X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FStats.c;h=5d4f772396c4fad5720e640b678621d4b0caf4eb;hb=5cb4bb13a817c44cdc4369c7f82949d9490d69a0;hp=ceaaaa770abca3d09a9a35764990c3abad237c20;hpb=45f650eab03f1b50fd5f8b1bad5635902ed5192a;p=ghc-hetmet.git diff --git a/ghc/rts/Stats.c b/ghc/rts/Stats.c index ceaaaa7..5d4f772 100644 --- a/ghc/rts/Stats.c +++ b/ghc/rts/Stats.c @@ -1,5 +1,5 @@ /* ----------------------------------------------------------------------------- - * $Id: Stats.c,v 1.23 2000/12/19 14:30:17 simonmar Exp $ + * $Id: Stats.c,v 1.46 2002/08/19 16:02:26 simonmar Exp $ * * (c) The GHC Team, 1998-1999 * @@ -7,7 +7,9 @@ * * ---------------------------------------------------------------------------*/ -#define NON_POSIX_SOURCE +/* Alas, no. This source is non-posix. + #include "PosixSource.h" +*/ #include "Rts.h" #include "RtsFlags.h" @@ -16,6 +18,8 @@ #include "MBlock.h" #include "Schedule.h" #include "Stats.h" +#include "ParTicky.h" /* ToDo: move into Rts.h */ +#include "Profiling.h" #ifdef HAVE_UNISTD_H #include @@ -55,6 +59,10 @@ #include #endif +#if defined(PAR) || !(!defined(HAVE_GETRUSAGE) || irix_TARGET_OS || defined(mingw32_TARGET_OS) || defined(cygwin32_TARGET_OS)) +#include +#endif + /* huh? */ #define BIG_STRING_LEN 512 @@ -86,12 +94,26 @@ static TICK_TYPE ExitElapsedTime = 0; static ullong GC_tot_alloc = 0; static ullong GC_tot_copied = 0; -static TICK_TYPE GC_start_time, GC_tot_time = 0; /* User GC Time */ -static TICK_TYPE GCe_start_time, GCe_tot_time = 0; /* Elapsed GC time */ +static TICK_TYPE GC_start_time = 0, GC_tot_time = 0; /* User GC Time */ +static TICK_TYPE GCe_start_time = 0, GCe_tot_time = 0; /* Elapsed GC time */ -lnat MaxResidency = 0; /* in words; for stats only */ -lnat AvgResidency = 0; -lnat ResidencySamples = 0; /* for stats only */ +#ifdef PROFILING +static TICK_TYPE RP_start_time = 0, RP_tot_time = 0; /* retainer prof user time */ +static TICK_TYPE RPe_start_time = 0, RPe_tot_time = 0; /* retainer prof elap time */ + +static TICK_TYPE HC_start_time, HC_tot_time = 0; // heap census prof user time +static TICK_TYPE HCe_start_time, HCe_tot_time = 0; // heap census prof elap time +#endif + +#ifdef PROFILING +#define PROF_VAL(x) (x) +#else +#define PROF_VAL(x) 0 +#endif + +static lnat MaxResidency = 0; // in words; for stats only +static lnat AvgResidency = 0; +static lnat ResidencySamples = 0; // for stats only static lnat GC_start_faults = 0, GC_end_faults = 0; @@ -102,53 +124,84 @@ static nat pageFaults(void); /* elapsedtime() -- The current elapsed time in seconds */ -#ifdef _WIN32 -#define NS_PER_SEC 10000000LL -/* Convert FILETIMEs into secs since the Epoch (Jan1-1970) */ +#if defined(mingw32_TARGET_OS) || defined(cygwin32_TARGET_OS) +#define HNS_PER_SEC 10000000LL /* FILETIMES are in units of 100ns */ +/* Convert FILETIMEs into secs */ #define FT2longlong(ll,ft) \ (ll)=(ft).dwHighDateTime; \ (ll) <<= 32; \ (ll) |= (ft).dwLowDateTime; \ - (ll) /= (unsigned long long) (NS_PER_SEC / CLOCKS_PER_SEC) + (ll) /= (unsigned long long) (HNS_PER_SEC / CLOCKS_PER_SEC) #endif -#ifdef _WIN32 +#if defined(mingw32_TARGET_OS) || defined(cygwin32_TARGET_OS) /* cygwin32 or mingw32 version */ static void getTimes(void) { - FILETIME creationTime, exitTime, kernelTime, userTime; + static int is_win9x = -1; + + FILETIME creationTime, exitTime, userTime, kernelTime = {0,0}; long long int kT, uT; + + if (is_win9x < 0) { + /* figure out whether we're on a Win9x box or not. */ + OSVERSIONINFO oi; + BOOL b; + + /* Need to init the size field first.*/ + oi.dwOSVersionInfoSize = sizeof(OSVERSIONINFO); + b = GetVersionEx(&oi); + + is_win9x = ( (b && (oi.dwPlatformId & VER_PLATFORM_WIN32_WINDOWS)) ? 1 : 0); + } - /* ToDo: pin down elapsed times to just the OS thread(s) that - are evaluating/managing Haskell code. - */ - if (!GetProcessTimes (GetCurrentProcess(), &creationTime, + if (is_win9x) { + /* On Win9x, just attribute all running time to the user. */ + SYSTEMTIME st; + + GetSystemTime(&st); + SystemTimeToFileTime(&st,&userTime); + } else { + /* ToDo: pin down elapsed times to just the OS thread(s) that + are evaluating/managing Haskell code. + */ + if (!GetProcessTimes (GetCurrentProcess(), &creationTime, &exitTime, &kernelTime, &userTime)) { /* Probably on a Win95 box..*/ - return 0; + CurrentElapsedTime = 0; + CurrentUserTime = 0; + return; + } } FT2longlong(kT,kernelTime); FT2longlong(uT,userTime); CurrentElapsedTime = uT + kT; CurrentUserTime = uT; + + if (is_win9x) { + /* Adjust for the fact that we're using system time & not + process time on Win9x. */ + CurrentUserTime -= ElapsedTimeStart; + CurrentElapsedTime -= ElapsedTimeStart; + } } -#else /* !_WIN32 */ +#else /* !win32 */ static void getTimes(void) { -# if !defined(HAVE_TIMES) +#ifndef HAVE_TIMES /* We will #ifdef around the fprintf for machines we *know* are unsupported. (WDP 94/05) */ fprintf(stderr, "NOTE: `getTimes' does nothing!\n"); return 0.0; -# else /* not stumped */ +#else /* not stumped */ struct tms t; clock_t r = times(&t); @@ -157,7 +210,7 @@ getTimes(void) #endif } -#endif /* !_WIN32 */ +#endif /* !win32 */ /* mut_user_time_during_GC() and mut_user_time() * @@ -172,24 +225,43 @@ getTimes(void) * stat_startGC() for details) */ double -mut_user_time_during_GC(void) +mut_user_time_during_GC( void ) { - return ((double)GC_start_time - (double)GC_tot_time); + return TICK_TO_DBL(GC_start_time - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time)); } double -mut_user_time(void) +mut_user_time( void ) { getTimes(); - return ((double)CurrentUserTime - (double)GC_tot_time); + return TICK_TO_DBL(CurrentUserTime - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time)); } +#ifdef PROFILING +/* + mut_user_time_during_RP() is similar to mut_user_time_during_GC(); + it returns the MUT time during retainer profiling. + The same is for mut_user_time_during_HC(); + */ +double +mut_user_time_during_RP( void ) +{ + return TICK_TO_DBL(RP_start_time - GC_tot_time - RP_tot_time - HC_tot_time); +} + +double +mut_user_time_during_heap_census( void ) +{ + return TICK_TO_DBL(HC_start_time - GC_tot_time - RP_tot_time - HC_tot_time); +} +#endif /* PROFILING */ + static nat pageFaults(void) { /* ToDo (on NT): better, get this via the performance data that's stored in the registry. */ -# if !defined(HAVE_GETRUSAGE) || irix_TARGET_OS || defined(_WIN32) +# if !defined(HAVE_GETRUSAGE) || irix_TARGET_OS || defined(mingw32_TARGET_OS) || defined(cygwin32_TARGET_OS) return 0; # else struct rusage t; @@ -237,20 +309,21 @@ stat_startInit(void) fprintf(stderr, "stat_init: bad call to 'sysconf'!\n"); stg_exit(EXIT_FAILURE); } - TicksPerSecond = (double) ticks; + TicksPerSecond = ticks; /* no "sysconf" or CLK_TCK; had better guess */ #elif defined(HZ) - TicksPerSecond = (StgDouble) (HZ); + TicksPerSecond = HZ; #elif defined(CLOCKS_PER_SEC) - TicksPerSecond = (StgDouble) (CLOCKS_PER_SEC); + TicksPerSecond = CLOCKS_PER_SEC; + #else /* had better guess wildly */ /* We will #ifdef around the fprintf for machines we *know* are unsupported. (WDP 94/05) */ fprintf(stderr, "NOTE: Guessing `TicksPerSecond = 60'!\n"); - TicksPerSecond = 60.0; + TicksPerSecond = 60; #endif getTimes(); @@ -281,9 +354,10 @@ stat_startExit(void) { getTimes(); MutElapsedStamp = CurrentElapsedTime; - MutElapsedTime = CurrentElapsedTime - GCe_tot_time - InitElapsedStamp; + MutElapsedTime = CurrentElapsedTime - GCe_tot_time - + PROF_VAL(RPe_tot_time + HCe_tot_time) - InitElapsedStamp; if (MutElapsedTime < 0) { MutElapsedTime = 0; } /* sometimes -0.00 */ - + /* for SMP, we don't know the mutator time yet, we have to inspect * all the running threads to find out, and they haven't stopped * yet. So we just timestamp MutUserTime at this point so we can @@ -293,7 +367,7 @@ stat_startExit(void) #ifdef SMP MutUserTime = CurrentUserTime; #else - MutUserTime = CurrentUserTime - GC_tot_time - InitUserTime; + MutUserTime = CurrentUserTime - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime; if (MutUserTime < 0) { MutUserTime = 0; } #endif } @@ -305,7 +379,7 @@ stat_endExit(void) #ifdef SMP ExitUserTime = CurrentUserTime - MutUserTime; #else - ExitUserTime = CurrentUserTime - MutUserTime - GC_tot_time - InitUserTime; + ExitUserTime = CurrentUserTime - MutUserTime - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime; #endif ExitElapsedTime = CurrentElapsedTime - MutElapsedStamp; if (ExitUserTime < 0) { @@ -432,23 +506,92 @@ stat_endGC(lnat alloc, lnat collect, lnat live, lnat copied, lnat gen) } /* ----------------------------------------------------------------------------- + Called at the beginning of each Retainer Profiliing + -------------------------------------------------------------------------- */ +#ifdef PROFILING +void +stat_startRP(void) +{ + getTimes(); + RP_start_time = CurrentUserTime; + RPe_start_time = CurrentElapsedTime; +} +#endif /* PROFILING */ + +/* ----------------------------------------------------------------------------- + Called at the end of each Retainer Profiliing + -------------------------------------------------------------------------- */ + +#ifdef PROFILING +void +stat_endRP( + nat retainerGeneration, +#ifdef DEBUG_RETAINER + nat maxCStackSize, + int maxStackSize, +#endif + double averageNumVisit) +{ + getTimes(); + RP_tot_time += CurrentUserTime - RP_start_time; + RPe_tot_time += CurrentElapsedTime - RPe_start_time; + + 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); +} +#endif /* PROFILING */ + +/* ----------------------------------------------------------------------------- + Called at the beginning of each heap census + -------------------------------------------------------------------------- */ +#ifdef PROFILING +void +stat_startHeapCensus(void) +{ + getTimes(); + HC_start_time = CurrentUserTime; + HCe_start_time = CurrentElapsedTime; +} +#endif /* PROFILING */ + +/* ----------------------------------------------------------------------------- + Called at the end of each heap census + -------------------------------------------------------------------------- */ +#ifdef PROFILING +void +stat_endHeapCensus(void) +{ + getTimes(); + HC_tot_time += CurrentUserTime - HC_start_time; + HCe_tot_time += CurrentElapsedTime - HCe_start_time; +} +#endif /* PROFILING */ + +/* ----------------------------------------------------------------------------- stat_workerStop Called under SMP when a worker thread finishes. We drop the timing stats for this thread into the task_ids struct for that thread. -------------------------------------------------------------------------- */ -#ifdef SMP +#if defined(SMP) void stat_workerStop(void) { nat i; pthread_t me = pthread_self(); + getTimes(); + for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) { if (task_ids[i].id == me) { - task_ids[i].mut_time = usertime() - task_ids[i].gc_time; - task_ids[i].mut_etime = elapsedtime() + task_ids[i].mut_time = CurrentUserTime - task_ids[i].gc_time; + task_ids[i].mut_etime = CurrentElapsedTime - GCe_tot_time - task_ids[i].elapsedtimestart; if (task_ids[i].mut_time < 0.0) { task_ids[i].mut_time = 0.0; } @@ -458,6 +601,14 @@ stat_workerStop(void) } #endif +#if defined(SMP) +long int stat_getElapsedTime () +{ + getTimes(); + return CurrentElapsedTime; +} +#endif + /* ----------------------------------------------------------------------------- Called at the end of execution @@ -559,45 +710,59 @@ stat_exit(int alloc) TICK_TO_DBL(MutUserTime), TICK_TO_DBL(MutElapsedTime)); fprintf(sf, " GC time %6.2fs (%6.2fs elapsed)\n", TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time)); +#ifdef PROFILING + fprintf(sf, " RP time %6.2fs (%6.2fs elapsed)\n", + TICK_TO_DBL(RP_tot_time), TICK_TO_DBL(RPe_tot_time)); + fprintf(sf, " PROF time %6.2fs (%6.2fs elapsed)\n", + TICK_TO_DBL(HC_tot_time), TICK_TO_DBL(HCe_tot_time)); +#endif fprintf(sf, " EXIT time %6.2fs (%6.2fs elapsed)\n", TICK_TO_DBL(ExitUserTime), TICK_TO_DBL(ExitElapsedTime)); fprintf(sf, " Total time %6.2fs (%6.2fs elapsed)\n\n", TICK_TO_DBL(time), TICK_TO_DBL(etime)); fprintf(sf, " %%GC time %5.1f%% (%.1f%% elapsed)\n\n", - TICK_TO_DBL(GC_tot_time)*100/time, - TICK_TO_DBL(GCe_tot_time)*100/etime); + TICK_TO_DBL(GC_tot_time)*100/TICK_TO_DBL(time), + TICK_TO_DBL(GCe_tot_time)*100/TICK_TO_DBL(etime)); - if (time - GC_tot_time == 0) + if (time - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time) == 0) ullong_format_string(0, temp, rtsTrue/*commas*/); else ullong_format_string( (ullong)((GC_tot_alloc*sizeof(W_))/ - TICK_TO_DBL(time - GC_tot_time)), + TICK_TO_DBL(time - GC_tot_time - + PROF_VAL(RP_tot_time + HC_tot_time))), temp, rtsTrue/*commas*/); fprintf(sf, " Alloc rate %s bytes per MUT second\n\n", temp); fprintf(sf, " Productivity %5.1f%% of total user, %.1f%% of total elapsed\n\n", - TICK_TO_DBL(time - GC_tot_time - InitUserTime) * 100 + TICK_TO_DBL(time - GC_tot_time - + PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime) * 100 / TICK_TO_DBL(time), - TICK_TO_DBL(time - GC_tot_time - InitUserTime) * 100 + TICK_TO_DBL(time - GC_tot_time - + PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime) * 100 / TICK_TO_DBL(etime)); } if (RtsFlags.GcFlags.giveStats == ONELINE_GC_STATS && sf != NULL) { - fprintf(sf, "<>\n", - GC_tot_alloc*sizeof(W_), total_collections, - AvgResidency*sizeof(W_)/ResidencySamples, + /* print the long long separately to avoid bugginess on mingwin (2001-07-02, mingw-0.5) */ + fprintf(sf, "<>\n", + total_collections, + ResidencySamples == 0 ? 0 : + AvgResidency*sizeof(W_)/ResidencySamples, MaxResidency*sizeof(W_), - ResidencySamples, - mblocks_allocated * MBLOCK_SIZE / (1024 * 1024), + ResidencySamples, + (unsigned long)(mblocks_allocated * MBLOCK_SIZE / (1024L * 1024L)), TICK_TO_DBL(InitUserTime), TICK_TO_DBL(InitElapsedTime), TICK_TO_DBL(MutUserTime), TICK_TO_DBL(MutElapsedTime), TICK_TO_DBL(GC_tot_time), TICK_TO_DBL(GCe_tot_time)); } fflush(sf); - fclose(sf); + if (sf != stderr) { + fclose(sf); + } } } @@ -606,15 +771,16 @@ stat_exit(int alloc) Produce some detailed info on the state of the generational GC. -------------------------------------------------------------------------- */ +#ifdef DEBUG void -stat_describe_gens(void) +statDescribeGens(void) { nat g, s, mut, mut_once, lge, live; StgMutClosure *m; bdescr *bd; step *step; - fprintf(stderr, " Gen Steps Max Mutable Mut-Once Step Blocks Live Large\n Blocks Closures Closures Objects\n"); + fprintf(stderr, " Gen Steps Max Mutable Mut-Once Step Blocks Live Large\n Blocks Closures Closures Objects\n"); for (g = 0; g < RtsFlags.GcFlags.generations; g++) { for (m = generations[g].mut_list, mut = 0; m != END_MUT_LIST; @@ -632,7 +798,7 @@ stat_describe_gens(void) lge++; live = 0; if (RtsFlags.GcFlags.generations == 1) { - bd = step->to_space; + bd = step->to_blocks; } else { bd = step->blocks; } @@ -648,6 +814,7 @@ stat_describe_gens(void) } fprintf(stderr,"\n"); } +#endif /* ----------------------------------------------------------------------------- Stats available via a programmatic interface, so eg. GHCi can time