#include "private.h"
#include "xbt/dict.h"
#include "xbt/sysdep.h"
+#include "xbt/ex.h"
+#include "surf/surf.h"
+
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <string.h>
+#include <stdio.h>
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi,
"Logging specific to SMPI (benchmarking)");
-xbt_dict_t allocs = NULL; /* Allocated on first use */
-xbt_dict_t samples = NULL; /* Allocated on first use */
+/* Shared allocations are handled through shared memory segments.
+ * Associated data and metadata are used as follows:
+ *
+ * mmap #1
+ * `allocs' dict ---- -.
+ * ---------- shared_data_t shared_metadata_t / | | |
+ * .->| <name> | ---> -------------------- <--. ----------------- | | | |
+ * | ---------- | fd of <name> | | | size of mmap | --| | | |
+ * | | count (2) | |-- | data | \ | | |
+ * `----------------- | <name> | | ----------------- ---- |
+ * -------------------- | ^ |
+ * | | |
+ * | | `allocs_metadata' dict |
+ * | | ---------------------- |
+ * | `-- | <addr of mmap #1> |<-'
+ * | .-- | <addr of mmap #2> |<-.
+ * | | ---------------------- |
+ * | | |
+ * | | |
+ * | | |
+ * | | mmap #2 |
+ * | v ---- -'
+ * | shared_metadata_t / | |
+ * | ----------------- | | |
+ * | | size of mmap | --| | |
+ * `-- | data | | | |
+ * ----------------- | | |
+ * \ | |
+ * ----
+ */
+
+#define PTR_STRLEN (2 + 2 * sizeof(void*) + 1)
+
+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 */
typedef struct {
- int count;
- char data[];
+ int fd;
+ int count;
+ char* loc;
} shared_data_t;
+typedef struct {
+ size_t size;
+ shared_data_t* data;
+} shared_metadata_t;
+
+static size_t shm_size(int fd) {
+ struct stat st;
+
+ if(fstat(fd, &st) < 0) {
+ xbt_die("Could not stat fd %d: %s", fd, strerror(errno));
+ }
+ return (size_t)st.st_size;
+}
+
+static void* shm_map(int fd, size_t size, shared_data_t* data) {
+ void* mem;
+ char loc[PTR_STRLEN];
+ shared_metadata_t* meta;
+
+ if(size > shm_size(fd)) {
+ if(ftruncate(fd, (off_t)size) < 0) {
+ 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));
+ }
+ if(!allocs_metadata) {
+ allocs_metadata = xbt_dict_new();
+ }
+ 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_DEBUG("MMAP %zu to %p", size, mem);
+ return mem;
+}
+
typedef struct {
- double time;
- int count;
- int max;
- int started;
+ int count;
+ double sum;
+ double sum_pow2;
+ double mean;
+ double relstderr;
+ int iters;
+ double threshold;
+ int started;
} local_data_t;
-void smpi_bench_destroy(void) {
- if (allocs) {
- xbt_dict_free(&allocs);
- }
- if (samples) {
- xbt_dict_free(&samples);
- }
+void smpi_bench_destroy(void)
+{
+ if (allocs) {
+ xbt_dict_free(&allocs);
+ }
+ if (samples) {
+ xbt_dict_free(&samples);
+ }
+ if(calls) {
+ xbt_dict_free(&calls);
+ }
}
-static void smpi_execute(double duration) {
- smx_host_t host;
+static void smpi_execute_flops(double flops)
+{
smx_action_t action;
- smx_mutex_t mutex;
- smx_cond_t cond;
- e_surf_action_state_t state;
-
- if(duration >= xbt_cfg_get_double(_surf_cfg_set, "smpi/cpu_threshold")) {
- host = SIMIX_host_self();
- mutex = SIMIX_mutex_init();
- cond = SIMIX_cond_init();
- DEBUG1("Sleep for %f to handle real computation time", duration);
- duration *= xbt_cfg_get_double(_surf_cfg_set, "smpi/running_power");
- action = SIMIX_action_execute(host, "computation", duration);
- SIMIX_mutex_lock(mutex);
- SIMIX_register_action_to_condition(action, cond);
- for(state = SIMIX_action_get_state(action);
- state == SURF_ACTION_READY ||
- state == SURF_ACTION_RUNNING; state = SIMIX_action_get_state(action)) {
- SIMIX_cond_wait(cond, mutex);
- }
- SIMIX_unregister_action_to_condition(action, cond);
- SIMIX_mutex_unlock(mutex);
- SIMIX_action_destroy(action);
- SIMIX_cond_destroy(cond);
- SIMIX_mutex_destroy(mutex);
- }
+ smx_host_t host;
+ host = SIMIX_host_self();
+
+ XBT_DEBUG("Handle real computation time: %f flops", flops);
+ action = SIMIX_req_host_execute("computation", host, flops, 1);
+#ifdef HAVE_TRACING
+ SIMIX_req_set_category (action, TRACE_internal_smpi_get_category());
+#endif
+ SIMIX_req_host_execution_wait(action);
}
-void smpi_bench_begin(int rank, const char* mpi_call) {
- if(mpi_call && rank >= 0 && xbt_cfg_get_int(_surf_cfg_set, "smpi/log_events")) {
- INFO3("SMPE: ts=%f rank=%d type=end et=%s", SIMIX_get_clock(), rank, mpi_call);
+static void smpi_execute(double duration)
+{
+ if (duration >= xbt_cfg_get_double(_surf_cfg_set, "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"));
}
+}
+
+void smpi_bench_begin(void)
+{
xbt_os_timer_start(smpi_process_timer());
+ smpi_current_rank = smpi_process_index();
}
-void smpi_bench_end(int rank, const char* mpi_call) {
+void smpi_bench_end(void)
+{
xbt_os_timer_t timer = smpi_process_timer();
xbt_os_timer_stop(timer);
smpi_execute(xbt_os_timer_elapsed(timer));
- if(mpi_call && rank >= 0 && xbt_cfg_get_int(_surf_cfg_set, "smpi/log_events")) {
- INFO3("SMPE: ts=%f rank=%d type=begin et=%s", SIMIX_get_clock(), rank, mpi_call);
+}
+
+unsigned int smpi_sleep(unsigned int secs)
+{
+ smpi_execute((double) secs);
+ return secs;
+}
+
+int smpi_gettimeofday(struct timeval *tv, struct timezone *tz)
+{
+ double now = SIMIX_get_clock();
+ if (tv) {
+ tv->tv_sec = (time_t) now;
+ tv->tv_usec = (suseconds_t)((now - tv->tv_sec) * 1e6);
}
+ return 0;
}
-unsigned int smpi_sleep(unsigned int secs) {
- smpi_execute((double)secs);
+static char *sample_location(int global, const char *file, int line)
+{
+ if (global) {
+ return bprintf("%s:%d", file, line);
+ } else {
+ return bprintf("%s:%d:%d", file, line, smpi_process_index());
+ }
}
-int smpi_gettimeofday(struct timeval* tv, struct timezone* tz) {
- double now = SIMIX_get_clock();
+int smpi_sample_1(int global, const char *file, int line, int iters, double threshold)
+{
+ char *loc = sample_location(global, file, line);
+ local_data_t *data;
- if(tv) {
- tv->tv_sec = (time_t)now;
- tv->tv_usec = (suseconds_t)(now * 1e6);
- }
- return 0;
+ smpi_bench_end(); /* Take time from previous MPI call into account */
+ if (!samples) {
+ samples = xbt_dict_new();
+ }
+ data = xbt_dict_get_or_null(samples, loc);
+ if (!data) {
+ data = (local_data_t *) xbt_new(local_data_t, 1);
+ data->count = 0;
+ data->sum = 0.0;
+ data->sum_pow2 = 0.0;
+ data->iters = iters;
+ data->threshold = threshold;
+ data->started = 0;
+ xbt_dict_set(samples, loc, data, &free);
+ return 0;
+ }
+ free(loc);
+ return 1;
}
-static char* sample_location(int global, const char* file, int line) {
- if(global) {
- return bprintf("%s:%d", file, line);
- } else {
- return bprintf("%s:%d:%d", file, line, smpi_process_index());
- }
+int smpi_sample_2(int global, const char *file, int line)
+{
+ char *loc = sample_location(global, file, line);
+ local_data_t *data;
+
+ xbt_assert(samples, "You did something very inconsistent, didn't you?");
+ data = xbt_dict_get_or_null(samples, loc);
+ if (!data) {
+ xbt_assert(data, "Please, do thing in order");
+ }
+ if (!data->started) {
+ if ((data->iters > 0 && data->count >= data->iters)
+ || (data->count > 1 && data->threshold > 0.0 && data->relstderr <= data->threshold)) {
+ XBT_DEBUG("Perform some wait of %f", data->mean);
+ smpi_execute(data->mean);
+ } else {
+ data->started = 1;
+ data->count++;
+ }
+ } else {
+ data->started = 0;
+ }
+ free(loc);
+ smpi_bench_begin();
+ smpi_process_simulated_start();
+ return data->started;
}
-void smpi_sample_1(int global, const char* file, int line, int max) {
- char* loc = sample_location(global, file, line);
- local_data_t* data;
+void smpi_sample_3(int global, const char *file, int line)
+{
+ char *loc = sample_location(global, file, line);
+ local_data_t *data;
+ double sample, n;
- smpi_bench_end(-1, NULL); /* Take time from previous MPI call into account */
- if (!samples) {
- samples = xbt_dict_new();
- }
- data = xbt_dict_get_or_null(samples, loc);
- if (!data) {
- data = (local_data_t*)xbt_new(local_data_t, 1);
- data->time = 0.0;
- data->count = 0;
- data->max = max;
- data->started = 0;
- xbt_dict_set(samples, loc, data, &free);
- }
- free(loc);
+ xbt_assert(samples, "You did something very inconsistent, didn't you?");
+ data = xbt_dict_get_or_null(samples, loc);
+ smpi_bench_end();
+ if(data && data->started && data->count < data->iters) {
+ sample = smpi_process_simulated_elapsed();
+ data->sum += sample;
+ data->sum_pow2 += sample * sample;
+ n = (double)data->count;
+ data->mean = data->sum / n;
+ data->relstderr = sqrt((data->sum_pow2 / n - data->mean * data->mean) / n) / data->mean;
+ XBT_DEBUG("Average mean after %d steps is %f, relative standard error is %f (sample was %f)", data->count,
+ data->mean, data->relstderr, sample);
+ }
+ free(loc);
}
-int smpi_sample_2(int global, const char* file, int line) {
- char* loc = sample_location(global, file, line);
- local_data_t* data;
- double* simu;
+void smpi_sample_flops(double flops)
+{
+ smpi_execute_flops(flops);
+}
- xbt_assert0(samples, "You did something very inconsistent, didn't you?");
- data = xbt_dict_get_or_null(samples, loc);
- if (!data) {
- xbt_assert0(data, "Please, do thing in order");
- }
- if (!data->started) {
- if (data->count < data->max) {
- data->started = 1;
- data->count++;
- } else {
- DEBUG1("Perform some wait of %f", data->time / (double)data->count);
- smpi_execute(data->time / (double)data->count);
+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;
+
+ 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();
+ }
+ 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));
}
- } else {
- data->started = 0;
- }
- free(loc);
- smpi_bench_begin(-1, NULL);
- smpi_process_simulated_start();
- return data->started;
+ }
+ 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, &free);
+ XBT_DEBUG("Mapping %s at %p through %d", loc, mem, fd);
+ } else {
+ mem = shm_map(data->fd, size, data);
+ data->count++;
+ }
+ XBT_DEBUG("Malloc %zu in %p (metadata at %p)", size, mem, data);
+ return mem;
}
-void smpi_sample_3(int global, const char* file, int line) {
- char* loc = sample_location(global, file, line);
- local_data_t* data;
- double spent;
+void smpi_shared_free(void *ptr)
+{
+ char loc[PTR_STRLEN];
+ shared_metadata_t* meta;
+ shared_data_t* data;
- xbt_assert0(samples, "You did something very inconsistent, didn't you?");
- data = xbt_dict_get_or_null(samples, loc);
- if (!data || !data->started || data->count >= data->max) {
- xbt_assert0(data, "Please, do thing in order");
- }
- smpi_bench_end(-1, NULL);
- data->time += smpi_process_simulated_elapsed();
- DEBUG2("Average mean after %d steps is %f", data->count, data->time / (double)data->count);
+ 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);
+ }
}
-void* smpi_shared_malloc(size_t size, const char* file, int line) {
- char* loc = bprintf("%s:%d:%zu", file, line, size);
- shared_data_t* data;
+int smpi_shared_known_call(const char* func, const char* input) {
+ char* loc = bprintf("%s:%s", func, input);
+ xbt_ex_t ex;
+ int known;
- if (!allocs) {
- allocs = xbt_dict_new();
+ if(!calls) {
+ calls = xbt_dict_new();
}
- data = xbt_dict_get_or_null(allocs, loc);
- if (!data) {
- data = (shared_data_t*)xbt_malloc0(sizeof(int) + size);
- data->count = 1;
- xbt_dict_set(allocs, loc, data, &free);
- } else {
- data->count++;
+ 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 data->data;
+ return known;
}
-void smpi_shared_free(void* ptr) {
- shared_data_t* data = (shared_data_t*)((int*)ptr - 1);
- char* loc;
+void* smpi_shared_get_call(const char* func, const char* input) {
+ char* loc = bprintf("%s:%s", func, input);
+ void* data;
- if (!allocs) {
- WARN0("Cannot free: nothing was allocated");
- return;
+ if(!calls) {
+ calls = xbt_dict_new();
}
- loc = xbt_dict_get_key(allocs, data);
- if (!loc) {
- WARN1("Cannot free: %p was not shared-allocated by SMPI", ptr);
- return;
- }
- data->count--;
- if (data->count <= 0) {
- xbt_dict_remove(allocs, loc);
+ data = xbt_dict_get(calls, loc);
+ free(loc);
+ return data;
+}
+
+void* smpi_shared_set_call(const char* func, const char* input, void* data) {
+ char* loc = bprintf("%s:%s", func, input);
+
+ if(!calls) {
+ calls = xbt_dict_new();
}
+ xbt_dict_set(calls, loc, data, NULL);
+ free(loc);
+ return data;
}
