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