+#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"
+
+#ifndef WIN32
+#include <sys/mman.h>
+#endif
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <math.h> // sqrt
+#include <unistd.h>
+#include <string.h>
+#include <stdio.h>
+
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi,
+ "Logging specific to SMPI (benchmarking)");
+
+/* 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 */
+
+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;
+ 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;
+}
+
+#ifndef WIN32
+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_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, NULL);
+ XBT_DEBUG("MMAP %zu to %p", size, mem);
+ return mem;
+}
+#endif
+
+void smpi_bench_destroy(void)
+{
+ xbt_dict_free(&allocs);
+ xbt_dict_free(&allocs_metadata);
+ xbt_dict_free(&samples);
+ 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_set_category (action, TRACE_internal_smpi_get_category());
+#endif
+ simcall_host_execution_wait(action);
+}
+
+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",
+ 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);
+// 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));
+}
+
+unsigned int smpi_sleep(unsigned int 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, 0, 0);
+ #ifdef HAVE_TRACING
+ simcall_set_category (action, TRACE_internal_smpi_get_category());
+ #endif
+ simcall_host_execution_wait(action);
+
+ smpi_bench_begin();
+ return secs;
+}
+
+int smpi_gettimeofday(struct timeval *tv)
+{
+ double now;
+ smpi_bench_end();
+ now = SIMIX_get_clock();
+ if (tv) {
+ tv->tv_sec = (time_t)now;
+#ifdef WIN32
+ tv->tv_usec = (useconds_t)((now - tv->tv_sec) * 1e6);
+#else
+ tv->tv_usec = (suseconds_t)((now - tv->tv_sec) * 1e6);
+#endif
+ }
+ smpi_bench_begin();
+ return 0;
+}
+
+extern double sg_maxmin_precision;
+unsigned long long smpi_rastro_resolution (void)
+{
+ smpi_bench_end();
+ double resolution = (1/sg_maxmin_precision);
+ smpi_bench_begin();
+ return (unsigned long long)resolution;
+}
+
+unsigned long long smpi_rastro_timestamp (void)
+{
+ smpi_bench_end();
+ double now = SIMIX_get_clock();
+
+ unsigned long long sec = (unsigned long long)now;
+ unsigned long long pre = (now - sec) * smpi_rastro_resolution();
+ smpi_bench_begin();
+ return (unsigned long long)sec * smpi_rastro_resolution() + pre;
+}
+
+/* ****************************** Functions related to the SMPI_SAMPLE_ macros ************************************/
+typedef struct {
+ 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;
+
+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());
+ }
+}
+static int sample_enough_benchs(local_data_t *data) {
+ int res = data->count >= data->iters;
+ if (data->threshold>0.0) {
+ if (data->count <2)
+ res = 0; // not enough data
+ if (data->relstderr > data->threshold)
+ res = 0; // stderr too high yet
+ }
+ XBT_DEBUG("%s (count:%d iter:%d stderr:%f thres:%f mean:%fs)",
+ (res?"enough benchs":"need more data"),
+ data->count, data->iters, data->relstderr, data->threshold, data->mean);
+ return res;
+}
+
+void 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;
+
+ smpi_bench_end(); /* Take time from previous, unrelated computation into account */
+ smpi_process_set_sampling(1);
+
+ if (!samples)
+ samples = xbt_dict_new_homogeneous(free);
+
+ data = xbt_dict_get_or_null(samples, loc);
+ if (!data) {
+ xbt_assert(threshold>0 || iters>0,
+ "You should provide either a positive amount of iterations to bench, or a positive maximal stderr (or both)");
+ 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->benching = 1; // If we have no data, we need at least one
+ data->mean = 0;
+ xbt_dict_set(samples, loc, data, NULL);
+ XBT_DEBUG("XXXXX First time ever on benched nest %s.",loc);
+ } else {
+ if (data->iters != iters || data->threshold != threshold) {
+ XBT_ERROR("Asked to bench block %s with different settings %d, %f is not %d, %f. How did you manage to give two numbers at the same line??",
+ loc, data->iters, data->threshold, iters,threshold);
+ THROW_IMPOSSIBLE;
+ }
+
+ // if we already have some data, check whether sample_2 should get one more bench or whether it should emulate the computation instead
+ data->benching = !sample_enough_benchs(data);
+ XBT_DEBUG("XXXX Re-entering the benched nest %s. %s",loc, (data->benching?"more benching needed":"we have enough data, skip computes"));
+ }
+ xbt_free(loc);
+}
+
+int smpi_sample_2(int global, const char *file, int line)
+{
+ 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);
+ XBT_DEBUG("sample2 %s",loc);
+ xbt_free(loc);
+
+ if (data->benching==1) {
+ // 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);
+ 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_process_set_sampling(0);
+ res = 0; // prepare to capture future, unrelated computations
+ }
+ smpi_bench_begin();
+ return res;
+}
+