New tracing interface
[ghc.git] / rts / Stats.c
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 1998-2005
4 *
5 * Statistics and timing-related functions.
6 *
7 * ---------------------------------------------------------------------------*/
8
9 #include "Rts.h"
10 #include "RtsFlags.h"
11 #include "RtsUtils.h"
12 #include "MBlock.h"
13 #include "Schedule.h"
14 #include "Stats.h"
15 #include "ParTicky.h" /* ToDo: move into Rts.h */
16 #include "Profiling.h"
17 #include "Storage.h"
18 #include "GetTime.h"
19
20 /* huh? */
21 #define BIG_STRING_LEN 512
22
23 #define TICK_TO_DBL(t) ((double)(t) / TICKS_PER_SECOND)
24
25 static Ticks ElapsedTimeStart = 0;
26
27 static Ticks InitUserTime = 0;
28 static Ticks InitElapsedTime = 0;
29 static Ticks InitElapsedStamp = 0;
30
31 static Ticks MutUserTime = 0;
32 static Ticks MutElapsedTime = 0;
33 static Ticks MutElapsedStamp = 0;
34
35 static Ticks ExitUserTime = 0;
36 static Ticks ExitElapsedTime = 0;
37
38 static ullong GC_tot_alloc = 0;
39 static ullong GC_tot_copied = 0;
40 static ullong GC_tot_scavd_copied = 0;
41
42 static Ticks GC_start_time = 0, GC_tot_time = 0; /* User GC Time */
43 static Ticks GCe_start_time = 0, GCe_tot_time = 0; /* Elapsed GC time */
44
45 #ifdef PROFILING
46 static Ticks RP_start_time = 0, RP_tot_time = 0; /* retainer prof user time */
47 static Ticks RPe_start_time = 0, RPe_tot_time = 0; /* retainer prof elap time */
48
49 static Ticks HC_start_time, HC_tot_time = 0; // heap census prof user time
50 static Ticks HCe_start_time, HCe_tot_time = 0; // heap census prof elap time
51 #endif
52
53 #ifdef PROFILING
54 #define PROF_VAL(x) (x)
55 #else
56 #define PROF_VAL(x) 0
57 #endif
58
59 static lnat MaxResidency = 0; // in words; for stats only
60 static lnat AvgResidency = 0;
61 static lnat ResidencySamples = 0; // for stats only
62
63 static lnat GC_start_faults = 0, GC_end_faults = 0;
64
65 static Ticks *GC_coll_times;
66
67 static void statsPrintf( char *s, ... )
68 GNUC3_ATTRIBUTE(format (printf, 1, 2));
69
70 static void statsFlush( void );
71 static void statsClose( void );
72
73 Ticks stat_getElapsedGCTime(void)
74 {
75 return GCe_tot_time;
76 }
77
78 Ticks stat_getElapsedTime(void)
79 {
80 return getProcessElapsedTime() - ElapsedTimeStart;
81 }
82
83 /* mut_user_time_during_GC() and mut_user_time()
84 *
85 * The former function can be used to get the current mutator time
86 * *during* a GC, i.e. between stat_startGC and stat_endGC. This is
87 * used in the heap profiler for accurately time stamping the heap
88 * sample.
89 *
90 * ATTENTION: mut_user_time_during_GC() relies on GC_start_time being
91 * defined in stat_startGC() - to minimise system calls,
92 * GC_start_time is, however, only defined when really needed (check
93 * stat_startGC() for details)
94 */
95 double
96 mut_user_time_during_GC( void )
97 {
98 return TICK_TO_DBL(GC_start_time - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time));
99 }
100
101 double
102 mut_user_time( void )
103 {
104 Ticks user;
105 user = getProcessCPUTime();
106 return TICK_TO_DBL(user - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time));
107 }
108
109 #ifdef PROFILING
110 /*
111 mut_user_time_during_RP() is similar to mut_user_time_during_GC();
112 it returns the MUT time during retainer profiling.
113 The same is for mut_user_time_during_HC();
114 */
115 double
116 mut_user_time_during_RP( void )
117 {
118 return TICK_TO_DBL(RP_start_time - GC_tot_time - RP_tot_time - HC_tot_time);
119 }
120
121 double
122 mut_user_time_during_heap_census( void )
123 {
124 return TICK_TO_DBL(HC_start_time - GC_tot_time - RP_tot_time - HC_tot_time);
125 }
126 #endif /* PROFILING */
127
128 void
129 initStats(void)
130 {
131 nat i;
132
133 if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
134 statsPrintf(" Alloc Copied Live GC GC TOT TOT Page Flts\n");
135 statsPrintf(" bytes bytes bytes user elap user elap\n");
136 }
137 GC_coll_times =
138 (Ticks *)stgMallocBytes(
139 sizeof(Ticks)*RtsFlags.GcFlags.generations,
140 "initStats");
141 for (i = 0; i < RtsFlags.GcFlags.generations; i++) {
142 GC_coll_times[i] = 0;
143 }
144 }
145
146 /* -----------------------------------------------------------------------------
147 Initialisation time...
148 -------------------------------------------------------------------------- */
149
150 void
151 stat_startInit(void)
152 {
153 Ticks elapsed;
154
155 elapsed = getProcessElapsedTime();
156 ElapsedTimeStart = elapsed;
157 }
158
159 void
160 stat_endInit(void)
161 {
162 Ticks user, elapsed;
163
164 getProcessTimes(&user, &elapsed);
165
166 InitUserTime = user;
167 InitElapsedStamp = elapsed;
168 if (ElapsedTimeStart > elapsed) {
169 InitElapsedTime = 0;
170 } else {
171 InitElapsedTime = elapsed - ElapsedTimeStart;
172 }
173 }
174
175 /* -----------------------------------------------------------------------------
176 stat_startExit and stat_endExit
177
178 These two measure the time taken in shutdownHaskell().
179 -------------------------------------------------------------------------- */
180
181 void
182 stat_startExit(void)
183 {
184 Ticks user, elapsed;
185
186 getProcessTimes(&user, &elapsed);
187
188 MutElapsedStamp = elapsed;
189 MutElapsedTime = elapsed - GCe_tot_time -
190 PROF_VAL(RPe_tot_time + HCe_tot_time) - InitElapsedStamp;
191 if (MutElapsedTime < 0) { MutElapsedTime = 0; } /* sometimes -0.00 */
192
193 MutUserTime = user - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime;
194 if (MutUserTime < 0) { MutUserTime = 0; }
195 }
196
197 void
198 stat_endExit(void)
199 {
200 Ticks user, elapsed;
201
202 getProcessTimes(&user, &elapsed);
203
204 ExitUserTime = user - MutUserTime - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime;
205 ExitElapsedTime = elapsed - MutElapsedStamp;
206 if (ExitUserTime < 0) {
207 ExitUserTime = 0;
208 }
209 if (ExitElapsedTime < 0) {
210 ExitElapsedTime = 0;
211 }
212 }
213
214 /* -----------------------------------------------------------------------------
215 Called at the beginning of each GC
216 -------------------------------------------------------------------------- */
217
218 static nat rub_bell = 0;
219
220 /* initialise global variables needed during GC
221 *
222 * * GC_start_time is read in mut_user_time_during_GC(), which in turn is
223 * needed if either PROFILING or DEBUGing is enabled
224 */
225 void
226 stat_startGC(void)
227 {
228 nat bell = RtsFlags.GcFlags.ringBell;
229
230 if (bell) {
231 if (bell > 1) {
232 debugBelch(" GC ");
233 rub_bell = 1;
234 } else {
235 debugBelch("\007");
236 }
237 }
238
239 #if defined(PROFILING) || defined(DEBUG)
240 GC_start_time = getProcessCPUTime(); // needed in mut_user_time_during_GC()
241 #endif
242
243 if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
244 #if !defined(PROFILING) && !defined(DEBUG)
245 GC_start_time = getProcessCPUTime();
246 #endif
247 GCe_start_time = getProcessElapsedTime();
248 if (RtsFlags.GcFlags.giveStats) {
249 GC_start_faults = getPageFaults();
250 }
251 }
252 }
253
254 /* -----------------------------------------------------------------------------
255 Called at the end of each GC
256 -------------------------------------------------------------------------- */
257
258 void
259 stat_endGC (lnat alloc, lnat live, lnat copied,
260 lnat scavd_copied, lnat gen)
261 {
262 if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
263 Ticks time, etime, gc_time, gc_etime;
264
265 getProcessTimes(&time, &etime);
266 gc_time = time - GC_start_time;
267 gc_etime = etime - GCe_start_time;
268
269 if (RtsFlags.GcFlags.giveStats == VERBOSE_GC_STATS) {
270 nat faults = getPageFaults();
271
272 statsPrintf("%9ld %9ld %9ld",
273 alloc*sizeof(W_), (copied+scavd_copied)*sizeof(W_),
274 live*sizeof(W_));
275 statsPrintf(" %5.2f %5.2f %7.2f %7.2f %4ld %4ld (Gen: %2ld)\n",
276 TICK_TO_DBL(gc_time),
277 TICK_TO_DBL(gc_etime),
278 TICK_TO_DBL(time),
279 TICK_TO_DBL(etime - ElapsedTimeStart),
280 faults - GC_start_faults,
281 GC_start_faults - GC_end_faults,
282 gen);
283
284 GC_end_faults = faults;
285 statsFlush();
286 }
287
288 GC_coll_times[gen] += gc_time;
289
290 GC_tot_copied += (ullong) copied;
291 GC_tot_scavd_copied += (ullong) scavd_copied;
292 GC_tot_alloc += (ullong) alloc;
293 GC_tot_time += gc_time;
294 GCe_tot_time += gc_etime;
295
296 #if defined(THREADED_RTS)
297 {
298 Task *task;
299 if ((task = myTask()) != NULL) {
300 task->gc_time += gc_time;
301 task->gc_etime += gc_etime;
302 }
303 }
304 #endif
305
306 if (gen == RtsFlags.GcFlags.generations-1) { /* major GC? */
307 if (live > MaxResidency) {
308 MaxResidency = live;
309 }
310 ResidencySamples++;
311 AvgResidency += live;
312 }
313 }
314
315 if (rub_bell) {
316 debugBelch("\b\b\b \b\b\b");
317 rub_bell = 0;
318 }
319 }
320
321 /* -----------------------------------------------------------------------------
322 Called at the beginning of each Retainer Profiliing
323 -------------------------------------------------------------------------- */
324 #ifdef PROFILING
325 void
326 stat_startRP(void)
327 {
328 Ticks user, elapsed;
329 getProcessTimes( &user, &elapsed );
330
331 RP_start_time = user;
332 RPe_start_time = elapsed;
333 }
334 #endif /* PROFILING */
335
336 /* -----------------------------------------------------------------------------
337 Called at the end of each Retainer Profiliing
338 -------------------------------------------------------------------------- */
339
340 #ifdef PROFILING
341 void
342 stat_endRP(
343 nat retainerGeneration,
344 #ifdef DEBUG_RETAINER
345 nat maxCStackSize,
346 int maxStackSize,
347 #endif
348 double averageNumVisit)
349 {
350 Ticks user, elapsed;
351 getProcessTimes( &user, &elapsed );
352
353 RP_tot_time += user - RP_start_time;
354 RPe_tot_time += elapsed - RPe_start_time;
355
356 fprintf(prof_file, "Retainer Profiling: %d, at %f seconds\n",
357 retainerGeneration, mut_user_time_during_RP());
358 #ifdef DEBUG_RETAINER
359 fprintf(prof_file, "\tMax C stack size = %u\n", maxCStackSize);
360 fprintf(prof_file, "\tMax auxiliary stack size = %u\n", maxStackSize);
361 #endif
362 fprintf(prof_file, "\tAverage number of visits per object = %f\n", averageNumVisit);
363 }
364 #endif /* PROFILING */
365
366 /* -----------------------------------------------------------------------------
367 Called at the beginning of each heap census
368 -------------------------------------------------------------------------- */
369 #ifdef PROFILING
370 void
371 stat_startHeapCensus(void)
372 {
373 Ticks user, elapsed;
374 getProcessTimes( &user, &elapsed );
375
376 HC_start_time = user;
377 HCe_start_time = elapsed;
378 }
379 #endif /* PROFILING */
380
381 /* -----------------------------------------------------------------------------
382 Called at the end of each heap census
383 -------------------------------------------------------------------------- */
384 #ifdef PROFILING
385 void
386 stat_endHeapCensus(void)
387 {
388 Ticks user, elapsed;
389 getProcessTimes( &user, &elapsed );
390
391 HC_tot_time += user - HC_start_time;
392 HCe_tot_time += elapsed - HCe_start_time;
393 }
394 #endif /* PROFILING */
395
396 /* -----------------------------------------------------------------------------
397 Called at the end of execution
398
399 NOTE: number of allocations is not entirely accurate: it doesn't
400 take into account the few bytes at the end of the heap that
401 were left unused when the heap-check failed.
402 -------------------------------------------------------------------------- */
403
404 void
405 stat_exit(int alloc)
406 {
407 if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
408
409 char temp[BIG_STRING_LEN];
410 Ticks time;
411 Ticks etime;
412 nat g, total_collections = 0;
413
414 getProcessTimes( &time, &etime );
415 etime -= ElapsedTimeStart;
416
417 GC_tot_alloc += alloc;
418
419 /* Count total garbage collections */
420 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
421 total_collections += generations[g].collections;
422
423 /* avoid divide by zero if time is measured as 0.00 seconds -- SDM */
424 if (time == 0.0) time = 1;
425 if (etime == 0.0) etime = 1;
426
427 if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
428 statsPrintf("%9ld %9.9s %9.9s", (lnat)alloc*sizeof(W_), "", "");
429 statsPrintf(" %5.2f %5.2f\n\n", 0.0, 0.0);
430 }
431
432 if (RtsFlags.GcFlags.giveStats >= SUMMARY_GC_STATS) {
433 ullong_format_string(GC_tot_alloc*sizeof(W_),
434 temp, rtsTrue/*commas*/);
435 statsPrintf("%11s bytes allocated in the heap\n", temp);
436
437 ullong_format_string(GC_tot_copied*sizeof(W_),
438 temp, rtsTrue/*commas*/);
439 statsPrintf("%11s bytes copied during GC (scavenged)\n", temp);
440
441 ullong_format_string(GC_tot_scavd_copied*sizeof(W_),
442 temp, rtsTrue/*commas*/);
443 statsPrintf("%11s bytes copied during GC (not scavenged)\n", temp);
444
445 if ( ResidencySamples > 0 ) {
446 ullong_format_string(MaxResidency*sizeof(W_),
447 temp, rtsTrue/*commas*/);
448 statsPrintf("%11s bytes maximum residency (%ld sample(s))\n",
449 temp, ResidencySamples);
450 }
451 statsPrintf("\n");
452
453 /* Print garbage collections in each gen */
454 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
455 statsPrintf("%11d collections in generation %d (%6.2fs)\n",
456 generations[g].collections, g,
457 TICK_TO_DBL(GC_coll_times[g]));
458 }
459
460 statsPrintf("\n%11ld Mb total memory in use\n\n",
461 mblocks_allocated * MBLOCK_SIZE / (1024 * 1024));
462
463 #if defined(THREADED_RTS)
464 {
465 nat i;
466 Task *task;
467 for (i = 0, task = all_tasks;
468 task != NULL;
469 i++, task = task->all_link) {
470 statsPrintf(" Task %2d %-8s : MUT time: %6.2fs (%6.2fs elapsed)\n"
471 " GC time: %6.2fs (%6.2fs elapsed)\n\n",
472 i,
473 (task->tso == NULL) ? "(worker)" : "(bound)",
474 TICK_TO_DBL(task->mut_time),
475 TICK_TO_DBL(task->mut_etime),
476 TICK_TO_DBL(task->gc_time),
477 TICK_TO_DBL(task->gc_etime));
478 }
479 }
480 #endif
481
482 statsPrintf(" INIT time %6.2fs (%6.2fs elapsed)\n",
483 TICK_TO_DBL(InitUserTime), TICK_TO_DBL(InitElapsedTime));
484 statsPrintf(" MUT time %6.2fs (%6.2fs elapsed)\n",
485 TICK_TO_DBL(MutUserTime), TICK_TO_DBL(MutElapsedTime));
486 statsPrintf(" GC time %6.2fs (%6.2fs elapsed)\n",
487 TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time));
488 #ifdef PROFILING
489 statsPrintf(" RP time %6.2fs (%6.2fs elapsed)\n",
490 TICK_TO_DBL(RP_tot_time), TICK_TO_DBL(RPe_tot_time));
491 statsPrintf(" PROF time %6.2fs (%6.2fs elapsed)\n",
492 TICK_TO_DBL(HC_tot_time), TICK_TO_DBL(HCe_tot_time));
493 #endif
494 statsPrintf(" EXIT time %6.2fs (%6.2fs elapsed)\n",
495 TICK_TO_DBL(ExitUserTime), TICK_TO_DBL(ExitElapsedTime));
496 statsPrintf(" Total time %6.2fs (%6.2fs elapsed)\n\n",
497 TICK_TO_DBL(time), TICK_TO_DBL(etime));
498 statsPrintf(" %%GC time %5.1f%% (%.1f%% elapsed)\n\n",
499 TICK_TO_DBL(GC_tot_time)*100/TICK_TO_DBL(time),
500 TICK_TO_DBL(GCe_tot_time)*100/TICK_TO_DBL(etime));
501
502 if (time - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time) == 0)
503 ullong_format_string(0, temp, rtsTrue/*commas*/);
504 else
505 ullong_format_string(
506 (ullong)((GC_tot_alloc*sizeof(W_))/
507 TICK_TO_DBL(time - GC_tot_time -
508 PROF_VAL(RP_tot_time + HC_tot_time))),
509 temp, rtsTrue/*commas*/);
510
511 statsPrintf(" Alloc rate %s bytes per MUT second\n\n", temp);
512
513 statsPrintf(" Productivity %5.1f%% of total user, %.1f%% of total elapsed\n\n",
514 TICK_TO_DBL(time - GC_tot_time -
515 PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime) * 100
516 / TICK_TO_DBL(time),
517 TICK_TO_DBL(time - GC_tot_time -
518 PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime) * 100
519 / TICK_TO_DBL(etime));
520 }
521
522 if (RtsFlags.GcFlags.giveStats == ONELINE_GC_STATS) {
523 /* print the long long separately to avoid bugginess on mingwin (2001-07-02, mingw-0.5) */
524 statsPrintf("<<ghc: %llu bytes, ", GC_tot_alloc*(ullong)sizeof(W_));
525 statsPrintf("%d GCs, %ld/%ld avg/max bytes residency (%ld samples), %luM in use, %.2f INIT (%.2f elapsed), %.2f MUT (%.2f elapsed), %.2f GC (%.2f elapsed) :ghc>>\n",
526 total_collections,
527 ResidencySamples == 0 ? 0 :
528 AvgResidency*sizeof(W_)/ResidencySamples,
529 MaxResidency*sizeof(W_),
530 ResidencySamples,
531 (unsigned long)(mblocks_allocated * MBLOCK_SIZE / (1024L * 1024L)),
532 TICK_TO_DBL(InitUserTime), TICK_TO_DBL(InitElapsedTime),
533 TICK_TO_DBL(MutUserTime), TICK_TO_DBL(MutElapsedTime),
534 TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time));
535 }
536
537 statsFlush();
538 statsClose();
539 }
540 }
541
542 /* -----------------------------------------------------------------------------
543 stat_describe_gens
544
545 Produce some detailed info on the state of the generational GC.
546 -------------------------------------------------------------------------- */
547 #ifdef DEBUG
548 void
549 statDescribeGens(void)
550 {
551 nat g, s, mut, lge;
552 lnat live;
553 bdescr *bd;
554 step *step;
555
556 debugBelch(
557 " Gen Steps Max Mutable Step Blocks Live Large\n"
558 " Blocks Closures Objects\n");
559
560 mut = 0;
561 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
562 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
563 mut += bd->free - bd->start;
564 }
565
566 debugBelch("%8d %8d %8d %9d", g, generations[g].n_steps,
567 generations[g].max_blocks, mut);
568
569 for (s = 0; s < generations[g].n_steps; s++) {
570 step = &generations[g].steps[s];
571 live = 0;
572 for (bd = step->large_objects, lge = 0; bd; bd = bd->link) {
573 lge++;
574 }
575 live = step->n_large_blocks * BLOCK_SIZE;
576 bd = step->blocks;
577 // This live figure will be slightly less that the "live" figure
578 // given by +RTS -Sstderr, because we take don't count the
579 // slop at the end of each block.
580 for (; bd; bd = bd->link) {
581 live += (bd->free - bd->start) * sizeof(W_);
582 }
583 if (s != 0) {
584 debugBelch("%36s","");
585 }
586 debugBelch("%6d %8d %8ld %8d\n", s, step->n_blocks,
587 live, lge);
588 }
589 }
590 debugBelch("\n");
591 }
592 #endif
593
594 /* -----------------------------------------------------------------------------
595 Stats available via a programmatic interface, so eg. GHCi can time
596 each compilation and expression evaluation.
597 -------------------------------------------------------------------------- */
598
599 extern HsInt64 getAllocations( void )
600 { return (HsInt64)total_allocated * sizeof(W_); }
601
602 /* -----------------------------------------------------------------------------
603 Dumping stuff in the stats file, or via the debug message interface
604 -------------------------------------------------------------------------- */
605
606 static void
607 statsPrintf( char *s, ... )
608 {
609 FILE *sf = RtsFlags.GcFlags.statsFile;
610 va_list ap;
611
612 va_start(ap,s);
613 if (sf == NULL) {
614 vdebugBelch(s,ap);
615 } else {
616 vfprintf(sf, s, ap);
617 }
618 va_end(ap);
619 }
620
621 static void
622 statsFlush( void )
623 {
624 FILE *sf = RtsFlags.GcFlags.statsFile;
625 if (sf != NULL) {
626 fflush(sf);
627 }
628 }
629
630 static void
631 statsClose( void )
632 {
633 FILE *sf = RtsFlags.GcFlags.statsFile;
634 if (sf != NULL) {
635 fclose(sf);
636 }
637 }