Stats.c 32 KB
Newer Older
1
/* -----------------------------------------------------------------------------
2
 *
3
 * (c) The GHC Team, 1998-2005
4 5 6 7 8
 *
 * Statistics and timing-related functions.
 *
 * ---------------------------------------------------------------------------*/

Simon Marlow's avatar
Simon Marlow committed
9
#include "PosixSource.h"
10
#include "Rts.h"
Simon Marlow's avatar
Simon Marlow committed
11

12
#include "RtsUtils.h"
13
#include "Schedule.h"
14
#include "Stats.h"
15
#include "Profiling.h"
16
#include "GetTime.h"
Simon Marlow's avatar
Simon Marlow committed
17
#include "sm/Storage.h"
18
#include "sm/GC.h" // gc_alloc_block_sync, whitehole_spin
Simon Marlow's avatar
Simon Marlow committed
19 20
#include "sm/GCThread.h"
#include "sm/BlockAlloc.h"
21

mrchebas@gmail.com's avatar
mrchebas@gmail.com committed
22 23 24 25
#if USE_PAPI
#include "Papi.h"
#endif

26 27 28
/* huh? */
#define BIG_STRING_LEN              512

Simon Marlow's avatar
Simon Marlow committed
29
#define TimeToSecondsDbl(t) ((double)(t) / TIME_RESOLUTION)
30

Simon Marlow's avatar
Simon Marlow committed
31
static Time
Simon Marlow's avatar
Simon Marlow committed
32 33 34 35
    start_init_cpu, start_init_elapsed,
    end_init_cpu,   end_init_elapsed,
    start_exit_cpu, start_exit_elapsed,
    end_exit_cpu,   end_exit_elapsed;
36

Simon Marlow's avatar
Simon Marlow committed
37
static Time GC_tot_cpu  = 0;
38

Simon Marlow's avatar
Simon Marlow committed
39 40
static StgWord64 GC_tot_alloc      = 0;
static StgWord64 GC_tot_copied     = 0;
41

Ian Lynagh's avatar
Ian Lynagh committed
42
static StgWord64 GC_par_max_copied = 0;
43
static StgWord64 GC_par_tot_copied = 0;
44

sof's avatar
sof committed
45
#ifdef PROFILING
Simon Marlow's avatar
Simon Marlow committed
46 47
static Time RP_start_time  = 0, RP_tot_time  = 0;  // retainer prof user time
static Time RPe_start_time = 0, RPe_tot_time = 0;  // retainer prof elap time
sof's avatar
sof committed
48

Simon Marlow's avatar
Simon Marlow committed
49 50
static Time HC_start_time, HC_tot_time = 0;     // heap census prof user time
static Time HCe_start_time, HCe_tot_time = 0;   // heap census prof elap time
sof's avatar
sof committed
51
#endif
52

sof's avatar
sof committed
53
#ifdef PROFILING
54
#define PROF_VAL(x)   (x)
sof's avatar
sof committed
55 56 57
#else
#define PROF_VAL(x)   0
#endif
58

59
// current = current as of last GC
60 61 62 63 64 65
static W_ current_residency = 0; // in words; for stats only
static W_ max_residency     = 0;
static W_ cumulative_residency = 0;
static W_ residency_samples = 0; // for stats only
static W_ current_slop      = 0;
static W_ max_slop          = 0;
66

67
static W_ GC_end_faults = 0;
68

Simon Marlow's avatar
Simon Marlow committed
69 70 71
static Time *GC_coll_cpu = NULL;
static Time *GC_coll_elapsed = NULL;
static Time *GC_coll_max_pause = NULL;
72

73 74 75
static void statsFlush( void );
static void statsClose( void );

Simon Marlow's avatar
Simon Marlow committed
76 77 78
/* -----------------------------------------------------------------------------
   Current elapsed time
   ------------------------------------------------------------------------- */
sof's avatar
sof committed
79

Simon Marlow's avatar
Simon Marlow committed
80
Time stat_getElapsedTime(void)
Simon Marlow's avatar
Simon Marlow committed
81
{
Simon Marlow's avatar
Simon Marlow committed
82
    return getProcessElapsedTime() - start_init_elapsed;
Simon Marlow's avatar
Simon Marlow committed
83 84
}

Simon Marlow's avatar
Simon Marlow committed
85 86 87
/* ---------------------------------------------------------------------------
   Measure the current MUT time, for profiling
   ------------------------------------------------------------------------ */
88

Ian Lynagh's avatar
Ian Lynagh committed
89
double
Simon Marlow's avatar
Simon Marlow committed
90
mut_user_time_until( Time t )
Ian Lynagh's avatar
Ian Lynagh committed
91
{
Simon Marlow's avatar
Simon Marlow committed
92
    return TimeToSecondsDbl(t - GC_tot_cpu);
93 94
    // heapCensus() time is included in GC_tot_cpu, so we don't need
    // to subtract it here.
Ian Lynagh's avatar
Ian Lynagh committed
95 96
}

97
double
98
mut_user_time( void )
99
{
Simon Marlow's avatar
Simon Marlow committed
100
    Time cpu;
Simon Marlow's avatar
Simon Marlow committed
101
    cpu = getProcessCPUTime();
Ian Lynagh's avatar
Ian Lynagh committed
102
    return mut_user_time_until(cpu);
103 104
}

105
#ifdef PROFILING
106
/*
Simon Marlow's avatar
Simon Marlow committed
107
  mut_user_time_during_RP() returns the MUT time during retainer profiling.
108
  The same is for mut_user_time_during_HC();
109 110 111 112
 */
double
mut_user_time_during_RP( void )
{
Simon Marlow's avatar
Simon Marlow committed
113
    return TimeToSecondsDbl(RP_start_time - GC_tot_cpu - RP_tot_time);
114 115 116
}

double
117
mut_user_time_during_heap_census( void )
118
{
Simon Marlow's avatar
Simon Marlow committed
119
    return TimeToSecondsDbl(HC_start_time - GC_tot_cpu - RP_tot_time);
120
}
sof's avatar
sof committed
121
#endif /* PROFILING */
122

Simon Marlow's avatar
Simon Marlow committed
123 124 125 126
/* ---------------------------------------------------------------------------
   initStats0() has no dependencies, it can be called right at the beginning
   ------------------------------------------------------------------------ */

127
void
128 129
initStats0(void)
{
Simon Marlow's avatar
Simon Marlow committed
130 131 132 133
    start_init_cpu    = 0;
    start_init_elapsed = 0;
    end_init_cpu     = 0;
    end_init_elapsed  = 0;
134

Simon Marlow's avatar
Simon Marlow committed
135 136 137 138
    start_exit_cpu    = 0;
    start_exit_elapsed = 0;
    end_exit_cpu     = 0;
    end_exit_elapsed  = 0;
139 140 141 142

    GC_tot_alloc     = 0;
    GC_tot_copied    = 0;
    GC_par_max_copied = 0;
143
    GC_par_tot_copied = 0;
Simon Marlow's avatar
Simon Marlow committed
144
    GC_tot_cpu  = 0;
145 146 147 148 149 150 151 152 153 154 155 156 157

#ifdef PROFILING
    RP_start_time  = 0;
    RP_tot_time  = 0;
    RPe_start_time = 0;
    RPe_tot_time = 0;

    HC_start_time = 0;
    HC_tot_time = 0;
    HCe_start_time = 0;
    HCe_tot_time = 0;
#endif

Simon Marlow's avatar
Simon Marlow committed
158
    max_residency = 0;
159
    cumulative_residency = 0;
Simon Marlow's avatar
Simon Marlow committed
160 161
    residency_samples = 0;
    max_slop = 0;
162 163 164 165

    GC_end_faults = 0;
}    

Simon Marlow's avatar
Simon Marlow committed
166 167 168 169
/* ---------------------------------------------------------------------------
   initStats1() can be called after setupRtsFlags()
   ------------------------------------------------------------------------ */

170 171
void
initStats1 (void)
172 173 174 175
{
    nat i;
  
    if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
176 177
        statsPrintf("    Alloc    Copied     Live     GC     GC      TOT      TOT  Page Flts\n");
        statsPrintf("    bytes     bytes     bytes   user   elap     user     elap\n");
178
    }
Simon Marlow's avatar
Simon Marlow committed
179
    GC_coll_cpu = 
Simon Marlow's avatar
Simon Marlow committed
180 181
	(Time *)stgMallocBytes(
            sizeof(Time)*RtsFlags.GcFlags.generations,
Simon Marlow's avatar
Simon Marlow committed
182 183
	    "initStats");
    GC_coll_elapsed = 
Simon Marlow's avatar
Simon Marlow committed
184 185
	(Time *)stgMallocBytes(
	    sizeof(Time)*RtsFlags.GcFlags.generations,
186
	    "initStats");
Simon Marlow's avatar
Simon Marlow committed
187
    GC_coll_max_pause =
Simon Marlow's avatar
Simon Marlow committed
188 189
	(Time *)stgMallocBytes(
	    sizeof(Time)*RtsFlags.GcFlags.generations,
190
	    "initStats");
191
    for (i = 0; i < RtsFlags.GcFlags.generations; i++) {
Simon Marlow's avatar
Simon Marlow committed
192 193 194
	GC_coll_cpu[i] = 0;
        GC_coll_elapsed[i] = 0;
        GC_coll_max_pause[i] = 0;
195
    }
196
}
197 198 199 200

/* -----------------------------------------------------------------------------
   Initialisation time...
   -------------------------------------------------------------------------- */
201 202

void
203
stat_startInit(void)
204
{
Simon Marlow's avatar
Simon Marlow committed
205
    getProcessTimes(&start_init_cpu, &start_init_elapsed);
206 207 208
}

void 
209
stat_endInit(void)
210
{
Simon Marlow's avatar
Simon Marlow committed
211
    getProcessTimes(&end_init_cpu, &end_init_elapsed);
212

mrchebas@gmail.com's avatar
mrchebas@gmail.com committed
213 214 215 216 217 218 219 220 221 222
#if USE_PAPI
    /* We start counting events for the mutator
     * when garbage collection starts
     * we switch to the GC event set. */
    papi_start_mutator_count();

    /* This flag is needed to avoid counting the last GC */
    papi_is_reporting = 1;

#endif
223 224
}

225 226 227 228 229 230
/* -----------------------------------------------------------------------------
   stat_startExit and stat_endExit
   
   These two measure the time taken in shutdownHaskell().
   -------------------------------------------------------------------------- */

231 232 233
void
stat_startExit(void)
{
Simon Marlow's avatar
Simon Marlow committed
234
    getProcessTimes(&start_exit_cpu, &start_exit_elapsed);
mrchebas@gmail.com's avatar
mrchebas@gmail.com committed
235 236 237 238 239 240 241 242 243

#if USE_PAPI
    /* We stop counting mutator events
     * GC events are not being counted at this point */
    papi_stop_mutator_count();

    /* This flag is needed, because GC is run once more after this function */
    papi_is_reporting = 0;
#endif
244 245 246 247 248
}

void
stat_endExit(void)
{
Simon Marlow's avatar
Simon Marlow committed
249
    getProcessTimes(&end_exit_cpu, &end_exit_elapsed);
250 251
}

252 253 254 255 256 257 258
/* -----------------------------------------------------------------------------
   Called at the beginning of each GC
   -------------------------------------------------------------------------- */

static nat rub_bell = 0;

void
259
stat_startGC (Capability *cap, gc_thread *gct)
260 261 262 263 264
{
    nat bell = RtsFlags.GcFlags.ringBell;

    if (bell) {
	if (bell > 1) {
265
	    debugBelch(" GC ");
266 267
	    rub_bell = 1;
	} else {
268
	    debugBelch("\007");
269 270 271
	}
    }

mrchebas@gmail.com's avatar
mrchebas@gmail.com committed
272 273 274 275 276 277 278 279
#if USE_PAPI
    if(papi_is_reporting) {
      /* Switch to counting GC events */
      papi_stop_mutator_count();
      papi_start_gc_count();
    }
#endif

Simon Marlow's avatar
Simon Marlow committed
280
    getProcessTimes(&gct->gc_start_cpu, &gct->gc_start_elapsed);
281 282 283 284 285 286 287 288 289

    // Post EVENT_GC_START with the same timestamp as used for stats
    // (though converted from Time=StgInt64 to EventTimestamp=StgWord64).
    // Here, as opposed to other places, the event is emitted on the cap
    // that initiates the GC and external tools expect it to have the same
    // timestamp as used in +RTS -s calculcations.
    traceEventGcStartAtT(cap,
                         TimeToNS(gct->gc_start_elapsed - start_init_elapsed));

Simon Marlow's avatar
Simon Marlow committed
290 291 292 293 294 295
    if (RtsFlags.GcFlags.giveStats != NO_GC_STATS)
    {
        gct->gc_start_faults = getPageFaults();
    }
}

296 297 298 299 300 301 302 303 304 305 306 307
/* -----------------------------------------------------------------------------
 * Calculate the total allocated memory since the start of the
 * program.  Also emits events reporting the per-cap allocation
 * totals.
 * -------------------------------------------------------------------------- */

static StgWord
calcTotalAllocated(void)
{
    W_ tot_alloc = 0;
    W_ n;
    for (n = 0; n < n_capabilities; n++) {
308 309
        tot_alloc += capabilities[n]->total_allocated;
        traceEventHeapAllocated(capabilities[n],
310
                                CAPSET_HEAP_DEFAULT,
311
                                capabilities[n]->total_allocated * sizeof(W_));
312 313 314 315 316
    }

    return tot_alloc;
}

317 318 319 320 321
/* -----------------------------------------------------------------------------
   Called at the end of each GC
   -------------------------------------------------------------------------- */

void
322
stat_endGC (Capability *cap, gc_thread *gct,
323
            W_ live, W_ copied, W_ slop, nat gen,
324
            nat par_n_threads, W_ par_max_copied, W_ par_tot_copied)
325
{
326 327 328
    W_ tot_alloc;
    W_ alloc;

329 330 331 332
    if (RtsFlags.GcFlags.giveStats != NO_GC_STATS ||
        RtsFlags.ProfFlags.doHeapProfile)
        // heap profiling needs GC_tot_time
    {
333
        Time cpu, elapsed, gc_cpu, gc_elapsed;
334

335 336 337 338 339 340 341
        // Has to be emitted while all caps stopped for GC, but before GC_END.
        // See trac.haskell.org/ThreadScope/wiki/RTSsummaryEvents
        // for a detailed design rationale of the current setup
        // of GC eventlog events.
        traceEventGcGlobalSync(cap);
	
        // Emitted before GC_END on all caps, which simplifies tools code.
342 343 344 345 346 347 348 349 350 351 352
        traceEventGcStats(cap,
                          CAPSET_HEAP_DEFAULT,
                          gen,
                          copied * sizeof(W_),
                          slop   * sizeof(W_),
                          /* current loss due to fragmentation */
                          (mblocks_allocated * BLOCKS_PER_MBLOCK - n_alloc_blocks)
                                 * BLOCK_SIZE_W * sizeof(W_),
                          par_n_threads,
                          par_max_copied * sizeof(W_),
                          par_tot_copied * sizeof(W_));
353

Simon Marlow's avatar
Simon Marlow committed
354 355
        getProcessTimes(&cpu, &elapsed);

356 357 358 359 360 361 362 363
        // Post EVENT_GC_END with the same timestamp as used for stats
        // (though converted from Time=StgInt64 to EventTimestamp=StgWord64).
        // Here, as opposed to other places, the event is emitted on the cap
        // that initiates the GC and external tools expect it to have the same
        // timestamp as used in +RTS -s calculcations.
        traceEventGcEndAtT(cap, TimeToNS(elapsed - start_init_elapsed));

        gc_elapsed = elapsed - gct->gc_start_elapsed;
Simon Marlow's avatar
Simon Marlow committed
364 365
        gc_cpu = cpu - gct->gc_start_cpu;

366 367 368 369 370 371 372 373 374 375 376
        /* For the moment we calculate both per-HEC and total allocation.
	 * There is thus redundancy here, but for the moment we will calculate
	 * it both the old and new way and assert they're the same.
	 * When we're sure it's working OK then we can simplify things.
         */
        tot_alloc = calcTotalAllocated();

        // allocated since the last GC
        alloc = tot_alloc - GC_tot_alloc;
        GC_tot_alloc = tot_alloc;

Simon Marlow's avatar
Simon Marlow committed
377
        if (RtsFlags.GcFlags.giveStats == VERBOSE_GC_STATS) {
Ian Lynagh's avatar
Ian Lynagh committed
378
	    W_ faults = getPageFaults();
379
	    
Ian Lynagh's avatar
Ian Lynagh committed
380
	    statsPrintf("%9" FMT_SizeT " %9" FMT_SizeT " %9" FMT_SizeT,
381
		    alloc*sizeof(W_), copied*sizeof(W_), 
382
			live*sizeof(W_));
383
            statsPrintf(" %6.3f %6.3f %8.3f %8.3f %4" FMT_Word " %4" FMT_Word "  (Gen: %2d)\n",
Simon Marlow's avatar
Simon Marlow committed
384 385 386 387
                    TimeToSecondsDbl(gc_cpu),
		    TimeToSecondsDbl(gc_elapsed),
		    TimeToSecondsDbl(cpu),
		    TimeToSecondsDbl(elapsed - start_init_elapsed),
Simon Marlow's avatar
Simon Marlow committed
388 389 390 391 392
		    faults - gct->gc_start_faults,
                        gct->gc_start_faults - GC_end_faults,
                    gen);

            GC_end_faults = faults;
393
	    statsFlush();
394 395
	}

Simon Marlow's avatar
Simon Marlow committed
396 397 398 399 400
        GC_coll_cpu[gen] += gc_cpu;
        GC_coll_elapsed[gen] += gc_elapsed;
        if (GC_coll_max_pause[gen] < gc_elapsed) {
            GC_coll_max_pause[gen] = gc_elapsed;
        }
401

Ian Lynagh's avatar
Ian Lynagh committed
402
	GC_tot_copied += (StgWord64) copied;
403 404
        GC_par_max_copied += (StgWord64) par_max_copied;
        GC_par_tot_copied += (StgWord64) par_tot_copied;
Simon Marlow's avatar
Simon Marlow committed
405
	GC_tot_cpu   += gc_cpu;
406
        
407 408 409
        traceEventHeapSize(cap,
	                   CAPSET_HEAP_DEFAULT,
			   mblocks_allocated * MBLOCK_SIZE_W * sizeof(W_));
410

411
	if (gen == RtsFlags.GcFlags.generations-1) { /* major GC? */
Simon Marlow's avatar
Simon Marlow committed
412 413
	    if (live > max_residency) {
		max_residency = live;
414
	    }
415
            current_residency = live;
Simon Marlow's avatar
Simon Marlow committed
416
	    residency_samples++;
417
	    cumulative_residency += live;
418 419 420
	    traceEventHeapLive(cap, 
    	    	    	       CAPSET_HEAP_DEFAULT,
	                       live * sizeof(W_));
421
	}
422

Simon Marlow's avatar
Simon Marlow committed
423
        if (slop > max_slop) max_slop = slop;
424 425 426
    }

    if (rub_bell) {
427
	debugBelch("\b\b\b  \b\b\b");
428 429
	rub_bell = 0;
    }
mrchebas@gmail.com's avatar
mrchebas@gmail.com committed
430 431 432 433

#if USE_PAPI
    if(papi_is_reporting) {
      /* Switch to counting mutator events */
434 435 436 437 438
      if (gen == 0) {
          papi_stop_gc0_count();
      } else {
          papi_stop_gc1_count();
      }
mrchebas@gmail.com's avatar
mrchebas@gmail.com committed
439 440 441
      papi_start_mutator_count();
    }
#endif
442 443
}

444 445 446
/* -----------------------------------------------------------------------------
   Called at the beginning of each Retainer Profiliing
   -------------------------------------------------------------------------- */
447 448 449
#ifdef PROFILING
void
stat_startRP(void)
450
{
Simon Marlow's avatar
Simon Marlow committed
451
    Time user, elapsed;
452
    getProcessTimes( &user, &elapsed );
453 454 455

    RP_start_time = user;
    RPe_start_time = elapsed;
456
}
sof's avatar
sof committed
457
#endif /* PROFILING */
458 459 460 461

/* -----------------------------------------------------------------------------
   Called at the end of each Retainer Profiliing
   -------------------------------------------------------------------------- */
462 463 464 465

#ifdef PROFILING
void
stat_endRP(
466 467 468 469 470
  nat retainerGeneration,
#ifdef DEBUG_RETAINER
  nat maxCStackSize,
  int maxStackSize,
#endif
471
  double averageNumVisit)
472
{
Simon Marlow's avatar
Simon Marlow committed
473
    Time user, elapsed;
474
    getProcessTimes( &user, &elapsed );
475 476 477

    RP_tot_time += user - RP_start_time;
    RPe_tot_time += elapsed - RPe_start_time;
478 479 480 481 482 483 484 485 486

  fprintf(prof_file, "Retainer Profiling: %d, at %f seconds\n", 
    retainerGeneration, mut_user_time_during_RP());
#ifdef DEBUG_RETAINER
  fprintf(prof_file, "\tMax C stack size = %u\n", maxCStackSize);
  fprintf(prof_file, "\tMax auxiliary stack size = %u\n", maxStackSize);
#endif
  fprintf(prof_file, "\tAverage number of visits per object = %f\n", averageNumVisit);
}
sof's avatar
sof committed
487
#endif /* PROFILING */
488 489

/* -----------------------------------------------------------------------------
490
   Called at the beginning of each heap census
491
   -------------------------------------------------------------------------- */
492 493
#ifdef PROFILING
void
494
stat_startHeapCensus(void)
495
{
Simon Marlow's avatar
Simon Marlow committed
496
    Time user, elapsed;
497
    getProcessTimes( &user, &elapsed );
498 499 500

    HC_start_time = user;
    HCe_start_time = elapsed;
501
}
sof's avatar
sof committed
502
#endif /* PROFILING */
503 504

/* -----------------------------------------------------------------------------
505
   Called at the end of each heap census
506
   -------------------------------------------------------------------------- */
507 508
#ifdef PROFILING
void
509
stat_endHeapCensus(void) 
510
{
Simon Marlow's avatar
Simon Marlow committed
511
    Time user, elapsed;
512
    getProcessTimes( &user, &elapsed );
513 514 515

    HC_tot_time += user - HC_start_time;
    HCe_tot_time += elapsed - HCe_start_time;
516
}
sof's avatar
sof committed
517
#endif /* PROFILING */
518

519 520 521 522 523 524 525 526
/* -----------------------------------------------------------------------------
   Called at the end of execution

   NOTE: number of allocations is not entirely accurate: it doesn't
   take into account the few bytes at the end of the heap that
   were left unused when the heap-check failed.
   -------------------------------------------------------------------------- */

Simon Marlow's avatar
Simon Marlow committed
527 528 529 530 531 532 533 534 535 536 537 538 539 540 541
#ifdef DEBUG
#define TICK_VAR_INI(arity) \
  StgInt SLOW_CALLS_##arity = 1; \
  StgInt RIGHT_ARITY_##arity = 1; \
  StgInt TAGGED_PTR_##arity = 0;

TICK_VAR_INI(1)
TICK_VAR_INI(2)

StgInt TOTAL_CALLS=1;
#endif

/* Report the value of a counter */
#define REPORT(counter) \
  { \
Ian Lynagh's avatar
Ian Lynagh committed
542
    showStgWord64(counter,temp,rtsTrue/*commas*/); \
Simon Marlow's avatar
Simon Marlow committed
543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561
    statsPrintf("  (" #counter ")  : %s\n",temp);				\
  }

/* Report the value of a counter as a percentage of another counter */
#define REPORT_PCT(counter,countertot) \
  statsPrintf("  (" #counter ") %% of (" #countertot ") : %.1f%%\n", \
	      counter*100.0/countertot)

#define TICK_PRINT(arity) \
  REPORT(SLOW_CALLS_##arity); \
  REPORT_PCT(RIGHT_ARITY_##arity,SLOW_CALLS_##arity); \
  REPORT_PCT(TAGGED_PTR_##arity,RIGHT_ARITY_##arity); \
  REPORT(RIGHT_ARITY_##arity); \
  REPORT(TAGGED_PTR_##arity)

#define TICK_PRINT_TOT(arity) \
  statsPrintf("  (SLOW_CALLS_" #arity ") %% of (TOTAL_CALLS) : %.1f%%\n", \
	      SLOW_CALLS_##arity * 100.0/TOTAL_CALLS)

Simon Marlow's avatar
Simon Marlow committed
562 563
static inline Time get_init_cpu(void) { return end_init_cpu - start_init_cpu; }
static inline Time get_init_elapsed(void) { return end_init_elapsed - start_init_elapsed; }
564

565

566
void
567
stat_exit (void)
568
{
Simon Marlow's avatar
Simon Marlow committed
569
    generation *gen;
Simon Marlow's avatar
Simon Marlow committed
570 571 572 573 574 575 576 577
    Time gc_cpu = 0;
    Time gc_elapsed = 0;
    Time init_cpu = 0;
    Time init_elapsed = 0;
    Time mut_cpu = 0;
    Time mut_elapsed = 0;
    Time exit_cpu = 0;
    Time exit_elapsed = 0;
578 579
    W_ tot_alloc;
    W_ alloc;
Simon Marlow's avatar
Simon Marlow committed
580

581
    if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
582 583

	char temp[BIG_STRING_LEN];
Simon Marlow's avatar
Simon Marlow committed
584 585
	Time tot_cpu;
	Time tot_elapsed;
Simon Marlow's avatar
Simon Marlow committed
586
	nat i, g, total_collections = 0;
587

Simon Marlow's avatar
Simon Marlow committed
588 589
	getProcessTimes( &tot_cpu, &tot_elapsed );
	tot_elapsed -= start_init_elapsed;
590

591
        tot_alloc = calcTotalAllocated();
592

593 594 595
        // allocated since the last GC
        alloc = tot_alloc - GC_tot_alloc;
        GC_tot_alloc = tot_alloc;
596

597 598 599
	/* Count total garbage collections */
	for (g = 0; g < RtsFlags.GcFlags.generations; g++)
	    total_collections += generations[g].collections;
600

Simon Marlow's avatar
Simon Marlow committed
601 602 603
	/* avoid divide by zero if tot_cpu is measured as 0.00 seconds -- SDM */
	if (tot_cpu  == 0.0)  tot_cpu = 1;
	if (tot_elapsed == 0.0) tot_elapsed = 1;
604
	
605
	if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
606
	    statsPrintf("%9" FMT_SizeT " %9.9s %9.9s", (W_)alloc*sizeof(W_), "", "");
607
	    statsPrintf(" %6.3f %6.3f\n\n", 0.0, 0.0);
608 609
	}

Simon Marlow's avatar
Simon Marlow committed
610 611 612 613 614
        for (i = 0; i < RtsFlags.GcFlags.generations; i++) {
            gc_cpu     += GC_coll_cpu[i];
            gc_elapsed += GC_coll_elapsed[i];
        }

615 616
        // heapCensus() is called by the GC, so RP and HC time are
        // included in the GC stats.  We therefore subtract them to
Ian Lynagh's avatar
Ian Lynagh committed
617
        // obtain the actual GC cpu time.
618 619
        gc_cpu     -= PROF_VAL(RP_tot_time + HC_tot_time);
        gc_elapsed -= PROF_VAL(RPe_tot_time + HCe_tot_time);
620

621 622
        init_cpu     = get_init_cpu();
        init_elapsed = get_init_elapsed();
dmp's avatar
dmp committed
623 624 625 626 627 628 629 630 631 632

        exit_cpu     = end_exit_cpu - start_exit_cpu;
        exit_elapsed = end_exit_elapsed - start_exit_elapsed;

        mut_elapsed = start_exit_elapsed - end_init_elapsed - gc_elapsed;

        mut_cpu = start_exit_cpu - end_init_cpu - gc_cpu
            - PROF_VAL(RP_tot_time + HC_tot_time);
        if (mut_cpu < 0) { mut_cpu = 0; }

633
	if (RtsFlags.GcFlags.giveStats >= SUMMARY_GC_STATS) {
Ian Lynagh's avatar
Ian Lynagh committed
634
	    showStgWord64(GC_tot_alloc*sizeof(W_), 
635
				 temp, rtsTrue/*commas*/);
636
	    statsPrintf("%16s bytes allocated in the heap\n", temp);
637

Ian Lynagh's avatar
Ian Lynagh committed
638
	    showStgWord64(GC_tot_copied*sizeof(W_), 
639
				 temp, rtsTrue/*commas*/);
640
	    statsPrintf("%16s bytes copied during GC\n", temp);
641

Simon Marlow's avatar
Simon Marlow committed
642 643
            if ( residency_samples > 0 ) {
		showStgWord64(max_residency*sizeof(W_), 
644
				     temp, rtsTrue/*commas*/);
Ian Lynagh's avatar
Ian Lynagh committed
645
		statsPrintf("%16s bytes maximum residency (%" FMT_Word " sample(s))\n",
Simon Marlow's avatar
Simon Marlow committed
646
			temp, residency_samples);
647
	    }
648

Simon Marlow's avatar
Simon Marlow committed
649
	    showStgWord64(max_slop*sizeof(W_), temp, rtsTrue/*commas*/);
650 651
	    statsPrintf("%16s bytes maximum slop\n", temp);

Ian Lynagh's avatar
Ian Lynagh committed
652
	    statsPrintf("%16" FMT_SizeT " MB total memory in use (%" FMT_SizeT " MB lost due to fragmentation)\n\n", 
653 654
                        (size_t)(peak_mblocks_allocated * MBLOCK_SIZE_W) / (1024 * 1024 / sizeof(W_)),
                        (size_t)(peak_mblocks_allocated * BLOCKS_PER_MBLOCK * BLOCK_SIZE_W - hw_alloc_blocks * BLOCK_SIZE_W) / (1024 * 1024 / sizeof(W_)));
655 656

	    /* Print garbage collections in each gen */
657
            statsPrintf("                                     Tot time (elapsed)  Avg pause  Max pause\n");
Simon Marlow's avatar
Simon Marlow committed
658 659
            for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
                gen = &generations[g];
660
                statsPrintf("  Gen %2d     %5d colls, %5d par   %6.3fs  %6.3fs     %3.4fs    %3.4fs\n",
Simon Marlow's avatar
Simon Marlow committed
661 662 663
                            gen->no,
                            gen->collections,
                            gen->par_collections,
Simon Marlow's avatar
Simon Marlow committed
664 665 666 667
                            TimeToSecondsDbl(GC_coll_cpu[g]),
                            TimeToSecondsDbl(GC_coll_elapsed[g]),
                            gen->collections == 0 ? 0 : TimeToSecondsDbl(GC_coll_elapsed[g] / gen->collections),
                            TimeToSecondsDbl(GC_coll_max_pause[g]));
Simon Marlow's avatar
Simon Marlow committed
668
            }
669

670
#if defined(THREADED_RTS)
671 672 673 674
            if (RtsFlags.ParFlags.parGcEnabled && n_capabilities > 1) {
                statsPrintf("\n  Parallel GC work balance: %.2f%% (serial 0%%, perfect 100%%)\n", 
                            100 * (((double)GC_par_tot_copied / (double)GC_par_max_copied) - 1)
                                / (n_capabilities - 1)
675 676 677
                    );
            }
#endif
Simon Marlow's avatar
Simon Marlow committed
678
            statsPrintf("\n");
679

680
#if defined(THREADED_RTS)
681 682 683 684
            statsPrintf("  TASKS: %d (%d bound, %d peak workers (%d total), using -N%d)\n",
                        taskCount, taskCount - workerCount,
                        peakWorkerCount, workerCount,
                        n_capabilities);
685

686 687
	    statsPrintf("\n");

688 689
            {
                nat i;
690
                SparkCounters sparks = { 0, 0, 0, 0, 0, 0};
691
                for (i = 0; i < n_capabilities; i++) {
692 693 694 695 696 697
                    sparks.created   += capabilities[i]->spark_stats.created;
                    sparks.dud       += capabilities[i]->spark_stats.dud;
                    sparks.overflowed+= capabilities[i]->spark_stats.overflowed;
                    sparks.converted += capabilities[i]->spark_stats.converted;
                    sparks.gcd       += capabilities[i]->spark_stats.gcd;
                    sparks.fizzled   += capabilities[i]->spark_stats.fizzled;
698 699
                }

Ian Lynagh's avatar
Ian Lynagh committed
700
                statsPrintf("  SPARKS: %" FMT_Word " (%" FMT_Word " converted, %" FMT_Word " overflowed, %" FMT_Word " dud, %" FMT_Word " GC'd, %" FMT_Word " fizzled)\n\n",
701 702
                            sparks.created + sparks.dud + sparks.overflowed,
                            sparks.converted, sparks.overflowed, sparks.dud,
703
                            sparks.gcd, sparks.fizzled);
704
            }
705 706
#endif

707
	    statsPrintf("  INIT    time  %7.3fs  (%7.3fs elapsed)\n",
Simon Marlow's avatar
Simon Marlow committed
708
                        TimeToSecondsDbl(init_cpu), TimeToSecondsDbl(init_elapsed));
Simon Marlow's avatar
Simon Marlow committed
709

710
            statsPrintf("  MUT     time  %7.3fs  (%7.3fs elapsed)\n",
Simon Marlow's avatar
Simon Marlow committed
711
                        TimeToSecondsDbl(mut_cpu), TimeToSecondsDbl(mut_elapsed));
712
            statsPrintf("  GC      time  %7.3fs  (%7.3fs elapsed)\n",
Simon Marlow's avatar
Simon Marlow committed
713
                        TimeToSecondsDbl(gc_cpu), TimeToSecondsDbl(gc_elapsed));
Simon Marlow's avatar
Simon Marlow committed
714

715
#ifdef PROFILING
716
	    statsPrintf("  RP      time  %7.3fs  (%7.3fs elapsed)\n",
Simon Marlow's avatar
Simon Marlow committed
717
		    TimeToSecondsDbl(RP_tot_time), TimeToSecondsDbl(RPe_tot_time));
718
	    statsPrintf("  PROF    time  %7.3fs  (%7.3fs elapsed)\n",
Simon Marlow's avatar
Simon Marlow committed
719
		    TimeToSecondsDbl(HC_tot_time), TimeToSecondsDbl(HCe_tot_time));
720
#endif 
721
	    statsPrintf("  EXIT    time  %7.3fs  (%7.3fs elapsed)\n",
Simon Marlow's avatar
Simon Marlow committed
722
		    TimeToSecondsDbl(exit_cpu), TimeToSecondsDbl(exit_elapsed));
723
	    statsPrintf("  Total   time  %7.3fs  (%7.3fs elapsed)\n\n",
Simon Marlow's avatar
Simon Marlow committed
724
		    TimeToSecondsDbl(tot_cpu), TimeToSecondsDbl(tot_elapsed));
Simon Marlow's avatar
Simon Marlow committed
725 726
#ifndef THREADED_RTS
	    statsPrintf("  %%GC     time     %5.1f%%  (%.1f%% elapsed)\n\n",
Simon Marlow's avatar
Simon Marlow committed
727 728
		    TimeToSecondsDbl(gc_cpu)*100/TimeToSecondsDbl(tot_cpu),
		    TimeToSecondsDbl(gc_elapsed)*100/TimeToSecondsDbl(tot_elapsed));
Simon Marlow's avatar
Simon Marlow committed
729 730
#endif

Simon Marlow's avatar
Simon Marlow committed
731
            if (mut_cpu == 0) {
Ian Lynagh's avatar
Ian Lynagh committed
732
		showStgWord64(0, temp, rtsTrue/*commas*/);
Simon Marlow's avatar
Simon Marlow committed
733
            } else {
Ian Lynagh's avatar
Ian Lynagh committed
734
		showStgWord64(
Simon Marlow's avatar
Simon Marlow committed
735 736 737 738
                    (StgWord64)((GC_tot_alloc*sizeof(W_)) / TimeToSecondsDbl(mut_cpu)),
                    temp, rtsTrue/*commas*/);
            }

739
	    statsPrintf("  Alloc rate    %s bytes per MUT second\n\n", temp);
740
	
741
	    statsPrintf("  Productivity %5.1f%% of total user, %.1f%% of total elapsed\n\n",
Simon Marlow's avatar
Simon Marlow committed
742
                    TimeToSecondsDbl(tot_cpu - gc_cpu -
Simon Marlow's avatar
Simon Marlow committed
743
				PROF_VAL(RP_tot_time + HC_tot_time) - init_cpu) * 100 
Simon Marlow's avatar
Simon Marlow committed
744
		    / TimeToSecondsDbl(tot_cpu), 
Simon Marlow's avatar
Simon Marlow committed
745 746 747
                    TimeToSecondsDbl(tot_cpu - gc_cpu -
                                PROF_VAL(RP_tot_time + HC_tot_time) - init_cpu) * 100
                    / TimeToSecondsDbl(tot_elapsed));
Simon Marlow's avatar
Simon Marlow committed
748 749 750 751 752 753 754 755 756

            /*
            TICK_PRINT(1);
            TICK_PRINT(2);
	    REPORT(TOTAL_CALLS);
            TICK_PRINT_TOT(1);
            TICK_PRINT_TOT(2);
            */

mrchebas@gmail.com's avatar
mrchebas@gmail.com committed
757
#if USE_PAPI
758
            papi_stats_report();
759 760 761
#endif
#if defined(THREADED_RTS) && defined(PROF_SPIN)
            {
Simon Marlow's avatar
Simon Marlow committed
762
                nat g;
763 764
                
                statsPrintf("gc_alloc_block_sync: %"FMT_Word64"\n", gc_alloc_block_sync.spin);
765
                statsPrintf("whitehole_spin: %"FMT_Word64"\n", whitehole_spin);
766
                for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
Simon Marlow's avatar
Simon Marlow committed
767
                    statsPrintf("gen[%d].sync: %"FMT_Word64"\n", g, generations[g].sync.spin);
768 769
                }
            }
mrchebas@gmail.com's avatar
mrchebas@gmail.com committed
770
#endif
771 772
	}

773
	if (RtsFlags.GcFlags.giveStats == ONELINE_GC_STATS) {
Ian Lynagh's avatar
Ian Lynagh committed
774 775 776 777 778 779 780 781
      char *fmt1, *fmt2;
      if (RtsFlags.MiscFlags.machineReadable) {
          fmt1 = " [(\"bytes allocated\", \"%llu\")\n";
          fmt2 = " ,(\"num_GCs\", \"%d\")\n"
                 " ,(\"average_bytes_used\", \"%ld\")\n"
                 " ,(\"max_bytes_used\", \"%ld\")\n"
                 " ,(\"num_byte_usage_samples\", \"%ld\")\n"
                 " ,(\"peak_megabytes_allocated\", \"%lu\")\n"
782 783 784 785 786 787
                 " ,(\"init_cpu_seconds\", \"%.3f\")\n"
                 " ,(\"init_wall_seconds\", \"%.3f\")\n"
                 " ,(\"mutator_cpu_seconds\", \"%.3f\")\n"
                 " ,(\"mutator_wall_seconds\", \"%.3f\")\n"
                 " ,(\"GC_cpu_seconds\", \"%.3f\")\n"
                 " ,(\"GC_wall_seconds\", \"%.3f\")\n"
Ian Lynagh's avatar
Ian Lynagh committed
788 789 790 791
                 " ]\n";
      }
      else {
          fmt1 = "<<ghc: %llu bytes, ";
792
          fmt2 = "%d GCs, %ld/%ld avg/max bytes residency (%ld samples), %luM in use, %.3f INIT (%.3f elapsed), %.3f MUT (%.3f elapsed), %.3f GC (%.3f elapsed) :ghc>>\n";
Ian Lynagh's avatar
Ian Lynagh committed
793
      }
rrt's avatar
rrt committed
794
	  /* print the long long separately to avoid bugginess on mingwin (2001-07-02, mingw-0.5) */
Ian Lynagh's avatar
Ian Lynagh committed
795
	  statsPrintf(fmt1, GC_tot_alloc*(StgWord64)sizeof(W_));
Ian Lynagh's avatar
Ian Lynagh committed
796
	  statsPrintf(fmt2,
rrt's avatar
rrt committed
797
		    total_collections,
Simon Marlow's avatar
Simon Marlow committed
798
		    residency_samples == 0 ? 0 : 
799
		        cumulative_residency*sizeof(W_)/residency_samples, 
Simon Marlow's avatar
Simon Marlow committed
800 801
		    max_residency*sizeof(W_), 
		    residency_samples,
802
		    (unsigned long)(peak_mblocks_allocated * MBLOCK_SIZE / (1024L * 1024L)),
Simon Marlow's avatar
Simon Marlow committed
803 804 805
		    TimeToSecondsDbl(init_cpu), TimeToSecondsDbl(init_elapsed),
		    TimeToSecondsDbl(mut_cpu), TimeToSecondsDbl(mut_elapsed),
		    TimeToSecondsDbl(gc_cpu), TimeToSecondsDbl(gc_elapsed));
806
	}
807

808 809
	statsFlush();
	statsClose();
810
    }
simonmarhaskell@gmail.com's avatar
simonmarhaskell@gmail.com committed
811

Simon Marlow's avatar
Simon Marlow committed
812
    if (GC_coll_cpu) {
Simon Marlow's avatar
Simon Marlow committed
813
      stgFree(GC_coll_cpu);
Simon Marlow's avatar
Simon Marlow committed
814 815 816
      GC_coll_cpu = NULL;
    }
    if (GC_coll_elapsed) {
Simon Marlow's avatar
Simon Marlow committed
817
      stgFree(GC_coll_elapsed);
Simon Marlow's avatar
Simon Marlow committed
818 819 820 821 822 823
      GC_coll_elapsed = NULL;
    }
    if (GC_coll_max_pause) {
      stgFree(GC_coll_max_pause);
      GC_coll_max_pause = NULL;
    }
824
}
825 826 827 828 829 830 831

/* -----------------------------------------------------------------------------
   stat_describe_gens

   Produce some detailed info on the state of the generational GC.
   -------------------------------------------------------------------------- */
void
832
statDescribeGens(void)
833
{
834
  nat g, mut, lge, i;
835 836 837
  W_ gen_slop;
  W_ tot_live, tot_slop;
  W_ gen_live, gen_blocks;
838
  bdescr *bd;
Simon Marlow's avatar
Simon Marlow committed
839 840
  generation *gen;
  
841
  debugBelch(
Simon Marlow's avatar
Simon Marlow committed
842 843 844 845
"----------------------------------------------------------\n"
"  Gen     Max  Mut-list  Blocks    Large     Live     Slop\n"
"       Blocks     Bytes          Objects                  \n"
"----------------------------------------------------------\n");
846

847 848
  tot_live = 0;
  tot_slop = 0;
Simon Marlow's avatar
Simon Marlow committed
849

850
  for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
Simon Marlow's avatar
Simon Marlow committed
851 852 853 854 855 856 857 858 859
      gen = &generations[g];

      for (bd = gen->large_objects, lge = 0; bd; bd = bd->link) {
          lge++;
      }

      gen_live   = genLiveWords(gen);
      gen_blocks = genLiveBlocks(gen);

860
      mut = 0;
861
      for (i = 0; i < n_capabilities; i++) {
862
          mut += countOccupied(capabilities[i]->mut_lists[g]);
863 864

          // Add the pinned object block.
865
          bd = capabilities[i]->pinned_object_block;
866 867 868 869 870
          if (bd != NULL) {
              gen_live   += bd->free - bd->start;
              gen_blocks += bd->blocks;
          }

Simon Marlow's avatar
Simon Marlow committed
871 872
          gen_live   += gcThreadLiveWords(i,g);
          gen_blocks += gcThreadLiveBlocks(i,g);
873 874
      }

Ian Lynagh's avatar
Ian Lynagh committed
875
      debugBelch("%5d %7" FMT_Word " %9d", g, (W_)gen->max_blocks, mut);
Simon Marlow's avatar
Simon Marlow committed
876

Simon Marlow's avatar
Simon Marlow committed
877 878
      gen_slop = gen_blocks * BLOCK_SIZE_W - gen_live;

Ian Lynagh's avatar
Ian Lynagh committed
879 880
      debugBelch("%8" FMT_Word " %8d %8" FMT_Word " %8" FMT_Word "\n", gen_blocks, lge,
                 gen_live*(W_)sizeof(W_), gen_slop*(W_)sizeof(W_));
Simon Marlow's avatar
Simon Marlow committed
881 882
      tot_live += gen_live;
      tot_slop += gen_slop;
883
  }
Simon Marlow's avatar
Simon Marlow committed
884
  debugBelch("----------------------------------------------------------\n");
Ian Lynagh's avatar
Ian Lynagh committed
885 886
  debugBelch("%41s%8" FMT_SizeT " %8" FMT_SizeT "\n",
             "",tot_live*sizeof(W_),tot_slop*sizeof(W_));
Simon Marlow's avatar
Simon Marlow committed
887
  debugBelch("----------------------------------------------------------\n");
888
  debugBelch("\n");
889
}
890 891 892 893 894 895

/* -----------------------------------------------------------------------------
   Stats available via a programmatic interface, so eg. GHCi can time
   each compilation and expression evaluation.
   -------------------------------------------------------------------------- */

896
extern HsInt64 getAllocations( void ) 
897
{ return (HsInt64)GC_tot_alloc * sizeof(W_); }
898

899 900
/* EZY: I'm not convinced I got all the casting right. */

pcapriotti's avatar
pcapriotti committed
901 902 903 904 905
extern rtsBool getGCStatsEnabled( void )
{
    return RtsFlags.GcFlags.giveStats != NO_GC_STATS;
}

906 907 908 909
extern void getGCStats( GCStats *s )
{
    nat total_collections = 0;
    nat g;
Simon Marlow's avatar
Simon Marlow committed
910 911 912 913
    Time gc_cpu = 0;
    Time gc_elapsed = 0;
    Time current_elapsed = 0;
    Time current_cpu = 0;
914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936

    getProcessTimes(&current_cpu, &current_elapsed);

    /* EZY: static inline'ify these */
    for (g = 0; g < RtsFlags.GcFlags.generations; g++)
        total_collections += generations[g].collections;

    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
        gc_cpu     += GC_coll_cpu[g];
        gc_elapsed += GC_coll_elapsed[g];
    }

    s->bytes_allocated = GC_tot_alloc*(StgWord64)sizeof(W_);
    s->num_gcs = total_collections;
    s->num_byte_usage_samples = residency_samples;
    s->max_bytes_used = max_residency*sizeof(W_);
    s->cumulative_bytes_used =