projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Adapt clat percentiles for min/max values
[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 "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
117/*
118 * Find and display the p-th percentile of clat
119 */
120static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
121 fio_fp64_t *plist)
122{
123 unsigned long sum = 0;
124 unsigned int len, i, j = 0, minv = -1U, maxv = 0;
125 unsigned int *ovals = NULL, oval_len = 0;
126 int is_last = 0, scale_down;
127
128 len = 0;
129 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
130 len++;
131
132 if (!len)
133 return;
134
135 /*
136 * Sort the percentile list. Note that it may already be sorted if
137 * we are using the default values, but since it's a short list this
138 * isn't a worry. Also note that this does not work for NaN values.
139 */
140 if (len > 1)
141 qsort((void*)plist, len, sizeof(plist[0]), double_cmp);
142
143 /*
144 * Calculate bucket values, note down max and min values
145 */
146 for (i = 0; i < FIO_IO_U_PLAT_NR; i++) {
147 sum += io_u_plat[i];
148 while (sum >= (plist[j].u.f / 100.0 * nr)) {
149 assert(plist[j].u.f <= 100.0);
150
151 if (j == oval_len) {
152 oval_len += 100;
153 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
154 }
155
156 ovals[j] = plat_idx_to_val(i);
157 if (ovals[j] < minv)
158 minv = ovals[j];
159 if (ovals[j] > maxv)
160 maxv = ovals[j];
161 j++;
162 }
163 }
164
165 /*
166 * We default to usecs, but if the value range is such that we
167 * should scale down to msecs, do that.
168 */
169 if (minv > 2000 && maxv > 99999) {
170 scale_down = 1;
171 log_info(" clat percentiles (msec):\n |");
172 } else {
173 scale_down = 0;
174 log_info(" clat percentiles (usec):\n |");
175 }
176
177 for (j = 0; j < len; j++) {
178 char fbuf[8];
179
180 /* for formatting */
181 if (j != 0 && (j % 4) == 0)
182 log_info(" |");
183
184 /* end of the list */
185 is_last = (j == len - 1);
186
187 if (plist[j].u.f < 10.0)
188 sprintf(fbuf, " %2.2f", plist[j].u.f);
189 else
190 sprintf(fbuf, "%2.2f", plist[j].u.f);
191
192 if (scale_down)
193 ovals[j] = (ovals[j] + 999) / 1000;
194
195 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
196
197 if (is_last)
198 break;
199
200 if (j % 4 == 3) /* for formatting */
201 log_info("\n");
202 }
203
204 if (ovals)
205 free(ovals);
206}
207
208static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
209 double *mean, double *dev)
210{
211 double n = is->samples;
212
213 if (is->samples == 0)
214 return 0;
215
216 *min = is->min_val;
217 *max = is->max_val;
218
219 n = (double) is->samples;
220 *mean = is->mean.u.f;
221
222 if (n > 1.0)
223 *dev = sqrt(is->S.u.f / (n - 1.0));
224 else
225 *dev = 0;
226
227 return 1;
228}
229
230void show_group_stats(struct group_run_stats *rs)
231{
232 char *p1, *p2, *p3, *p4;
233 const char *ddir_str[] = { " READ", " WRITE" };
234 int i;
235
236 log_info("Run status group %d (all jobs):\n", rs->groupid);
237
238 for (i = 0; i <= DDIR_WRITE; i++) {
239 const int i2p = is_power_of_2(rs->kb_base);
240
241 if (!rs->max_run[i])
242 continue;
243
244 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p);
245 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p);
246 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p);
247 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p);
248
249 log_info("%s: io=%sB, aggrb=%sB/s, minb=%sB/s, maxb=%sB/s,"
250 " mint=%llumsec, maxt=%llumsec\n", ddir_str[i], p1, p2,
251 p3, p4, rs->min_run[i],
252 rs->max_run[i]);
253
254 free(p1);
255 free(p2);
256 free(p3);
257 free(p4);
258 }
259}
260
261#define ts_total_io_u(ts) \
262 ((ts)->total_io_u[0] + (ts)->total_io_u[1])
263
264static void stat_calc_dist(unsigned int *map, unsigned long total,
265 double *io_u_dist)
266{
267 int i;
268
269 /*
270 * Do depth distribution calculations
271 */
272 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
273 if (total) {
274 io_u_dist[i] = (double) map[i] / (double) total;
275 io_u_dist[i] *= 100.0;
276 if (io_u_dist[i] < 0.1 && map[i])
277 io_u_dist[i] = 0.1;
278 } else
279 io_u_dist[i] = 0.0;
280 }
281}
282
283static void stat_calc_lat(struct thread_stat *ts, double *dst,
284 unsigned int *src, int nr)
285{
286 unsigned long total = ts_total_io_u(ts);
287 int i;
288
289 /*
290 * Do latency distribution calculations
291 */
292 for (i = 0; i < nr; i++) {
293 if (total) {
294 dst[i] = (double) src[i] / (double) total;
295 dst[i] *= 100.0;
296 if (dst[i] < 0.01 && src[i])
297 dst[i] = 0.01;
298 } else
299 dst[i] = 0.0;
300 }
301}
302
303static void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
304{
305 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
306}
307
308static void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
309{
310 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
311}
312
313static int usec_to_msec(unsigned long *min, unsigned long *max, double *mean,
314 double *dev)
315{
316 if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) {
317 *min /= 1000;
318 *max /= 1000;
319 *mean /= 1000.0;
320 *dev /= 1000.0;
321 return 0;
322 }
323
324 return 1;
325}
326
327static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
328 int ddir)
329{
330 const char *ddir_str[] = { "read ", "write" };
331 unsigned long min, max, runt;
332 unsigned long long bw, iops;
333 double mean, dev;
334 char *io_p, *bw_p, *iops_p;
335 int i2p;
336
337 assert(ddir_rw(ddir));
338
339 if (!ts->runtime[ddir])
340 return;
341
342 i2p = is_power_of_2(rs->kb_base);
343 runt = ts->runtime[ddir];
344
345 bw = (1000 * ts->io_bytes[ddir]) / runt;
346 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p);
347 bw_p = num2str(bw, 6, 1, i2p);
348
349 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
350 iops_p = num2str(iops, 6, 1, 0);
351
352 log_info(" %s: io=%sB, bw=%sB/s, iops=%s, runt=%6llumsec\n",
353 ddir_str[ddir], io_p, bw_p, iops_p,
354 ts->runtime[ddir]);
355
356 free(io_p);
357 free(bw_p);
358 free(iops_p);
359
360 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
361 const char *base = "(usec)";
362 char *minp, *maxp;
363
364 if (!usec_to_msec(&min, &max, &mean, &dev))
365 base = "(msec)";
366
367 minp = num2str(min, 6, 1, 0);
368 maxp = num2str(max, 6, 1, 0);
369
370 log_info(" slat %s: min=%s, max=%s, avg=%5.02f,"
371 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
372
373 free(minp);
374 free(maxp);
375 }
376 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
377 const char *base = "(usec)";
378 char *minp, *maxp;
379
380 if (!usec_to_msec(&min, &max, &mean, &dev))
381 base = "(msec)";
382
383 minp = num2str(min, 6, 1, 0);
384 maxp = num2str(max, 6, 1, 0);
385
386 log_info(" clat %s: min=%s, max=%s, avg=%5.02f,"
387 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
388
389 free(minp);
390 free(maxp);
391 }
392 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
393 const char *base = "(usec)";
394 char *minp, *maxp;
395
396 if (!usec_to_msec(&min, &max, &mean, &dev))
397 base = "(msec)";
398
399 minp = num2str(min, 6, 1, 0);
400 maxp = num2str(max, 6, 1, 0);
401
402 log_info(" lat %s: min=%s, max=%s, avg=%5.02f,"
403 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
404
405 free(minp);
406 free(maxp);
407 }
408 if (ts->clat_percentiles) {
409 show_clat_percentiles(ts->io_u_plat[ddir],
410 ts->clat_stat[ddir].samples,
411 ts->percentile_list);
412 }
413 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
414 double p_of_agg;
415
416 p_of_agg = mean * 100 / (double) rs->agg[ddir];
417 log_info(" bw (KB/s) : min=%5lu, max=%5lu, per=%3.2f%%,"
418 " avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg,
419 mean, dev);
420 }
421}
422
423static void show_lat(double *io_u_lat, int nr, const char **ranges,
424 const char *msg)
425{
426 int new_line = 1, i, line = 0;
427
428 for (i = 0; i < nr; i++) {
429 if (io_u_lat[i] <= 0.0)
430 continue;
431 if (new_line) {
432 if (line)
433 log_info("\n");
434 log_info(" lat (%s): ", msg);
435 new_line = 0;
436 line = 0;
437 }
438 if (line)
439 log_info(", ");
440 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
441 line++;
442 if (line == 5)
443 new_line = 1;
444 }
445}
446
447static void show_lat_u(double *io_u_lat_u)
448{
449 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
450 "250=", "500=", "750=", "1000=", };
451
452 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
453}
454
455static void show_lat_m(double *io_u_lat_m)
456{
457 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
458 "250=", "500=", "750=", "1000=", "2000=",
459 ">=2000=", };
460
461 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
462}
463
464static void show_latencies(double *io_u_lat_u, double *io_u_lat_m)
465{
466 show_lat_u(io_u_lat_u);
467 log_info("\n");
468 show_lat_m(io_u_lat_m);
469 log_info("\n");
470}
471
472void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
473{
474 double usr_cpu, sys_cpu;
475 unsigned long runtime;
476 double io_u_dist[FIO_IO_U_MAP_NR];
477 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
478 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
479
480 if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
481 !(ts->total_io_u[0] + ts->total_io_u[1]))
482 return;
483
484 if (!ts->error) {
485 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
486 ts->name, ts->groupid, ts->members,
487 ts->error, (int) ts->pid);
488 } else {
489 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
490 ts->name, ts->groupid, ts->members,
491 ts->error, ts->verror, (int) ts->pid);
492 }
493
494 if (ts->description)
495 log_info(" Description : [%s]\n", ts->description);
496
497 if (ts->io_bytes[DDIR_READ])
498 show_ddir_status(rs, ts, DDIR_READ);
499 if (ts->io_bytes[DDIR_WRITE])
500 show_ddir_status(rs, ts, DDIR_WRITE);
501
502 runtime = ts->total_run_time;
503 if (runtime) {
504 double runt = (double) runtime;
505
506 usr_cpu = (double) ts->usr_time * 100 / runt;
507 sys_cpu = (double) ts->sys_time * 100 / runt;
508 } else {
509 usr_cpu = 0;
510 sys_cpu = 0;
511 }
512
513 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
514 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
515
516 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
517 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
518 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
519 io_u_dist[1], io_u_dist[2],
520 io_u_dist[3], io_u_dist[4],
521 io_u_dist[5], io_u_dist[6]);
522
523 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
524 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
525 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
526 io_u_dist[1], io_u_dist[2],
527 io_u_dist[3], io_u_dist[4],
528 io_u_dist[5], io_u_dist[6]);
529 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
530 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
531 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
532 io_u_dist[1], io_u_dist[2],
533 io_u_dist[3], io_u_dist[4],
534 io_u_dist[5], io_u_dist[6]);
535 log_info(" issued r/w/d: total=%lu/%lu/%lu, short=%lu/%lu/%lu\n",
536 ts->total_io_u[0], ts->total_io_u[1],
537 ts->total_io_u[2],
538 ts->short_io_u[0], ts->short_io_u[1],
539 ts->short_io_u[2]);
540 stat_calc_lat_u(ts, io_u_lat_u);
541 stat_calc_lat_m(ts, io_u_lat_m);
542 show_latencies(io_u_lat_u, io_u_lat_m);
543 if (ts->continue_on_error) {
544 log_info(" errors : total=%lu, first_error=%d/<%s>\n",
545 ts->total_err_count,
546 ts->first_error,
547 strerror(ts->first_error));
548 }
549}
550
551static void show_ddir_status_terse(struct thread_stat *ts,
552 struct group_run_stats *rs, int ddir)
553{
554 unsigned long min, max;
555 unsigned long long bw;
556 double mean, dev;
557
558 assert(ddir_rw(ddir));
559
560 bw = 0;
561 if (ts->runtime[ddir])
562 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
563
564 log_info(";%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw,
565 ts->runtime[ddir]);
566
567 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
568 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
569 else
570 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
571
572 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
573 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
574 else
575 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
576
577 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
578 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
579 else
580 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
581
582 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
583 double p_of_agg;
584
585 p_of_agg = mean * 100 / (double) rs->agg[ddir];
586 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
587 } else
588 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
589}
590
591#define FIO_TERSE_VERSION "2"
592
593static void show_thread_status_terse(struct thread_stat *ts,
594 struct group_run_stats *rs)
595{
596 double io_u_dist[FIO_IO_U_MAP_NR];
597 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
598 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
599 double usr_cpu, sys_cpu;
600 int i;
601
602 /* General Info */
603 log_info("%s;%s;%d;%d", FIO_TERSE_VERSION, ts->name, ts->groupid,
604 ts->error);
605 /* Log Read Status */
606 show_ddir_status_terse(ts, rs, 0);
607 /* Log Write Status */
608 show_ddir_status_terse(ts, rs, 1);
609
610 /* CPU Usage */
611 if (ts->total_run_time) {
612 double runt = (double) ts->total_run_time;
613
614 usr_cpu = (double) ts->usr_time * 100 / runt;
615 sys_cpu = (double) ts->sys_time * 100 / runt;
616 } else {
617 usr_cpu = 0;
618 sys_cpu = 0;
619 }
620
621 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
622 ts->minf);
623
624 /* Calc % distribution of IO depths, usecond, msecond latency */
625 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
626 stat_calc_lat_u(ts, io_u_lat_u);
627 stat_calc_lat_m(ts, io_u_lat_m);
628
629 /* Only show fixed 7 I/O depth levels*/
630 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
631 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
632 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
633
634 /* Microsecond latency */
635 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
636 log_info(";%3.2f%%", io_u_lat_u[i]);
637 /* Millisecond latency */
638 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
639 log_info(";%3.2f%%", io_u_lat_m[i]);
640 /* Additional output if continue_on_error set - default off*/
641 if (ts->continue_on_error)
642 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
643 log_info("\n");
644
645 /* Additional output if description is set */
646 if (ts->description)
647 log_info(";%s", ts->description);
648
649 log_info("\n");
650}
651
652static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
653{
654 double mean, S;
655
656 if (src->samples == 0)
657 return;
658
659 dst->min_val = min(dst->min_val, src->min_val);
660 dst->max_val = max(dst->max_val, src->max_val);
661
662 /*
663 * Compute new mean and S after the merge
664 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
665 * #Parallel_algorithm>
666 */
667 if (nr == 1) {
668 mean = src->mean.u.f;
669 S = src->S.u.f;
670 } else {
671 double delta = src->mean.u.f - dst->mean.u.f;
672
673 mean = ((src->mean.u.f * src->samples) +
674 (dst->mean.u.f * dst->samples)) /
675 (dst->samples + src->samples);
676
677 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
678 (dst->samples * src->samples) /
679 (dst->samples + src->samples);
680 }
681
682 dst->samples += src->samples;
683 dst->mean.u.f = mean;
684 dst->S.u.f = S;
685}
686
687void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
688{
689 int i;
690
691 for (i = 0; i < 2; i++) {
692 if (dst->max_run[i] < src->max_run[i])
693 dst->max_run[i] = src->max_run[i];
694 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
695 dst->min_run[i] = src->min_run[i];
696 if (dst->max_bw[i] < src->max_bw[i])
697 dst->max_bw[i] = src->max_bw[i];
698 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
699 dst->min_bw[i] = src->min_bw[i];
700
701 dst->io_kb[i] += src->io_kb[i];
702 dst->agg[i] += src->agg[i];
703 }
704
705}
706
707void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
708{
709 int l, k;
710
711 for (l = 0; l <= DDIR_WRITE; l++) {
712 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
713 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
714 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
715 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
716
717 dst->io_bytes[l] += src->io_bytes[l];
718
719 if (dst->runtime[l] < src->runtime[l])
720 dst->runtime[l] = src->runtime[l];
721 }
722
723 dst->usr_time += src->usr_time;
724 dst->sys_time += src->sys_time;
725 dst->ctx += src->ctx;
726 dst->majf += src->majf;
727 dst->minf += src->minf;
728
729 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
730 dst->io_u_map[k] += src->io_u_map[k];
731 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
732 dst->io_u_submit[k] += src->io_u_submit[k];
733 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
734 dst->io_u_complete[k] += src->io_u_complete[k];
735 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
736 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
737 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
738 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
739
740 for (k = 0; k <= 2; k++) {
741 dst->total_io_u[k] += src->total_io_u[k];
742 dst->short_io_u[k] += src->short_io_u[k];
743 }
744
745 for (k = 0; k <= DDIR_WRITE; k++) {
746 int m;
747 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
748 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
749 }
750
751 dst->total_run_time += src->total_run_time;
752 dst->total_submit += src->total_submit;
753 dst->total_complete += src->total_complete;
754}
755
756void init_group_run_stat(struct group_run_stats *gs)
757{
758 memset(gs, 0, sizeof(*gs));
759 gs->min_bw[0] = gs->min_run[0] = ~0UL;
760 gs->min_bw[1] = gs->min_run[1] = ~0UL;
761}
762
763void init_thread_stat(struct thread_stat *ts)
764{
765 int j;
766
767 memset(ts, 0, sizeof(*ts));
768
769 for (j = 0; j <= DDIR_WRITE; j++) {
770 ts->lat_stat[j].min_val = -1UL;
771 ts->clat_stat[j].min_val = -1UL;
772 ts->slat_stat[j].min_val = -1UL;
773 ts->bw_stat[j].min_val = -1UL;
774 }
775 ts->groupid = -1;
776}
777
778void show_run_stats(void)
779{
780 struct group_run_stats *runstats, *rs;
781 struct thread_data *td;
782 struct thread_stat *threadstats, *ts;
783 int i, j, nr_ts, last_ts, idx;
784 int kb_base_warned = 0;
785
786 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
787
788 for (i = 0; i < groupid + 1; i++)
789 init_group_run_stat(&runstats[i]);
790
791 /*
792 * find out how many threads stats we need. if group reporting isn't
793 * enabled, it's one-per-td.
794 */
795 nr_ts = 0;
796 last_ts = -1;
797 for_each_td(td, i) {
798 if (!td->o.group_reporting) {
799 nr_ts++;
800 continue;
801 }
802 if (last_ts == td->groupid)
803 continue;
804
805 last_ts = td->groupid;
806 nr_ts++;
807 }
808
809 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
810
811 for (i = 0; i < nr_ts; i++)
812 init_thread_stat(&threadstats[i]);
813
814 j = 0;
815 last_ts = -1;
816 idx = 0;
817 for_each_td(td, i) {
818 if (idx && (!td->o.group_reporting ||
819 (td->o.group_reporting && last_ts != td->groupid))) {
820 idx = 0;
821 j++;
822 }
823
824 last_ts = td->groupid;
825
826 ts = &threadstats[j];
827
828 ts->clat_percentiles = td->o.clat_percentiles;
829 if (td->o.overwrite_plist)
830 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
831 else
832 memcpy(ts->percentile_list, def_percentile_list, sizeof(def_percentile_list));
833
834 idx++;
835 ts->members++;
836
837 if (ts->groupid == -1) {
838 /*
839 * These are per-group shared already
840 */
841 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
842 if (td->o.description)
843 strncpy(ts->description, td->o.description,
844 FIO_JOBNAME_SIZE);
845 else
846 memset(ts->description, 0, FIO_JOBNAME_SIZE);
847
848 ts->groupid = td->groupid;
849
850 /*
851 * first pid in group, not very useful...
852 */
853 ts->pid = td->pid;
854
855 ts->kb_base = td->o.kb_base;
856 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
857 log_info("fio: kb_base differs for jobs in group, using"
858 " %u as the base\n", ts->kb_base);
859 kb_base_warned = 1;
860 }
861
862 ts->continue_on_error = td->o.continue_on_error;
863 ts->total_err_count += td->total_err_count;
864 ts->first_error = td->first_error;
865 if (!ts->error) {
866 if (!td->error && td->o.continue_on_error &&
867 td->first_error) {
868 ts->error = td->first_error;
869 strcpy(ts->verror, td->verror);
870 } else if (td->error) {
871 ts->error = td->error;
872 strcpy(ts->verror, td->verror);
873 }
874 }
875
876 sum_thread_stats(ts, &td->ts, idx);
877 }
878
879 for (i = 0; i < nr_ts; i++) {
880 unsigned long long bw;
881
882 ts = &threadstats[i];
883 rs = &runstats[ts->groupid];
884 rs->kb_base = ts->kb_base;
885
886 for (j = 0; j <= DDIR_WRITE; j++) {
887 if (!ts->runtime[j])
888 continue;
889 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
890 rs->min_run[j] = ts->runtime[j];
891 if (ts->runtime[j] > rs->max_run[j])
892 rs->max_run[j] = ts->runtime[j];
893
894 bw = 0;
895 if (ts->runtime[j]) {
896 unsigned long runt;
897
898 runt = ts->runtime[j];
899 bw = ts->io_bytes[j] / runt;
900 }
901 if (bw < rs->min_bw[j])
902 rs->min_bw[j] = bw;
903 if (bw > rs->max_bw[j])
904 rs->max_bw[j] = bw;
905
906 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
907 }
908 }
909
910 for (i = 0; i < groupid + 1; i++) {
911 unsigned long max_run[2];
912
913 rs = &runstats[i];
914 max_run[0] = rs->max_run[0];
915 max_run[1] = rs->max_run[1];
916
917 if (rs->max_run[0])
918 rs->agg[0] = (rs->io_kb[0] * 1000) / max_run[0];
919 if (rs->max_run[1])
920 rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1];
921 }
922
923 /*
924 * don't overwrite last signal output
925 */
926 if (!terse_output)
927 log_info("\n");
928
929 for (i = 0; i < nr_ts; i++) {
930 ts = &threadstats[i];
931 rs = &runstats[ts->groupid];
932
933 if (is_backend)
934 fio_server_send_ts(ts, rs);
935 else if (terse_output)
936 show_thread_status_terse(ts, rs);
937 else
938 show_thread_status(ts, rs);
939 }
940
941 if (!terse_output) {
942 for (i = 0; i < groupid + 1; i++) {
943 rs = &runstats[i];
944
945 rs->groupid = i;
946 if (is_backend)
947 fio_server_send_gs(rs);
948 else
949 show_group_stats(rs);
950 }
951
952 show_disk_util();
953 }
954
955 free(runstats);
956 free(threadstats);
957}
958
959static inline void add_stat_sample(struct io_stat *is, unsigned long data)
960{
961 double val = data;
962 double delta;
963
964 if (data > is->max_val)
965 is->max_val = data;
966 if (data < is->min_val)
967 is->min_val = data;
968
969 delta = val - is->mean.u.f;
970 if (delta) {
971 is->mean.u.f += delta / (is->samples + 1.0);
972 is->S.u.f += delta * (val - is->mean.u.f);
973 }
974
975 is->samples++;
976}
977
978static void __add_log_sample(struct io_log *iolog, unsigned long val,
979 enum fio_ddir ddir, unsigned int bs,
980 unsigned long t)
981{
982 const int nr_samples = iolog->nr_samples;
983
984 if (iolog->nr_samples == iolog->max_samples) {
985 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
986
987 iolog->log = realloc(iolog->log, new_size);
988 iolog->max_samples <<= 1;
989 }
990
991 iolog->log[nr_samples].val = val;
992 iolog->log[nr_samples].time = t;
993 iolog->log[nr_samples].ddir = ddir;
994 iolog->log[nr_samples].bs = bs;
995 iolog->nr_samples++;
996}
997
998static void add_log_sample(struct thread_data *td, struct io_log *iolog,
999 unsigned long val, enum fio_ddir ddir,
1000 unsigned int bs)
1001{
1002 if (!ddir_rw(ddir))
1003 return;
1004
1005 __add_log_sample(iolog, val, ddir, bs, mtime_since_now(&td->epoch));
1006}
1007
1008void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1009{
1010 struct io_log *iolog;
1011
1012 if (!ddir_rw(ddir))
1013 return;
1014
1015 iolog = agg_io_log[ddir];
1016 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
1017}
1018
1019static void add_clat_percentile_sample(struct thread_stat *ts,
1020 unsigned long usec, enum fio_ddir ddir)
1021{
1022 unsigned int idx = plat_val_to_idx(usec);
1023 assert(idx < FIO_IO_U_PLAT_NR);
1024
1025 ts->io_u_plat[ddir][idx]++;
1026}
1027
1028void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1029 unsigned long usec, unsigned int bs)
1030{
1031 struct thread_stat *ts = &td->ts;
1032
1033 if (!ddir_rw(ddir))
1034 return;
1035
1036 add_stat_sample(&ts->clat_stat[ddir], usec);
1037
1038 if (td->clat_log)
1039 add_log_sample(td, td->clat_log, usec, ddir, bs);
1040
1041 if (ts->clat_percentiles)
1042 add_clat_percentile_sample(ts, usec, ddir);
1043}
1044
1045void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1046 unsigned long usec, unsigned int bs)
1047{
1048 struct thread_stat *ts = &td->ts;
1049
1050 if (!ddir_rw(ddir))
1051 return;
1052
1053 add_stat_sample(&ts->slat_stat[ddir], usec);
1054
1055 if (td->slat_log)
1056 add_log_sample(td, td->slat_log, usec, ddir, bs);
1057}
1058
1059void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1060 unsigned long usec, unsigned int bs)
1061{
1062 struct thread_stat *ts = &td->ts;
1063
1064 if (!ddir_rw(ddir))
1065 return;
1066
1067 add_stat_sample(&ts->lat_stat[ddir], usec);
1068
1069 if (td->lat_log)
1070 add_log_sample(td, td->lat_log, usec, ddir, bs);
1071}
1072
1073void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1074 struct timeval *t)
1075{
1076 struct thread_stat *ts = &td->ts;
1077 unsigned long spent, rate;
1078
1079 if (!ddir_rw(ddir))
1080 return;
1081
1082 spent = mtime_since(&td->bw_sample_time, t);
1083 if (spent < td->o.bw_avg_time)
1084 return;
1085
1086 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) *
1087 1000 / spent / 1024;
1088 add_stat_sample(&ts->bw_stat[ddir], rate);
1089
1090 if (td->bw_log)
1091 add_log_sample(td, td->bw_log, rate, ddir, bs);
1092
1093 fio_gettime(&td->bw_sample_time, NULL);
1094 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1095}
1096
1097void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
1098 struct timeval *t)
1099{
1100 struct thread_stat *ts = &td->ts;
1101 unsigned long spent, iops;
1102
1103 if (!ddir_rw(ddir))
1104 return;
1105
1106 spent = mtime_since(&td->iops_sample_time, t);
1107 if (spent < td->o.iops_avg_time)
1108 return;
1109
1110 iops = ((td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]) * 1000) / spent;
1111
1112 add_stat_sample(&ts->iops_stat[ddir], iops);
1113
1114 if (td->iops_log) {
1115 assert(iops);
1116 add_log_sample(td, td->iops_log, iops, ddir, 0);
1117 }
1118
1119 fio_gettime(&td->iops_sample_time, NULL);
1120 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1121}
fio - Flexible IO Tester