projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Fio 1.99.9
[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
237static int 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 int usec_to_msec(unsigned long *min, unsigned long *max, double *mean,
343 double *dev)
344{
345 if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) {
346 *min /= 1000;
347 *max /= 1000;
348 *mean /= 1000.0;
349 *dev /= 1000.0;
350 return 0;
351 }
352
353 return 1;
354}
355
356static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
357 int ddir)
358{
359 const char *ddir_str[] = { "read ", "write" };
360 unsigned long min, max, runt;
361 unsigned long long bw, iops;
362 double mean, dev;
363 char *io_p, *bw_p, *iops_p;
364 int i2p;
365
366 assert(ddir_rw(ddir));
367
368 if (!ts->runtime[ddir])
369 return;
370
371 i2p = is_power_of_2(rs->kb_base);
372 runt = ts->runtime[ddir];
373
374 bw = (1000 * ts->io_bytes[ddir]) / runt;
375 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p);
376 bw_p = num2str(bw, 6, 1, i2p);
377
378 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
379 iops_p = num2str(iops, 6, 1, 0);
380
381 log_info(" %s: io=%sB, bw=%sB/s, iops=%s, runt=%6llumsec\n",
382 ddir_str[ddir], io_p, bw_p, iops_p,
383 ts->runtime[ddir]);
384
385 free(io_p);
386 free(bw_p);
387 free(iops_p);
388
389 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
390 const char *base = "(usec)";
391 char *minp, *maxp;
392
393 if (!usec_to_msec(&min, &max, &mean, &dev))
394 base = "(msec)";
395
396 minp = num2str(min, 6, 1, 0);
397 maxp = num2str(max, 6, 1, 0);
398
399 log_info(" slat %s: min=%s, max=%s, avg=%5.02f,"
400 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
401
402 free(minp);
403 free(maxp);
404 }
405 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
406 const char *base = "(usec)";
407 char *minp, *maxp;
408
409 if (!usec_to_msec(&min, &max, &mean, &dev))
410 base = "(msec)";
411
412 minp = num2str(min, 6, 1, 0);
413 maxp = num2str(max, 6, 1, 0);
414
415 log_info(" clat %s: min=%s, max=%s, avg=%5.02f,"
416 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
417
418 free(minp);
419 free(maxp);
420 }
421 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
422 const char *base = "(usec)";
423 char *minp, *maxp;
424
425 if (!usec_to_msec(&min, &max, &mean, &dev))
426 base = "(msec)";
427
428 minp = num2str(min, 6, 1, 0);
429 maxp = num2str(max, 6, 1, 0);
430
431 log_info(" lat %s: min=%s, max=%s, avg=%5.02f,"
432 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
433
434 free(minp);
435 free(maxp);
436 }
437 if (ts->clat_percentiles) {
438 show_clat_percentiles(ts->io_u_plat[ddir],
439 ts->clat_stat[ddir].samples,
440 ts->percentile_list);
441 }
442 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
443 double p_of_agg;
444 const char *bw_str = "KB";
445
446 p_of_agg = mean * 100 / (double) rs->agg[ddir];
447 if (p_of_agg > 100.0)
448 p_of_agg = 100.0;
449
450 if (mean > 999999.9) {
451 min /= 1000.0;
452 max /= 1000.0;
453 mean /= 1000.0;
454 dev /= 1000.0;
455 bw_str = "MB";
456 }
457
458 log_info(" bw (%s/s) : min=%5lu, max=%5lu, per=%3.2f%%,"
459 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
460 p_of_agg, mean, dev);
461 }
462}
463
464static int show_lat(double *io_u_lat, int nr, const char **ranges,
465 const char *msg)
466{
467 int new_line = 1, i, line = 0, shown = 0;
468
469 for (i = 0; i < nr; i++) {
470 if (io_u_lat[i] <= 0.0)
471 continue;
472 shown = 1;
473 if (new_line) {
474 if (line)
475 log_info("\n");
476 log_info(" lat (%s) : ", msg);
477 new_line = 0;
478 line = 0;
479 }
480 if (line)
481 log_info(", ");
482 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
483 line++;
484 if (line == 5)
485 new_line = 1;
486 }
487
488 if (shown)
489 log_info("\n");
490
491 return shown;
492}
493
494static void show_lat_u(double *io_u_lat_u)
495{
496 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
497 "250=", "500=", "750=", "1000=", };
498
499 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
500}
501
502static void show_lat_m(double *io_u_lat_m)
503{
504 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
505 "250=", "500=", "750=", "1000=", "2000=",
506 ">=2000=", };
507
508 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
509}
510
511static void show_latencies(double *io_u_lat_u, double *io_u_lat_m)
512{
513 show_lat_u(io_u_lat_u);
514 show_lat_m(io_u_lat_m);
515}
516
517void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
518{
519 double usr_cpu, sys_cpu;
520 unsigned long runtime;
521 double io_u_dist[FIO_IO_U_MAP_NR];
522 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
523 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
524
525 if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
526 !(ts->total_io_u[0] + ts->total_io_u[1]))
527 return;
528
529 if (!ts->error) {
530 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
531 ts->name, ts->groupid, ts->members,
532 ts->error, (int) ts->pid);
533 } else {
534 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
535 ts->name, ts->groupid, ts->members,
536 ts->error, ts->verror, (int) ts->pid);
537 }
538
539 if (strlen(ts->description))
540 log_info(" Description : [%s]\n", ts->description);
541
542 if (ts->io_bytes[DDIR_READ])
543 show_ddir_status(rs, ts, DDIR_READ);
544 if (ts->io_bytes[DDIR_WRITE])
545 show_ddir_status(rs, ts, DDIR_WRITE);
546
547 stat_calc_lat_u(ts, io_u_lat_u);
548 stat_calc_lat_m(ts, io_u_lat_m);
549 show_latencies(io_u_lat_u, io_u_lat_m);
550
551 runtime = ts->total_run_time;
552 if (runtime) {
553 double runt = (double) runtime;
554
555 usr_cpu = (double) ts->usr_time * 100 / runt;
556 sys_cpu = (double) ts->sys_time * 100 / runt;
557 } else {
558 usr_cpu = 0;
559 sys_cpu = 0;
560 }
561
562 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
563 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
564
565 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
566 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
567 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
568 io_u_dist[1], io_u_dist[2],
569 io_u_dist[3], io_u_dist[4],
570 io_u_dist[5], io_u_dist[6]);
571
572 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
573 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
574 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
575 io_u_dist[1], io_u_dist[2],
576 io_u_dist[3], io_u_dist[4],
577 io_u_dist[5], io_u_dist[6]);
578 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
579 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
580 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
581 io_u_dist[1], io_u_dist[2],
582 io_u_dist[3], io_u_dist[4],
583 io_u_dist[5], io_u_dist[6]);
584 log_info(" issued : total=r=%lu/w=%lu/d=%lu,"
585 " short=r=%lu/w=%lu/d=%lu\n",
586 ts->total_io_u[0], ts->total_io_u[1],
587 ts->total_io_u[2],
588 ts->short_io_u[0], ts->short_io_u[1],
589 ts->short_io_u[2]);
590 if (ts->continue_on_error) {
591 log_info(" errors : total=%lu, first_error=%d/<%s>\n",
592 ts->total_err_count,
593 ts->first_error,
594 strerror(ts->first_error));
595 }
596}
597
598static void show_ddir_status_terse(struct thread_stat *ts,
599 struct group_run_stats *rs, int ddir)
600{
601 unsigned long min, max;
602 unsigned long long bw, iops;
603 unsigned int *ovals = NULL;
604 double mean, dev;
605 unsigned int len, minv, maxv;
606 int i;
607
608 assert(ddir_rw(ddir));
609
610 iops = bw = 0;
611 if (ts->runtime[ddir]) {
612 uint64_t runt = ts->runtime[ddir];
613
614 bw = ts->io_bytes[ddir] / runt;
615 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
616 }
617
618 log_info(";%llu;%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, iops,
619 ts->runtime[ddir]);
620
621 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
622 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
623 else
624 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
625
626 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
627 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
628 else
629 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
630
631 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
632 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
633 else
634 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
635
636 if (ts->clat_percentiles) {
637 len = calc_clat_percentiles(ts->io_u_plat[ddir],
638 ts->clat_stat[ddir].samples,
639 ts->percentile_list, &ovals, &maxv,
640 &minv);
641 } else
642 len = 0;
643
644 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
645 if (i >= len) {
646 log_info(";0%%=0");
647 continue;
648 }
649 log_info(";%2.2f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
650 }
651 if (ovals)
652 free(ovals);
653
654 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
655 double p_of_agg;
656
657 p_of_agg = mean * 100 / (double) rs->agg[ddir];
658 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
659 } else
660 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
661}
662
663static void show_thread_status_terse_v2(struct thread_stat *ts,
664 struct group_run_stats *rs)
665{
666 double io_u_dist[FIO_IO_U_MAP_NR];
667 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
668 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
669 double usr_cpu, sys_cpu;
670 int i;
671
672 /* General Info */
673 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
674 /* Log Read Status */
675 show_ddir_status_terse(ts, rs, 0);
676 /* Log Write Status */
677 show_ddir_status_terse(ts, rs, 1);
678
679 /* CPU Usage */
680 if (ts->total_run_time) {
681 double runt = (double) ts->total_run_time;
682
683 usr_cpu = (double) ts->usr_time * 100 / runt;
684 sys_cpu = (double) ts->sys_time * 100 / runt;
685 } else {
686 usr_cpu = 0;
687 sys_cpu = 0;
688 }
689
690 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
691 ts->minf);
692
693 /* Calc % distribution of IO depths, usecond, msecond latency */
694 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
695 stat_calc_lat_u(ts, io_u_lat_u);
696 stat_calc_lat_m(ts, io_u_lat_m);
697
698 /* Only show fixed 7 I/O depth levels*/
699 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
700 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
701 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
702
703 /* Microsecond latency */
704 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
705 log_info(";%3.2f%%", io_u_lat_u[i]);
706 /* Millisecond latency */
707 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
708 log_info(";%3.2f%%", io_u_lat_m[i]);
709 /* Additional output if continue_on_error set - default off*/
710 if (ts->continue_on_error)
711 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
712 log_info("\n");
713
714 /* Additional output if description is set */
715 if (ts->description)
716 log_info(";%s", ts->description);
717
718 log_info("\n");
719}
720
721#define FIO_TERSE_VERSION "3"
722
723static void show_thread_status_terse_v3(struct thread_stat *ts,
724 struct group_run_stats *rs)
725{
726 double io_u_dist[FIO_IO_U_MAP_NR];
727 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
728 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
729 double usr_cpu, sys_cpu;
730 int i;
731
732 /* General Info */
733 log_info("%s;%s;%s;%d;%d", FIO_TERSE_VERSION, fio_version_string,
734 ts->name, ts->groupid, ts->error);
735 /* Log Read Status */
736 show_ddir_status_terse(ts, rs, 0);
737 /* Log Write Status */
738 show_ddir_status_terse(ts, rs, 1);
739
740 /* CPU Usage */
741 if (ts->total_run_time) {
742 double runt = (double) ts->total_run_time;
743
744 usr_cpu = (double) ts->usr_time * 100 / runt;
745 sys_cpu = (double) ts->sys_time * 100 / runt;
746 } else {
747 usr_cpu = 0;
748 sys_cpu = 0;
749 }
750
751 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
752 ts->minf);
753
754 /* Calc % distribution of IO depths, usecond, msecond latency */
755 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
756 stat_calc_lat_u(ts, io_u_lat_u);
757 stat_calc_lat_m(ts, io_u_lat_m);
758
759 /* Only show fixed 7 I/O depth levels*/
760 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
761 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
762 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
763
764 /* Microsecond latency */
765 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
766 log_info(";%3.2f%%", io_u_lat_u[i]);
767 /* Millisecond latency */
768 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
769 log_info(";%3.2f%%", io_u_lat_m[i]);
770
771 /* disk util stats, if any */
772 show_disk_util(1);
773
774 /* Additional output if continue_on_error set - default off*/
775 if (ts->continue_on_error)
776 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
777 log_info("\n");
778
779 /* Additional output if description is set */
780 if (strlen(ts->description))
781 log_info(";%s", ts->description);
782}
783
784static void show_thread_status_terse(struct thread_stat *ts,
785 struct group_run_stats *rs)
786{
787 if (terse_version == 2)
788 show_thread_status_terse_v2(ts, rs);
789 else if (terse_version == 3)
790 show_thread_status_terse_v3(ts, rs);
791 else
792 log_err("fio: bad terse version!? %d\n", terse_version);
793}
794
795static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
796{
797 double mean, S;
798
799 if (src->samples == 0)
800 return;
801
802 dst->min_val = min(dst->min_val, src->min_val);
803 dst->max_val = max(dst->max_val, src->max_val);
804
805 /*
806 * Compute new mean and S after the merge
807 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
808 * #Parallel_algorithm>
809 */
810 if (nr == 1) {
811 mean = src->mean.u.f;
812 S = src->S.u.f;
813 } else {
814 double delta = src->mean.u.f - dst->mean.u.f;
815
816 mean = ((src->mean.u.f * src->samples) +
817 (dst->mean.u.f * dst->samples)) /
818 (dst->samples + src->samples);
819
820 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
821 (dst->samples * src->samples) /
822 (dst->samples + src->samples);
823 }
824
825 dst->samples += src->samples;
826 dst->mean.u.f = mean;
827 dst->S.u.f = S;
828}
829
830void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
831{
832 int i;
833
834 for (i = 0; i < 2; i++) {
835 if (dst->max_run[i] < src->max_run[i])
836 dst->max_run[i] = src->max_run[i];
837 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
838 dst->min_run[i] = src->min_run[i];
839 if (dst->max_bw[i] < src->max_bw[i])
840 dst->max_bw[i] = src->max_bw[i];
841 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
842 dst->min_bw[i] = src->min_bw[i];
843
844 dst->io_kb[i] += src->io_kb[i];
845 dst->agg[i] += src->agg[i];
846 }
847
848}
849
850void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
851{
852 int l, k;
853
854 for (l = 0; l <= DDIR_WRITE; l++) {
855 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
856 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
857 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
858 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
859
860 dst->io_bytes[l] += src->io_bytes[l];
861
862 if (dst->runtime[l] < src->runtime[l])
863 dst->runtime[l] = src->runtime[l];
864 }
865
866 dst->usr_time += src->usr_time;
867 dst->sys_time += src->sys_time;
868 dst->ctx += src->ctx;
869 dst->majf += src->majf;
870 dst->minf += src->minf;
871
872 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
873 dst->io_u_map[k] += src->io_u_map[k];
874 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
875 dst->io_u_submit[k] += src->io_u_submit[k];
876 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
877 dst->io_u_complete[k] += src->io_u_complete[k];
878 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
879 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
880 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
881 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
882
883 for (k = 0; k <= 2; k++) {
884 dst->total_io_u[k] += src->total_io_u[k];
885 dst->short_io_u[k] += src->short_io_u[k];
886 }
887
888 for (k = 0; k <= DDIR_WRITE; k++) {
889 int m;
890 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
891 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
892 }
893
894 dst->total_run_time += src->total_run_time;
895 dst->total_submit += src->total_submit;
896 dst->total_complete += src->total_complete;
897}
898
899void init_group_run_stat(struct group_run_stats *gs)
900{
901 memset(gs, 0, sizeof(*gs));
902 gs->min_bw[0] = gs->min_run[0] = ~0UL;
903 gs->min_bw[1] = gs->min_run[1] = ~0UL;
904}
905
906void init_thread_stat(struct thread_stat *ts)
907{
908 int j;
909
910 memset(ts, 0, sizeof(*ts));
911
912 for (j = 0; j <= DDIR_WRITE; j++) {
913 ts->lat_stat[j].min_val = -1UL;
914 ts->clat_stat[j].min_val = -1UL;
915 ts->slat_stat[j].min_val = -1UL;
916 ts->bw_stat[j].min_val = -1UL;
917 }
918 ts->groupid = -1;
919}
920
921void show_run_stats(void)
922{
923 struct group_run_stats *runstats, *rs;
924 struct thread_data *td;
925 struct thread_stat *threadstats, *ts;
926 int i, j, nr_ts, last_ts, idx;
927 int kb_base_warned = 0;
928
929 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
930
931 for (i = 0; i < groupid + 1; i++)
932 init_group_run_stat(&runstats[i]);
933
934 /*
935 * find out how many threads stats we need. if group reporting isn't
936 * enabled, it's one-per-td.
937 */
938 nr_ts = 0;
939 last_ts = -1;
940 for_each_td(td, i) {
941 if (!td->o.group_reporting) {
942 nr_ts++;
943 continue;
944 }
945 if (last_ts == td->groupid)
946 continue;
947
948 last_ts = td->groupid;
949 nr_ts++;
950 }
951
952 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
953
954 for (i = 0; i < nr_ts; i++)
955 init_thread_stat(&threadstats[i]);
956
957 j = 0;
958 last_ts = -1;
959 idx = 0;
960 for_each_td(td, i) {
961 if (idx && (!td->o.group_reporting ||
962 (td->o.group_reporting && last_ts != td->groupid))) {
963 idx = 0;
964 j++;
965 }
966
967 last_ts = td->groupid;
968
969 ts = &threadstats[j];
970
971 ts->clat_percentiles = td->o.clat_percentiles;
972 if (td->o.overwrite_plist)
973 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
974 else
975 memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list));
976
977 idx++;
978 ts->members++;
979
980 if (ts->groupid == -1) {
981 /*
982 * These are per-group shared already
983 */
984 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
985 if (td->o.description)
986 strncpy(ts->description, td->o.description,
987 FIO_JOBNAME_SIZE);
988 else
989 memset(ts->description, 0, FIO_JOBNAME_SIZE);
990
991 ts->groupid = td->groupid;
992
993 /*
994 * first pid in group, not very useful...
995 */
996 ts->pid = td->pid;
997
998 ts->kb_base = td->o.kb_base;
999 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1000 log_info("fio: kb_base differs for jobs in group, using"
1001 " %u as the base\n", ts->kb_base);
1002 kb_base_warned = 1;
1003 }
1004
1005 ts->continue_on_error = td->o.continue_on_error;
1006 ts->total_err_count += td->total_err_count;
1007 ts->first_error = td->first_error;
1008 if (!ts->error) {
1009 if (!td->error && td->o.continue_on_error &&
1010 td->first_error) {
1011 ts->error = td->first_error;
1012 strcpy(ts->verror, td->verror);
1013 } else if (td->error) {
1014 ts->error = td->error;
1015 strcpy(ts->verror, td->verror);
1016 }
1017 }
1018
1019 sum_thread_stats(ts, &td->ts, idx);
1020 }
1021
1022 for (i = 0; i < nr_ts; i++) {
1023 unsigned long long bw;
1024
1025 ts = &threadstats[i];
1026 rs = &runstats[ts->groupid];
1027 rs->kb_base = ts->kb_base;
1028
1029 for (j = 0; j <= DDIR_WRITE; j++) {
1030 if (!ts->runtime[j])
1031 continue;
1032 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1033 rs->min_run[j] = ts->runtime[j];
1034 if (ts->runtime[j] > rs->max_run[j])
1035 rs->max_run[j] = ts->runtime[j];
1036
1037 bw = 0;
1038 if (ts->runtime[j]) {
1039 unsigned long runt;
1040
1041 runt = ts->runtime[j];
1042 bw = ts->io_bytes[j] / runt;
1043 }
1044 if (bw < rs->min_bw[j])
1045 rs->min_bw[j] = bw;
1046 if (bw > rs->max_bw[j])
1047 rs->max_bw[j] = bw;
1048
1049 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1050 }
1051 }
1052
1053 for (i = 0; i < groupid + 1; i++) {
1054 unsigned long max_run[2];
1055
1056 rs = &runstats[i];
1057 max_run[0] = rs->max_run[0];
1058 max_run[1] = rs->max_run[1];
1059
1060 if (rs->max_run[0])
1061 rs->agg[0] = (rs->io_kb[0] * 1000) / max_run[0];
1062 if (rs->max_run[1])
1063 rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1];
1064 }
1065
1066 /*
1067 * don't overwrite last signal output
1068 */
1069 if (!terse_output)
1070 log_info("\n");
1071
1072 for (i = 0; i < nr_ts; i++) {
1073 ts = &threadstats[i];
1074 rs = &runstats[ts->groupid];
1075
1076 if (is_backend)
1077 fio_server_send_ts(ts, rs);
1078 else if (terse_output)
1079 show_thread_status_terse(ts, rs);
1080 else
1081 show_thread_status(ts, rs);
1082 }
1083
1084 for (i = 0; i < groupid + 1; i++) {
1085 rs = &runstats[i];
1086
1087 rs->groupid = i;
1088 if (is_backend)
1089 fio_server_send_gs(rs);
1090 else if (!terse_output)
1091 show_group_stats(rs);
1092 }
1093
1094 if (is_backend)
1095 fio_server_send_du();
1096 else if (!terse_output)
1097 show_disk_util(0);
1098
1099 free_disk_util();
1100
1101 free(runstats);
1102 free(threadstats);
1103}
1104
1105static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1106{
1107 double val = data;
1108 double delta;
1109
1110 if (data > is->max_val)
1111 is->max_val = data;
1112 if (data < is->min_val)
1113 is->min_val = data;
1114
1115 delta = val - is->mean.u.f;
1116 if (delta) {
1117 is->mean.u.f += delta / (is->samples + 1.0);
1118 is->S.u.f += delta * (val - is->mean.u.f);
1119 }
1120
1121 is->samples++;
1122}
1123
1124static void __add_log_sample(struct io_log *iolog, unsigned long val,
1125 enum fio_ddir ddir, unsigned int bs,
1126 unsigned long t)
1127{
1128 const int nr_samples = iolog->nr_samples;
1129
1130 if (iolog->nr_samples == iolog->max_samples) {
1131 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
1132
1133 iolog->log = realloc(iolog->log, new_size);
1134 iolog->max_samples <<= 1;
1135 }
1136
1137 iolog->log[nr_samples].val = val;
1138 iolog->log[nr_samples].time = t;
1139 iolog->log[nr_samples].ddir = ddir;
1140 iolog->log[nr_samples].bs = bs;
1141 iolog->nr_samples++;
1142}
1143
1144static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1145 unsigned long val, enum fio_ddir ddir,
1146 unsigned int bs)
1147{
1148 if (!ddir_rw(ddir))
1149 return;
1150
1151 __add_log_sample(iolog, val, ddir, bs, mtime_since_now(&td->epoch));
1152}
1153
1154void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1155{
1156 struct io_log *iolog;
1157
1158 if (!ddir_rw(ddir))
1159 return;
1160
1161 iolog = agg_io_log[ddir];
1162 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
1163}
1164
1165static void add_clat_percentile_sample(struct thread_stat *ts,
1166 unsigned long usec, enum fio_ddir ddir)
1167{
1168 unsigned int idx = plat_val_to_idx(usec);
1169 assert(idx < FIO_IO_U_PLAT_NR);
1170
1171 ts->io_u_plat[ddir][idx]++;
1172}
1173
1174void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1175 unsigned long usec, unsigned int bs)
1176{
1177 struct thread_stat *ts = &td->ts;
1178
1179 if (!ddir_rw(ddir))
1180 return;
1181
1182 add_stat_sample(&ts->clat_stat[ddir], usec);
1183
1184 if (td->clat_log)
1185 add_log_sample(td, td->clat_log, usec, ddir, bs);
1186
1187 if (ts->clat_percentiles)
1188 add_clat_percentile_sample(ts, usec, ddir);
1189}
1190
1191void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1192 unsigned long usec, unsigned int bs)
1193{
1194 struct thread_stat *ts = &td->ts;
1195
1196 if (!ddir_rw(ddir))
1197 return;
1198
1199 add_stat_sample(&ts->slat_stat[ddir], usec);
1200
1201 if (td->slat_log)
1202 add_log_sample(td, td->slat_log, usec, ddir, bs);
1203}
1204
1205void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1206 unsigned long usec, unsigned int bs)
1207{
1208 struct thread_stat *ts = &td->ts;
1209
1210 if (!ddir_rw(ddir))
1211 return;
1212
1213 add_stat_sample(&ts->lat_stat[ddir], usec);
1214
1215 if (td->lat_log)
1216 add_log_sample(td, td->lat_log, usec, ddir, bs);
1217}
1218
1219void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1220 struct timeval *t)
1221{
1222 struct thread_stat *ts = &td->ts;
1223 unsigned long spent, rate;
1224
1225 if (!ddir_rw(ddir))
1226 return;
1227
1228 spent = mtime_since(&td->bw_sample_time, t);
1229 if (spent < td->o.bw_avg_time)
1230 return;
1231
1232 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) *
1233 1000 / spent / 1024;
1234 add_stat_sample(&ts->bw_stat[ddir], rate);
1235
1236 if (td->bw_log)
1237 add_log_sample(td, td->bw_log, rate, ddir, bs);
1238
1239 fio_gettime(&td->bw_sample_time, NULL);
1240 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1241}
1242
1243void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
1244 struct timeval *t)
1245{
1246 struct thread_stat *ts = &td->ts;
1247 unsigned long spent, iops;
1248
1249 if (!ddir_rw(ddir))
1250 return;
1251
1252 spent = mtime_since(&td->iops_sample_time, t);
1253 if (spent < td->o.iops_avg_time)
1254 return;
1255
1256 iops = ((td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]) * 1000) / spent;
1257
1258 add_stat_sample(&ts->iops_stat[ddir], iops);
1259
1260 if (td->iops_log) {
1261 assert(iops);
1262 add_log_sample(td, td->iops_log, iops, ddir, 0);
1263 }
1264
1265 fio_gettime(&td->iops_sample_time, NULL);
1266 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1267}
fio - Flexible IO Tester