta * 1000 / spent / 1024;
1270 add_stat_sample(&ts->bw_stat[ddir], rate);
1271
1272 if (td->bw_log)
1273 add_log_sample(td, td->bw_log, rate, ddir, bs);
1274
1275 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1276 }
1277
1278 fio_gettime(&td->bw_sample_time, NULL);
1279}
1280
1281void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
1282 struct timeval *t)
1283{
1284 struct thread_stat *ts = &td->ts;
1285 unsigned long spent, iops;
1286
1287 if (!ddir_rw(ddir))
1288 return;
1289
1290 spent = mtime_since(&td->iops_sample_time, t);
1291 if (spent < td->o.iops_avg_time)
1292 return;
1293
1294 /*
1295 * Compute both read and write rates for the interval.
1296 */
1297 for (ddir = DDIR_READ; ddir <= DDIR_WRITE; ddir++) {
1298 uint64_t delta;
1299
1300 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1301 if (!delta)
1302 continue; /* No entries for interval */
1303
1304 iops = (delta * 1000) / spent;
1305 add_stat_sample(&ts->iops_stat[ddir], iops);
1306
1307 if (td->iops_log)
1308 add_log_sample(td, td->iops_log, iops, ddir, 0);
1309
1310 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1311 }
1312
1313 fio_gettime(&td->iops_sample_time, NULL);
1314}
fio - Flexible IO Tester