1 /* -----------------------------------------------------------------------------
2 * $Id: Stats.c,v 1.18 1999/11/11 13:17:36 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"
25 # ifdef HAVE_SYS_TIMES_H
26 # include <sys/times.h>
30 #ifdef HAVE_SYS_TIME_H
40 #if ! irix_TARGET_OS && ! defined(__MINGW32__)
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 static double ElapsedTimeStart = 0.0;
62 static double TicksPerSecond = 0.0;
64 static double InitUserTime = 0.0;
65 static double InitElapsedTime = 0.0;
66 static double InitElapsedStamp = 0.0;
68 static double MutUserTime = 0.0;
69 static double MutElapsedTime = 0.0;
70 static double MutElapsedStamp = 0.0;
72 static double ExitUserTime = 0.0;
73 static double ExitElapsedTime = 0.0;
75 static ullong GC_tot_alloc = 0;
76 static ullong GC_tot_copied = 0;
78 static double GC_start_time, GC_tot_time = 0; /* User GC Time */
79 static double GCe_start_time, GCe_tot_time = 0; /* Elapsed GC time */
81 lnat MaxResidency = 0; /* in words; for stats only */
82 lnat ResidencySamples = 0; /* for stats only */
84 static lnat GC_start_faults = 0, GC_end_faults = 0;
86 static double *GC_coll_times;
88 /* ToDo: convert this to use integers? --SDM */
90 /* elapsedtime() -- The current elapsed time in seconds */
93 #define NS_PER_SEC 10000000LL
94 /* Convert FILETIMEs into secs since the Epoch (Jan1-1970) */
95 #define FT2longlong(ll,ft) \
96 (ll)=(ft).dwHighDateTime; \
98 (ll) |= (ft).dwLowDateTime; \
99 (ll) /= (unsigned long long) (NS_PER_SEC / CLOCKS_PER_SEC)
103 /* cygwin32 or mingw32 version */
107 FILETIME creationTime, exitTime, kernelTime, userTime;
108 long long int kT, uT;
111 /* ToDo: pin down elapsed times to just the OS thread(s) that
112 are evaluating/managing Haskell code.
114 if (!GetProcessTimes (GetCurrentProcess(), &creationTime,
115 &exitTime, &kernelTime, &userTime)) {
116 /* Probably on a Win95 box..*/
120 FT2longlong(kT,kernelTime);
121 FT2longlong(uT,userTime);
122 return (((StgDouble)(uT + kT))/TicksPerSecond);
131 # if ! (defined(HAVE_TIMES) || defined(HAVE_FTIME))
132 /* We will #ifdef around the fprintf for machines
133 we *know* are unsupported. (WDP 94/05)
135 fprintf(stderr, "NOTE: `elapsedtime' does nothing!\n");
138 # else /* not stumped */
140 /* "ftime" may be nicer, but "times" is more standard;
141 but, on a Sun, if you do not get the SysV one, you are *hosed*...
144 # if defined(HAVE_TIMES) && ! sunos4_TARGET_OS
146 clock_t r = times(&t);
148 return (((double)r)/TicksPerSecond);
150 # else /* HAVE_FTIME */
154 return (fabs(t.time + 1e-3*t.millitm));
156 # endif /* HAVE_FTIME */
157 # endif /* not stumped */
161 /* mut_user_time_during_GC() and mut_user_time()
163 * This function can be used to get the current mutator time *during*
164 * a GC, i.e. between stat_startGC and stat_endGC. This is used in
165 * the heap profiler for accurately time stamping the heap sample.
168 mut_user_time_during_GC(void)
170 return (GC_start_time - GC_tot_time);
176 return (usertime() - GC_tot_time);
183 /* ToDo (on NT): better, get this via the performance data
184 that's stored in the registry. */
185 # if !defined(HAVE_GETRUSAGE) || irix_TARGET_OS || defined(_WIN32)
190 getrusage(RUSAGE_SELF, &t);
195 /* ToDo: use gettimeofday on systems that support it (u-sec accuracy) */
202 /* Determine TicksPerSecond ... */
204 ticks = sysconf(_SC_CLK_TCK);
206 fprintf(stderr, "stat_init: bad call to 'sysconf'!\n");
207 stg_exit(EXIT_FAILURE);
209 TicksPerSecond = (double) ticks;
211 /* no "sysconf"; had better guess */
213 TicksPerSecond = (StgDouble) (HZ);
215 #elif defined(CLOCKS_PER_SEC)
216 TicksPerSecond = (StgDouble) (CLOCKS_PER_SEC);
217 #else /* had better guess wildly */
218 /* We will #ifdef around the fprintf for machines
219 we *know* are unsupported. (WDP 94/05)
221 fprintf(stderr, "NOTE: Guessing `TicksPerSecond = 60'!\n");
222 TicksPerSecond = 60.0;
225 ElapsedTimeStart = elapsedtime();
233 FILE *sf = RtsFlags.GcFlags.statsFile;
235 if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
236 fprintf(sf, " Alloc Collect Live GC GC TOT TOT Page Flts\n");
237 fprintf(sf, " bytes bytes bytes user elap user elap\n");
240 (double *)stgMallocBytes(sizeof(double) * RtsFlags.GcFlags.generations,
242 for (i = 0; i < RtsFlags.GcFlags.generations; i++) {
243 GC_coll_times[i] = 0.0;
251 FILETIME creationTime, exitTime, kernelTime, userTime;
254 /* Convert FILETIMEs into long longs */
256 if (!GetProcessTimes (GetCurrentProcess(), &creationTime,
257 &exitTime, &kernelTime, &userTime)) {
258 /* Probably exec'ing this on a Win95 box..*/
262 FT2longlong(uT,userTime);
263 return (((StgDouble)uT)/TicksPerSecond);
270 # if ! (defined(HAVE_GETRUSAGE) || defined(HAVE_TIMES))
271 /* We will #ifdef around the fprintf for machines
272 we *know* are unsupported. (WDP 94/05)
274 fprintf(stderr, "NOTE: `usertime' does nothing!\n");
277 # else /* not stumped */
279 # if defined(HAVE_TIMES)
283 return(((double)(t.tms_utime))/TicksPerSecond);
285 # else /* HAVE_GETRUSAGE */
288 getrusage(RUSAGE_SELF, &t);
289 return(t.ru_utime.tv_sec + 1e-6*t.ru_utime.tv_usec);
291 # endif /* HAVE_GETRUSAGE */
292 # endif /* not stumped */
294 #endif /* ! _WIN32 */
299 InitUserTime = usertime();
300 InitElapsedStamp = elapsedtime();
301 InitElapsedTime = InitElapsedStamp - ElapsedTimeStart;
302 if (InitElapsedTime < 0.0) {
303 InitElapsedTime = 0.0;
307 /* -----------------------------------------------------------------------------
308 stat_startExit and stat_endExit
310 These two measure the time taken in shutdownHaskell().
311 -------------------------------------------------------------------------- */
316 MutElapsedStamp = elapsedtime();
317 MutElapsedTime = MutElapsedStamp - GCe_tot_time - InitElapsedStamp;
318 if (MutElapsedTime < 0) { MutElapsedTime = 0; } /* sometimes -0.00 */
320 /* for SMP, we don't know the mutator time yet, we have to inspect
321 * all the running threads to find out, and they haven't stopped
322 * yet. So we just timestamp MutUserTime at this point so we can
323 * calculate the EXIT time. The real MutUserTime is calculated
324 * in stat_exit below.
327 MutUserTime = usertime();
329 MutUserTime = usertime() - GC_tot_time - InitUserTime;
330 if (MutUserTime < 0) { MutUserTime = 0; }
338 ExitUserTime = usertime() - MutUserTime;
340 ExitUserTime = usertime() - MutUserTime - GC_tot_time - InitUserTime;
342 ExitElapsedTime = elapsedtime() - MutElapsedStamp;
343 if (ExitUserTime < 0.0) {
346 if (ExitElapsedTime < 0.0) {
347 ExitElapsedTime = 0.0;
351 /* -----------------------------------------------------------------------------
352 Called at the beginning of each GC
353 -------------------------------------------------------------------------- */
355 static nat rub_bell = 0;
360 FILE *sf = RtsFlags.GcFlags.statsFile;
362 nat bell = RtsFlags.GcFlags.ringBell;
366 fprintf(stderr, " GC ");
369 fprintf(stderr, "\007");
374 GC_start_time = usertime();
375 GCe_start_time = elapsedtime();
376 if (RtsFlags.GcFlags.giveStats) {
377 GC_start_faults = pagefaults();
382 /* -----------------------------------------------------------------------------
383 Called at the end of each GC
384 -------------------------------------------------------------------------- */
387 stat_endGC(lnat alloc, lnat collect, lnat live, lnat copied, lnat gen)
389 FILE *sf = RtsFlags.GcFlags.statsFile;
392 double time = usertime();
393 double etime = elapsedtime();
394 double gc_time = time-GC_start_time;
395 double gc_etime = etime-GCe_start_time;
397 if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
398 nat faults = pagefaults();
400 fprintf(sf, "%9ld %9ld %9ld",
401 alloc*sizeof(W_), collect*sizeof(W_), live*sizeof(W_));
402 fprintf(sf, " %5.2f %5.2f %7.2f %7.2f %4ld %4ld (Gen: %2ld)\n",
407 faults - GC_start_faults,
408 GC_start_faults - GC_end_faults,
411 GC_end_faults = faults;
415 GC_coll_times[gen] += time-GC_start_time;
417 GC_tot_copied += (ullong) copied;
418 GC_tot_alloc += (ullong) alloc;
419 GC_tot_time += gc_time;
420 GCe_tot_time += gc_etime;
425 pthread_t me = pthread_self();
427 for (i = 0; i < RtsFlags.ConcFlags.nNodes; i++) {
428 if (me == task_ids[i].id) {
429 task_ids[i].gc_time += gc_time;
430 task_ids[i].gc_etime += gc_etime;
437 if (gen == RtsFlags.GcFlags.generations-1) { /* major GC? */
438 if (live > MaxResidency) {
446 fprintf(stderr, "\b\b\b \b\b\b");
451 /* -----------------------------------------------------------------------------
454 Called under SMP when a worker thread finishes. We drop the timing
455 stats for this thread into the task_ids struct for that thread.
456 -------------------------------------------------------------------------- */
460 stat_workerStop(void)
463 pthread_t me = pthread_self();
465 for (i = 0; i < RtsFlags.ConcFlags.nNodes; i++) {
466 if (task_ids[i].id == me) {
467 task_ids[i].mut_time = usertime() - task_ids[i].gc_time;
468 task_ids[i].mut_etime = elapsedtime()
470 - task_ids[i].elapsedtimestart;
471 if (task_ids[i].mut_time < 0.0) { task_ids[i].mut_time = 0.0; }
472 if (task_ids[i].mut_etime < 0.0) { task_ids[i].mut_etime = 0.0; }
478 /* -----------------------------------------------------------------------------
479 Called at the end of execution
481 NOTE: number of allocations is not entirely accurate: it doesn't
482 take into account the few bytes at the end of the heap that
483 were left unused when the heap-check failed.
484 -------------------------------------------------------------------------- */
489 FILE *sf = RtsFlags.GcFlags.statsFile;
492 char temp[BIG_STRING_LEN];
493 double time = usertime();
494 double etime = elapsedtime() - ElapsedTimeStart;
496 /* avoid divide by zero if time is measured as 0.00 seconds -- SDM */
497 if (time == 0.0) time = 0.0001;
498 if (etime == 0.0) etime = 0.0001;
500 if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
501 fprintf(sf, "%9ld %9.9s %9.9s", (lnat)alloc*sizeof(W_), "", "");
502 fprintf(sf, " %5.2f %5.2f\n\n", 0.0, 0.0);
505 GC_tot_alloc += alloc;
507 ullong_format_string(GC_tot_alloc*sizeof(W_), temp, rtsTrue/*commas*/);
508 fprintf(sf, "%11s bytes allocated in the heap\n", temp);
510 ullong_format_string(GC_tot_copied*sizeof(W_), temp, rtsTrue/*commas*/);
511 fprintf(sf, "%11s bytes copied during GC\n", temp);
513 if ( ResidencySamples > 0 ) {
514 ullong_format_string(MaxResidency*sizeof(W_), temp, rtsTrue/*commas*/);
515 fprintf(sf, "%11s bytes maximum residency (%ld sample(s))\n",
521 { /* Count garbage collections */
523 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
524 fprintf(sf, "%11d collections in generation %d (%6.2fs)\n",
525 generations[g].collections, g, GC_coll_times[g]);
528 fprintf(sf,"\n%11ld Mb total memory in use\n\n",
529 mblocks_allocated * MBLOCK_SIZE / (1024 * 1024));
531 /* For SMP, we have to get the user time from each thread
532 * and try to work out the total time.
538 for (i = 0; i < RtsFlags.ConcFlags.nNodes; i++) {
539 MutUserTime += task_ids[i].mut_time;
540 fprintf(sf, " Task %2d: MUT time: %6.2fs (%6.2fs elapsed)\n"
541 " GC time: %6.2fs (%6.2fs elapsed)\n\n",
543 task_ids[i].mut_time, task_ids[i].mut_etime,
544 task_ids[i].gc_time, task_ids[i].gc_etime);
547 time = MutUserTime + GC_tot_time + InitUserTime + ExitUserTime;
548 if (MutUserTime < 0) { MutUserTime = 0; }
551 fprintf(sf, " INIT time %6.2fs (%6.2fs elapsed)\n",
552 InitUserTime, InitElapsedTime);
553 fprintf(sf, " MUT time %6.2fs (%6.2fs elapsed)\n",
554 MutUserTime, MutElapsedTime);
555 fprintf(sf, " GC time %6.2fs (%6.2fs elapsed)\n",
556 GC_tot_time, GCe_tot_time);
557 fprintf(sf, " EXIT time %6.2fs (%6.2fs elapsed)\n",
558 ExitUserTime, ExitElapsedTime);
559 fprintf(sf, " Total time %6.2fs (%6.2fs elapsed)\n\n",
562 fprintf(sf, " %%GC time %5.1f%% (%.1f%% elapsed)\n\n",
563 GC_tot_time*100./time, GCe_tot_time*100./etime);
565 if (time - GC_tot_time == 0.0)
566 ullong_format_string(0, temp, rtsTrue/*commas*/);
568 ullong_format_string((ullong)(GC_tot_alloc*sizeof(W_)/
569 (time - GC_tot_time)),
570 temp, rtsTrue/*commas*/);
572 fprintf(sf, " Alloc rate %s bytes per MUT second\n\n", temp);
574 fprintf(sf, " Productivity %5.1f%% of total user, %.1f%% of total elapsed\n\n",
575 (time - GC_tot_time - InitUserTime) * 100. / time,
576 (time - GC_tot_time - InitUserTime) * 100. / etime);
582 /* -----------------------------------------------------------------------------
585 Produce some detailed info on the state of the generational GC.
586 -------------------------------------------------------------------------- */
588 stat_describe_gens(void)
590 nat g, s, mut, mut_once, lge, live;
595 fprintf(stderr, " Gen Steps Max Mutable Mut-Once Step Blocks Live Large\n Blocks Closures Closures Objects\n");
597 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
598 for (m = generations[g].mut_list, mut = 0; m != END_MUT_LIST;
601 for (m = generations[g].mut_once_list, mut_once = 0; m != END_MUT_LIST;
604 fprintf(stderr, "%8d %8d %8d %9d %9d", g, generations[g].n_steps,
605 generations[g].max_blocks, mut, mut_once);
607 for (s = 0; s < generations[g].n_steps; s++) {
608 step = &generations[g].steps[s];
609 for (bd = step->large_objects, lge = 0; bd; bd = bd->link)
612 if (RtsFlags.GcFlags.generations == 1) {
617 for (; bd; bd = bd->link) {
618 live += (bd->free - bd->start) * sizeof(W_);
621 fprintf(stderr,"%46s","");
623 fprintf(stderr,"%6d %8d %8d %8d\n", s, step->n_blocks,
627 fprintf(stderr,"\n");