X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/3f87b9e5de9b132f1141e59f76377db534ad6801..a5e649421a14d09f23290b79ebe2e61066f8efde:/src/smpi/smpi_bench.c diff --git a/src/smpi/smpi_bench.c b/src/smpi/smpi_bench.c index ad115e170e..089275297c 100644 --- a/src/smpi/smpi_bench.c +++ b/src/smpi/smpi_bench.c @@ -4,18 +4,21 @@ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ -#include // sqrt #include "private.h" #include "xbt/dict.h" #include "xbt/sysdep.h" #include "xbt/ex.h" #include "surf/surf.h" +#include "simgrid/sg_config.h" +#ifndef WIN32 #include +#endif #include #include #include #include +#include // sqrt #include #include #include @@ -82,6 +85,7 @@ static size_t shm_size(int fd) { return (size_t)st.st_size; } +#ifndef WIN32 static void* shm_map(int fd, size_t size, shared_data_t* data) { void* mem; char loc[PTR_STRLEN]; @@ -92,6 +96,7 @@ static void* shm_map(int fd, size_t size, shared_data_t* data) { xbt_die("Could not truncate fd %d to %zu: %s", fd, size, strerror(errno)); } } + mem = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); if(mem == MAP_FAILED) { xbt_die("Could not map fd %d: %s", fd, strerror(errno)); @@ -107,6 +112,7 @@ static void* shm_map(int fd, size_t size, shared_data_t* data) { XBT_DEBUG("MMAP %zu to %p", size, mem); return mem; } +#endif void smpi_bench_destroy(void) { @@ -115,8 +121,7 @@ void smpi_bench_destroy(void) xbt_dict_free(&calls); } -static void smpi_execute_flops(double flops) -{ +void smpi_execute_flops(double flops) { smx_action_t action; smx_host_t host; host = SIMIX_host_self(); @@ -132,20 +137,19 @@ static void smpi_execute_flops(double flops) static void smpi_execute(double duration) { /* FIXME: a global variable would be less expensive to consult than a call to xbt_cfg_get_double() right on the critical path */ - if (duration >= xbt_cfg_get_double(_surf_cfg_set, "smpi/cpu_threshold")) { + if (duration >= sg_cfg_get_double("smpi/cpu_threshold")) { XBT_DEBUG("Sleep for %f to handle real computation time", duration); smpi_execute_flops(duration * - xbt_cfg_get_double(_surf_cfg_set, - "smpi/running_power")); + sg_cfg_get_double("smpi/running_power")); } else { XBT_DEBUG("Real computation took %f while option smpi/cpu_threshold is set to %f => ignore it", - duration, xbt_cfg_get_double(_surf_cfg_set, "smpi/cpu_threshold")); + duration, sg_cfg_get_double("smpi/cpu_threshold")); } } void smpi_bench_begin(void) { - xbt_os_timer_start(smpi_process_timer()); + xbt_os_threadtimer_start(smpi_process_timer()); smpi_current_rank = smpi_process_index(); } @@ -153,26 +157,30 @@ void smpi_bench_end(void) { xbt_os_timer_t timer = smpi_process_timer(); - xbt_os_timer_stop(timer); + xbt_os_threadtimer_stop(timer); smpi_execute(xbt_os_timer_elapsed(timer)); } unsigned int smpi_sleep(unsigned int secs) { smpi_bench_end(); - smpi_execute((double) secs); + smpi_execute_flops((double) secs*simcall_host_get_speed(SIMIX_host_self())); smpi_bench_begin(); return secs; } -int smpi_gettimeofday(struct timeval *tv, struct timezone *tz) +int smpi_gettimeofday(struct timeval *tv) { double now; smpi_bench_end(); now = SIMIX_get_clock(); if (tv) { tv->tv_sec = (time_t)now; +#ifdef WIN32 + tv->tv_usec = (useconds_t)((now - tv->tv_sec) * 1e6); +#else tv->tv_usec = (suseconds_t)((now - tv->tv_sec) * 1e6); +#endif } smpi_bench_begin(); return 0; @@ -218,8 +226,13 @@ static char *sample_location(int global, const char *file, int line) { } } static int sample_enough_benchs(local_data_t *data) { - int res = (data->iters > 0 && data->count >= data->iters) - || (data->count > 2 && data->threshold > 0.0 && data->relstderr <= data->threshold); + int res = data->count >= data->iters; + if (data->threshold>0.0) { + if (data->count <2) + res = 0; // not enough data + if (data->relstderr > data->threshold) + res = 0; // stderr too high yet + } XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)", (res?"enough benchs":"need more data"), data->count, data->iters, data->relstderr, data->threshold, data->mean); @@ -237,6 +250,8 @@ void smpi_sample_1(int global, const char *file, int line, int iters, double thr data = xbt_dict_get_or_null(samples, loc); if (!data) { + xbt_assert(threshold>0 || iters>0, + "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)"); data = (local_data_t *) xbt_new(local_data_t, 1); data->count = 0; data->sum = 0.0; @@ -258,7 +273,7 @@ void smpi_sample_1(int global, const char *file, int line, int iters, double thr data->benching = !sample_enough_benchs(data); XBT_DEBUG("XXXX Re-entering the benched nest %s. %s",loc, (data->benching?"more benching needed":"we have enough data, skip computes")); } - free(loc); + xbt_free(loc); } int smpi_sample_2(int global, const char *file, int line) @@ -269,7 +284,7 @@ int smpi_sample_2(int global, const char *file, int line) xbt_assert(samples, "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!"); data = xbt_dict_get(samples, loc); XBT_DEBUG("sample2 %s",loc); - free(loc); + xbt_free(loc); if (data->benching==1) { // we need to run a new bench @@ -298,13 +313,14 @@ void smpi_sample_3(int global, const char *file, int line) xbt_assert(samples, "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!"); data = xbt_dict_get(samples, loc); XBT_DEBUG("sample3 %s",loc); + xbt_free(loc); if (data->benching==0) { THROW_IMPOSSIBLE; } // ok, benchmarking this loop is over - xbt_os_timer_stop(smpi_process_timer()); + xbt_os_threadtimer_stop(smpi_process_timer()); // update the stats double sample, n; @@ -325,9 +341,12 @@ void smpi_sample_3(int global, const char *file, int line) data->benching = 0; } -void smpi_sample_flops(double flops) +#ifndef WIN32 +static void smpi_shared_alloc_free(void *p) { - smpi_execute_flops(flops); + shared_data_t *data = p; + xbt_free(data->loc); + xbt_free(data); } void *smpi_shared_malloc(size_t size, const char *file, int line) @@ -338,79 +357,91 @@ void *smpi_shared_malloc(size_t size, const char *file, int line) int fd; void* mem; shared_data_t *data; - - for(i = 0; i < len; i++) { - /* Make the 'loc' ID be a flat filename */ - if(loc[i] == '/') { - loc[i] = '_'; - } - } - if (!allocs) { - allocs = xbt_dict_new_homogeneous(free); - } - data = xbt_dict_get_or_null(allocs, loc); - if(!data) { - fd = shm_open(loc, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH); - if(fd < 0) { - switch(errno) { - case EEXIST: - xbt_die("Please cleanup /dev/shm/%s", loc); - default: - xbt_die("An unhandled error occured while opening %s: %s", loc, strerror(errno)); + if (sg_cfg_get_boolean("smpi/use_shared_malloc")){ + for(i = 0; i < len; i++) { + /* Make the 'loc' ID be a flat filename */ + if(loc[i] == '/') { + loc[i] = '_'; } } - data = xbt_new(shared_data_t, 1); - data->fd = fd; - data->count = 1; - data->loc = loc; - mem = shm_map(fd, size, data); - if(shm_unlink(loc) < 0) { - XBT_WARN("Could not early unlink %s: %s", loc, strerror(errno)); + if (!allocs) { + allocs = xbt_dict_new_homogeneous(smpi_shared_alloc_free); } - xbt_dict_set(allocs, loc, data, NULL); - XBT_DEBUG("Mapping %s at %p through %d", loc, mem, fd); - } else { - mem = shm_map(data->fd, size, data); - data->count++; + data = xbt_dict_get_or_null(allocs, loc); + if(!data) { + fd = shm_open(loc, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH); + if(fd < 0) { + switch(errno) { + case EEXIST: + xbt_die("Please cleanup /dev/shm/%s", loc); + default: + xbt_die("An unhandled error occured while opening %s: %s", loc, strerror(errno)); + } + } + data = xbt_new(shared_data_t, 1); + data->fd = fd; + data->count = 1; + data->loc = loc; + mem = shm_map(fd, size, data); + if(shm_unlink(loc) < 0) { + XBT_WARN("Could not early unlink %s: %s", loc, strerror(errno)); + } + xbt_dict_set(allocs, loc, data, NULL); + XBT_DEBUG("Mapping %s at %p through %d", loc, mem, fd); + } else { + xbt_free(loc); + mem = shm_map(data->fd, size, data); + data->count++; + } + XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, data); + }else{ + mem = xbt_malloc(size); + XBT_DEBUG("Classic malloc %zu in %p", size, mem); } - XBT_DEBUG("Malloc %zu in %p (metadata at %p)", size, mem, data); + return mem; } - void smpi_shared_free(void *ptr) { char loc[PTR_STRLEN]; shared_metadata_t* meta; shared_data_t* data; - - if (!allocs) { - XBT_WARN("Cannot free: nothing was allocated"); - return; - } - if(!allocs_metadata) { - XBT_WARN("Cannot free: no metadata was allocated"); - } - snprintf(loc, PTR_STRLEN, "%p", ptr); - meta = (shared_metadata_t*)xbt_dict_get_or_null(allocs_metadata, loc); - if (!meta) { - XBT_WARN("Cannot free: %p was not shared-allocated by SMPI", ptr); - return; - } - data = meta->data; - if(!data) { - XBT_WARN("Cannot free: something is broken in the metadata link"); - return; - } - if(munmap(ptr, meta->size) < 0) { - XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno)); - } - data->count--; - if (data->count <= 0) { - close(data->fd); - xbt_dict_remove(allocs, data->loc); - free(data->loc); + if (sg_cfg_get_boolean("smpi/use_shared_malloc")){ + + if (!allocs) { + XBT_WARN("Cannot free: nothing was allocated"); + return; + } + if(!allocs_metadata) { + XBT_WARN("Cannot free: no metadata was allocated"); + } + snprintf(loc, PTR_STRLEN, "%p", ptr); + meta = (shared_metadata_t*)xbt_dict_get_or_null(allocs_metadata, loc); + if (!meta) { + XBT_WARN("Cannot free: %p was not shared-allocated by SMPI", ptr); + return; + } + data = meta->data; + if(!data) { + XBT_WARN("Cannot free: something is broken in the metadata link"); + return; + } + if(munmap(ptr, meta->size) < 0) { + XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno)); + } + data->count--; + XBT_DEBUG("Shared free - no removal - of %p, count = %d", ptr, data->count); + if (data->count <= 0) { + close(data->fd); + xbt_dict_remove(allocs, data->loc); + XBT_DEBUG("Shared free - with removal - of %p", ptr); + } + }else{ + XBT_DEBUG("Classic free of %p", ptr); + xbt_free(ptr); } } +#endif int smpi_shared_known_call(const char* func, const char* input) { char* loc = bprintf("%s:%s", func, input);