projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Add IEEE 754 pack and unpack helpers
[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
13void update_rusage_stat(struct thread_data *td)
14{
15 struct thread_stat *ts = &td->ts;
16
17 getrusage(RUSAGE_SELF, &td->ru_end);
18
19 ts->usr_time += mtime_since(&td->ru_start.ru_utime,
20 &td->ru_end.ru_utime);
21 ts->sys_time += mtime_since(&td->ru_start.ru_stime,
22 &td->ru_end.ru_stime);
23 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
24 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
25 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
26 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
27
28 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
29}
30
31/*
32 * Given a latency, return the index of the corresponding bucket in
33 * the structure tracking percentiles.
34 *
35 * (1) find the group (and error bits) that the value (latency)
36 * belongs to by looking at its MSB. (2) find the bucket number in the
37 * group by looking at the index bits.
38 *
39 */
40static unsigned int plat_val_to_idx(unsigned int val)
41{
42 unsigned int msb, error_bits, base, offset, idx;
43
44 /* Find MSB starting from bit 0 */
45 if (val == 0)
46 msb = 0;
47 else
48 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
49
50 /*
51 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
52 * all bits of the sample as index
53 */
54 if (msb <= FIO_IO_U_PLAT_BITS)
55 return val;
56
57 /* Compute the number of error bits to discard*/
58 error_bits = msb - FIO_IO_U_PLAT_BITS;
59
60 /* Compute the number of buckets before the group */
61 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
62
63 /*
64 * Discard the error bits and apply the mask to find the
65 * index for the buckets in the group
66 */
67 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
68
69 /* Make sure the index does not exceed (array size - 1) */
70 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1)?
71 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
72
73 return idx;
74}
75
76/*
77 * Convert the given index of the bucket array to the value
78 * represented by the bucket
79 */
80static unsigned int plat_idx_to_val(unsigned int idx)
81{
82 unsigned int error_bits, k, base;
83
84 assert(idx < FIO_IO_U_PLAT_NR);
85
86 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
87 * all bits of the sample as index */
88 if (idx < (FIO_IO_U_PLAT_VAL << 1) )
89 return idx;
90
91 /* Find the group and compute the minimum value of that group */
92 error_bits = (idx >> FIO_IO_U_PLAT_BITS) -1;
93 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
94
95 /* Find its bucket number of the group */
96 k = idx % FIO_IO_U_PLAT_VAL;
97
98 /* Return the mean of the range of the bucket */
99 return base + ((k + 0.5) * (1 << error_bits));
100}
101
102static int double_cmp(const void *a, const void *b)
103{
104 const double fa = *(const double *)a;
105 const double fb = *(const double *)b;
106 int cmp = 0;
107
108 if (fa > fb)
109 cmp = 1;
110 else if (fa < fb)
111 cmp = -1;
112
113 return cmp;
114}
115
116/*
117 * Find and display the p-th percentile of clat
118 */
119static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
120 double *user_list)
121{
122 unsigned long sum = 0;
123 unsigned int len, i, j = 0;
124 const double *plist;
125 int is_last = 0;
126 static const double def_list[FIO_IO_U_LIST_MAX_LEN] = {
127 1.0, 5.0, 10.0, 20.0, 30.0,
128 40.0, 50.0, 60.0, 70.0, 80.0,
129 90.0, 95.0, 99.0, 99.5, 99.9};
130
131 plist = user_list;
132 if (!plist)
133 plist = def_list;
134
135 for (len = 0; len <FIO_IO_U_LIST_MAX_LEN && plist[len] != 0; len++)
136 ;
137
138 /*
139 * Sort the user-specified list. Note that this does not work
140 * for NaN values
141 */
142 if (user_list && len > 1)
143 qsort((void*)user_list, len, sizeof(user_list[0]), double_cmp);
144
145 log_info(" clat percentiles (usec) :");
146
147 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
148 sum += io_u_plat[i];
149 while (sum >= (plist[j] / 100 * nr)) {
150 assert(plist[j] <= 100.0);
151
152 /* for formatting */
153 if (j != 0 && (j % 4) == 0)
154 log_info(" ");
155
156 /* end of the list */
157 is_last = (j == len - 1);
158
159 log_info(" %2.2fth=%u%c", plist[j], plat_idx_to_val(i),
160 (is_last? '\n' : ','));
161
162 if (is_last)
163 break;
164
165 if (j % 4 == 3) /* for formatting */
166 log_info("\n");
167 j++;
168 }
169 }
170}
171
172static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
173 double *mean, double *dev)
174{
175 double n = is->samples;
176
177 if (is->samples == 0)
178 return 0;
179
180 *min = is->min_val;
181 *max = is->max_val;
182
183 n = (double) is->samples;
184 *mean = is->mean;
185
186 if (n > 1.0)
187 *dev = sqrt(is->S / (n - 1.0));
188 else
189 *dev = 0;
190
191 return 1;
192}
193
194void show_group_stats(struct group_run_stats *rs)
195{
196 char *p1, *p2, *p3, *p4;
197 const char *ddir_str[] = { " READ", " WRITE" };
198 int i;
199
200 log_info("\nRun status group %d (all jobs):\n", rs->groupid);
201
202 for (i = 0; i <= DDIR_WRITE; i++) {
203 const int i2p = is_power_of_2(rs->kb_base);
204
205 if (!rs->max_run[i])
206 continue;
207
208 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p);
209 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p);
210 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p);
211 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p);
212
213 log_info("%s: io=%sB, aggrb=%sB/s, minb=%sB/s, maxb=%sB/s,"
214 " mint=%llumsec, maxt=%llumsec\n", ddir_str[i], p1, p2,
215 p3, p4, rs->min_run[i],
216 rs->max_run[i]);
217
218 free(p1);
219 free(p2);
220 free(p3);
221 free(p4);
222 }
223}
224
225#define ts_total_io_u(ts) \
226 ((ts)->total_io_u[0] + (ts)->total_io_u[1])
227
228static void stat_calc_dist(unsigned int *map, unsigned long total,
229 double *io_u_dist)
230{
231 int i;
232
233 /*
234 * Do depth distribution calculations
235 */
236 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
237 if (total) {
238 io_u_dist[i] = (double) map[i] / (double) total;
239 io_u_dist[i] *= 100.0;
240 if (io_u_dist[i] < 0.1 && map[i])
241 io_u_dist[i] = 0.1;
242 } else
243 io_u_dist[i] = 0.0;
244 }
245}
246
247static void stat_calc_lat(struct thread_stat *ts, double *dst,
248 unsigned int *src, int nr)
249{
250 unsigned long total = ts_total_io_u(ts);
251 int i;
252
253 /*
254 * Do latency distribution calculations
255 */
256 for (i = 0; i < nr; i++) {
257 if (total) {
258 dst[i] = (double) src[i] / (double) total;
259 dst[i] *= 100.0;
260 if (dst[i] < 0.01 && src[i])
261 dst[i] = 0.01;
262 } else
263 dst[i] = 0.0;
264 }
265}
266
267static void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
268{
269 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
270}
271
272static void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
273{
274 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
275}
276
277static int usec_to_msec(unsigned long *min, unsigned long *max, double *mean,
278 double *dev)
279{
280 if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) {
281 *min /= 1000;
282 *max /= 1000;
283 *mean /= 1000.0;
284 *dev /= 1000.0;
285 return 0;
286 }
287
288 return 1;
289}
290
291static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
292 int ddir)
293{
294 const char *ddir_str[] = { "read ", "write" };
295 unsigned long min, max, runt;
296 unsigned long long bw, iops;
297 double mean, dev;
298 char *io_p, *bw_p, *iops_p;
299 int i2p;
300
301 assert(ddir_rw(ddir));
302
303 if (!ts->runtime[ddir])
304 return;
305
306 i2p = is_power_of_2(rs->kb_base);
307 runt = ts->runtime[ddir];
308
309 bw = (1000 * ts->io_bytes[ddir]) / runt;
310 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p);
311 bw_p = num2str(bw, 6, 1, i2p);
312
313 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
314 iops_p = num2str(iops, 6, 1, 0);
315
316 log_info(" %s: io=%sB, bw=%sB/s, iops=%s, runt=%6llumsec\n",
317 ddir_str[ddir], io_p, bw_p, iops_p,
318 ts->runtime[ddir]);
319
320 free(io_p);
321 free(bw_p);
322 free(iops_p);
323
324 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
325 const char *base = "(usec)";
326 char *minp, *maxp;
327
328 if (!usec_to_msec(&min, &max, &mean, &dev))
329 base = "(msec)";
330
331 minp = num2str(min, 6, 1, 0);
332 maxp = num2str(max, 6, 1, 0);
333
334 log_info(" slat %s: min=%s, max=%s, avg=%5.02f,"
335 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
336
337 free(minp);
338 free(maxp);
339 }
340 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
341 const char *base = "(usec)";
342 char *minp, *maxp;
343
344 if (!usec_to_msec(&min, &max, &mean, &dev))
345 base = "(msec)";
346
347 minp = num2str(min, 6, 1, 0);
348 maxp = num2str(max, 6, 1, 0);
349
350 log_info(" clat %s: min=%s, max=%s, avg=%5.02f,"
351 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
352
353 free(minp);
354 free(maxp);
355 }
356 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
357 const char *base = "(usec)";
358 char *minp, *maxp;
359
360 if (!usec_to_msec(&min, &max, &mean, &dev))
361 base = "(msec)";
362
363 minp = num2str(min, 6, 1, 0);
364 maxp = num2str(max, 6, 1, 0);
365
366 log_info(" lat %s: min=%s, max=%s, avg=%5.02f,"
367 " stdev=%5.02f\n", base, minp, maxp, mean, dev);
368
369 free(minp);
370 free(maxp);
371 }
372 if (ts->clat_percentiles) {
373 show_clat_percentiles(ts->io_u_plat[ddir],
374 ts->clat_stat[ddir].samples,
375 ts->percentile_list);
376 }
377 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
378 double p_of_agg;
379
380 p_of_agg = mean * 100 / (double) rs->agg[ddir];
381 log_info(" bw (KB/s) : min=%5lu, max=%5lu, per=%3.2f%%,"
382 " avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg,
383 mean, dev);
384 }
385}
386
387static void show_lat(double *io_u_lat, int nr, const char **ranges,
388 const char *msg)
389{
390 int new_line = 1, i, line = 0;
391
392 for (i = 0; i < nr; i++) {
393 if (io_u_lat[i] <= 0.0)
394 continue;
395 if (new_line) {
396 if (line)
397 log_info("\n");
398 log_info(" lat (%s): ", msg);
399 new_line = 0;
400 line = 0;
401 }
402 if (line)
403 log_info(", ");
404 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
405 line++;
406 if (line == 5)
407 new_line = 1;
408 }
409}
410
411static void show_lat_u(double *io_u_lat_u)
412{
413 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
414 "250=", "500=", "750=", "1000=", };
415
416 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
417}
418
419static void show_lat_m(double *io_u_lat_m)
420{
421 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
422 "250=", "500=", "750=", "1000=", "2000=",
423 ">=2000=", };
424
425 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
426}
427
428static void show_latencies(double *io_u_lat_u, double *io_u_lat_m)
429{
430 show_lat_u(io_u_lat_u);
431 log_info("\n");
432 show_lat_m(io_u_lat_m);
433 log_info("\n");
434}
435
436void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
437{
438 double usr_cpu, sys_cpu;
439 unsigned long runtime;
440 double io_u_dist[FIO_IO_U_MAP_NR];
441 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
442 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
443
444 if (!(ts->io_bytes[0] + ts->io_bytes[1]) &&
445 !(ts->total_io_u[0] + ts->total_io_u[1]))
446 return;
447
448 if (!ts->error) {
449 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n",
450 ts->name, ts->groupid, ts->members,
451 ts->error, (int) ts->pid);
452 } else {
453 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n",
454 ts->name, ts->groupid, ts->members,
455 ts->error, ts->verror, (int) ts->pid);
456 }
457
458 if (ts->description)
459 log_info(" Description : [%s]\n", ts->description);
460
461 if (ts->io_bytes[DDIR_READ])
462 show_ddir_status(rs, ts, DDIR_READ);
463 if (ts->io_bytes[DDIR_WRITE])
464 show_ddir_status(rs, ts, DDIR_WRITE);
465
466 runtime = ts->total_run_time;
467 if (runtime) {
468 double runt = (double) runtime;
469
470 usr_cpu = (double) ts->usr_time * 100 / runt;
471 sys_cpu = (double) ts->sys_time * 100 / runt;
472 } else {
473 usr_cpu = 0;
474 sys_cpu = 0;
475 }
476
477 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu,"
478 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf);
479
480 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
481 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
482 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
483 io_u_dist[1], io_u_dist[2],
484 io_u_dist[3], io_u_dist[4],
485 io_u_dist[5], io_u_dist[6]);
486
487 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
488 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
489 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
490 io_u_dist[1], io_u_dist[2],
491 io_u_dist[3], io_u_dist[4],
492 io_u_dist[5], io_u_dist[6]);
493 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
494 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
495 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
496 io_u_dist[1], io_u_dist[2],
497 io_u_dist[3], io_u_dist[4],
498 io_u_dist[5], io_u_dist[6]);
499 log_info(" issued r/w/d: total=%lu/%lu/%lu, short=%lu/%lu/%lu\n",
500 ts->total_io_u[0], ts->total_io_u[1],
501 ts->total_io_u[2],
502 ts->short_io_u[0], ts->short_io_u[1],
503 ts->short_io_u[2]);
504 stat_calc_lat_u(ts, io_u_lat_u);
505 stat_calc_lat_m(ts, io_u_lat_m);
506 show_latencies(io_u_lat_u, io_u_lat_m);
507 if (ts->continue_on_error) {
508 log_info(" errors : total=%lu, first_error=%d/<%s>\n",
509 ts->total_err_count,
510 ts->first_error,
511 strerror(ts->first_error));
512 }
513}
514
515static void show_ddir_status_terse(struct thread_stat *ts,
516 struct group_run_stats *rs, int ddir)
517{
518 unsigned long min, max;
519 unsigned long long bw;
520 double mean, dev;
521
522 assert(ddir_rw(ddir));
523
524 bw = 0;
525 if (ts->runtime[ddir])
526 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
527
528 log_info(";%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw,
529 ts->runtime[ddir]);
530
531 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
532 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
533 else
534 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
535
536 if (calc_lat(&ts->clat_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->lat_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->bw_stat[ddir], &min, &max, &mean, &dev)) {
547 double p_of_agg;
548
549 p_of_agg = mean * 100 / (double) rs->agg[ddir];
550 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
551 } else
552 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
553}
554
555#define FIO_TERSE_VERSION "2"
556
557static void show_thread_status_terse(struct thread_stat *ts,
558 struct group_run_stats *rs)
559{
560 double io_u_dist[FIO_IO_U_MAP_NR];
561 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
562 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
563 double usr_cpu, sys_cpu;
564 int i;
565
566 /* General Info */
567 log_info("%s;%s;%d;%d", FIO_TERSE_VERSION, ts->name, ts->groupid,
568 ts->error);
569 /* Log Read Status */
570 show_ddir_status_terse(ts, rs, 0);
571 /* Log Write Status */
572 show_ddir_status_terse(ts, rs, 1);
573
574 /* CPU Usage */
575 if (ts->total_run_time) {
576 double runt = (double) ts->total_run_time;
577
578 usr_cpu = (double) ts->usr_time * 100 / runt;
579 sys_cpu = (double) ts->sys_time * 100 / runt;
580 } else {
581 usr_cpu = 0;
582 sys_cpu = 0;
583 }
584
585 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf,
586 ts->minf);
587
588 /* Calc % distribution of IO depths, usecond, msecond latency */
589 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist);
590 stat_calc_lat_u(ts, io_u_lat_u);
591 stat_calc_lat_m(ts, io_u_lat_m);
592
593 /* Only show fixed 7 I/O depth levels*/
594 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
595 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
596 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
597
598 /* Microsecond latency */
599 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
600 log_info(";%3.2f%%", io_u_lat_u[i]);
601 /* Millisecond latency */
602 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
603 log_info(";%3.2f%%", io_u_lat_m[i]);
604 /* Additional output if continue_on_error set - default off*/
605 if (ts->continue_on_error)
606 log_info(";%lu;%d", ts->total_err_count, ts->first_error);
607 log_info("\n");
608
609 /* Additional output if description is set */
610 if (ts->description)
611 log_info(";%s", ts->description);
612
613 log_info("\n");
614}
615
616static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
617{
618 double mean, S;
619
620 if (src->samples == 0)
621 return;
622
623 dst->min_val = min(dst->min_val, src->min_val);
624 dst->max_val = max(dst->max_val, src->max_val);
625
626 /*
627 * Compute new mean and S after the merge
628 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
629 * #Parallel_algorithm>
630 */
631 if (nr == 1) {
632 mean = src->mean;
633 S = src->S;
634 } else {
635 double delta = src->mean - dst->mean;
636
637 mean = ((src->mean * src->samples) +
638 (dst->mean * dst->samples)) /
639 (dst->samples + src->samples);
640
641 S = src->S + dst->S + pow(delta, 2.0) *
642 (dst->samples * src->samples) /
643 (dst->samples + src->samples);
644 }
645
646 dst->samples += src->samples;
647 dst->mean = mean;
648 dst->S = S;
649}
650
651void show_run_stats(void)
652{
653 struct group_run_stats *runstats, *rs;
654 struct thread_data *td;
655 struct thread_stat *threadstats, *ts;
656 int i, j, k, l, nr_ts, last_ts, idx;
657 int kb_base_warned = 0;
658
659 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
660
661 for (i = 0; i < groupid + 1; i++) {
662 rs = &runstats[i];
663
664 memset(rs, 0, sizeof(*rs));
665 rs->min_bw[0] = rs->min_run[0] = ~0UL;
666 rs->min_bw[1] = rs->min_run[1] = ~0UL;
667 }
668
669 /*
670 * find out how many threads stats we need. if group reporting isn't
671 * enabled, it's one-per-td.
672 */
673 nr_ts = 0;
674 last_ts = -1;
675 for_each_td(td, i) {
676 if (!td->o.group_reporting) {
677 nr_ts++;
678 continue;
679 }
680 if (last_ts == td->groupid)
681 continue;
682
683 last_ts = td->groupid;
684 nr_ts++;
685 }
686
687 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
688
689 for (i = 0; i < nr_ts; i++) {
690 ts = &threadstats[i];
691
692 memset(ts, 0, sizeof(*ts));
693 for (j = 0; j <= DDIR_WRITE; j++) {
694 ts->lat_stat[j].min_val = -1UL;
695 ts->clat_stat[j].min_val = -1UL;
696 ts->slat_stat[j].min_val = -1UL;
697 ts->bw_stat[j].min_val = -1UL;
698 }
699 ts->groupid = -1;
700 }
701
702 j = 0;
703 last_ts = -1;
704 idx = 0;
705 for_each_td(td, i) {
706 if (idx && (!td->o.group_reporting ||
707 (td->o.group_reporting && last_ts != td->groupid))) {
708 idx = 0;
709 j++;
710 }
711
712 last_ts = td->groupid;
713
714 ts = &threadstats[j];
715
716 ts->clat_percentiles = td->o.clat_percentiles;
717 if (td->o.overwrite_plist)
718 ts->percentile_list = td->o.percentile_list;
719 else
720 ts->percentile_list = NULL;
721
722 idx++;
723 ts->members++;
724
725 if (ts->groupid == -1) {
726 /*
727 * These are per-group shared already
728 */
729 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
730 if (td->o.description)
731 strncpy(ts->description, td->o.description,
732 FIO_JOBNAME_SIZE);
733 else
734 memset(ts->description, 0, FIO_JOBNAME_SIZE);
735
736 ts->groupid = td->groupid;
737
738 /*
739 * first pid in group, not very useful...
740 */
741 ts->pid = td->pid;
742
743 ts->kb_base = td->o.kb_base;
744 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
745 log_info("fio: kb_base differs for jobs in group, using"
746 " %u as the base\n", ts->kb_base);
747 kb_base_warned = 1;
748 }
749
750 ts->continue_on_error = td->o.continue_on_error;
751 ts->total_err_count += td->total_err_count;
752 ts->first_error = td->first_error;
753 if (!ts->error) {
754 if (!td->error && td->o.continue_on_error &&
755 td->first_error) {
756 ts->error = td->first_error;
757 strcpy(ts->verror, td->verror);
758 } else if (td->error) {
759 ts->error = td->error;
760 strcpy(ts->verror, td->verror);
761 }
762 }
763
764 for (l = 0; l <= DDIR_WRITE; l++) {
765 sum_stat(&ts->clat_stat[l], &td->ts.clat_stat[l], idx);
766 sum_stat(&ts->slat_stat[l], &td->ts.slat_stat[l], idx);
767 sum_stat(&ts->lat_stat[l], &td->ts.lat_stat[l], idx);
768 sum_stat(&ts->bw_stat[l], &td->ts.bw_stat[l], idx);
769
770 ts->io_bytes[l] += td->ts.io_bytes[l];
771
772 if (ts->runtime[l] < td->ts.runtime[l])
773 ts->runtime[l] = td->ts.runtime[l];
774 }
775
776 ts->usr_time += td->ts.usr_time;
777 ts->sys_time += td->ts.sys_time;
778 ts->ctx += td->ts.ctx;
779 ts->majf += td->ts.majf;
780 ts->minf += td->ts.minf;
781
782 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
783 ts->io_u_map[k] += td->ts.io_u_map[k];
784 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
785 ts->io_u_submit[k] += td->ts.io_u_submit[k];
786 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
787 ts->io_u_complete[k] += td->ts.io_u_complete[k];
788 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
789 ts->io_u_lat_u[k] += td->ts.io_u_lat_u[k];
790 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
791 ts->io_u_lat_m[k] += td->ts.io_u_lat_m[k];
792
793
794 for (k = 0; k <= 2; k++) {
795 ts->total_io_u[k] += td->ts.total_io_u[k];
796 ts->short_io_u[k] += td->ts.short_io_u[k];
797 }
798
799 for (k = 0; k <= DDIR_WRITE; k++) {
800 int m;
801 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
802 ts->io_u_plat[k][m] += td->ts.io_u_plat[k][m];
803 }
804
805 ts->total_run_time += td->ts.total_run_time;
806 ts->total_submit += td->ts.total_submit;
807 ts->total_complete += td->ts.total_complete;
808 }
809
810 for (i = 0; i < nr_ts; i++) {
811 unsigned long long bw;
812
813 ts = &threadstats[i];
814 rs = &runstats[ts->groupid];
815 rs->kb_base = ts->kb_base;
816
817 for (j = 0; j <= DDIR_WRITE; j++) {
818 if (!ts->runtime[j])
819 continue;
820 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
821 rs->min_run[j] = ts->runtime[j];
822 if (ts->runtime[j] > rs->max_run[j])
823 rs->max_run[j] = ts->runtime[j];
824
825 bw = 0;
826 if (ts->runtime[j]) {
827 unsigned long runt;
828
829 runt = ts->runtime[j];
830 bw = ts->io_bytes[j] / runt;
831 }
832 if (bw < rs->min_bw[j])
833 rs->min_bw[j] = bw;
834 if (bw > rs->max_bw[j])
835 rs->max_bw[j] = bw;
836
837 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
838 }
839 }
840
841 for (i = 0; i < groupid + 1; i++) {
842 unsigned long max_run[2];
843
844 rs = &runstats[i];
845 max_run[0] = rs->max_run[0];
846 max_run[1] = rs->max_run[1];
847
848 if (rs->max_run[0])
849 rs->agg[0] = (rs->io_kb[0] * 1000) / max_run[0];
850 if (rs->max_run[1])
851 rs->agg[1] = (rs->io_kb[1] * 1000) / max_run[1];
852 }
853
854 /*
855 * don't overwrite last signal output
856 */
857 if (!terse_output)
858 log_info("\n");
859
860 for (i = 0; i < nr_ts; i++) {
861 ts = &threadstats[i];
862 rs = &runstats[ts->groupid];
863
864 if (is_backend)
865 fio_server_send_ts(ts, rs);
866 else if (terse_output)
867 show_thread_status_terse(ts, rs);
868 else
869 show_thread_status(ts, rs);
870 }
871
872 if (!terse_output) {
873 for (i = 0; i < groupid + 1; i++) {
874 rs = &runstats[i];
875
876 rs->groupid = i;
877 if (is_backend)
878 fio_server_send_gs(rs);
879 else
880 show_group_stats(rs);
881 }
882
883 show_disk_util();
884 }
885
886 free(runstats);
887 free(threadstats);
888}
889
890static inline void add_stat_sample(struct io_stat *is, unsigned long data)
891{
892 double val = data;
893 double delta;
894
895 if (data > is->max_val)
896 is->max_val = data;
897 if (data < is->min_val)
898 is->min_val = data;
899
900 delta = val - is->mean;
901 if (delta) {
902 is->mean += delta / (is->samples + 1.0);
903 is->S += delta * (val - is->mean);
904 }
905
906 is->samples++;
907}
908
909static void __add_log_sample(struct io_log *iolog, unsigned long val,
910 enum fio_ddir ddir, unsigned int bs,
911 unsigned long t)
912{
913 const int nr_samples = iolog->nr_samples;
914
915 if (iolog->nr_samples == iolog->max_samples) {
916 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
917
918 iolog->log = realloc(iolog->log, new_size);
919 iolog->max_samples <<= 1;
920 }
921
922 iolog->log[nr_samples].val = val;
923 iolog->log[nr_samples].time = t;
924 iolog->log[nr_samples].ddir = ddir;
925 iolog->log[nr_samples].bs = bs;
926 iolog->nr_samples++;
927}
928
929static void add_log_sample(struct thread_data *td, struct io_log *iolog,
930 unsigned long val, enum fio_ddir ddir,
931 unsigned int bs)
932{
933 if (!ddir_rw(ddir))
934 return;
935
936 __add_log_sample(iolog, val, ddir, bs, mtime_since_now(&td->epoch));
937}
938
939void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
940{
941 struct io_log *iolog;
942
943 if (!ddir_rw(ddir))
944 return;
945
946 iolog = agg_io_log[ddir];
947 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
948}
949
950static void add_clat_percentile_sample(struct thread_stat *ts,
951 unsigned long usec, enum fio_ddir ddir)
952{
953 unsigned int idx = plat_val_to_idx(usec);
954 assert(idx < FIO_IO_U_PLAT_NR);
955
956 ts->io_u_plat[ddir][idx]++;
957}
958
959void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
960 unsigned long usec, unsigned int bs)
961{
962 struct thread_stat *ts = &td->ts;
963
964 if (!ddir_rw(ddir))
965 return;
966
967 add_stat_sample(&ts->clat_stat[ddir], usec);
968
969 if (td->clat_log)
970 add_log_sample(td, td->clat_log, usec, ddir, bs);
971
972 if (ts->clat_percentiles)
973 add_clat_percentile_sample(ts, usec, ddir);
974}
975
976void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
977 unsigned long usec, unsigned int bs)
978{
979 struct thread_stat *ts = &td->ts;
980
981 if (!ddir_rw(ddir))
982 return;
983
984 add_stat_sample(&ts->slat_stat[ddir], usec);
985
986 if (td->slat_log)
987 add_log_sample(td, td->slat_log, usec, ddir, bs);
988}
989
990void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
991 unsigned long usec, unsigned int bs)
992{
993 struct thread_stat *ts = &td->ts;
994
995 if (!ddir_rw(ddir))
996 return;
997
998 add_stat_sample(&ts->lat_stat[ddir], usec);
999
1000 if (td->lat_log)
1001 add_log_sample(td, td->lat_log, usec, ddir, bs);
1002}
1003
1004void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1005 struct timeval *t)
1006{
1007 struct thread_stat *ts = &td->ts;
1008 unsigned long spent, rate;
1009
1010 if (!ddir_rw(ddir))
1011 return;
1012
1013 spent = mtime_since(&td->stat_sample_time[ddir], t);
1014 if (spent < td->o.bw_avg_time)
1015 return;
1016
1017 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) *
1018 1000 / spent / 1024;
1019 add_stat_sample(&ts->bw_stat[ddir], rate);
1020
1021 if (td->bw_log)
1022 add_log_sample(td, td->bw_log, rate, ddir, bs);
1023
1024 fio_gettime(&td->stat_sample_time[ddir], NULL);
1025 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1026}
fio - Flexible IO Tester