-/* 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"
#include <string.h>
#include <stdio.h>
+#ifndef MAP_ANONYMOUS
+#define MAP_ANONYMOUS MAP_ANON
+#endif
+
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi,
"Logging specific to SMPI (benchmarking)");
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* fds;
+void** mappings;
+int loaded_page = -1;
+char* start_data_exe = NULL;
+int size_data_exe = 0;
+int smpi_privatize_global_variables;
+
typedef struct {
int fd;
int count;
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_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_host_execution_wait(action);
}
-static void smpi_execute(double duration)
+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 >= sg_cfg_get_double("smpi/cpu_threshold")) {
- XBT_DEBUG("Sleep for %f to handle real computation time", duration);
- smpi_execute_flops(duration *
- sg_cfg_get_double("smpi/running_power"));
+ 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 %f while option smpi/cpu_threshold is set to %f => ignore it",
- duration, sg_cfg_get_double("smpi/cpu_threshold"));
+ XBT_DEBUG("Real computation took %g while option smpi/cpu_threshold is set to %g => ignore it",
+ duration, smpi_cpu_threshold);
}
}
+void switch_data_segment(int dest);
+
void smpi_bench_begin(void)
{
+ switch_data_segment(smpi_process_index());
xbt_os_threadtimer_start(smpi_process_timer());
smpi_current_rank = smpi_process_index();
}
void smpi_bench_end(void)
{
xbt_os_timer_t timer = smpi_process_timer();
-
xbt_os_threadtimer_stop(timer);
- smpi_execute(xbt_os_timer_elapsed(timer));
+// 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.");
+ }
+ // 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));
+ }
}
-unsigned int smpi_sleep(unsigned int secs)
+/* Private sleep function used by smpi_sleep() and smpi_usleep() */
+static unsigned int private_sleep(double secs)
{
smpi_bench_end();
- smpi_execute_flops((double) secs*simcall_host_get_speed(SIMIX_host_self()));
+
+ 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 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();
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;
}
/* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
typedef struct {
- int iters; /* amount of requested iterations */
- int count; /* amount of iterations done so far */
double threshold; /* maximal stderr requested (if positive) */
double relstderr; /* observed stderr so far */
double mean; /* mean of benched times, to be used if the block is disabled */
double sum; /* sum of benched times (to compute the mean and stderr) */
double sum_pow2; /* sum of the square of the benched times (to compute the stderr) */
+ int iters; /* amount of requested iterations */
+ int count; /* amount of iterations done so far */
int benching; /* 1: we are benchmarking; 0: we have enough data, no bench anymore */
} local_data_t;
local_data_t *data;
smpi_bench_end(); /* Take time from previous, unrelated computation into account */
+ smpi_process_set_sampling(1);
+
if (!samples)
samples = xbt_dict_new_homogeneous(free);
{
char *loc = sample_location(global, file, line);
local_data_t *data;
+ int res;
xbt_assert(samples, "Y U NO use SMPI_SAMPLE_* macros? Stop messing directly with smpi_sample_* functions!");
data = xbt_dict_get(samples, loc);
// we need to run a new bench
XBT_DEBUG("benchmarking: count:%d iter:%d stderr:%f thres:%f; mean:%f",
data->count, data->iters, data->relstderr, data->threshold, data->mean);
- smpi_bench_begin();
- return 1;
+ res = 1;
} else {
// Enough data, no more bench (either we got enough data from previous visits to this benched nest, or we just ran one bench and need to bail out now that our job is done).
// Just sleep instead
XBT_DEBUG("No benchmark (either no need, or just ran one): count >= iter (%d >= %d) or stderr<thres (%f<=%f). apply the %fs delay instead",
data->count, data->iters, data->relstderr, data->threshold, data->mean);
smpi_execute(data->mean);
-
- smpi_bench_begin(); // prepare to capture future, unrelated computations
- return 0;
+ smpi_process_set_sampling(0);
+ res = 0; // prepare to capture future, unrelated computations
}
+ smpi_bench_begin();
+ return res;
}
case EEXIST:
xbt_die("Please cleanup /dev/shm/%s", loc);
default:
- xbt_die("An unhandled error occured while opening %s: %s", loc, strerror(errno));
+ xbt_die("An unhandled error occured while opening %s. shm_open: %s", loc, strerror(errno));
}
}
data = xbt_new(shared_data_t, 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_WARN("Could not early unlink %s. shm_unlink: %s", loc, strerror(errno));
}
xbt_dict_set(allocs, loc, data, NULL);
XBT_DEBUG("Mapping %s at %p through %d", loc, mem, fd);
}
#endif
-int smpi_shared_known_call(const char* func, const char* input) {
- char* loc = bprintf("%s:%s", func, input);
- xbt_ex_t ex;
- int known;
+int smpi_shared_known_call(const char* func, const char* input)
+{
+ char* loc = bprintf("%s:%s", func, input);
+ xbt_ex_t ex;
+ int known = 0;
- if(!calls) {
- calls = xbt_dict_new_homogeneous(NULL);
- }
- TRY {
- xbt_dict_get(calls, loc); /* Succeed or throw */
- known = 1;
- }
- CATCH(ex) {
- if(ex.category == not_found_error) {
- known = 0;
- xbt_ex_free(ex);
- } else {
- RETHROW;
- }
- }
- free(loc);
- return known;
+ if (!calls) {
+ calls = xbt_dict_new_homogeneous(NULL);
+ }
+ TRY {
+ xbt_dict_get(calls, loc); /* Succeed or throw */
+ known = 1;
+ }
+ TRY_CLEANUP {
+ xbt_free(loc);
+ }
+ CATCH(ex) {
+ if (ex.category != not_found_error)
+ RETHROW;
+ xbt_ex_free(ex);
+ }
+ return known;
}
void* smpi_shared_get_call(const char* func, const char* input) {
free(loc);
return 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
+ *
+ */
+
+
+
+void switch_data_segment(int dest){
+
+ if(size_data_exe == 0)//no need to switch
+ return;
+
+ if (loaded_page==dest)//no need to switch either
+ return;
+
+
+#ifdef HAVE_MMAP
+ int i;
+ if(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);
+ }
+ }
+ 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))
+ xbt_die("Couldn't map the new region");
+ loaded_page=dest;
+#endif
+}
+
+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);
+ 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))
+ + strtoul(lfields[2], NULL, 16);
+ found++;
+ }
+
+ }
+
+ }
+
+ size_data_exe =(unsigned long)start_data_exe - (unsigned long)TOPAGE(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",size_data_exe, start_data_exe );
+
+ if(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*));
+
+
+ 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;
+ 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");
+ status = unlink (path);
+ if (status)
+ xbt_die("Impossible to unlink temporary file for memory mapping");
+
+ status = ftruncate(file_descriptor, 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);
+
+ 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);
+
+ //store the address of the mapping for further switches
+ fds[i]=file_descriptor;
+ mappings[i]= address;
+ }
+
+#endif
+
+}
+
+void smpi_destroy_global_memory_segments(){
+ if(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));
+ }
+ close(fds[i]);
+ }
+ xbt_free(mappings);
+ xbt_free(fds);
+
+#endif
+
+}
+
