1 /* -----------------------------------------------------------------------------
2 * $Id: Stats.c,v 1.29 2001/07/03 16:26:03 rrt Exp $
4 * (c) The GHC Team, 1998-1999
6 * Statistics and timing-related functions.
8 * ---------------------------------------------------------------------------*/
10 #define NON_POSIX_SOURCE
15 #include "StoragePriv.h"
19 #include "ParTicky.h" // ToDo: move into Rts.h
25 #ifndef mingw32_TARGET_OS
26 # ifdef HAVE_SYS_TIMES_H
27 # include <sys/times.h>
31 #ifdef HAVE_SYS_TIME_H
41 #if ! irix_TARGET_OS && ! defined(mingw32_TARGET_OS)
42 # if defined(HAVE_SYS_RESOURCE_H)
43 # include <sys/resource.h>
47 #ifdef HAVE_SYS_TIMEB_H
48 #include <sys/timeb.h>
59 #if defined(PAR) || !(!defined(HAVE_GETRUSAGE) || irix_TARGET_OS || defined(mingw32_TARGET_OS) || defined(cygwin32_TARGET_OS))
60 #include <sys/resource.h>
64 #define BIG_STRING_LEN 512
66 /* We're not trying to be terribly accurate here, using the
67 * basic times() function to get a resolution of about 100ths of a
68 * second, depending on the OS. A long int will do fine for holding
71 #define TICK_TYPE long int
72 #define TICK_TO_DBL(t) ((double)(t) / TicksPerSecond)
74 static int TicksPerSecond = 0;
76 static TICK_TYPE ElapsedTimeStart = 0;
77 static TICK_TYPE CurrentElapsedTime = 0;
78 static TICK_TYPE CurrentUserTime = 0;
80 static TICK_TYPE InitUserTime = 0;
81 static TICK_TYPE InitElapsedTime = 0;
82 static TICK_TYPE InitElapsedStamp = 0;
84 static TICK_TYPE MutUserTime = 0;
85 static TICK_TYPE MutElapsedTime = 0;
86 static TICK_TYPE MutElapsedStamp = 0;
88 static TICK_TYPE ExitUserTime = 0;
89 static TICK_TYPE ExitElapsedTime = 0;
91 static ullong GC_tot_alloc = 0;
92 static ullong GC_tot_copied = 0;
94 static TICK_TYPE GC_start_time, GC_tot_time = 0; /* User GC Time */
95 static TICK_TYPE GCe_start_time, GCe_tot_time = 0; /* Elapsed GC time */
97 lnat MaxResidency = 0; /* in words; for stats only */
98 lnat AvgResidency = 0;
99 lnat ResidencySamples = 0; /* for stats only */
101 static lnat GC_start_faults = 0, GC_end_faults = 0;
103 static TICK_TYPE *GC_coll_times;
105 static void getTimes(void);
106 static nat pageFaults(void);
108 /* elapsedtime() -- The current elapsed time in seconds */
110 #if defined(mingw32_TARGET_OS) || defined(cygwin32_TARGET_OS)
111 #define HNS_PER_SEC 10000000LL /* FILETIMES are in units of 100ns */
112 /* Convert FILETIMEs into secs */
113 #define FT2longlong(ll,ft) \
114 (ll)=(ft).dwHighDateTime; \
116 (ll) |= (ft).dwLowDateTime; \
117 (ll) /= (unsigned long long) (HNS_PER_SEC / CLOCKS_PER_SEC)
120 #if defined(mingw32_TARGET_OS) || defined(cygwin32_TARGET_OS)
121 /* cygwin32 or mingw32 version */
125 FILETIME creationTime, exitTime, kernelTime, userTime;
126 long long int kT, uT;
128 /* ToDo: pin down elapsed times to just the OS thread(s) that
129 are evaluating/managing Haskell code.
131 if (!GetProcessTimes (GetCurrentProcess(), &creationTime,
132 &exitTime, &kernelTime, &userTime)) {
133 /* Probably on a Win95 box..*/
137 FT2longlong(kT,kernelTime);
138 FT2longlong(uT,userTime);
139 CurrentElapsedTime = uT + kT;
140 CurrentUserTime = uT;
150 /* We will #ifdef around the fprintf for machines
151 we *know* are unsupported. (WDP 94/05)
153 fprintf(stderr, "NOTE: `getTimes' does nothing!\n");
156 #else /* not stumped */
158 clock_t r = times(&t);
160 CurrentElapsedTime = r;
161 CurrentUserTime = t.tms_utime;
167 /* mut_user_time_during_GC() and mut_user_time()
169 * The former function can be used to get the current mutator time
170 * *during* a GC, i.e. between stat_startGC and stat_endGC. This is
171 * used in the heap profiler for accurately time stamping the heap
174 * ATTENTION: mut_user_time_during_GC() relies on GC_start_time being
175 * defined in stat_startGC() - to minimise system calls,
176 * GC_start_time is, however, only defined when really needed (check
177 * stat_startGC() for details)
180 mut_user_time_during_GC(void)
182 return TICK_TO_DBL(GC_start_time - GC_tot_time);
189 return TICK_TO_DBL(CurrentUserTime - GC_tot_time);
195 /* ToDo (on NT): better, get this via the performance data
196 that's stored in the registry. */
197 # if !defined(HAVE_GETRUSAGE) || irix_TARGET_OS || defined(mingw32_TARGET_OS) || defined(cygwin32_TARGET_OS)
202 getrusage(RUSAGE_SELF, &t);
211 FILE *sf = RtsFlags.GcFlags.statsFile;
213 if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
214 fprintf(sf, " Alloc Collect Live GC GC TOT TOT Page Flts\n");
215 fprintf(sf, " bytes bytes bytes user elap user elap\n");
218 (TICK_TYPE *)stgMallocBytes(
219 sizeof(TICK_TYPE)*RtsFlags.GcFlags.generations,
221 for (i = 0; i < RtsFlags.GcFlags.generations; i++) {
222 GC_coll_times[i] = 0;
226 /* -----------------------------------------------------------------------------
227 Initialisation time...
228 -------------------------------------------------------------------------- */
233 /* Determine TicksPerSecond ... */
234 #if defined(CLK_TCK) /* defined by POSIX */
235 TicksPerSecond = CLK_TCK;
237 #elif defined(HAVE_SYSCONF)
240 ticks = sysconf(_SC_CLK_TCK);
242 fprintf(stderr, "stat_init: bad call to 'sysconf'!\n");
243 stg_exit(EXIT_FAILURE);
245 TicksPerSecond = ticks;
247 /* no "sysconf" or CLK_TCK; had better guess */
251 #elif defined(CLOCKS_PER_SEC)
252 TicksPerSecond = CLOCKS_PER_SEC;
254 #else /* had better guess wildly */
255 /* We will #ifdef around the fprintf for machines
256 we *know* are unsupported. (WDP 94/05)
258 fprintf(stderr, "NOTE: Guessing `TicksPerSecond = 60'!\n");
263 ElapsedTimeStart = CurrentElapsedTime;
270 InitUserTime = CurrentUserTime;
271 InitElapsedStamp = CurrentElapsedTime;
272 if (ElapsedTimeStart > CurrentElapsedTime) {
275 InitElapsedTime = CurrentElapsedTime - ElapsedTimeStart;
279 /* -----------------------------------------------------------------------------
280 stat_startExit and stat_endExit
282 These two measure the time taken in shutdownHaskell().
283 -------------------------------------------------------------------------- */
289 MutElapsedStamp = CurrentElapsedTime;
290 MutElapsedTime = CurrentElapsedTime - GCe_tot_time - InitElapsedStamp;
291 if (MutElapsedTime < 0) { MutElapsedTime = 0; } /* sometimes -0.00 */
293 /* for SMP, we don't know the mutator time yet, we have to inspect
294 * all the running threads to find out, and they haven't stopped
295 * yet. So we just timestamp MutUserTime at this point so we can
296 * calculate the EXIT time. The real MutUserTime is calculated
297 * in stat_exit below.
300 MutUserTime = CurrentUserTime;
302 MutUserTime = CurrentUserTime - GC_tot_time - InitUserTime;
303 if (MutUserTime < 0) { MutUserTime = 0; }
312 ExitUserTime = CurrentUserTime - MutUserTime;
314 ExitUserTime = CurrentUserTime - MutUserTime - GC_tot_time - InitUserTime;
316 ExitElapsedTime = CurrentElapsedTime - MutElapsedStamp;
317 if (ExitUserTime < 0) {
320 if (ExitElapsedTime < 0) {
325 /* -----------------------------------------------------------------------------
326 Called at the beginning of each GC
327 -------------------------------------------------------------------------- */
329 static nat rub_bell = 0;
331 /* initialise global variables needed during GC
333 * * GC_start_time is read in mut_user_time_during_GC(), which in turn is
334 * needed if either PROFILING or DEBUGing is enabled
339 nat bell = RtsFlags.GcFlags.ringBell;
343 fprintf(stderr, " GC ");
346 fprintf(stderr, "\007");
350 #if defined(PROFILING) || defined(DEBUG)
352 GC_start_time = CurrentUserTime; /* needed in mut_user_time_during_GC() */
355 if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
356 #if !defined(PROFILING) && !defined(DEBUG)
358 GC_start_time = CurrentUserTime;
360 GCe_start_time = CurrentElapsedTime;
361 if (RtsFlags.GcFlags.giveStats) {
362 GC_start_faults = pageFaults();
367 /* -----------------------------------------------------------------------------
368 Called at the end of each GC
369 -------------------------------------------------------------------------- */
372 stat_endGC(lnat alloc, lnat collect, lnat live, lnat copied, lnat gen)
374 FILE *sf = RtsFlags.GcFlags.statsFile;
376 if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
377 TICK_TYPE time, etime, gc_time, gc_etime;
380 time = CurrentUserTime;
381 etime = CurrentElapsedTime;
382 gc_time = time - GC_start_time;
383 gc_etime = etime - GCe_start_time;
385 if (RtsFlags.GcFlags.giveStats == VERBOSE_GC_STATS && sf != NULL) {
386 nat faults = pageFaults();
388 fprintf(sf, "%9ld %9ld %9ld",
389 alloc*sizeof(W_), collect*sizeof(W_), live*sizeof(W_));
390 fprintf(sf, " %5.2f %5.2f %7.2f %7.2f %4ld %4ld (Gen: %2ld)\n",
391 TICK_TO_DBL(gc_time),
392 TICK_TO_DBL(gc_etime),
394 TICK_TO_DBL(etime - ElapsedTimeStart),
395 faults - GC_start_faults,
396 GC_start_faults - GC_end_faults,
399 GC_end_faults = faults;
403 GC_coll_times[gen] += gc_time;
405 GC_tot_copied += (ullong) copied;
406 GC_tot_alloc += (ullong) alloc;
407 GC_tot_time += gc_time;
408 GCe_tot_time += gc_etime;
413 pthread_t me = pthread_self();
415 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
416 if (me == task_ids[i].id) {
417 task_ids[i].gc_time += gc_time;
418 task_ids[i].gc_etime += gc_etime;
425 if (gen == RtsFlags.GcFlags.generations-1) { /* major GC? */
426 if (live > MaxResidency) {
430 AvgResidency += live;
435 fprintf(stderr, "\b\b\b \b\b\b");
440 /* -----------------------------------------------------------------------------
443 Called under SMP when a worker thread finishes. We drop the timing
444 stats for this thread into the task_ids struct for that thread.
445 -------------------------------------------------------------------------- */
449 stat_workerStop(void)
452 pthread_t me = pthread_self();
454 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
455 if (task_ids[i].id == me) {
456 task_ids[i].mut_time = usertime() - task_ids[i].gc_time;
457 task_ids[i].mut_etime = elapsedtime()
459 - task_ids[i].elapsedtimestart;
460 if (task_ids[i].mut_time < 0.0) { task_ids[i].mut_time = 0.0; }
461 if (task_ids[i].mut_etime < 0.0) { task_ids[i].mut_etime = 0.0; }
467 /* -----------------------------------------------------------------------------
468 Called at the end of execution
470 NOTE: number of allocations is not entirely accurate: it doesn't
471 take into account the few bytes at the end of the heap that
472 were left unused when the heap-check failed.
473 -------------------------------------------------------------------------- */
478 FILE *sf = RtsFlags.GcFlags.statsFile;
480 if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
482 char temp[BIG_STRING_LEN];
485 nat g, total_collections = 0;
488 time = CurrentUserTime;
489 etime = CurrentElapsedTime - ElapsedTimeStart;
491 GC_tot_alloc += alloc;
493 /* avoid divide by zero if time is measured as 0.00 seconds -- SDM */
494 if (time == 0.0) time = 1;
495 if (etime == 0.0) etime = 1;
497 /* Count total garbage collections */
498 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
499 total_collections += generations[g].collections;
501 /* For SMP, we have to get the user time from each thread
502 * and try to work out the total time.
507 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
508 MutUserTime += task_ids[i].mut_time;
511 time = MutUserTime + GC_tot_time + InitUserTime + ExitUserTime;
512 if (MutUserTime < 0) { MutUserTime = 0; }
515 if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS && sf != NULL) {
516 fprintf(sf, "%9ld %9.9s %9.9s", (lnat)alloc*sizeof(W_), "", "");
517 fprintf(sf, " %5.2f %5.2f\n\n", 0.0, 0.0);
520 if (RtsFlags.GcFlags.giveStats >= SUMMARY_GC_STATS && sf != NULL) {
521 ullong_format_string(GC_tot_alloc*sizeof(W_),
522 temp, rtsTrue/*commas*/);
523 fprintf(sf, "%11s bytes allocated in the heap\n", temp);
525 ullong_format_string(GC_tot_copied*sizeof(W_),
526 temp, rtsTrue/*commas*/);
527 fprintf(sf, "%11s bytes copied during GC\n", temp);
529 if ( ResidencySamples > 0 ) {
530 ullong_format_string(MaxResidency*sizeof(W_),
531 temp, rtsTrue/*commas*/);
532 fprintf(sf, "%11s bytes maximum residency (%ld sample(s))\n",
533 temp, ResidencySamples);
537 /* Print garbage collections in each gen */
538 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
539 fprintf(sf, "%11d collections in generation %d (%6.2fs)\n",
540 generations[g].collections, g,
541 TICK_TO_DBL(GC_coll_times[g]));
544 fprintf(sf,"\n%11ld Mb total memory in use\n\n",
545 mblocks_allocated * MBLOCK_SIZE / (1024 * 1024));
550 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
551 fprintf(sf, " Task %2d: MUT time: %6.2fs (%6.2fs elapsed)\n"
552 " GC time: %6.2fs (%6.2fs elapsed)\n\n",
554 TICK_TO_DBL(task_ids[i].mut_time),
555 TICK_TO_DBL(task_ids[i].mut_etime),
556 TICK_TO_DBL(task_ids[i].gc_time),
557 TICK_TO_DBL(task_ids[i].gc_etime));
562 fprintf(sf, " INIT time %6.2fs (%6.2fs elapsed)\n",
563 TICK_TO_DBL(InitUserTime), TICK_TO_DBL(InitElapsedTime));
564 fprintf(sf, " MUT time %6.2fs (%6.2fs elapsed)\n",
565 TICK_TO_DBL(MutUserTime), TICK_TO_DBL(MutElapsedTime));
566 fprintf(sf, " GC time %6.2fs (%6.2fs elapsed)\n",
567 TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time));
568 fprintf(sf, " EXIT time %6.2fs (%6.2fs elapsed)\n",
569 TICK_TO_DBL(ExitUserTime), TICK_TO_DBL(ExitElapsedTime));
570 fprintf(sf, " Total time %6.2fs (%6.2fs elapsed)\n\n",
571 TICK_TO_DBL(time), TICK_TO_DBL(etime));
572 fprintf(sf, " %%GC time %5.1f%% (%.1f%% elapsed)\n\n",
573 TICK_TO_DBL(GC_tot_time)*100/time,
574 TICK_TO_DBL(GCe_tot_time)*100/etime);
576 if (time - GC_tot_time == 0)
577 ullong_format_string(0, temp, rtsTrue/*commas*/);
579 ullong_format_string(
580 (ullong)((GC_tot_alloc*sizeof(W_))/
581 TICK_TO_DBL(time - GC_tot_time)),
582 temp, rtsTrue/*commas*/);
584 fprintf(sf, " Alloc rate %s bytes per MUT second\n\n", temp);
586 fprintf(sf, " Productivity %5.1f%% of total user, %.1f%% of total elapsed\n\n",
587 TICK_TO_DBL(time - GC_tot_time - InitUserTime) * 100
589 TICK_TO_DBL(time - GC_tot_time - InitUserTime) * 100
590 / TICK_TO_DBL(etime));
593 if (RtsFlags.GcFlags.giveStats == ONELINE_GC_STATS && sf != NULL) {
594 /* print the long long separately to avoid bugginess on mingwin (2001-07-02, mingw-0.5) */
595 fprintf(sf, "<<ghc: %llu bytes, ", GC_tot_alloc*sizeof(W_));
596 fprintf(sf, "%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",
598 AvgResidency*sizeof(W_)/ResidencySamples,
599 MaxResidency*sizeof(W_),
601 (unsigned long)(mblocks_allocated * MBLOCK_SIZE / (1024L * 1024L)),
602 TICK_TO_DBL(InitUserTime), TICK_TO_DBL(InitElapsedTime),
603 TICK_TO_DBL(MutUserTime), TICK_TO_DBL(MutElapsedTime),
604 TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time));
612 /* -----------------------------------------------------------------------------
615 Produce some detailed info on the state of the generational GC.
616 -------------------------------------------------------------------------- */
618 stat_describe_gens(void)
620 nat g, s, mut, mut_once, lge, live;
625 fprintf(stderr, " Gen Steps Max Mutable Mut-Once Step Blocks Live Large\n Blocks Closures Closures Objects\n");
627 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
628 for (m = generations[g].mut_list, mut = 0; m != END_MUT_LIST;
631 for (m = generations[g].mut_once_list, mut_once = 0; m != END_MUT_LIST;
634 fprintf(stderr, "%8d %8d %8d %9d %9d", g, generations[g].n_steps,
635 generations[g].max_blocks, mut, mut_once);
637 for (s = 0; s < generations[g].n_steps; s++) {
638 step = &generations[g].steps[s];
639 for (bd = step->large_objects, lge = 0; bd; bd = bd->link)
642 if (RtsFlags.GcFlags.generations == 1) {
647 for (; bd; bd = bd->link) {
648 live += (bd->free - bd->start) * sizeof(W_);
651 fprintf(stderr,"%46s","");
653 fprintf(stderr,"%6d %8d %8d %8d\n", s, step->n_blocks,
657 fprintf(stderr,"\n");
660 /* -----------------------------------------------------------------------------
661 Stats available via a programmatic interface, so eg. GHCi can time
662 each compilation and expression evaluation.
663 -------------------------------------------------------------------------- */
665 extern HsInt getAllocations( void )
666 { return (HsInt)(total_allocated * sizeof(W_)); }