X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/afcb6663eebe4a13bf9ffed05192253f1b6e5297..e9401099487c0170193c87c12126cf719051a1c1:/src/smpi/smpi_bench.c diff --git a/src/smpi/smpi_bench.c b/src/smpi/smpi_bench.c index cc313b7407..d8630166a6 100644 --- a/src/smpi/smpi_bench.c +++ b/src/smpi/smpi_bench.c @@ -1,9 +1,10 @@ -/* Copyright (c) 2007, 2009-2013. The SimGrid Team. +/* Copyright (c) 2007, 2009-2014. The SimGrid Team. * All rights reserved. */ /* 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. */ + * under the terms of the license (GNU LGPL) which comes with this package. */ +#include "internal_config.h" #include "private.h" #include "xbt/dict.h" #include "xbt/sysdep.h" @@ -11,6 +12,8 @@ #include "xbt/hash.h" #include "surf/surf.h" #include "simgrid/sg_config.h" +#include "simgrid/modelchecker.h" +#include "mc/mc_replay.h" #ifndef WIN32 #include @@ -24,6 +27,10 @@ #include #include +#ifndef MAP_ANONYMOUS +#define MAP_ANONYMOUS MAP_ANON +#endif + XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi, "Logging specific to SMPI (benchmarking)"); @@ -64,11 +71,17 @@ xbt_dict_t allocs = NULL; /* Allocated on first use */ xbt_dict_t allocs_metadata = NULL; /* Allocated on first use */ xbt_dict_t samples = NULL; /* Allocated on first use */ xbt_dict_t calls = NULL; /* Allocated on first use */ -__thread int smpi_current_rank = 0; /* Updated after each MPI call */ double smpi_cpu_threshold; double smpi_running_power; +int smpi_loaded_page = -1; +char* smpi_start_data_exe = NULL; +int smpi_size_data_exe = 0; +int smpi_privatize_global_variables; +double smpi_total_benched_time = 0; +smpi_privatisation_region_t smpi_privatisation_regions; + typedef struct { int fd; int count; @@ -106,7 +119,7 @@ static void* shm_map(int fd, size_t size, shared_data_t* data) { xbt_die("Could not map fd %d: %s", fd, strerror(errno)); } if(!allocs_metadata) { - allocs_metadata = xbt_dict_new_homogeneous(xbt_free); + allocs_metadata = xbt_dict_new_homogeneous(xbt_free_f); } snprintf(loc, PTR_STRLEN, "%p", mem); meta = xbt_new(shared_metadata_t, 1); @@ -126,16 +139,27 @@ void smpi_bench_destroy(void) xbt_dict_free(&calls); } +XBT_PUBLIC(void) smpi_execute_flops_(double *flops); +void smpi_execute_flops_(double *flops) +{ + smpi_execute_flops(*flops); +} + +XBT_PUBLIC(void) smpi_execute_(double *duration); +void smpi_execute_(double *duration) +{ + smpi_execute(*duration); +} + void smpi_execute_flops(double flops) { - smx_action_t action; + smx_synchro_t action; smx_host_t host; host = SIMIX_host_self(); - XBT_DEBUG("Handle real computation time: %g flops", flops); - action = simcall_host_execute("computation", host, flops, 1); -#ifdef HAVE_TRACING + XBT_DEBUG("Handle real computation time: %f flops", flops); + action = simcall_host_execute("computation", host, flops, 1, 0, 0); simcall_set_category (action, TRACE_internal_smpi_get_category()); -#endif simcall_host_execution_wait(action); + smpi_switch_data_segment(smpi_process_index()); } void smpi_execute(double duration) @@ -143,18 +167,14 @@ void smpi_execute(double duration) if (duration >= smpi_cpu_threshold) { XBT_DEBUG("Sleep for %g to handle real computation time", duration); double flops = duration * smpi_running_power; -#ifdef HAVE_TRACING int rank = smpi_process_index(); instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); extra->type=TRACING_COMPUTING; extra->comp_size=flops; TRACE_smpi_computing_in(rank, extra); -#endif smpi_execute_flops(flops); -#ifdef HAVE_TRACING TRACE_smpi_computing_out(rank); -#endif } else { XBT_DEBUG("Real computation took %g while option smpi/cpu_threshold is set to %g => ignore it", @@ -162,44 +182,73 @@ void smpi_execute(double duration) } } +void smpi_switch_data_segment(int dest); + void smpi_bench_begin(void) { + smpi_switch_data_segment(smpi_process_index()); + + if (MC_is_active() || MC_record_replay_is_active()) + return; + xbt_os_threadtimer_start(smpi_process_timer()); - smpi_current_rank = smpi_process_index(); } void smpi_bench_end(void) { + + if (MC_is_active() || MC_record_replay_is_active()) + return; + xbt_os_timer_t timer = smpi_process_timer(); xbt_os_threadtimer_stop(timer); +// smpi_switch_data_segment(smpi_process_count()); if (smpi_process_get_sampling()) { XBT_CRITICAL("Cannot do recursive benchmarks."); XBT_CRITICAL("Are you trying to make a call to MPI within a SMPI_SAMPLE_ block?"); xbt_backtrace_display_current(); xbt_die("Aborting."); } - smpi_execute(xbt_os_timer_elapsed(timer)); + // Simulate the benchmarked computation unless disabled via command-line argument + if (sg_cfg_get_boolean("smpi/simulate_computation")) { + smpi_execute(xbt_os_timer_elapsed(timer)); + } + + smpi_total_benched_time += xbt_os_timer_elapsed(timer); } -unsigned int smpi_sleep(unsigned int secs) +/* Private sleep function used by smpi_sleep() and smpi_usleep() */ +static unsigned int private_sleep(double secs) { - smx_action_t action; - smpi_bench_end(); - double flops = (double) secs*simcall_host_get_speed(SIMIX_host_self()); - XBT_DEBUG("Sleep for: %f flops", flops); - action = simcall_host_execute("computation", SIMIX_host_self(), flops, 1); - #ifdef HAVE_TRACING - simcall_set_category (action, TRACE_internal_smpi_get_category()); - #endif - simcall_host_execution_wait(action); + XBT_DEBUG("Sleep for: %lf secs", secs); + int rank = smpi_comm_rank(MPI_COMM_WORLD); + instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1); + extra->type=TRACING_SLEEPING; + extra->sleep_duration=secs; + TRACE_smpi_sleeping_in(rank, extra); + + simcall_process_sleep(secs); + + TRACE_smpi_sleeping_out(rank); smpi_bench_begin(); - return secs; + return 0; +} + +unsigned int smpi_sleep(unsigned int secs) +{ + return private_sleep((double)secs); } -int smpi_gettimeofday(struct timeval *tv) +int smpi_usleep(useconds_t usecs) +{ + return (int)private_sleep((double)usecs / 1000000.0); +} + + +int smpi_gettimeofday(struct timeval *tv, void* tz) { double now; smpi_bench_end(); @@ -216,11 +265,11 @@ int smpi_gettimeofday(struct timeval *tv) return 0; } -extern double sg_maxmin_precision; +extern double sg_surf_precision; unsigned long long smpi_rastro_resolution (void) { smpi_bench_end(); - double resolution = (1/sg_maxmin_precision); + double resolution = (1/sg_surf_precision); smpi_bench_begin(); return (unsigned long long)resolution; } @@ -411,7 +460,7 @@ void *smpi_shared_malloc(size_t size, const char *file, int line) { void* mem; if (sg_cfg_get_boolean("smpi/use_shared_malloc")){ - char *loc = bprintf("%zu_%s_%d", (size_t)getpid(), file, line); + char *loc = bprintf("/%zu_%s_%d", (size_t)getpid(), file, line); int fd; shared_data_t *data; loc = smpi_shared_alloc_hash(loc); /* hash loc, in order to have something @@ -460,7 +509,7 @@ void smpi_shared_free(void *ptr) shared_metadata_t* meta; shared_data_t* data; if (sg_cfg_get_boolean("smpi/use_shared_malloc")){ - + if (!allocs) { XBT_WARN("Cannot free: nothing was allocated"); return; @@ -542,3 +591,223 @@ void* smpi_shared_set_call(const char* func, const char* input, void* data) { free(loc); return data; } + + + + +#define TOPAGE(addr) (void *)(((unsigned long)(addr) / xbt_pagesize) * xbt_pagesize) + + +/** Map a given SMPI privatization segment (make a SMPI process active) + */ +void smpi_switch_data_segment(int dest){ + + if (smpi_loaded_page==dest)//no need to switch either + return; + + // So the job: + smpi_really_switch_data_segment(dest); +} + +/** Map a given SMPI privatization segment (make a SMPI process active) + * even if SMPI thinks it is already active + * + * When doing a state restoration, the state of the restored variables + * might not be consistent with the state of the virtual memory. + * In this case, we to change the data segment. + */ +void smpi_really_switch_data_segment(int dest) { + + if(smpi_size_data_exe == 0)//no need to switch + return; + +#ifdef HAVE_MMAP + int i; + if(smpi_loaded_page==-1){//initial switch, do the copy from the real page here + for (i=0; i< SIMIX_process_count(); i++){ + memcpy(smpi_privatisation_regions[i].address, + TOPAGE(smpi_start_data_exe), smpi_size_data_exe); + } + } + + // FIXME, cross-process support (mmap across process when necessary) + int current = smpi_privatisation_regions[dest].file_descriptor; + XBT_DEBUG("Switching data frame to the one of process %d", dest); + void* tmp = mmap (TOPAGE(smpi_start_data_exe), smpi_size_data_exe, + PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED, current, 0); + if (tmp != TOPAGE(smpi_start_data_exe)) + xbt_die("Couldn't map the new region"); + smpi_loaded_page = dest; +#endif +} + +int smpi_is_privatisation_file(char* file) +{ + return strncmp("/dev/shm/my-buffer-", file, 19) == 0; +} + +void smpi_get_executable_global_size(){ + int size_bss_binary=0; + int size_data_binary=0; + FILE *fp; + char *line = NULL; /* Temporal storage for each line that is readed */ + ssize_t read; /* Number of bytes readed */ + size_t n = 0; /* Amount of bytes to read by xbt_getline */ + + char *lfields[7]; + int i, found = 0; + + char *command = bprintf("objdump --section-headers %s", xbt_binary_name); + + fp = popen(command, "r"); + + if(fp == NULL){ + perror("popen failed"); + xbt_abort(); + } + + while ((read = xbt_getline(&line, &n, fp)) != -1 && found != 2) { + + if(n == 0) + continue; + + /* Wipeout the new line character */ + line[read - 1] = '\0'; + + lfields[0] = strtok(line, " "); + + if(lfields[0] == NULL) + continue; + + if(strcmp(lfields[0], "Sections:") == 0 + || strcmp(lfields[0], "Idx") == 0 + || strncmp(lfields[0], xbt_binary_name, strlen(xbt_binary_name)) == 0) + continue; + + for (i = 1; i < 7 && lfields[i - 1] != NULL; i++) { + lfields[i] = strtok(NULL, " "); + } + + /* + * we are looking for these fields + 23 .data 02625a20 00000000006013e0 00000000006013e0 000013e0 2**5 + CONTENTS, ALLOC, LOAD, DATA + 24 .bss 02625a40 0000000002c26e00 0000000002c26e00 02626e00 2**5 + ALLOC + */ + + if(i>=6){ + if(strcmp(lfields[1], ".data") == 0){ + size_data_binary = strtoul(lfields[2], NULL, 16); + smpi_start_data_exe = (char*) strtoul(lfields[4], NULL, 16); + found++; + }else if(strcmp(lfields[1], ".bss") == 0){ + //the beginning of bss is not exactly the end of data if not aligned, grow bss reported size accordingly + //TODO : check if this is OK, as some segments may be inserted between them.. + size_bss_binary = ((char*) strtoul(lfields[4], NULL, 16) - (smpi_start_data_exe + size_data_binary)) + + strtoul(lfields[2], NULL, 16); + found++; + } + + } + + } + + smpi_size_data_exe = (unsigned long) smpi_start_data_exe + - (unsigned long) TOPAGE(smpi_start_data_exe) + + size_data_binary+size_bss_binary; + xbt_free(command); + xbt_free(line); + pclose(fp); + +} + +void smpi_initialize_global_memory_segments(){ + +#ifndef HAVE_MMAP + smpi_privatize_global_variables=0; + return; +#else + + unsigned int i = 0; + smpi_get_executable_global_size(); + + XBT_DEBUG ("bss+data segment found : size %d starting at %p", + smpi_size_data_exe, smpi_start_data_exe ); + + if (smpi_size_data_exe == 0){//no need to switch + smpi_privatize_global_variables=0; + return; + } + + smpi_privatisation_regions = (smpi_privatisation_region_t) malloc( + smpi_process_count() * sizeof(struct s_smpi_privatisation_region)); + + for (i=0; i< SIMIX_process_count(); i++){ + //create SIMIX_process_count() mappings of this size with the same data inside + void *address = NULL; + char path[] = "/dev/shm/my-buffer-XXXXXX"; + int status; + + int file_descriptor= mkstemp (path); + if (file_descriptor < 0) { + if (errno==EMFILE) { + xbt_die("Impossible to create temporary file for memory mapping: %s\n\ +The open() system call failed with the EMFILE error code (too many files). \n\n\ +This means that you reached the system limits concerning the amount of files per process. \ +This is not a surprise if you are trying to virtualize many processes on top of SMPI. \ +Don't panic -- you should simply increase your system limits and try again. \n\n\ +First, check what your limits are:\n\ + cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\ + ulimit -Hn # Gives you the per process hard limit\n\ + ulimit -Sn # Gives you the per process soft limit\n\ + cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\ +If one of these values is less than the amount of MPI processes that you try to run, then you got the explanation of this error. \ +Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/", + strerror(errno)); + } + xbt_die("Impossible to create temporary file for memory mapping: %s", + strerror(errno)); + } + + status = unlink (path); + if (status) + xbt_die("Impossible to unlink temporary file for memory mapping"); + + status = ftruncate(file_descriptor, smpi_size_data_exe); + if(status) + xbt_die("Impossible to set the size of the temporary file for memory mapping"); + + /* Ask for a free region */ + address = mmap (NULL, smpi_size_data_exe, PROT_READ | PROT_WRITE, MAP_SHARED, file_descriptor, 0); + if (address == MAP_FAILED) + xbt_die("Couldn't find a free region for memory mapping"); + + //initialize the values + memcpy(address, TOPAGE(smpi_start_data_exe), smpi_size_data_exe); + + //store the address of the mapping for further switches + smpi_privatisation_regions[i].file_descriptor = file_descriptor; + smpi_privatisation_regions[i].address = address; + } + +#endif + +} + +void smpi_destroy_global_memory_segments(){ + if (smpi_size_data_exe == 0)//no need to switch + return; +#ifdef HAVE_MMAP + int i; + for (i=0; i< smpi_process_count(); i++){ + if(munmap(smpi_privatisation_regions[i].address, smpi_size_data_exe) < 0) { + XBT_WARN("Unmapping of fd %d failed: %s", + smpi_privatisation_regions[i].file_descriptor, strerror(errno)); + } + close(smpi_privatisation_regions[i].file_descriptor); + } + xbt_free(smpi_privatisation_regions); +#endif + +}