1 /* -----------------------------------------------------------------------------
2 * $Id: Stats.c,v 1.23 2000/12/19 14:30:17 simonmar Exp $
4 * (c) The GHC Team, 1998-1999
6 * Statistics and timing-related functions.
8 * ---------------------------------------------------------------------------*/
10 #define NON_POSIX_SOURCE
15 #include "StoragePriv.h"
24 #ifndef mingw32_TARGET_OS
25 # ifdef HAVE_SYS_TIMES_H
26 # include <sys/times.h>
30 #ifdef HAVE_SYS_TIME_H
40 #if ! irix_TARGET_OS && ! defined(mingw32_TARGET_OS)
41 # if defined(HAVE_SYS_RESOURCE_H)
42 # include <sys/resource.h>
46 #ifdef HAVE_SYS_TIMEB_H
47 #include <sys/timeb.h>
59 #define BIG_STRING_LEN 512
61 /* We're not trying to be terribly accurate here, using the
62 * basic times() function to get a resolution of about 100ths of a
63 * second, depending on the OS. A long int will do fine for holding
66 #define TICK_TYPE long int
67 #define TICK_TO_DBL(t) ((double)(t) / TicksPerSecond)
69 static int TicksPerSecond = 0;
71 static TICK_TYPE ElapsedTimeStart = 0;
72 static TICK_TYPE CurrentElapsedTime = 0;
73 static TICK_TYPE CurrentUserTime = 0;
75 static TICK_TYPE InitUserTime = 0;
76 static TICK_TYPE InitElapsedTime = 0;
77 static TICK_TYPE InitElapsedStamp = 0;
79 static TICK_TYPE MutUserTime = 0;
80 static TICK_TYPE MutElapsedTime = 0;
81 static TICK_TYPE MutElapsedStamp = 0;
83 static TICK_TYPE ExitUserTime = 0;
84 static TICK_TYPE ExitElapsedTime = 0;
86 static ullong GC_tot_alloc = 0;
87 static ullong GC_tot_copied = 0;
89 static TICK_TYPE GC_start_time, GC_tot_time = 0; /* User GC Time */
90 static TICK_TYPE GCe_start_time, GCe_tot_time = 0; /* Elapsed GC time */
92 lnat MaxResidency = 0; /* in words; for stats only */
93 lnat AvgResidency = 0;
94 lnat ResidencySamples = 0; /* for stats only */
96 static lnat GC_start_faults = 0, GC_end_faults = 0;
98 static TICK_TYPE *GC_coll_times;
100 static void getTimes(void);
101 static nat pageFaults(void);
103 /* elapsedtime() -- The current elapsed time in seconds */
106 #define NS_PER_SEC 10000000LL
107 /* Convert FILETIMEs into secs since the Epoch (Jan1-1970) */
108 #define FT2longlong(ll,ft) \
109 (ll)=(ft).dwHighDateTime; \
111 (ll) |= (ft).dwLowDateTime; \
112 (ll) /= (unsigned long long) (NS_PER_SEC / CLOCKS_PER_SEC)
116 /* cygwin32 or mingw32 version */
120 FILETIME creationTime, exitTime, kernelTime, userTime;
121 long long int kT, uT;
123 /* ToDo: pin down elapsed times to just the OS thread(s) that
124 are evaluating/managing Haskell code.
126 if (!GetProcessTimes (GetCurrentProcess(), &creationTime,
127 &exitTime, &kernelTime, &userTime)) {
128 /* Probably on a Win95 box..*/
132 FT2longlong(kT,kernelTime);
133 FT2longlong(uT,userTime);
134 CurrentElapsedTime = uT + kT;
135 CurrentUserTime = uT;
144 # if !defined(HAVE_TIMES)
145 /* We will #ifdef around the fprintf for machines
146 we *know* are unsupported. (WDP 94/05)
148 fprintf(stderr, "NOTE: `getTimes' does nothing!\n");
151 # else /* not stumped */
153 clock_t r = times(&t);
155 CurrentElapsedTime = r;
156 CurrentUserTime = t.tms_utime;
162 /* mut_user_time_during_GC() and mut_user_time()
164 * The former function can be used to get the current mutator time
165 * *during* a GC, i.e. between stat_startGC and stat_endGC. This is
166 * used in the heap profiler for accurately time stamping the heap
169 * ATTENTION: mut_user_time_during_GC() relies on GC_start_time being
170 * defined in stat_startGC() - to minimise system calls,
171 * GC_start_time is, however, only defined when really needed (check
172 * stat_startGC() for details)
175 mut_user_time_during_GC(void)
177 return ((double)GC_start_time - (double)GC_tot_time);
184 return ((double)CurrentUserTime - (double)GC_tot_time);
190 /* ToDo (on NT): better, get this via the performance data
191 that's stored in the registry. */
192 # if !defined(HAVE_GETRUSAGE) || irix_TARGET_OS || defined(_WIN32)
197 getrusage(RUSAGE_SELF, &t);
206 FILE *sf = RtsFlags.GcFlags.statsFile;
208 if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
209 fprintf(sf, " Alloc Collect Live GC GC TOT TOT Page Flts\n");
210 fprintf(sf, " bytes bytes bytes user elap user elap\n");
213 (TICK_TYPE *)stgMallocBytes(
214 sizeof(TICK_TYPE)*RtsFlags.GcFlags.generations,
216 for (i = 0; i < RtsFlags.GcFlags.generations; i++) {
217 GC_coll_times[i] = 0;
221 /* -----------------------------------------------------------------------------
222 Initialisation time...
223 -------------------------------------------------------------------------- */
228 /* Determine TicksPerSecond ... */
229 #if defined(CLK_TCK) /* defined by POSIX */
230 TicksPerSecond = CLK_TCK;
232 #elif defined(HAVE_SYSCONF)
235 ticks = sysconf(_SC_CLK_TCK);
237 fprintf(stderr, "stat_init: bad call to 'sysconf'!\n");
238 stg_exit(EXIT_FAILURE);
240 TicksPerSecond = (double) ticks;
242 /* no "sysconf" or CLK_TCK; had better guess */
244 TicksPerSecond = (StgDouble) (HZ);
246 #elif defined(CLOCKS_PER_SEC)
247 TicksPerSecond = (StgDouble) (CLOCKS_PER_SEC);
248 #else /* had better guess wildly */
249 /* We will #ifdef around the fprintf for machines
250 we *know* are unsupported. (WDP 94/05)
252 fprintf(stderr, "NOTE: Guessing `TicksPerSecond = 60'!\n");
253 TicksPerSecond = 60.0;
257 ElapsedTimeStart = CurrentElapsedTime;
264 InitUserTime = CurrentUserTime;
265 InitElapsedStamp = CurrentElapsedTime;
266 if (ElapsedTimeStart > CurrentElapsedTime) {
269 InitElapsedTime = CurrentElapsedTime - ElapsedTimeStart;
273 /* -----------------------------------------------------------------------------
274 stat_startExit and stat_endExit
276 These two measure the time taken in shutdownHaskell().
277 -------------------------------------------------------------------------- */
283 MutElapsedStamp = CurrentElapsedTime;
284 MutElapsedTime = CurrentElapsedTime - GCe_tot_time - InitElapsedStamp;
285 if (MutElapsedTime < 0) { MutElapsedTime = 0; } /* sometimes -0.00 */
287 /* for SMP, we don't know the mutator time yet, we have to inspect
288 * all the running threads to find out, and they haven't stopped
289 * yet. So we just timestamp MutUserTime at this point so we can
290 * calculate the EXIT time. The real MutUserTime is calculated
291 * in stat_exit below.
294 MutUserTime = CurrentUserTime;
296 MutUserTime = CurrentUserTime - GC_tot_time - InitUserTime;
297 if (MutUserTime < 0) { MutUserTime = 0; }
306 ExitUserTime = CurrentUserTime - MutUserTime;
308 ExitUserTime = CurrentUserTime - MutUserTime - GC_tot_time - InitUserTime;
310 ExitElapsedTime = CurrentElapsedTime - MutElapsedStamp;
311 if (ExitUserTime < 0) {
314 if (ExitElapsedTime < 0) {
319 /* -----------------------------------------------------------------------------
320 Called at the beginning of each GC
321 -------------------------------------------------------------------------- */
323 static nat rub_bell = 0;
325 /* initialise global variables needed during GC
327 * * GC_start_time is read in mut_user_time_during_GC(), which in turn is
328 * needed if either PROFILING or DEBUGing is enabled
333 nat bell = RtsFlags.GcFlags.ringBell;
337 fprintf(stderr, " GC ");
340 fprintf(stderr, "\007");
344 #if defined(PROFILING) || defined(DEBUG)
346 GC_start_time = CurrentUserTime; /* needed in mut_user_time_during_GC() */
349 if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
350 #if !defined(PROFILING) && !defined(DEBUG)
352 GC_start_time = CurrentUserTime;
354 GCe_start_time = CurrentElapsedTime;
355 if (RtsFlags.GcFlags.giveStats) {
356 GC_start_faults = pageFaults();
361 /* -----------------------------------------------------------------------------
362 Called at the end of each GC
363 -------------------------------------------------------------------------- */
366 stat_endGC(lnat alloc, lnat collect, lnat live, lnat copied, lnat gen)
368 FILE *sf = RtsFlags.GcFlags.statsFile;
370 if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
371 TICK_TYPE time, etime, gc_time, gc_etime;
374 time = CurrentUserTime;
375 etime = CurrentElapsedTime;
376 gc_time = time - GC_start_time;
377 gc_etime = etime - GCe_start_time;
379 if (RtsFlags.GcFlags.giveStats == VERBOSE_GC_STATS && sf != NULL) {
380 nat faults = pageFaults();
382 fprintf(sf, "%9ld %9ld %9ld",
383 alloc*sizeof(W_), collect*sizeof(W_), live*sizeof(W_));
384 fprintf(sf, " %5.2f %5.2f %7.2f %7.2f %4ld %4ld (Gen: %2ld)\n",
385 TICK_TO_DBL(gc_time),
386 TICK_TO_DBL(gc_etime),
388 TICK_TO_DBL(etime - ElapsedTimeStart),
389 faults - GC_start_faults,
390 GC_start_faults - GC_end_faults,
393 GC_end_faults = faults;
397 GC_coll_times[gen] += gc_time;
399 GC_tot_copied += (ullong) copied;
400 GC_tot_alloc += (ullong) alloc;
401 GC_tot_time += gc_time;
402 GCe_tot_time += gc_etime;
407 pthread_t me = pthread_self();
409 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
410 if (me == task_ids[i].id) {
411 task_ids[i].gc_time += gc_time;
412 task_ids[i].gc_etime += gc_etime;
419 if (gen == RtsFlags.GcFlags.generations-1) { /* major GC? */
420 if (live > MaxResidency) {
424 AvgResidency += live;
429 fprintf(stderr, "\b\b\b \b\b\b");
434 /* -----------------------------------------------------------------------------
437 Called under SMP when a worker thread finishes. We drop the timing
438 stats for this thread into the task_ids struct for that thread.
439 -------------------------------------------------------------------------- */
443 stat_workerStop(void)
446 pthread_t me = pthread_self();
448 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
449 if (task_ids[i].id == me) {
450 task_ids[i].mut_time = usertime() - task_ids[i].gc_time;
451 task_ids[i].mut_etime = elapsedtime()
453 - task_ids[i].elapsedtimestart;
454 if (task_ids[i].mut_time < 0.0) { task_ids[i].mut_time = 0.0; }
455 if (task_ids[i].mut_etime < 0.0) { task_ids[i].mut_etime = 0.0; }
461 /* -----------------------------------------------------------------------------
462 Called at the end of execution
464 NOTE: number of allocations is not entirely accurate: it doesn't
465 take into account the few bytes at the end of the heap that
466 were left unused when the heap-check failed.
467 -------------------------------------------------------------------------- */
472 FILE *sf = RtsFlags.GcFlags.statsFile;
474 if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
476 char temp[BIG_STRING_LEN];
479 nat g, total_collections = 0;
482 time = CurrentUserTime;
483 etime = CurrentElapsedTime - ElapsedTimeStart;
485 GC_tot_alloc += alloc;
487 /* avoid divide by zero if time is measured as 0.00 seconds -- SDM */
488 if (time == 0.0) time = 1;
489 if (etime == 0.0) etime = 1;
491 /* Count total garbage collections */
492 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
493 total_collections += generations[g].collections;
495 /* For SMP, we have to get the user time from each thread
496 * and try to work out the total time.
501 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
502 MutUserTime += task_ids[i].mut_time;
505 time = MutUserTime + GC_tot_time + InitUserTime + ExitUserTime;
506 if (MutUserTime < 0) { MutUserTime = 0; }
509 if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS && sf != NULL) {
510 fprintf(sf, "%9ld %9.9s %9.9s", (lnat)alloc*sizeof(W_), "", "");
511 fprintf(sf, " %5.2f %5.2f\n\n", 0.0, 0.0);
514 if (RtsFlags.GcFlags.giveStats >= SUMMARY_GC_STATS && sf != NULL) {
515 ullong_format_string(GC_tot_alloc*sizeof(W_),
516 temp, rtsTrue/*commas*/);
517 fprintf(sf, "%11s bytes allocated in the heap\n", temp);
519 ullong_format_string(GC_tot_copied*sizeof(W_),
520 temp, rtsTrue/*commas*/);
521 fprintf(sf, "%11s bytes copied during GC\n", temp);
523 if ( ResidencySamples > 0 ) {
524 ullong_format_string(MaxResidency*sizeof(W_),
525 temp, rtsTrue/*commas*/);
526 fprintf(sf, "%11s bytes maximum residency (%ld sample(s))\n",
527 temp, ResidencySamples);
531 /* Print garbage collections in each gen */
532 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
533 fprintf(sf, "%11d collections in generation %d (%6.2fs)\n",
534 generations[g].collections, g,
535 TICK_TO_DBL(GC_coll_times[g]));
538 fprintf(sf,"\n%11ld Mb total memory in use\n\n",
539 mblocks_allocated * MBLOCK_SIZE / (1024 * 1024));
544 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
545 fprintf(sf, " Task %2d: MUT time: %6.2fs (%6.2fs elapsed)\n"
546 " GC time: %6.2fs (%6.2fs elapsed)\n\n",
548 TICK_TO_DBL(task_ids[i].mut_time),
549 TICK_TO_DBL(task_ids[i].mut_etime),
550 TICK_TO_DBL(task_ids[i].gc_time),
551 TICK_TO_DBL(task_ids[i].gc_etime));
556 fprintf(sf, " INIT time %6.2fs (%6.2fs elapsed)\n",
557 TICK_TO_DBL(InitUserTime), TICK_TO_DBL(InitElapsedTime));
558 fprintf(sf, " MUT time %6.2fs (%6.2fs elapsed)\n",
559 TICK_TO_DBL(MutUserTime), TICK_TO_DBL(MutElapsedTime));
560 fprintf(sf, " GC time %6.2fs (%6.2fs elapsed)\n",
561 TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time));
562 fprintf(sf, " EXIT time %6.2fs (%6.2fs elapsed)\n",
563 TICK_TO_DBL(ExitUserTime), TICK_TO_DBL(ExitElapsedTime));
564 fprintf(sf, " Total time %6.2fs (%6.2fs elapsed)\n\n",
565 TICK_TO_DBL(time), TICK_TO_DBL(etime));
566 fprintf(sf, " %%GC time %5.1f%% (%.1f%% elapsed)\n\n",
567 TICK_TO_DBL(GC_tot_time)*100/time,
568 TICK_TO_DBL(GCe_tot_time)*100/etime);
570 if (time - GC_tot_time == 0)
571 ullong_format_string(0, temp, rtsTrue/*commas*/);
573 ullong_format_string(
574 (ullong)((GC_tot_alloc*sizeof(W_))/
575 TICK_TO_DBL(time - GC_tot_time)),
576 temp, rtsTrue/*commas*/);
578 fprintf(sf, " Alloc rate %s bytes per MUT second\n\n", temp);
580 fprintf(sf, " Productivity %5.1f%% of total user, %.1f%% of total elapsed\n\n",
581 TICK_TO_DBL(time - GC_tot_time - InitUserTime) * 100
583 TICK_TO_DBL(time - GC_tot_time - InitUserTime) * 100
584 / TICK_TO_DBL(etime));
587 if (RtsFlags.GcFlags.giveStats == ONELINE_GC_STATS && sf != NULL) {
588 fprintf(sf, "<<ghc: %lld bytes, %d GCs, %ld/%ld avg/max bytes residency (%ld samples), %ldM in use, %.2f INIT (%.2f elapsed), %.2f MUT (%.2f elapsed), %.2f GC (%.2f elapsed) :ghc>>\n",
589 GC_tot_alloc*sizeof(W_), total_collections,
590 AvgResidency*sizeof(W_)/ResidencySamples,
591 MaxResidency*sizeof(W_),
593 mblocks_allocated * MBLOCK_SIZE / (1024 * 1024),
594 TICK_TO_DBL(InitUserTime), TICK_TO_DBL(InitElapsedTime),
595 TICK_TO_DBL(MutUserTime), TICK_TO_DBL(MutElapsedTime),
596 TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time));
604 /* -----------------------------------------------------------------------------
607 Produce some detailed info on the state of the generational GC.
608 -------------------------------------------------------------------------- */
610 stat_describe_gens(void)
612 nat g, s, mut, mut_once, lge, live;
617 fprintf(stderr, " Gen Steps Max Mutable Mut-Once Step Blocks Live Large\n Blocks Closures Closures Objects\n");
619 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
620 for (m = generations[g].mut_list, mut = 0; m != END_MUT_LIST;
623 for (m = generations[g].mut_once_list, mut_once = 0; m != END_MUT_LIST;
626 fprintf(stderr, "%8d %8d %8d %9d %9d", g, generations[g].n_steps,
627 generations[g].max_blocks, mut, mut_once);
629 for (s = 0; s < generations[g].n_steps; s++) {
630 step = &generations[g].steps[s];
631 for (bd = step->large_objects, lge = 0; bd; bd = bd->link)
634 if (RtsFlags.GcFlags.generations == 1) {
639 for (; bd; bd = bd->link) {
640 live += (bd->free - bd->start) * sizeof(W_);
643 fprintf(stderr,"%46s","");
645 fprintf(stderr,"%6d %8d %8d %8d\n", s, step->n_blocks,
649 fprintf(stderr,"\n");
652 /* -----------------------------------------------------------------------------
653 Stats available via a programmatic interface, so eg. GHCi can time
654 each compilation and expression evaluation.
655 -------------------------------------------------------------------------- */
657 extern HsInt getAllocations( void )
658 { return (HsInt)(total_allocated * sizeof(W_)); }