projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Fix assumption that pointers fits in a 'long'
[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 = 100.0;
444 const char *bw_str = "KB";
445
446 if (rs->agg[ddir]) {
447 p_of_agg = mean * 100 / (double) rs->agg[ddir];
448 if (p_of_agg > 100.0)
449 p_of_agg = 100.0;
450 }
451
452 if (mean > 999999.9) {
453 min /= 1000.0;
454 max /= 1000.0;
455 mean /= 1000.0;
456 dev /= 1000.0;
457 bw_str = "MB";
458 }
459
460 log_info(" bw (%s/s) : min=%5lu, max=%5lu, per=%3.2f%%,"
461 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
462 p_of_agg, mean, dev);
463 }
464}
465
466static int show_lat(double *io_u_lat, int nr, const char **ranges,
467 const char *msg)
468{
469 int new_line = 1, i, line = 0, shown = 0;
470
471 for (i = 0; i < nr; i++) {
472 if (io_u_lat[i] <= 0.0)
473 continue;
474 shown = 1;
475 if (new_line) {
476 if (line)
477 log_info("\n");
478 log_info(" lat (%s) : ", msg);
479 new_line = 0;
480 line = 0;
481 }
482 if (line)
483 log_info(", ");
484 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
485 line++;
486 if (line == 5)
487 new_line = 1;
488 }
489
490 if (shown)
491 log_info("\n");
492
493 return shown;
494}
495
496static void show_lat_u(double *io_u_lat_u)
497{
498 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
499 "250=", "500=", "750=", "1000=", };
500
501 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
502}
503
504static void show_lat_m(double *io_u_lat_m)
505{
506 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
507 "250=", "500=", "750=", "1000=", "2000=",
508 ">=2000=", };
509
510 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
511}
512
513static void show_latencies(double *io_u_lat_u, double *io_u_lat_m)
514{
515 show_lat_u(io_u_lat_u);
516 show_lat_m(io_u_lat_m);
517}
518
519void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
520{
521 double usr_cpu, sys_cpu;
522 unsigned long runtime;
523 double io_u_dist[FIO_IO_U_MAP_NR];
524 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
525 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
526
527 if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
528 !(ts->total_io_u[0] + ts->total_io_u[1]))
529 return;
530
531 if (!ts->error) {
532 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
533 ts->name, ts->groupid, ts->members,
534 ts->error, (int) ts->pid);
535 } else {
536 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
537 ts->name, ts->groupid, ts->members,
538 ts->error, ts->verror, (int) ts->pid);
539 }
540
541 if (strlen(ts->description))
542 log_info(" Description : [%s]\n", ts->description);
543
544 if (ts->io_bytes[DDIR_READ])
545 show_ddir_status(rs, ts, DDIR_READ);
546 if (ts->io_bytes[DDIR_WRITE])
547 show_ddir_status(rs, ts, DDIR_WRITE);
548
549 stat_calc_lat_u(ts, io_u_lat_u);
550 stat_calc_lat_m(ts, io_u_lat_m);
551 show_latencies(io_u_lat_u, io_u_lat_m);
552
553 runtime = ts->total_run_time;
554 if (runtime) {
555 double runt = (double) runtime;
556
557 usr_cpu = (double) ts->usr_time * 100 / runt;
558 sys_cpu = (double) ts->sys_time * 100 / runt;
559 } else {
560 usr_cpu = 0;
561 sys_cpu = 0;
562 }
563
564 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
565 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
566
567 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
568 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
569 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
570 io_u_dist[1], io_u_dist[2],
571 io_u_dist[3], io_u_dist[4],
572 io_u_dist[5], io_u_dist[6]);
573
574 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
575 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
576 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
577 io_u_dist[1], io_u_dist[2],
578 io_u_dist[3], io_u_dist[4],
579 io_u_dist[5], io_u_dist[6]);
580 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
581 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
582 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
583 io_u_dist[1], io_u_dist[2],
584 io_u_dist[3], io_u_dist[4],
585 io_u_dist[5], io_u_dist[6]);
586 log_info(" issued : total=r=%lu/w=%lu/d=%lu,"
587 " short=r=%lu/w=%lu/d=%lu\n",
588 ts->total_io_u[0], ts->total_io_u[1],
589 ts->total_io_u[2],
590 ts->short_io_u[0], ts->short_io_u[1],
591 ts->short_io_u[2]);
592 if (ts->continue_on_error) {
593 log_info(" errors : total=%lu, first_error=%d/<%s>\n",
594 ts->total_err_count,
595 ts->first_error,
596 strerror(ts->first_error));
597 }
598}
599
600static void show_ddir_status_terse(struct thread_stat *ts,
601 struct group_run_stats *rs, int ddir)
602{
603 unsigned long min, max;
604 unsigned long long bw, iops;
605 unsigned int *ovals = NULL;
606 double mean, dev;
607 unsigned int len, minv, maxv;
608 int i;
609
610 assert(ddir_rw(ddir));
611
612 iops = bw = 0;
613 if (ts->runtime[ddir]) {
614 uint64_t runt = ts->runtime[ddir];
615
616 bw = ts->io_bytes[ddir] / runt;
617 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
618 }
619
620 log_info(";%llu;%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, iops,
621 ts->runtime[ddir]);
622
623 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
624 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
625 else
626 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
627
628 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
629 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
630 else
631 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
632
633 if (ts->clat_percentiles) {
634 len = calc_clat_percentiles(ts->io_u_plat[ddir],
635 ts->clat_stat[ddir].samples,
636 ts->percentile_list, &ovals, &maxv,
637 &minv);
638 } else
639 len = 0;
640
641 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
642 if (i >= len) {
643 log_info(";0%%=0");
644 continue;
645 }
646 log_info(";%2.2f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
647 }
648
649 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
650 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
651 else
652 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
653
654 if (ovals)
655 free(ovals);
656
657 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
658 double p_of_agg = 100.0;
659
660 if (rs->agg[ddir]) {
661 p_of_agg = mean * 100 / (double) rs->agg[ddir];
662 if (p_of_agg > 100.0)
663 p_of_agg = 100.0;
664 }
665
666 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
667 } else
668 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
669}
670
671static void show_thread_status_terse_v2(struct thread_stat *ts,
672 struct group_run_stats *rs)
673{
674 double io_u_dist[FIO_IO_U_MAP_NR];
675 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
676 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
677 double usr_cpu, sys_cpu;
678 int i;
679
680 /* General Info */
681 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
682 /* Log Read Status */
683 show_ddir_status_terse(ts, rs, 0);
684 /* Log Write Status */
685 show_ddir_status_terse(ts, rs, 1);
686
687 /* CPU Usage */
688 if (ts->total_run_time) {
689 double runt = (double) ts->total_run_time;
690
691 usr_cpu = (double) ts->usr_time * 100 / runt;
692 sys_cpu = (double) ts->sys_time * 100 / runt;
693 } else {
694 usr_cpu = 0;
695 sys_cpu = 0;
696 }
697
698 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
699 ts->minf);
700
701 /* Calc % distribution of IO depths, usecond, msecond latency */
702 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
703 stat_calc_lat_u(ts, io_u_lat_u);
704 stat_calc_lat_m(ts, io_u_lat_m);
705
706 /* Only show fixed 7 I/O depth levels*/
707 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
708 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
709 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
710
711 /* Microsecond latency */
712 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
713 log_info(";%3.2f%%", io_u_lat_u[i]);
714 /* Millisecond latency */
715 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
716 log_info(";%3.2f%%", io_u_lat_m[i]);
717 /* Additional output if continue_on_error set - default off*/
718 if (ts->continue_on_error)
719 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
720 log_info("\n");
721
722 /* Additional output if description is set */
723 if (ts->description)
724 log_info(";%s", ts->description);
725
726 log_info("\n");
727}
728
729#define FIO_TERSE_VERSION "3"
730
731static void show_thread_status_terse_v3(struct thread_stat *ts,
732 struct group_run_stats *rs)
733{
734 double io_u_dist[FIO_IO_U_MAP_NR];
735 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
736 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
737 double usr_cpu, sys_cpu;
738 int i;
739
740 /* General Info */
741 log_info("%s;%s;%s;%d;%d", FIO_TERSE_VERSION, fio_version_string,
742 ts->name, ts->groupid, ts->error);
743 /* Log Read Status */
744 show_ddir_status_terse(ts, rs, 0);
745 /* Log Write Status */
746 show_ddir_status_terse(ts, rs, 1);
747
748 /* CPU Usage */
749 if (ts->total_run_time) {
750 double runt = (double) ts->total_run_time;
751
752 usr_cpu = (double) ts->usr_time * 100 / runt;
753 sys_cpu = (double) ts->sys_time * 100 / runt;
754 } else {
755 usr_cpu = 0;
756 sys_cpu = 0;
757 }
758
759 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
760 ts->minf);
761
762 /* Calc % distribution of IO depths, usecond, msecond latency */
763 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
764 stat_calc_lat_u(ts, io_u_lat_u);
765 stat_calc_lat_m(ts, io_u_lat_m);
766
767 /* Only show fixed 7 I/O depth levels*/
768 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
769 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
770 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
771
772 /* Microsecond latency */
773 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
774 log_info(";%3.2f%%", io_u_lat_u[i]);
775 /* Millisecond latency */
776 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
777 log_info(";%3.2f%%", io_u_lat_m[i]);
778
779 /* disk util stats, if any */
780 show_disk_util(1);
781
782 /* Additional output if continue_on_error set - default off*/
783 if (ts->continue_on_error)
784 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
785 log_info("\n");
786
787 /* Additional output if description is set */
788 if (strlen(ts->description))
789 log_info(";%s", ts->description);
790}
791
792static void show_thread_status_terse(struct thread_stat *ts,
793 struct group_run_stats *rs)
794{
795 if (terse_version == 2)
796 show_thread_status_terse_v2(ts, rs);
797 else if (terse_version == 3)
798 show_thread_status_terse_v3(ts, rs);
799 else
800 log_err("fio: bad terse version!? %d\n", terse_version);
801}
802
803static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
804{
805 double mean, S;
806
807 if (src->samples == 0)
808 return;
809
810 dst->min_val = min(dst->min_val, src->min_val);
811 dst->max_val = max(dst->max_val, src->max_val);
812
813 /*
814 * Compute new mean and S after the merge
815 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
816 * #Parallel_algorithm>
817 */
818 if (nr == 1) {
819 mean = src->mean.u.f;
820 S = src->S.u.f;
821 } else {
822 double delta = src->mean.u.f - dst->mean.u.f;
823
824 mean = ((src->mean.u.f * src->samples) +
825 (dst->mean.u.f * dst->samples)) /
826 (dst->samples + src->samples);
827
828 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
829 (dst->samples * src->samples) /
830 (dst->samples + src->samples);
831 }
832
833 dst->samples += src->samples;
834 dst->mean.u.f = mean;
835 dst->S.u.f = S;
836}
837
838void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
839{
840 int i;
841
842 for (i = 0; i < 2; i++) {
843 if (dst->max_run[i] < src->max_run[i])
844 dst->max_run[i] = src->max_run[i];
845 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
846 dst->min_run[i] = src->min_run[i];
847 if (dst->max_bw[i] < src->max_bw[i])
848 dst->max_bw[i] = src->max_bw[i];
849 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
850 dst->min_bw[i] = src->min_bw[i];
851
852 dst->io_kb[i] += src->io_kb[i];
853 dst->agg[i] += src->agg[i];
854 }
855
856}
857
858void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
859{
860 int l, k;
861
862 for (l = 0; l <= DDIR_WRITE; l++) {
863 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
864 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
865 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
866 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
867
868 dst->io_bytes[l] += src->io_bytes[l];
869
870 if (dst->runtime[l] < src->runtime[l])
871 dst->runtime[l] = src->runtime[l];
872 }
873
874 dst->usr_time += src->usr_time;
875 dst->sys_time += src->sys_time;
876 dst->ctx += src->ctx;
877 dst->majf += src->majf;
878 dst->minf += src->minf;
879
880 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
881 dst->io_u_map[k] += src->io_u_map[k];
882 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
883 dst->io_u_submit[k] += src->io_u_submit[k];
884 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
885 dst->io_u_complete[k] += src->io_u_complete[k];
886 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
887 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
888 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
889 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
890
891 for (k = 0; k <= 2; k++) {
892 dst->total_io_u[k] += src->total_io_u[k];
893 dst->short_io_u[k] += src->short_io_u[k];
894 }
895
896 for (k = 0; k <= DDIR_WRITE; k++) {
897 int m;
898 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
899 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
900 }
901
902 dst->total_run_time += src->total_run_time;
903 dst->total_submit += src->total_submit;
904 dst->total_complete += src->total_complete;
905}
906
907void init_group_run_stat(struct group_run_stats *gs)
908{
909 memset(gs, 0, sizeof(*gs));
910 gs->min_bw[0] = gs->min_run[0] = ~0UL;
911 gs->min_bw[1] = gs->min_run[1] = ~0UL;
912}
913
914void init_thread_stat(struct thread_stat *ts)
915{
916 int j;
917
918 memset(ts, 0, sizeof(*ts));
919
920 for (j = 0; j <= DDIR_WRITE; j++) {
921 ts->lat_stat[j].min_val = -1UL;
922 ts->clat_stat[j].min_val = -1UL;
923 ts->slat_stat[j].min_val = -1UL;
924 ts->bw_stat[j].min_val = -1UL;
925 }
926 ts->groupid = -1;
927}
928
929void show_run_stats(void)
930{
931 struct group_run_stats *runstats, *rs;
932 struct thread_data *td;
933 struct thread_stat *threadstats, *ts;
934 int i, j, nr_ts, last_ts, idx;
935 int kb_base_warned = 0;
936
937 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
938
939 for (i = 0; i < groupid + 1; i++)
940 init_group_run_stat(&runstats[i]);
941
942 /*
943 * find out how many threads stats we need. if group reporting isn't
944 * enabled, it's one-per-td.
945 */
946 nr_ts = 0;
947 last_ts = -1;
948 for_each_td(td, i) {
949 if (!td->o.group_reporting) {
950 nr_ts++;
951 continue;
952 }
953 if (last_ts == td->groupid)
954 continue;
955
956 last_ts = td->groupid;
957 nr_ts++;
958 }
959
960 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
961
962 for (i = 0; i < nr_ts; i++)
963 init_thread_stat(&threadstats[i]);
964
965 j = 0;
966 last_ts = -1;
967 idx = 0;
968 for_each_td(td, i) {
969 if (idx && (!td->o.group_reporting ||
970 (td->o.group_reporting && last_ts != td->groupid))) {
971 idx = 0;
972 j++;
973 }
974
975 last_ts = td->groupid;
976
977 ts = &threadstats[j];
978
979 ts->clat_percentiles = td->o.clat_percentiles;
980 if (td->o.overwrite_plist)
981 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
982 else
983 memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list));
984
985 idx++;
986 ts->members++;
987
988 if (ts->groupid == -1) {
989 /*
990 * These are per-group shared already
991 */
992 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
993 if (td->o.description)
994 strncpy(ts->description, td->o.description,
995 FIO_JOBNAME_SIZE);
996 else
997 memset(ts->description, 0, FIO_JOBNAME_SIZE);
998
999 ts->groupid = td->groupid;
1000
1001 /*
1002 * first pid in group, not very useful...
1003 */
1004 ts->pid = td->pid;
1005
1006 ts->kb_base = td->o.kb_base;
1007 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1008 log_info("fio: kb_base differs for jobs in group, using"
1009 " %u as the base\n", ts->kb_base);
1010 kb_base_warned = 1;
1011 }
1012
1013 ts->continue_on_error = td->o.continue_on_error;
1014 ts->total_err_count += td->total_err_count;
1015 ts->first_error = td->first_error;
1016 if (!ts->error) {
1017 if (!td->error && td->o.continue_on_error &&
1018 td->first_error) {
1019 ts->error = td->first_error;
1020 strcpy(ts->verror, td->verror);
1021 } else if (td->error) {
1022 ts->error = td->error;
1023 strcpy(ts->verror, td->verror);
1024 }
1025 }
1026
1027 sum_thread_stats(ts, &td->ts, idx);
1028 }
1029
1030 for (i = 0; i < nr_ts; i++) {
1031 unsigned long long bw;
1032
1033 ts = &threadstats[i];
1034 rs = &runstats[ts->groupid];
1035 rs->kb_base = ts->kb_base;
1036
1037 for (j = 0; j <= DDIR_WRITE; j++) {
1038 if (!ts->runtime[j])
1039 continue;
1040 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1041 rs->min_run[j] = ts->runtime[j];
1042 if (ts->runtime[j] > rs->max_run[j])
1043 rs->max_run[j] = ts->runtime[j];
1044
1045 bw = 0;
1046 if (ts->runtime[j]) {
1047 unsigned long runt;
1048
1049 runt = ts->runtime[j];
1050 bw = ts->io_bytes[j] / runt;
1051 }
1052 if (bw < rs->min_bw[j])
1053 rs->min_bw[j] = bw;
1054 if (bw > rs->max_bw[j])
1055 rs->max_bw[j] = bw;
1056
1057 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1058 }
1059 }
1060
1061 for (i = 0; i < groupid + 1; i++) {
1062 unsigned long max_run[2];
1063
1064 rs = &runstats[i];
1065 max_run[0] = rs->max_run[0];
1066 max_run[1] = rs->max_run[1];
1067
1068 if (rs->max_run[0])
1069 rs->agg[0] = (rs->io_kb[0] * 1000) / max_run[0];
1070 if (rs->max_run[1])
1071 rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1];
1072 }
1073
1074 /*
1075 * don't overwrite last signal output
1076 */
1077 if (!terse_output)
1078 log_info("\n");
1079
1080 for (i = 0; i < nr_ts; i++) {
1081 ts = &threadstats[i];
1082 rs = &runstats[ts->groupid];
1083
1084 if (is_backend)
1085 fio_server_send_ts(ts, rs);
1086 else if (terse_output)
1087 show_thread_status_terse(ts, rs);
1088 else
1089 show_thread_status(ts, rs);
1090 }
1091
1092 for (i = 0; i < groupid + 1; i++) {
1093 rs = &runstats[i];
1094
1095 rs->groupid = i;
1096 if (is_backend)
1097 fio_server_send_gs(rs);
1098 else if (!terse_output)
1099 show_group_stats(rs);
1100 }
1101
1102 if (is_backend)
1103 fio_server_send_du();
1104 else if (!terse_output)
1105 show_disk_util(0);
1106
1107 free_disk_util();
1108
1109 free(runstats);
1110 free(threadstats);
1111}
1112
1113static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1114{
1115 double val = data;
1116 double delta;
1117
1118 if (data > is->max_val)
1119 is->max_val = data;
1120 if (data < is->min_val)
1121 is->min_val = data;
1122
1123 delta = val - is->mean.u.f;
1124 if (delta) {
1125 is->mean.u.f += delta / (is->samples + 1.0);
1126 is->S.u.f += delta * (val - is->mean.u.f);
1127 }
1128
1129 is->samples++;
1130}
1131
1132static void __add_log_sample(struct io_log *iolog, unsigned long val,
1133 enum fio_ddir ddir, unsigned int bs,
1134 unsigned long t)
1135{
1136 const int nr_samples = iolog->nr_samples;
1137
1138 if (!iolog->nr_samples)
1139 iolog->avg_last = t;
1140
1141 if (iolog->nr_samples == iolog->max_samples) {
1142 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
1143
1144 iolog->log = realloc(iolog->log, new_size);
1145 iolog->max_samples <<= 1;
1146 }
1147
1148 iolog->log[nr_samples].val = val;
1149 iolog->log[nr_samples].time = t;
1150 iolog->log[nr_samples].ddir = ddir;
1151 iolog->log[nr_samples].bs = bs;
1152 iolog->nr_samples++;
1153}
1154
1155static inline void reset_io_stat(struct io_stat *ios)
1156{
1157 ios->max_val = ios->min_val = ios->samples = 0;
1158 ios->mean.u.f = ios->S.u.f = 0;
1159}
1160
1161static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1162 unsigned long val, enum fio_ddir ddir,
1163 unsigned int bs)
1164{
1165 unsigned long elapsed, this_window;
1166
1167 if (!ddir_rw(ddir))
1168 return;
1169
1170 elapsed = mtime_since_now(&td->epoch);
1171
1172 /*
1173 * If no time averaging, just add the log sample.
1174 */
1175 if (!iolog->avg_msec) {
1176 __add_log_sample(iolog, val, ddir, bs, elapsed);
1177 return;
1178 }
1179
1180 /*
1181 * Add the sample. If the time period has passed, then
1182 * add that entry to the log and clear.
1183 */
1184 add_stat_sample(&iolog->avg_window[ddir], val);
1185
1186 /*
1187 * If period hasn't passed, adding the above sample is all we
1188 * need to do.
1189 */
1190 this_window = elapsed - iolog->avg_last;
1191 if (this_window < iolog->avg_msec)
1192 return;
1193
1194 /*
1195 * Note an entry in the log. Use the mean from the logged samples,
1196 * making sure to properly round up. Only write a log entry if we
1197 * had actual samples done.
1198 */
1199 if (iolog->avg_window[DDIR_READ].samples) {
1200 unsigned long mr;
1201
1202 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1203 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed);
1204 }
1205 if (iolog->avg_window[DDIR_WRITE].samples) {
1206 unsigned long mw;
1207
1208 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1209 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed);
1210 }
1211
1212 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1213 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1214 iolog->avg_last = elapsed;
1215}
1216
1217void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1218{
1219 struct io_log *iolog;
1220
1221 if (!ddir_rw(ddir))
1222 return;
1223
1224 iolog = agg_io_log[ddir];
1225 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
1226}
1227
1228static void add_clat_percentile_sample(struct thread_stat *ts,
1229 unsigned long usec, enum fio_ddir ddir)
1230{
1231 unsigned int idx = plat_val_to_idx(usec);
1232 assert(idx < FIO_IO_U_PLAT_NR);
1233
1234 ts->io_u_plat[ddir][idx]++;
1235}
1236
1237void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1238 unsigned long usec, unsigned int bs)
1239{
1240 struct thread_stat *ts = &td->ts;
1241
1242 if (!ddir_rw(ddir))
1243 return;
1244
1245 add_stat_sample(&ts->clat_stat[ddir], usec);
1246
1247 if (td->clat_log)
1248 add_log_sample(td, td->clat_log, usec, ddir, bs);
1249
1250 if (ts->clat_percentiles)
1251 add_clat_percentile_sample(ts, usec, ddir);
1252}
1253
1254void add_slat_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->slat_stat[ddir], usec);
1263
1264 if (td->slat_log)
1265 add_log_sample(td, td->slat_log, usec, ddir, bs);
1266}
1267
1268void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1269 unsigned long usec, unsigned int bs)
1270{
1271 struct thread_stat *ts = &td->ts;
1272
1273 if (!ddir_rw(ddir))
1274 return;
1275
1276 add_stat_sample(&ts->lat_stat[ddir], usec);
1277
1278 if (td->lat_log)
1279 add_log_sample(td, td->lat_log, usec, ddir, bs);
1280}
1281
1282void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1283 struct timeval *t)
1284{
1285 struct thread_stat *ts = &td->ts;
1286 unsigned long spent, rate;
1287
1288 if (!ddir_rw(ddir))
1289 return;
1290
1291 spent = mtime_since(&td->bw_sample_time, t);
1292 if (spent < td->o.bw_avg_time)
1293 return;
1294
1295 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) *
1296 1000 / spent / 1024;
1297 add_stat_sample(&ts->bw_stat[ddir], rate);
1298
1299 if (td->bw_log)
1300 add_log_sample(td, td->bw_log, rate, ddir, bs);
1301
1302 fio_gettime(&td->bw_sample_time, NULL);
1303 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1304}
1305
1306void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
1307 struct timeval *t)
1308{
1309 struct thread_stat *ts = &td->ts;
1310 unsigned long spent, iops;
1311
1312 if (!ddir_rw(ddir))
1313 return;
1314
1315 spent = mtime_since(&td->iops_sample_time, t);
1316 if (spent < td->o.iops_avg_time)
1317 return;
1318
1319 iops = ((td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]) * 1000) / spent;
1320
1321 add_stat_sample(&ts->iops_stat[ddir], iops);
1322
1323 if (td->iops_log)
1324 add_log_sample(td, td->iops_log, iops, ddir, 0);
1325
1326 fio_gettime(&td->iops_sample_time, NULL);
1327 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1328}
fio - Flexible IO Tester