Improve latency_target runs
[fio.git] / eta.c
1 /*
2  * Status and ETA code
3  */
4 #include <unistd.h>
5 #include <fcntl.h>
6 #include <string.h>
7
8 #include "fio.h"
9
10 static char run_str[REAL_MAX_JOBS + 1];
11
12 /*
13  * Sets the status of the 'td' in the printed status map.
14  */
15 static void check_str_update(struct thread_data *td)
16 {
17         char c = run_str[td->thread_number - 1];
18
19         switch (td->runstate) {
20         case TD_REAPED:
21                 if (td->error)
22                         c = 'X';
23                 else if (td->sig)
24                         c = 'K';
25                 else
26                         c = '_';
27                 break;
28         case TD_EXITED:
29                 c = 'E';
30                 break;
31         case TD_RAMP:
32                 c = '/';
33                 break;
34         case TD_RUNNING:
35                 if (td_rw(td)) {
36                         if (td_random(td)) {
37                                 if (td->o.rwmix[DDIR_READ] == 100)
38                                         c = 'r';
39                                 else if (td->o.rwmix[DDIR_WRITE] == 100)
40                                         c = 'w';
41                                 else
42                                         c = 'm';
43                         } else {
44                                 if (td->o.rwmix[DDIR_READ] == 100)
45                                         c = 'R';
46                                 else if (td->o.rwmix[DDIR_WRITE] == 100)
47                                         c = 'W';
48                                 else
49                                         c = 'M';
50                         }
51                 } else if (td_read(td)) {
52                         if (td_random(td))
53                                 c = 'r';
54                         else
55                                 c = 'R';
56                 } else if (td_write(td)) {
57                         if (td_random(td))
58                                 c = 'w';
59                         else
60                                 c = 'W';
61                 } else {
62                         if (td_random(td))
63                                 c = 'd';
64                         else
65                                 c = 'D';
66                 }
67                 break;
68         case TD_PRE_READING:
69                 c = 'p';
70                 break;
71         case TD_VERIFYING:
72                 c = 'V';
73                 break;
74         case TD_FSYNCING:
75                 c = 'F';
76                 break;
77         case TD_CREATED:
78                 c = 'C';
79                 break;
80         case TD_INITIALIZED:
81         case TD_SETTING_UP:
82                 c = 'I';
83                 break;
84         case TD_NOT_CREATED:
85                 c = 'P';
86                 break;
87         default:
88                 log_err("state %d\n", td->runstate);
89         }
90
91         run_str[td->thread_number - 1] = c;
92 }
93
94 /*
95  * Convert seconds to a printable string.
96  */
97 void eta_to_str(char *str, unsigned long eta_sec)
98 {
99         unsigned int d, h, m, s;
100         int disp_hour = 0;
101
102         s = eta_sec % 60;
103         eta_sec /= 60;
104         m = eta_sec % 60;
105         eta_sec /= 60;
106         h = eta_sec % 24;
107         eta_sec /= 24;
108         d = eta_sec;
109
110         if (d) {
111                 disp_hour = 1;
112                 str += sprintf(str, "%02ud:", d);
113         }
114
115         if (h || disp_hour)
116                 str += sprintf(str, "%02uh:", h);
117
118         str += sprintf(str, "%02um:", m);
119         str += sprintf(str, "%02us", s);
120 }
121
122 /*
123  * Best effort calculation of the estimated pending runtime of a job.
124  */
125 static int thread_eta(struct thread_data *td)
126 {
127         unsigned long long bytes_total, bytes_done;
128         unsigned long eta_sec = 0;
129         unsigned long elapsed;
130
131         elapsed = (mtime_since_now(&td->epoch) + 999) / 1000;
132
133         bytes_total = td->total_io_size;
134
135         if (td->o.fill_device && td->o.size  == -1ULL) {
136                 if (!td->fill_device_size || td->fill_device_size == -1ULL)
137                         return 0;
138
139                 bytes_total = td->fill_device_size;
140         }
141
142         if (td->o.zone_size && td->o.zone_skip && bytes_total) {
143                 unsigned int nr_zones;
144                 uint64_t zone_bytes;
145
146                 zone_bytes = bytes_total + td->o.zone_size + td->o.zone_skip;
147                 nr_zones = (zone_bytes - 1) / (td->o.zone_size + td->o.zone_skip);
148                 bytes_total -= nr_zones * td->o.zone_skip;
149         }
150
151         /*
152          * if writing and verifying afterwards, bytes_total will be twice the
153          * size. In a mixed workload, verify phase will be the size of the
154          * first stage writes.
155          */
156         if (td->o.do_verify && td->o.verify && td_write(td)) {
157                 if (td_rw(td)) {
158                         unsigned int perc = 50;
159
160                         if (td->o.rwmix[DDIR_WRITE])
161                                 perc = td->o.rwmix[DDIR_WRITE];
162
163                         bytes_total += (bytes_total * perc) / 100;
164                 } else
165                         bytes_total <<= 1;
166         }
167
168         if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
169                 double perc, perc_t;
170
171                 bytes_done = ddir_rw_sum(td->io_bytes);
172                 perc = (double) bytes_done / (double) bytes_total;
173                 if (perc > 1.0)
174                         perc = 1.0;
175
176                 if (td->o.time_based) {
177                         perc_t = (double) elapsed /
178                                         (double) (td->o.timeout / 1000);
179                         if (perc_t < perc)
180                                 perc = perc_t;
181                 }
182
183                 eta_sec = (unsigned long) (elapsed * (1.0 / perc)) - elapsed;
184
185                 if (td->o.timeout &&
186                     eta_sec > ( (td->o.timeout / 1000) + done_secs - elapsed))
187                         eta_sec = (td->o.timeout / 1000)  + done_secs
188                                         - elapsed;
189         } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
190                         || td->runstate == TD_INITIALIZED
191                         || td->runstate == TD_SETTING_UP
192                         || td->runstate == TD_RAMP
193                         || td->runstate == TD_PRE_READING) {
194                 int t_eta = 0, r_eta = 0;
195                 unsigned long long rate_bytes;
196
197                 /*
198                  * We can only guess - assume it'll run the full timeout
199                  * if given, otherwise assume it'll run at the specified rate.
200                  */
201                 if (td->o.timeout) {
202                         t_eta = (td->o.timeout + td->o.start_delay  +
203                                         td->o.ramp_time ) / 1000;
204
205                         if (in_ramp_time(td)) {
206                                 unsigned long ramp_left;
207
208                                 ramp_left = mtime_since_now(&td->epoch);
209                                 ramp_left = (ramp_left + 999) / 1000;
210                                 if (ramp_left <= t_eta)
211                                         t_eta -= ramp_left;
212                         }
213                 }
214                 rate_bytes = ddir_rw_sum(td->o.rate);
215                 if (rate_bytes) {
216                         r_eta = (bytes_total / 1024) / rate_bytes;
217                         r_eta += td->o.start_delay / 1000;
218                 }
219
220                 if (r_eta && t_eta)
221                         eta_sec = min(r_eta, t_eta);
222                 else if (r_eta)
223                         eta_sec = r_eta;
224                 else if (t_eta)
225                         eta_sec = t_eta;
226                 else
227                         eta_sec = 0;
228         } else {
229                 /*
230                  * thread is already done or waiting for fsync
231                  */
232                 eta_sec = 0;
233         }
234
235         return eta_sec;
236 }
237
238 static void calc_rate(int unified_rw_rep, unsigned long mtime,
239                       unsigned long long *io_bytes,
240                       unsigned long long *prev_io_bytes, unsigned int *rate)
241 {
242         int i;
243
244         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
245                 unsigned long long diff;
246
247                 diff = io_bytes[i] - prev_io_bytes[i];
248                 if (unified_rw_rep) {
249                         rate[i] = 0;
250                         rate[0] += ((1000 * diff) / mtime) / 1024;
251                 } else
252                         rate[i] = ((1000 * diff) / mtime) / 1024;
253
254                 prev_io_bytes[i] = io_bytes[i];
255         }
256 }
257
258 static void calc_iops(int unified_rw_rep, unsigned long mtime,
259                       unsigned long long *io_iops,
260                       unsigned long long *prev_io_iops, unsigned int *iops)
261 {
262         int i;
263
264         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
265                 unsigned long long diff;
266
267                 diff = io_iops[i] - prev_io_iops[i];
268                 if (unified_rw_rep) {
269                         iops[i] = 0;
270                         iops[0] += (diff * 1000) / mtime;
271                 } else
272                         iops[i] = (diff * 1000) / mtime;
273
274                 prev_io_iops[i] = io_iops[i];
275         }
276 }
277
278 /*
279  * Print status of the jobs we know about. This includes rate estimates,
280  * ETA, thread state, etc.
281  */
282 int calc_thread_status(struct jobs_eta *je, int force)
283 {
284         struct thread_data *td;
285         int i, unified_rw_rep;
286         unsigned long rate_time, disp_time, bw_avg_time, *eta_secs;
287         unsigned long long io_bytes[DDIR_RWDIR_CNT];
288         unsigned long long io_iops[DDIR_RWDIR_CNT];
289         struct timeval now;
290
291         static unsigned long long rate_io_bytes[DDIR_RWDIR_CNT];
292         static unsigned long long disp_io_bytes[DDIR_RWDIR_CNT];
293         static unsigned long long disp_io_iops[DDIR_RWDIR_CNT];
294         static struct timeval rate_prev_time, disp_prev_time;
295
296         if (!force) {
297                 if (output_format != FIO_OUTPUT_NORMAL &&
298                     f_out == stdout)
299                         return 0;
300                 if (temp_stall_ts || eta_print == FIO_ETA_NEVER)
301                         return 0;
302
303                 if (!isatty(STDOUT_FILENO) && (eta_print != FIO_ETA_ALWAYS))
304                         return 0;
305         }
306
307         if (!ddir_rw_sum(rate_io_bytes))
308                 fill_start_time(&rate_prev_time);
309         if (!ddir_rw_sum(disp_io_bytes))
310                 fill_start_time(&disp_prev_time);
311
312         eta_secs = malloc(thread_number * sizeof(unsigned long));
313         memset(eta_secs, 0, thread_number * sizeof(unsigned long));
314
315         je->elapsed_sec = (mtime_since_genesis() + 999) / 1000;
316
317         io_bytes[DDIR_READ] = io_bytes[DDIR_WRITE] = io_bytes[DDIR_TRIM] = 0;
318         io_iops[DDIR_READ] = io_iops[DDIR_WRITE] = io_iops[DDIR_TRIM] = 0;
319         bw_avg_time = ULONG_MAX;
320         unified_rw_rep = 0;
321         for_each_td(td, i) {
322                 unified_rw_rep += td->o.unified_rw_rep;
323                 if (is_power_of_2(td->o.kb_base))
324                         je->is_pow2 = 1;
325                 je->unit_base = td->o.unit_base;
326                 if (td->o.bw_avg_time < bw_avg_time)
327                         bw_avg_time = td->o.bw_avg_time;
328                 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING
329                     || td->runstate == TD_FSYNCING
330                     || td->runstate == TD_PRE_READING) {
331                         je->nr_running++;
332                         if (td_read(td)) {
333                                 je->t_rate[0] += td->o.rate[DDIR_READ];
334                                 je->t_iops[0] += td->o.rate_iops[DDIR_READ];
335                                 je->m_rate[0] += td->o.ratemin[DDIR_READ];
336                                 je->m_iops[0] += td->o.rate_iops_min[DDIR_READ];
337                         }
338                         if (td_write(td)) {
339                                 je->t_rate[1] += td->o.rate[DDIR_WRITE];
340                                 je->t_iops[1] += td->o.rate_iops[DDIR_WRITE];
341                                 je->m_rate[1] += td->o.ratemin[DDIR_WRITE];
342                                 je->m_iops[1] += td->o.rate_iops_min[DDIR_WRITE];
343                         }
344                         if (td_trim(td)) {
345                                 je->t_rate[2] += td->o.rate[DDIR_TRIM];
346                                 je->t_iops[2] += td->o.rate_iops[DDIR_TRIM];
347                                 je->m_rate[2] += td->o.ratemin[DDIR_TRIM];
348                                 je->m_iops[2] += td->o.rate_iops_min[DDIR_TRIM];
349                         }
350
351                         je->files_open += td->nr_open_files;
352                 } else if (td->runstate == TD_RAMP) {
353                         je->nr_running++;
354                         je->nr_ramp++;
355                 } else if (td->runstate == TD_SETTING_UP) {
356                         je->nr_running++;
357                         je->nr_setting_up++;
358                 } else if (td->runstate < TD_RUNNING)
359                         je->nr_pending++;
360
361                 if (je->elapsed_sec >= 3)
362                         eta_secs[i] = thread_eta(td);
363                 else
364                         eta_secs[i] = INT_MAX;
365
366                 check_str_update(td);
367
368                 if (td->runstate > TD_SETTING_UP) {
369                         int ddir;
370
371                         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
372                                 if (unified_rw_rep) {
373                                         io_bytes[0] += td->io_bytes[ddir];
374                                         io_iops[0] += td->io_blocks[ddir];
375                                 } else {
376                                         io_bytes[ddir] += td->io_bytes[ddir];
377                                         io_iops[ddir] += td->io_blocks[ddir];
378                                 }
379                         }
380                 }
381         }
382
383         if (exitall_on_terminate)
384                 je->eta_sec = INT_MAX;
385         else
386                 je->eta_sec = 0;
387
388         for_each_td(td, i) {
389                 if (exitall_on_terminate) {
390                         if (eta_secs[i] < je->eta_sec)
391                                 je->eta_sec = eta_secs[i];
392                 } else {
393                         if (eta_secs[i] > je->eta_sec)
394                                 je->eta_sec = eta_secs[i];
395                 }
396         }
397
398         free(eta_secs);
399
400         fio_gettime(&now, NULL);
401         rate_time = mtime_since(&rate_prev_time, &now);
402
403         if (write_bw_log && rate_time > bw_avg_time && !in_ramp_time(td)) {
404                 calc_rate(unified_rw_rep, rate_time, io_bytes, rate_io_bytes,
405                                 je->rate);
406                 memcpy(&rate_prev_time, &now, sizeof(now));
407                 add_agg_sample(je->rate[DDIR_READ], DDIR_READ, 0);
408                 add_agg_sample(je->rate[DDIR_WRITE], DDIR_WRITE, 0);
409                 add_agg_sample(je->rate[DDIR_TRIM], DDIR_TRIM, 0);
410         }
411
412         disp_time = mtime_since(&disp_prev_time, &now);
413
414         /*
415          * Allow a little slack, the target is to print it every 1000 msecs
416          */
417         if (!force && disp_time < 900)
418                 return 0;
419
420         calc_rate(unified_rw_rep, disp_time, io_bytes, disp_io_bytes, je->rate);
421         calc_iops(unified_rw_rep, disp_time, io_iops, disp_io_iops, je->iops);
422
423         memcpy(&disp_prev_time, &now, sizeof(now));
424
425         if (!force && !je->nr_running && !je->nr_pending)
426                 return 0;
427
428         je->nr_threads = thread_number;
429         memcpy(je->run_str, run_str, thread_number * sizeof(char));
430         return 1;
431 }
432
433 void display_thread_status(struct jobs_eta *je)
434 {
435         static struct timeval disp_eta_new_line;
436         static int eta_new_line_init, eta_new_line_pending;
437         static int linelen_last;
438         static int eta_good;
439         char output[REAL_MAX_JOBS + 512], *p = output;
440         char eta_str[128];
441         double perc = 0.0;
442
443         if (je->eta_sec != INT_MAX && je->elapsed_sec) {
444                 perc = (double) je->elapsed_sec / (double) (je->elapsed_sec + je->eta_sec);
445                 eta_to_str(eta_str, je->eta_sec);
446         }
447
448         if (eta_new_line_pending) {
449                 eta_new_line_pending = 0;
450                 p += sprintf(p, "\n");
451         }
452
453         p += sprintf(p, "Jobs: %d (f=%d)", je->nr_running, je->files_open);
454         if (je->m_rate[0] || je->m_rate[1] || je->t_rate[0] || je->t_rate[1]) {
455                 char *tr, *mr;
456
457                 mr = num2str(je->m_rate[0] + je->m_rate[1], 4, 0, je->is_pow2, 8);
458                 tr = num2str(je->t_rate[0] + je->t_rate[1], 4, 0, je->is_pow2, 8);
459                 p += sprintf(p, ", CR=%s/%s KB/s", tr, mr);
460                 free(tr);
461                 free(mr);
462         } else if (je->m_iops[0] || je->m_iops[1] || je->t_iops[0] || je->t_iops[1]) {
463                 p += sprintf(p, ", CR=%d/%d IOPS",
464                                         je->t_iops[0] + je->t_iops[1],
465                                         je->m_iops[0] + je->m_iops[1]);
466         }
467         if (je->eta_sec != INT_MAX && je->nr_running) {
468                 char perc_str[32];
469                 char *iops_str[DDIR_RWDIR_CNT];
470                 char *rate_str[DDIR_RWDIR_CNT];
471                 size_t left;
472                 int l;
473                 int ddir;
474
475                 if ((!je->eta_sec && !eta_good) || je->nr_ramp == je->nr_running)
476                         strcpy(perc_str, "-.-% done");
477                 else {
478                         double mult = 100.0;
479
480                         if (je->nr_setting_up && je->nr_running)
481                                 mult *= (1.0 - (double) je->nr_setting_up / (double) je->nr_running);
482
483                         eta_good = 1;
484                         perc *= mult;
485                         sprintf(perc_str, "%3.1f%% done", perc);
486                 }
487
488                 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
489                         rate_str[ddir] = num2str(je->rate[ddir], 5,
490                                                 1024, je->is_pow2, je->unit_base);
491                         iops_str[ddir] = num2str(je->iops[ddir], 4, 1, 0, 0);
492                 }
493
494                 left = sizeof(output) - (p - output) - 1;
495
496                 l = snprintf(p, left, ": [%s] [%s] [%s/%s/%s /s] [%s/%s/%s iops] [eta %s]",
497                                 je->run_str, perc_str, rate_str[DDIR_READ],
498                                 rate_str[DDIR_WRITE], rate_str[DDIR_TRIM],
499                                 iops_str[DDIR_READ], iops_str[DDIR_WRITE],
500                                 iops_str[DDIR_TRIM], eta_str);
501                 p += l;
502                 if (l >= 0 && l < linelen_last)
503                         p += sprintf(p, "%*s", linelen_last - l, "");
504                 linelen_last = l;
505
506                 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
507                         free(rate_str[ddir]);
508                         free(iops_str[ddir]);
509                 }
510         }
511         p += sprintf(p, "\r");
512
513         printf("%s", output);
514
515         if (!eta_new_line_init) {
516                 fio_gettime(&disp_eta_new_line, NULL);
517                 eta_new_line_init = 1;
518         } else if (eta_new_line &&
519                    mtime_since_now(&disp_eta_new_line) > eta_new_line * 1000) {
520                 fio_gettime(&disp_eta_new_line, NULL);
521                 eta_new_line_pending = 1;
522         }
523
524         fflush(stdout);
525 }
526
527 void print_thread_status(void)
528 {
529         struct jobs_eta *je;
530         size_t size;
531
532         if (!thread_number)
533                 return;
534
535         size = sizeof(*je) + thread_number * sizeof(char) + 1;
536         je = malloc(size);
537         memset(je, 0, size);
538
539         if (calc_thread_status(je, 0))
540                 display_thread_status(je);
541
542         free(je);
543 }
544
545 void print_status_init(int thr_number)
546 {
547         run_str[thr_number] = 'P';
548 }