net: 'nodelay' HOWTO/man page update
[fio.git] / eta.c
... / ...
CommitLineData
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
10static char run_str[REAL_MAX_JOBS + 1];
11
12/*
13 * Sets the status of the 'td' in the printed status map.
14 */
15static 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 */
97static 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 */
125static 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 /*
143 * if writing and verifying afterwards, bytes_total will be twice the
144 * size. In a mixed workload, verify phase will be the size of the
145 * first stage writes.
146 */
147 if (td->o.do_verify && td->o.verify && td_write(td)) {
148 if (td_rw(td)) {
149 unsigned int perc = 50;
150
151 if (td->o.rwmix[DDIR_WRITE])
152 perc = td->o.rwmix[DDIR_WRITE];
153
154 bytes_total += (bytes_total * perc) / 100;
155 } else
156 bytes_total <<= 1;
157 }
158
159 if (td->o.zone_size && td->o.zone_skip)
160 bytes_total /= (td->o.zone_skip / td->o.zone_size);
161
162 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
163 double perc, perc_t;
164
165 bytes_done = ddir_rw_sum(td->io_bytes);
166 perc = (double) bytes_done / (double) bytes_total;
167 if (perc > 1.0)
168 perc = 1.0;
169
170 if (td->o.time_based) {
171 perc_t = (double) elapsed / (double) td->o.timeout;
172 if (perc_t < perc)
173 perc = perc_t;
174 }
175
176 eta_sec = (unsigned long) (elapsed * (1.0 / perc)) - elapsed;
177
178 if (td->o.timeout &&
179 eta_sec > (td->o.timeout + done_secs - elapsed))
180 eta_sec = td->o.timeout + done_secs - elapsed;
181 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
182 || td->runstate == TD_INITIALIZED
183 || td->runstate == TD_RAMP
184 || td->runstate == TD_PRE_READING) {
185 int t_eta = 0, r_eta = 0;
186 unsigned long long rate_bytes;
187
188 /*
189 * We can only guess - assume it'll run the full timeout
190 * if given, otherwise assume it'll run at the specified rate.
191 */
192 if (td->o.timeout) {
193 t_eta = td->o.timeout + td->o.start_delay +
194 td->o.ramp_time;
195
196 if (in_ramp_time(td)) {
197 unsigned long ramp_left;
198
199 ramp_left = mtime_since_now(&td->epoch);
200 ramp_left = (ramp_left + 999) / 1000;
201 if (ramp_left <= t_eta)
202 t_eta -= ramp_left;
203 }
204 }
205 rate_bytes = ddir_rw_sum(td->o.rate);
206 if (rate_bytes) {
207 r_eta = (bytes_total / 1024) / rate_bytes;
208 r_eta += td->o.start_delay;
209 }
210
211 if (r_eta && t_eta)
212 eta_sec = min(r_eta, t_eta);
213 else if (r_eta)
214 eta_sec = r_eta;
215 else if (t_eta)
216 eta_sec = t_eta;
217 else
218 eta_sec = 0;
219 } else {
220 /*
221 * thread is already done or waiting for fsync
222 */
223 eta_sec = 0;
224 }
225
226 return eta_sec;
227}
228
229static void calc_rate(int unified_rw_rep, unsigned long mtime,
230 unsigned long long *io_bytes,
231 unsigned long long *prev_io_bytes, unsigned int *rate)
232{
233 int i;
234
235 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
236 unsigned long long diff;
237
238 diff = io_bytes[i] - prev_io_bytes[i];
239 if (unified_rw_rep) {
240 rate[i] = 0;
241 rate[0] += ((1000 * diff) / mtime) / 1024;
242 } else
243 rate[i] = ((1000 * diff) / mtime) / 1024;
244
245 prev_io_bytes[i] = io_bytes[i];
246 }
247}
248
249static void calc_iops(int unified_rw_rep, unsigned long mtime,
250 unsigned long long *io_iops,
251 unsigned long long *prev_io_iops, unsigned int *iops)
252{
253 int i;
254
255 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
256 unsigned long long diff;
257
258 diff = io_iops[i] - prev_io_iops[i];
259 if (unified_rw_rep) {
260 iops[i] = 0;
261 iops[0] += (diff * 1000) / mtime;
262 } else
263 iops[i] = (diff * 1000) / mtime;
264
265 prev_io_iops[i] = io_iops[i];
266 }
267}
268
269/*
270 * Print status of the jobs we know about. This includes rate estimates,
271 * ETA, thread state, etc.
272 */
273int calc_thread_status(struct jobs_eta *je, int force)
274{
275 struct thread_data *td;
276 int i, unified_rw_rep;
277 unsigned long rate_time, disp_time, bw_avg_time, *eta_secs;
278 unsigned long long io_bytes[DDIR_RWDIR_CNT];
279 unsigned long long io_iops[DDIR_RWDIR_CNT];
280 struct timeval now;
281
282 static unsigned long long rate_io_bytes[DDIR_RWDIR_CNT];
283 static unsigned long long disp_io_bytes[DDIR_RWDIR_CNT];
284 static unsigned long long disp_io_iops[DDIR_RWDIR_CNT];
285 static struct timeval rate_prev_time, disp_prev_time;
286
287 if (!force) {
288 if (output_format != FIO_OUTPUT_NORMAL)
289 return 0;
290 if (temp_stall_ts || eta_print == FIO_ETA_NEVER)
291 return 0;
292
293 if (!isatty(STDOUT_FILENO) && (eta_print != FIO_ETA_ALWAYS))
294 return 0;
295 }
296
297 if (!ddir_rw_sum(rate_io_bytes))
298 fill_start_time(&rate_prev_time);
299 if (!ddir_rw_sum(disp_io_bytes))
300 fill_start_time(&disp_prev_time);
301
302 eta_secs = malloc(thread_number * sizeof(unsigned long));
303 memset(eta_secs, 0, thread_number * sizeof(unsigned long));
304
305 je->elapsed_sec = (mtime_since_genesis() + 999) / 1000;
306
307 io_bytes[DDIR_READ] = io_bytes[DDIR_WRITE] = io_bytes[DDIR_TRIM] = 0;
308 io_iops[DDIR_READ] = io_iops[DDIR_WRITE] = io_iops[DDIR_TRIM] = 0;
309 bw_avg_time = ULONG_MAX;
310 unified_rw_rep = 0;
311 for_each_td(td, i) {
312 unified_rw_rep += td->o.unified_rw_rep;
313 if (is_power_of_2(td->o.kb_base))
314 je->is_pow2 = 1;
315 if (td->o.bw_avg_time < bw_avg_time)
316 bw_avg_time = td->o.bw_avg_time;
317 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING
318 || td->runstate == TD_FSYNCING
319 || td->runstate == TD_PRE_READING) {
320 je->nr_running++;
321 if (td_read(td)) {
322 je->t_rate += td->o.rate[DDIR_READ];
323 je->t_iops += td->o.rate_iops[DDIR_READ];
324 je->m_rate += td->o.ratemin[DDIR_READ];
325 je->m_iops += td->o.rate_iops_min[DDIR_READ];
326 }
327 if (td_write(td)) {
328 je->t_rate += td->o.rate[DDIR_WRITE];
329 je->t_iops += td->o.rate_iops[DDIR_WRITE];
330 je->m_rate += td->o.ratemin[DDIR_WRITE];
331 je->m_iops += td->o.rate_iops_min[DDIR_WRITE];
332 }
333 if (td_trim(td)) {
334 je->t_rate += td->o.rate[DDIR_TRIM];
335 je->t_iops += td->o.rate_iops[DDIR_TRIM];
336 je->m_rate += td->o.ratemin[DDIR_TRIM];
337 je->m_iops += td->o.rate_iops_min[DDIR_TRIM];
338 }
339
340 je->files_open += td->nr_open_files;
341 } else if (td->runstate == TD_RAMP) {
342 je->nr_running++;
343 je->nr_ramp++;
344 } else if (td->runstate == TD_SETTING_UP)
345 je->nr_running++;
346 else if (td->runstate < TD_RUNNING)
347 je->nr_pending++;
348
349 if (je->elapsed_sec >= 3)
350 eta_secs[i] = thread_eta(td);
351 else
352 eta_secs[i] = INT_MAX;
353
354 check_str_update(td);
355
356 if (td->runstate > TD_RAMP) {
357 int ddir;
358
359 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
360 if (unified_rw_rep) {
361 io_bytes[0] += td->io_bytes[ddir];
362 io_iops[0] += td->io_blocks[ddir];
363 } else {
364 io_bytes[ddir] += td->io_bytes[ddir];
365 io_iops[ddir] += td->io_blocks[ddir];
366 }
367 }
368 }
369 }
370
371 if (exitall_on_terminate)
372 je->eta_sec = INT_MAX;
373 else
374 je->eta_sec = 0;
375
376 for_each_td(td, i) {
377 if (exitall_on_terminate) {
378 if (eta_secs[i] < je->eta_sec)
379 je->eta_sec = eta_secs[i];
380 } else {
381 if (eta_secs[i] > je->eta_sec)
382 je->eta_sec = eta_secs[i];
383 }
384 }
385
386 free(eta_secs);
387
388 fio_gettime(&now, NULL);
389 rate_time = mtime_since(&rate_prev_time, &now);
390
391 if (write_bw_log && rate_time > bw_avg_time && !in_ramp_time(td)) {
392 calc_rate(unified_rw_rep, rate_time, io_bytes, rate_io_bytes,
393 je->rate);
394 memcpy(&rate_prev_time, &now, sizeof(now));
395 add_agg_sample(je->rate[DDIR_READ], DDIR_READ, 0);
396 add_agg_sample(je->rate[DDIR_WRITE], DDIR_WRITE, 0);
397 add_agg_sample(je->rate[DDIR_TRIM], DDIR_TRIM, 0);
398 }
399
400 disp_time = mtime_since(&disp_prev_time, &now);
401
402 /*
403 * Allow a little slack, the target is to print it every 1000 msecs
404 */
405 if (!force && disp_time < 900)
406 return 0;
407
408 calc_rate(unified_rw_rep, disp_time, io_bytes, disp_io_bytes, je->rate);
409 calc_iops(unified_rw_rep, disp_time, io_iops, disp_io_iops, je->iops);
410
411 memcpy(&disp_prev_time, &now, sizeof(now));
412
413 if (!force && !je->nr_running && !je->nr_pending)
414 return 0;
415
416 je->nr_threads = thread_number;
417 memcpy(je->run_str, run_str, thread_number * sizeof(char));
418
419 return 1;
420}
421
422void display_thread_status(struct jobs_eta *je)
423{
424 static int linelen_last;
425 static int eta_good;
426 char output[REAL_MAX_JOBS + 512], *p = output;
427 char eta_str[128];
428 double perc = 0.0;
429
430 if (je->eta_sec != INT_MAX && je->elapsed_sec) {
431 perc = (double) je->elapsed_sec / (double) (je->elapsed_sec + je->eta_sec);
432 eta_to_str(eta_str, je->eta_sec);
433 }
434
435 p += sprintf(p, "Jobs: %d (f=%d)", je->nr_running, je->files_open);
436 if (je->m_rate || je->t_rate) {
437 char *tr, *mr;
438
439 mr = num2str(je->m_rate, 4, 0, je->is_pow2);
440 tr = num2str(je->t_rate, 4, 0, je->is_pow2);
441 p += sprintf(p, ", CR=%s/%s KB/s", tr, mr);
442 free(tr);
443 free(mr);
444 } else if (je->m_iops || je->t_iops)
445 p += sprintf(p, ", CR=%d/%d IOPS", je->t_iops, je->m_iops);
446 if (je->eta_sec != INT_MAX && je->nr_running) {
447 char perc_str[32];
448 char *iops_str[DDIR_RWDIR_CNT];
449 char *rate_str[DDIR_RWDIR_CNT];
450 size_t left;
451 int l;
452 int ddir;
453
454 if ((!je->eta_sec && !eta_good) || je->nr_ramp == je->nr_running)
455 strcpy(perc_str, "-.-% done");
456 else {
457 eta_good = 1;
458 perc *= 100.0;
459 sprintf(perc_str, "%3.1f%% done", perc);
460 }
461
462 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
463 rate_str[ddir] = num2str(je->rate[ddir], 5,
464 1024, je->is_pow2);
465 iops_str[ddir] = num2str(je->iops[ddir], 4, 1, 0);
466 }
467
468 left = sizeof(output) - (p - output) - 1;
469
470 l = snprintf(p, left, ": [%s] [%s] [%s/%s/%s /s] [%s/%s/%s iops] [eta %s]",
471 je->run_str, perc_str, rate_str[DDIR_READ],
472 rate_str[DDIR_WRITE], rate_str[DDIR_TRIM],
473 iops_str[DDIR_READ], iops_str[DDIR_WRITE],
474 iops_str[DDIR_TRIM], eta_str);
475 p += l;
476 if (l >= 0 && l < linelen_last)
477 p += sprintf(p, "%*s", linelen_last - l, "");
478 linelen_last = l;
479
480 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
481 free(rate_str[ddir]);
482 free(iops_str[ddir]);
483 }
484 }
485 p += sprintf(p, "\r");
486
487 printf("%s", output);
488 fflush(stdout);
489}
490
491void print_thread_status(void)
492{
493 struct jobs_eta *je;
494 size_t size;
495
496 if (!thread_number)
497 return;
498
499 size = sizeof(*je) + thread_number * sizeof(char) + 1;
500 je = malloc(size);
501 memset(je, 0, size);
502
503 if (calc_thread_status(je, 0))
504 display_thread_status(je);
505
506 free(je);
507}
508
509void print_status_init(int thr_number)
510{
511 run_str[thr_number] = 'P';
512}