#include "xbt/dict.h"
#include "xbt/sysdep.h"
#include "xbt/ex.h"
+#include "xbt/hash.h"
#include "surf/surf.h"
#include "simgrid/sg_config.h"
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;
+
typedef struct {
int fd;
int count;
xbt_die("Could not map fd %d: %s", fd, strerror(errno));
}
if(!allocs_metadata) {
- allocs_metadata = xbt_dict_new();
+ allocs_metadata = xbt_dict_new_homogeneous(xbt_free);
}
snprintf(loc, PTR_STRLEN, "%p", mem);
meta = xbt_new(shared_metadata_t, 1);
meta->size = size;
meta->data = data;
- xbt_dict_set(allocs_metadata, loc, meta, &free);
+ xbt_dict_set(allocs_metadata, loc, meta, NULL);
XBT_DEBUG("MMAP %zu to %p", size, mem);
return mem;
}
smx_action_t action;
smx_host_t host;
host = SIMIX_host_self();
-
- XBT_DEBUG("Handle real computation time: %f flops", flops);
+ XBT_DEBUG("Handle real computation time: %g flops", flops);
action = simcall_host_execute("computation", host, flops, 1);
#ifdef HAVE_TRACING
simcall_set_category (action, TRACE_internal_smpi_get_category());
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 >= 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);
}
}
unsigned int smpi_sleep(unsigned int secs)
{
+ smx_action_t action;
+
smpi_bench_end();
- smpi_execute_flops((double) secs*simcall_host_get_speed(SIMIX_host_self()));
+
+ 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);
+
smpi_bench_begin();
return secs;
}
/* ****************************** 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;
+int smpi_sample_is_running = 0;
+
static char *sample_location(int global, const char *file, int line) {
if (global) {
return bprintf("%s:%d", file, line);
local_data_t *data;
smpi_bench_end(); /* Take time from previous, unrelated computation into account */
+ smpi_sample_is_running++;
+
if (!samples)
samples = xbt_dict_new_homogeneous(free);
data->count, data->iters, data->relstderr, data->threshold, data->mean);
smpi_execute(data->mean);
+ smpi_sample_is_running--;
smpi_bench_begin(); // prepare to capture future, unrelated computations
return 0;
}
xbt_free(data);
}
+static char *smpi_shared_alloc_hash(char *loc)
+{
+ char hash[42];
+ char s[7];
+ unsigned val;
+ int i, j;
+
+ xbt_sha(loc, hash);
+ hash[41] = '\0';
+ s[6] = '\0';
+ loc = xbt_realloc(loc, 30);
+ loc[0] = '/';
+ for (i = 0; i < 40; i += 6) { /* base64 encode */
+ memcpy(s, hash + i, 6);
+ val = strtoul(s, NULL, 16);
+ for (j = 0; j < 4; j++) {
+ unsigned char x = (val >> (18 - 3 * j)) & 0x3f;
+ loc[1 + 4 * i / 6 + j] =
+ "ABCDEFGHIJKLMNOPQRSTUVZXYZabcdefghijklmnopqrstuvzxyz0123456789-_"[x];
+ }
+ }
+ loc[29] = '\0';
+ return loc;
+}
+
void *smpi_shared_malloc(size_t size, const char *file, int line)
{
- char *loc = bprintf("%zu_%s_%d", (size_t)getpid(), file, line);
- size_t len = strlen(loc);
- size_t i;
- int fd;
void* mem;
- shared_data_t *data;
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] = '_';
- }
- }
+ 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
+ * not too long */
if (!allocs) {
allocs = xbt_dict_new_homogeneous(smpi_shared_alloc_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) {
+ 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));
+ xbt_die("An unhandled error occured while opening %s. shm_open: %s", loc, strerror(errno));
}
}
data = xbt_new(shared_data_t, 1);
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 (shm_unlink(loc) < 0) {
+ 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);
data->count++;
}
XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, data);
- }else{
+ } else {
mem = xbt_malloc(size);
XBT_DEBUG("Classic malloc %zu in %p", size, mem);
}