# include <valgrind/valgrind.h>
#endif /* HAVE_VALGRIND_VALGRIND_H */
-#ifdef _XBT_WIN32
-#include "win32_ucontext.h"
-#include "win32_ucontext.c"
-#else
-#include "ucontext.h"
-#endif
-
-/* lower this if you want to reduce the memory consumption */
-#ifndef CONTEXT_STACK_SIZE /* allow lua to override this */
-#define CONTEXT_STACK_SIZE 128*1024
-#endif /*CONTEXT_STACK_SIZE */
-
-
typedef char * raw_stack_t;
typedef void (*rawctx_entry_point_t)(void *);
#ifdef HAVE_VALGRIND_VALGRIND_H
unsigned int valgrind_stack_id; /* the valgrind stack id */
#endif
+#ifdef TIME_BENCH
+ unsigned int thread; /* Just for measuring purposes */
+#endif
} s_smx_ctx_raw_t, *smx_ctx_raw_t;
smx_ctx_raw_t maestro_raw_context;
XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simix_context);
+#ifdef CONTEXT_THREADS
static xbt_parmap_t parmap;
+#endif
+
+static smx_context_factory_t raw_factory;
+
+#ifdef TIME_BENCH
+#include "xbt/xbt_os_time.h"
+#define NUM_THREADS 4
+static xbt_os_timer_t timer;
+static double time_thread_sr[NUM_THREADS];
+static double time_thread_ssr[NUM_THREADS];
+static double time_wasted_sr = 0;
+static double time_wasted_ssr = 0;
+static unsigned int sr_count = 0;
+static unsigned int ssr_count = 0;
+static char new_sr = 0;
+#endif
static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory)
-{
+{
+#ifdef TIME_BENCH
+ XBT_CRITICAL("Total wasted time in %u SR: %lf", sr_count, time_wasted_sr);
+ XBT_CRITICAL("Total wasted time in %u SSR: %lf", ssr_count, time_wasted_ssr);
+#endif
+
+#ifdef CONTEXT_THREADS
if(parmap)
xbt_parmap_destroy(parmap);
+#endif
return smx_ctx_base_factory_finalize(factory);
}
/* If the user provided a function for the process then use it
otherwise is the context for maestro */
if (code) {
- context->malloced_stack = xbt_malloc0(CONTEXT_STACK_SIZE);
+ context->malloced_stack = xbt_malloc0(smx_context_stack_size);
context->stack_top =
- raw_makecontext(context->malloced_stack,CONTEXT_STACK_SIZE,
+ raw_makecontext(context->malloced_stack, smx_context_stack_size,
(void(*)(void*))smx_ctx_raw_wrapper,context);
#ifdef HAVE_VALGRIND_VALGRIND_H
context->valgrind_stack_id =
VALGRIND_STACK_REGISTER(context->malloced_stack,
- context->malloced_stack + CONTEXT_STACK_SIZE);
+ context->malloced_stack + smx_context_stack_size);
#endif /* HAVE_VALGRIND_VALGRIND_H */
}else{
context)->valgrind_stack_id);
#endif /* HAVE_VALGRIND_VALGRIND_H */
+ free(((smx_ctx_raw_t)context)->malloced_stack);
}
smx_ctx_base_free(context);
}
smx_ctx_raw_stop((smx_context_t) context);
}
-static void smx_ctx_raw_resume(smx_context_t context)
+static void smx_ctx_raw_resume(smx_process_t process)
{
- smx_current_context = context;
+ smx_ctx_raw_t context = (smx_ctx_raw_t)process->context;
+ smx_current_context = (smx_context_t)context;
raw_swapcontext(
&((smx_ctx_raw_t) context)->old_stack_top,
((smx_ctx_raw_t) context)->stack_top);
}
-static void smx_ctx_raw_runall(xbt_dynar_t processes)
+#ifdef TIME_BENCH
+static void smx_ctx_raw_runall_serial(xbt_dynar_t processes)
{
smx_process_t process;
unsigned int cursor;
+ double elapsed = 0;
+ double tmax = 0;
+ unsigned long num_proc = xbt_dynar_length(processes);
+ unsigned int t=0;
+ unsigned int data_size = (num_proc / NUM_THREADS) + ((num_proc % NUM_THREADS) ? 1 : 0);
+
+ ssr_count++;
+ time_thread_ssr[0] = 0;
xbt_dynar_foreach(processes, cursor, process) {
- DEBUG2("Schedule item %u of %lu",cursor,xbt_dynar_length(processes));
- smx_ctx_raw_resume(process->context);
+ XBT_DEBUG("Schedule item %u of %lu",cursor,xbt_dynar_length(processes));
+ if(cursor >= t * data_size + data_size){
+ if(time_thread_ssr[t] > tmax)
+ tmax = time_thread_ssr[t];
+ t++;
+ time_thread_ssr[t] = 0;
+ }
+
+ if(new_sr){
+ ((smx_ctx_raw_t)process->context)->thread = t;
+ time_thread_sr[t] = 0;
+ }
+
+ xbt_os_timer_start(timer);
+ smx_ctx_raw_resume(process);
+ xbt_os_timer_stop(timer);
+ elapsed = xbt_os_timer_elapsed(timer);
+ time_thread_ssr[t] += elapsed;
+ time_thread_sr[((smx_ctx_raw_t)process->context)->thread] += elapsed;
+ }
+
+ if(new_sr)
+ new_sr = FALSE;
+
+ if(time_thread_ssr[t] > tmax)
+ tmax = time_thread_ssr[t];
+
+ for(cursor=0; cursor <= t; cursor++){
+ XBT_VERB("Time SSR thread %u = %lf (max %lf)", cursor, time_thread_ssr[cursor], tmax);
+ time_wasted_ssr += tmax - time_thread_ssr[cursor];
}
+
xbt_dynar_reset(processes);
}
-static void smx_ctx_raw_resume_parallel(smx_process_t process)
+void smx_ctx_raw_new_sr(void);
+void smx_ctx_raw_new_sr(void)
{
- smx_ctx_raw_t context = (smx_ctx_raw_t)process->context;
- xbt_os_thread_set_extra_data(context);
- raw_swapcontext(
- &context->old_stack_top,
- context->stack_top);
- xbt_os_thread_set_extra_data(NULL);
+ int i;
+ double tmax = 0;
+ new_sr = TRUE;
+ sr_count++;
+ for(i=0; i < NUM_THREADS; i++){
+ if(time_thread_sr[i] > tmax)
+ tmax = time_thread_sr[i];
+ }
+
+ for(i=0; i < NUM_THREADS; i++){
+ XBT_VERB("Time SR thread %u = %lf (max %lf)", i, time_thread_sr[i], tmax);
+ time_wasted_sr += tmax - time_thread_sr[i];
+ }
+
+ XBT_VERB("New scheduling round");
+}
+#else
+static void smx_ctx_raw_runall_serial(xbt_dynar_t processes)
+{
+ smx_process_t process;
+ unsigned int cursor;
+
+ xbt_dynar_foreach(processes, cursor, process) {
+ XBT_DEBUG("Schedule item %u of %lu",cursor,xbt_dynar_length(processes));
+ smx_ctx_raw_resume(process);
+ }
+ xbt_dynar_reset(processes);
}
+#endif
static void smx_ctx_raw_runall_parallel(xbt_dynar_t processes)
{
- xbt_parmap_apply(parmap, (void_f_pvoid_t)smx_ctx_raw_resume_parallel, processes);
+#ifdef CONTEXT_THREADS
+ xbt_parmap_apply(parmap, (void_f_pvoid_t)smx_ctx_raw_resume, processes);
+#endif
xbt_dynar_reset(processes);
}
static smx_context_t smx_ctx_raw_self_parallel(void)
{
- smx_context_t self_context = (smx_context_t) xbt_os_thread_get_extra_data();
- return self_context ? self_context : (smx_context_t) maestro_raw_context;
+ return smx_current_context;
+}
+
+static int smx_ctx_raw_get_thread_id(){
+ return (int)(unsigned long)xbt_os_thread_get_extra_data();
+}
+
+static void smx_ctx_raw_runall(xbt_dynar_t processes)
+{
+ if (xbt_dynar_length(processes) >= SIMIX_context_get_parallel_threshold()) {
+ XBT_DEBUG("Runall // %lu", xbt_dynar_length(processes));
+ raw_factory->self = smx_ctx_raw_self_parallel;
+ raw_factory->get_thread_id = smx_ctx_raw_get_thread_id;
+ smx_ctx_raw_runall_parallel(processes);
+ } else {
+ XBT_DEBUG("Runall serial %lu", xbt_dynar_length(processes));
+ raw_factory->self = smx_ctx_base_self;
+ raw_factory->get_thread_id = smx_ctx_base_get_thread_id;
+ smx_ctx_raw_runall_serial(processes);
+ }
}
void SIMIX_ctx_raw_factory_init(smx_context_factory_t *factory)
{
- VERB0("Using raw contexts. Because the glibc is just not good enough for us.");
+ XBT_VERB("Using raw contexts. Because the glibc is just not good enough for us.");
smx_ctx_base_factory_init(factory);
(*factory)->finalize = smx_ctx_raw_factory_finalize;
(*factory)->suspend = smx_ctx_raw_suspend;
(*factory)->name = "smx_raw_context_factory";
- if(_surf_parallel_contexts){
-#ifdef CONTEXT_THREADS /* To use parallel ucontexts a thread pool is needed */
- parmap = xbt_parmap_new(2);
- (*factory)->runall = smx_ctx_raw_runall_parallel;
- (*factory)->self = smx_ctx_raw_self_parallel;
-#else
- THROW0(arg_error, 0, "No thread support for parallel context execution");
+ if (SIMIX_context_is_parallel()) {
+#ifdef CONTEXT_THREADS
+ parmap = xbt_parmap_new(SIMIX_context_get_nthreads());
#endif
- }else{
- (*factory)->runall = smx_ctx_raw_runall;
+ if (SIMIX_context_get_parallel_threshold() > 1) {
+ /* choose dynamically */
+ (*factory)->runall = smx_ctx_raw_runall;
+ }
+ else {
+ /* always parallel */
+ (*factory)->self = smx_ctx_raw_self_parallel;
+ (*factory)->get_thread_id = smx_ctx_raw_get_thread_id;
+ (*factory)->runall = smx_ctx_raw_runall_parallel;
+ }
+ }
+ else {
+ /* always serial */
+ (*factory)->self = smx_ctx_base_self;
+ (*factory)->get_thread_id = smx_ctx_base_get_thread_id;
+ (*factory)->runall = smx_ctx_raw_runall_serial;
}
+ raw_factory = *factory;
+#ifdef TIME_BENCH
+ timer = xbt_os_timer_new();
+#endif
}
