13 * Cheesy number->string conversion, complete with carry rounding error.
15 static char *num2str(unsigned long num, int maxlen, int base, int pow2)
17 char postfix[] = { ' ', 'K', 'M', 'G', 'P', 'E' };
18 unsigned int thousand;
29 for (i = 0; base > 1; i++)
35 len = sprintf(buf, "%'lu", num);
38 buf[len] = postfix[i];
44 if ((num % thousand) >= (thousand / 2))
55 void update_rusage_stat(struct thread_data *td)
57 struct thread_stat *ts = &td->ts;
59 getrusage(RUSAGE_SELF, &ts->ru_end);
61 ts->usr_time += mtime_since(&ts->ru_start.ru_utime, &ts->ru_end.ru_utime);
62 ts->sys_time += mtime_since(&ts->ru_start.ru_stime, &ts->ru_end.ru_stime);
63 ts->ctx += ts->ru_end.ru_nvcsw + ts->ru_end.ru_nivcsw - (ts->ru_start.ru_nvcsw + ts->ru_start.ru_nivcsw);
65 memcpy(&ts->ru_start, &ts->ru_end, sizeof(ts->ru_end));
68 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
69 double *mean, double *dev)
71 double n = is->samples;
79 n = (double) is->samples;
83 *dev = sqrt(is->S / (n - 1.0));
90 static void show_group_stats(struct group_run_stats *rs, int id)
92 char *p1, *p2, *p3, *p4;
93 const char *ddir_str[] = { " READ", " WRITE" };
96 log_info("\nRun status group %d (all jobs):\n", id);
98 for (i = 0; i <= DDIR_WRITE; i++) {
102 p1 = num2str(rs->io_kb[i], 6, 1000, 1);
103 p2 = num2str(rs->agg[i], 6, 1000, 1);
104 p3 = num2str(rs->min_bw[i], 6, 1000, 1);
105 p4 = num2str(rs->max_bw[i], 6, 1000, 1);
107 log_info("%s: io=%siB, aggrb=%siB/s, minb=%siB/s, maxb=%siB/s, mint=%llumsec, maxt=%llumsec\n", ddir_str[i], p1, p2, p3, p4, rs->min_run[i], rs->max_run[i]);
116 #define ts_total_io_u(ts) \
117 ((ts)->total_io_u[0] + (ts)->total_io_u[1])
119 static void stat_calc_dist(struct thread_stat *ts, double *io_u_dist)
124 * Do depth distribution calculations
126 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
127 io_u_dist[i] = (double) ts->io_u_map[i] / (double) ts_total_io_u(ts);
128 io_u_dist[i] *= 100.0;
129 if (io_u_dist[i] < 0.1 && ts->io_u_map[i])
134 static void stat_calc_lat(struct thread_stat *ts, double *io_u_lat)
139 * Do latency distribution calculations
141 for (i = 0; i < FIO_IO_U_LAT_NR; i++) {
142 io_u_lat[i] = (double) ts->io_u_lat[i] / (double) ts_total_io_u(ts);
143 io_u_lat[i] *= 100.0;
144 if (io_u_lat[i] < 0.01 && ts->io_u_lat[i])
149 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
152 const char *ddir_str[] = { "read ", "write" };
153 unsigned long min, max;
154 unsigned long long bw, iops;
156 char *io_p, *bw_p, *iops_p;
158 if (!ts->runtime[ddir])
161 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
162 iops = (1000 * ts->total_io_u[ddir]) / ts->runtime[ddir];
163 io_p = num2str(ts->io_bytes[ddir] >> 10, 6, 1000, 1);
164 bw_p = num2str(bw, 6, 1000, 1);
165 iops_p = num2str(iops, 6, 1, 0);
167 log_info(" %s: io=%siB, bw=%siB/s, iops=%s, runt=%6lumsec\n", ddir_str[ddir], io_p, bw_p, iops_p, ts->runtime[ddir]);
173 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
174 log_info(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, stdev=%5.02f\n", min, max, mean, dev);
176 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
177 log_info(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, stdev=%5.02f\n", min, max, mean, dev);
179 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
182 p_of_agg = mean * 100 / (double) rs->agg[ddir];
183 log_info(" bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg, mean, dev);
187 static void show_thread_status(struct thread_stat *ts,
188 struct group_run_stats *rs)
190 double usr_cpu, sys_cpu;
191 unsigned long runtime;
192 double io_u_dist[FIO_IO_U_MAP_NR];
193 double io_u_lat[FIO_IO_U_LAT_NR];
195 if (!(ts->io_bytes[0] + ts->io_bytes[1]))
199 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n", ts->name, ts->groupid, ts->members, ts->error, ts->pid);
201 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n", ts->name, ts->groupid, ts->members, ts->error, ts->verror, ts->pid);
204 log_info(" Description : [%s]\n", ts->description);
206 if (ts->io_bytes[DDIR_READ])
207 show_ddir_status(rs, ts, DDIR_READ);
208 if (ts->io_bytes[DDIR_WRITE])
209 show_ddir_status(rs, ts, DDIR_WRITE);
211 runtime = ts->total_run_time;
213 double runt = (double) runtime;
215 usr_cpu = (double) ts->usr_time * 100 / runt;
216 sys_cpu = (double) ts->sys_time * 100 / runt;
222 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, ts->ctx);
224 stat_calc_dist(ts, io_u_dist);
225 stat_calc_lat(ts, io_u_lat);
227 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], io_u_dist[4], io_u_dist[5], io_u_dist[6]);
228 log_info(" issued r/w: total=%lu/%lu, short=%lu/%lu\n", ts->total_io_u[0], ts->total_io_u[1], ts->short_io_u[0], ts->short_io_u[1]);
230 log_info(" lat (msec): 2=%3.2f%%, 4=%3.2f%%, 10=%3.2f%%, 20=%3.2f%%, 50=%3.2f%%, 100=%3.2f%%\n", io_u_lat[0], io_u_lat[1], io_u_lat[2], io_u_lat[3], io_u_lat[4], io_u_lat[5]);
231 log_info(" lat (msec): 250=%3.2f%%, 500=%3.2f%%, 750=%3.2f%%, 1000=%3.2f%%, >=2000=%3.2f%%\n", io_u_lat[6], io_u_lat[7], io_u_lat[8], io_u_lat[9], io_u_lat[10]);
234 static void show_ddir_status_terse(struct thread_stat *ts,
235 struct group_run_stats *rs, int ddir)
237 unsigned long min, max;
238 unsigned long long bw;
242 if (ts->runtime[ddir])
243 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
245 log_info(";%llu;%llu;%lu", ts->io_bytes[ddir] >> 10, bw, ts->runtime[ddir]);
247 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
248 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
250 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
252 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
253 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
255 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
257 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
260 p_of_agg = mean * 100 / (double) rs->agg[ddir];
261 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
263 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
267 static void show_thread_status_terse(struct thread_stat *ts,
268 struct group_run_stats *rs)
270 double io_u_dist[FIO_IO_U_MAP_NR];
271 double io_u_lat[FIO_IO_U_LAT_NR];
272 double usr_cpu, sys_cpu;
274 log_info("%s;%d;%d", ts->name, ts->groupid, ts->error);
276 show_ddir_status_terse(ts, rs, 0);
277 show_ddir_status_terse(ts, rs, 1);
279 if (ts->total_run_time) {
280 double runt = (double) ts->total_run_time;
282 usr_cpu = (double) ts->usr_time * 100 / runt;
283 sys_cpu = (double) ts->sys_time * 100 / runt;
289 log_info(";%f%%;%f%%;%lu", usr_cpu, sys_cpu, ts->ctx);
291 stat_calc_dist(ts, io_u_dist);
292 stat_calc_lat(ts, io_u_lat);
294 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], io_u_dist[4], io_u_dist[5], io_u_dist[6]);
296 log_info(";%3.2f%%;%3.2f%%;%3.2f%%;%3.2f%%;%3.2f%%;%3.2f%%\n", io_u_lat[0], io_u_lat[1], io_u_lat[2], io_u_lat[3], io_u_lat[4], io_u_lat[5]);
297 log_info(";%3.2f%%;%3.2f%%;%3.2f%%;%3.2f%%;%3.2f%%", io_u_lat[6], io_u_lat[7], io_u_lat[8], io_u_lat[9], io_u_lat[10]);
300 log_info(";%s", ts->description);
305 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
309 dst->min_val = min(dst->min_val, src->min_val);
310 dst->max_val = max(dst->max_val, src->max_val);
311 dst->samples += src->samples;
314 * Needs a new method for calculating stddev, we cannot just
315 * average them we do below for nr > 1
321 mean = ((src->mean * (double) (nr - 1)) + dst->mean) / ((double) nr);
322 S = ((src->S * (double) (nr - 1)) + dst->S) / ((double) nr);
329 void show_run_stats(void)
331 struct group_run_stats *runstats, *rs;
332 struct thread_data *td;
333 struct thread_stat *threadstats, *ts;
334 int i, j, k, l, nr_ts, last_ts, idx;
336 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
338 for (i = 0; i < groupid + 1; i++) {
341 memset(rs, 0, sizeof(*rs));
342 rs->min_bw[0] = rs->min_run[0] = ~0UL;
343 rs->min_bw[1] = rs->min_run[1] = ~0UL;
347 * find out how many threads stats we need. if group reporting isn't
348 * enabled, it's one-per-td.
353 if (!td->o.group_reporting) {
357 if (last_ts == td->groupid)
360 last_ts = td->groupid;
364 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
366 for (i = 0; i < nr_ts; i++) {
367 ts = &threadstats[i];
369 memset(ts, 0, sizeof(*ts));
370 for (j = 0; j <= DDIR_WRITE; j++) {
371 ts->clat_stat[j].min_val = -1UL;
372 ts->slat_stat[j].min_val = -1UL;
373 ts->bw_stat[j].min_val = -1UL;
382 if (idx && (!td->o.group_reporting ||
383 (td->o.group_reporting && last_ts != td->groupid))) {
388 last_ts = td->groupid;
390 ts = &threadstats[j];
395 if (ts->groupid == -1) {
397 * These are per-group shared already
399 ts->name = td->o.name;
400 ts->description = td->o.description;
401 ts->groupid = td->groupid;
404 * first pid in group, not very useful...
409 if (td->error && !ts->error) {
410 ts->error = td->error;
411 ts->verror = td->verror;
414 for (l = 0; l <= DDIR_WRITE; l++) {
415 sum_stat(&ts->clat_stat[l], &td->ts.clat_stat[l], idx);
416 sum_stat(&ts->slat_stat[l], &td->ts.slat_stat[l], idx);
417 sum_stat(&ts->bw_stat[l], &td->ts.bw_stat[l], idx);
419 ts->stat_io_bytes[l] += td->ts.stat_io_bytes[l];
420 ts->io_bytes[l] += td->ts.io_bytes[l];
422 if (ts->runtime[l] < td->ts.runtime[l])
423 ts->runtime[l] = td->ts.runtime[l];
426 ts->usr_time += td->ts.usr_time;
427 ts->sys_time += td->ts.sys_time;
428 ts->ctx += td->ts.ctx;
430 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
431 ts->io_u_map[k] += td->ts.io_u_map[k];
432 for (k = 0; k < FIO_IO_U_LAT_NR; k++)
433 ts->io_u_lat[k] += td->ts.io_u_lat[k];
435 for (k = 0; k <= DDIR_WRITE; k++) {
436 ts->total_io_u[k] += td->ts.total_io_u[k];
437 ts->short_io_u[k] += td->ts.short_io_u[k];
440 ts->total_run_time += td->ts.total_run_time;
443 for (i = 0; i < nr_ts; i++) {
444 unsigned long long bw;
446 ts = &threadstats[i];
447 rs = &runstats[ts->groupid];
449 for (j = 0; j <= DDIR_WRITE; j++) {
452 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
453 rs->min_run[j] = ts->runtime[j];
454 if (ts->runtime[j] > rs->max_run[j])
455 rs->max_run[j] = ts->runtime[j];
459 bw = ts->io_bytes[j] / (unsigned long long) ts->runtime[j];
460 if (bw < rs->min_bw[j])
462 if (bw > rs->max_bw[j])
465 rs->io_kb[j] += ts->io_bytes[j] >> 10;
469 for (i = 0; i < groupid + 1; i++) {
473 rs->agg[0] = (rs->io_kb[0]*1024) / rs->max_run[0];
475 rs->agg[1] = (rs->io_kb[1]*1024) / rs->max_run[1];
479 * don't overwrite last signal output
484 for (i = 0; i < nr_ts; i++) {
485 ts = &threadstats[i];
486 rs = &runstats[ts->groupid];
489 show_thread_status_terse(ts, rs);
491 show_thread_status(ts, rs);
495 for (i = 0; i < groupid + 1; i++)
496 show_group_stats(&runstats[i], i);
505 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
510 if (data > is->max_val)
512 if (data < is->min_val)
515 delta = val - is->mean;
517 is->mean += delta / (is->samples + 1.0);
518 is->S += delta * (val - is->mean);
524 static void __add_log_sample(struct io_log *iolog, unsigned long val,
525 enum fio_ddir ddir, unsigned long time)
527 if (iolog->nr_samples == iolog->max_samples) {
528 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
530 iolog->log = realloc(iolog->log, new_size);
531 iolog->max_samples <<= 1;
534 iolog->log[iolog->nr_samples].val = val;
535 iolog->log[iolog->nr_samples].time = time;
536 iolog->log[iolog->nr_samples].ddir = ddir;
540 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
541 unsigned long val, enum fio_ddir ddir)
543 __add_log_sample(iolog, val, ddir, mtime_since_now(&td->epoch));
546 void add_agg_sample(unsigned long val, enum fio_ddir ddir)
548 struct io_log *iolog = agg_io_log[ddir];
550 __add_log_sample(iolog, val, ddir, mtime_since_genesis());
553 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
556 struct thread_stat *ts = &td->ts;
558 add_stat_sample(&ts->clat_stat[ddir], msec);
561 add_log_sample(td, ts->clat_log, msec, ddir);
564 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
567 struct thread_stat *ts = &td->ts;
569 add_stat_sample(&ts->slat_stat[ddir], msec);
572 add_log_sample(td, ts->slat_log, msec, ddir);
575 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir,
578 struct thread_stat *ts = &td->ts;
579 unsigned long spent = mtime_since(&ts->stat_sample_time[ddir], t);
582 if (spent < td->o.bw_avg_time)
585 rate = (td->this_io_bytes[ddir] - ts->stat_io_bytes[ddir]) / spent;
586 add_stat_sample(&ts->bw_stat[ddir], rate);
589 add_log_sample(td, ts->bw_log, rate, ddir);
591 fio_gettime(&ts->stat_sample_time[ddir], NULL);
592 ts->stat_io_bytes[ddir] = td->this_io_bytes[ddir];