X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/ce1289d4421b08ce6f69d8203784ee648671972e..e9401099487c0170193c87c12126cf719051a1c1:/src/smpi/smpi_bench.c diff --git a/src/smpi/smpi_bench.c b/src/smpi/smpi_bench.c index e33d659375..d8630166a6 100644 --- a/src/smpi/smpi_bench.c +++ b/src/smpi/smpi_bench.c @@ -12,6 +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 @@ -73,12 +75,12 @@ xbt_dict_t calls = NULL; /* Allocated on first use */ double smpi_cpu_threshold; double smpi_running_power; -int* fds; -void** mappings; -int loaded_page = -1; -char* start_data_exe = NULL; -int size_data_exe = 0; +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; @@ -150,15 +152,14 @@ void smpi_execute_(double *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: %f flops", flops); action = simcall_host_execute("computation", host, flops, 1, 0, 0); -#ifdef HAVE_TRACING 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) @@ -166,18 +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", @@ -185,19 +182,27 @@ void smpi_execute(double duration) } } -void switch_data_segment(int dest); +void smpi_switch_data_segment(int dest); void smpi_bench_begin(void) { - switch_data_segment(smpi_process_index()); + smpi_switch_data_segment(smpi_process_index()); + + if (MC_is_active() || MC_record_replay_is_active()) + return; + xbt_os_threadtimer_start(smpi_process_timer()); } 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); -// switch_data_segment(smpi_process_count()); +// 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?"); @@ -208,6 +213,8 @@ void smpi_bench_end(void) if (sg_cfg_get_boolean("smpi/simulate_computation")) { smpi_execute(xbt_os_timer_elapsed(timer)); } + + smpi_total_benched_time += xbt_os_timer_elapsed(timer); } /* Private sleep function used by smpi_sleep() and smpi_usleep() */ @@ -216,17 +223,15 @@ static unsigned int private_sleep(double secs) smpi_bench_end(); XBT_DEBUG("Sleep for: %lf secs", secs); - #ifdef HAVE_TRACING 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); -#endif + simcall_process_sleep(secs); -#ifdef HAVE_TRACING + TRACE_smpi_sleeping_out(rank); -#endif smpi_bench_begin(); return 0; @@ -455,7 +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 @@ -504,7 +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; @@ -593,40 +598,54 @@ void* smpi_shared_set_call(const char* func, const char* input, void* data) { #define TOPAGE(addr) (void *)(((unsigned long)(addr) / xbt_pagesize) * xbt_pagesize) -/* - * - read the executable data+bss section addresses and sizes - * - for each process create a copy of these sections with mmap - * - store them in a dynar - * +/** 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); +} -void switch_data_segment(int dest){ - - if(size_data_exe == 0)//no need to switch - return; +/** 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 (loaded_page==dest)//no need to switch either + if(smpi_size_data_exe == 0)//no need to switch return; - #ifdef HAVE_MMAP int i; - if(loaded_page==-1){//initial switch, do the copy from the real page here + if(smpi_loaded_page==-1){//initial switch, do the copy from the real page here for (i=0; i< SIMIX_process_count(); i++){ - memcpy(mappings[i],TOPAGE(start_data_exe),size_data_exe); + memcpy(smpi_privatisation_regions[i].address, + TOPAGE(smpi_start_data_exe), smpi_size_data_exe); } } - int current= fds[dest]; - XBT_VERB("Switching data frame to the one of process %d", dest); - void* tmp = mmap (TOPAGE(start_data_exe), size_data_exe, PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED, current, 0); - if (tmp != TOPAGE(start_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"); - loaded_page=dest; + 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; @@ -680,12 +699,12 @@ void smpi_get_executable_global_size(){ if(i>=6){ if(strcmp(lfields[1], ".data") == 0){ size_data_binary = strtoul(lfields[2], NULL, 16); - start_data_exe = (char*) strtoul(lfields[4], 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) - (start_data_exe + size_data_binary)) + //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++; } @@ -694,7 +713,9 @@ void smpi_get_executable_global_size(){ } - size_data_exe =(unsigned long)start_data_exe - (unsigned long)TOPAGE(start_data_exe)+ size_data_binary+size_bss_binary; + 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); @@ -711,49 +732,63 @@ void smpi_initialize_global_memory_segments(){ unsigned int i = 0; smpi_get_executable_global_size(); - XBT_DEBUG ("bss+data segment found : size %d starting at %p",size_data_exe, start_data_exe ); + XBT_DEBUG ("bss+data segment found : size %d starting at %p", + smpi_size_data_exe, smpi_start_data_exe ); - if(size_data_exe == 0){//no need to switch + if (smpi_size_data_exe == 0){//no need to switch smpi_privatize_global_variables=0; return; } - fds= (int*)xbt_malloc((smpi_process_count())*sizeof(int)); - mappings= (void**)xbt_malloc((smpi_process_count())*sizeof(void*)); - + 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, *tmp = NULL; + void *address = NULL; char path[] = "/dev/shm/my-buffer-XXXXXX"; int status; + int file_descriptor= mkstemp (path); - if (file_descriptor < 0) - xbt_die("Impossible to create temporary file for memory mapping"); + 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, size_data_exe); + 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, size_data_exe, PROT_NONE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0); - + 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"); - tmp = mmap (address, size_data_exe, PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED, file_descriptor, 0); - - if (tmp != address) - xbt_die("Couldn't obtain the right address"); //initialize the values - memcpy(address,TOPAGE(start_data_exe),size_data_exe); + memcpy(address, TOPAGE(smpi_start_data_exe), smpi_size_data_exe); //store the address of the mapping for further switches - fds[i]=file_descriptor; - mappings[i]= address; + smpi_privatisation_regions[i].file_descriptor = file_descriptor; + smpi_privatisation_regions[i].address = address; } #endif @@ -761,20 +796,18 @@ void smpi_initialize_global_memory_segments(){ } void smpi_destroy_global_memory_segments(){ - if(size_data_exe == 0)//no need to switch + 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(mappings[i],size_data_exe) < 0) { - XBT_WARN("Unmapping of fd %d failed: %s", fds[i], strerror(errno)); + 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(fds[i]); + close(smpi_privatisation_regions[i].file_descriptor); } - xbt_free(mappings); - xbt_free(fds); - + xbt_free(smpi_privatisation_regions); #endif } -