/* context_raw - fast context switching inspired from System V ucontexts */
-/* Copyright (c) 2009-2013. The SimGrid Team.
+/* Copyright (c) 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. */
-
+#include <math.h>
#include "smx_private.h"
#include "xbt/parmap.h"
+#include "xbt/dynar.h"
#include "mc/mc.h"
-#ifdef HAVE_VALGRIND_VALGRIND_H
-# include <valgrind/valgrind.h>
-#endif /* HAVE_VALGRIND_VALGRIND_H */
-
typedef char * raw_stack_t;
typedef void (*rawctx_entry_point_t)(void *);
s_smx_ctx_base_t super; /* Fields of super implementation */
char *malloced_stack; /* malloced area containing the stack */
raw_stack_t stack_top; /* pointer to stack top (within previous area) */
-#ifdef HAVE_VALGRIND_VALGRIND_H
- unsigned int valgrind_stack_id; /* the valgrind stack id */
-#endif
#ifdef TIME_BENCH_PER_SR
unsigned int thread; /* Just for measuring purposes */
#endif
static smx_ctx_raw_t* raw_workers_context; /* space to save the worker context in each thread */
static unsigned long raw_threads_working; /* number of threads that have started their work */
static xbt_os_thread_key_t raw_worker_id_key; /* thread-specific storage for the thread id */
+#endif
+#ifdef ADAPTIVE_THRESHOLD
+#define SCHED_ROUND_LIMIT 5
+xbt_os_timer_t round_time;
+double par_time,seq_time;
+double par_ratio,seq_ratio;
+int reached_seq_limit, reached_par_limit;
+static unsigned int par_proc_that_ran = 0,seq_proc_that_ran = 0; /* Counters of processes that have run in SCHED_ROUND_LIMIT scheduling rounds */
+static unsigned int seq_sched_round=0, par_sched_round=0; /* Amount of SR that ran serial/parallel*/
+/*Varables used to calculate running variance and mean*/
+double prev_avg_par_proc=0,prev_avg_seq_proc=0;
+double delta=0;
+double s_par_proc=0,s_seq_proc=0; /*Standard deviation of number of processes computed in par/seq during the current simulation*/
+double avg_par_proc=0,sd_par_proc=0;
+double avg_seq_proc=0,sd_seq_proc=0;
+long long par_window=(long long)HUGE_VAL,seq_window=0;
#endif
static unsigned long raw_process_index = 0; /* index of the next process to run in the
* list of runnable processes */
static smx_ctx_raw_t raw_maestro_context;
-
extern raw_stack_t raw_makecontext(char* malloced_stack, int stack_size,
rawctx_entry_point_t entry_point, void* arg);
extern void raw_swapcontext(raw_stack_t* old, raw_stack_t new);
#endif
" mov %rdi,%rax\n" /* stack */
" add %rsi,%rax\n" /* size */
+ " andq $-16, %rax\n" /* align stack */
" movq $0, -8(%rax)\n" /* @return for func */
" mov %rdx,-16(%rax)\n" /* func */
" mov %rcx,-24(%rax)\n" /* arg/rdi */
#endif
" movl 4(%esp),%eax\n" /* stack */
" addl 8(%esp),%eax\n" /* size */
+ " andl $-16, %eax\n" /* align stack */
" movl 12(%esp),%ecx\n" /* func */
" movl 16(%esp),%edx\n" /* arg */
" movl %edx, -4(%eax)\n"
static char new_sr = 0;
#endif
+#ifdef TIME_BENCH_ENTIRE_SRS
+static unsigned int sr_count = 0;
+static xbt_os_timer_t timer;
+#endif
+
static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory);
static smx_context_t smx_ctx_raw_create_context(xbt_main_func_t code, int argc,
- char **argv, void_pfn_smxprocess_t cleanup_func, void *data);
+ char **argv, void_pfn_smxprocess_t cleanup_func, smx_process_t process);
static void smx_ctx_raw_free(smx_context_t context);
static void smx_ctx_raw_wrapper(smx_ctx_raw_t context);
static void smx_ctx_raw_stop(smx_context_t context);
static void smx_ctx_raw_suspend_serial(smx_context_t context);
static void smx_ctx_raw_resume_serial(smx_process_t first_process);
+#ifdef TIME_BENCH_PER_SR
+static void smx_ctx_raw_runall_serial(xbt_dynar_t processes);
+void smx_ctx_raw_new_sr(void);
+#else
static void smx_ctx_raw_runall_serial(void);
+#endif
static void smx_ctx_raw_suspend_parallel(smx_context_t context);
static void smx_ctx_raw_resume_parallel(smx_process_t first_process);
static void smx_ctx_raw_runall_parallel(void);
(*factory)->runall = smx_ctx_raw_runall_serial;
(*factory)->suspend = smx_ctx_raw_suspend_serial;
}
+#ifdef TIME_BENCH_ENTIRE_SRS
+ (*factory)->runall = smx_ctx_raw_runall;
+ (*factory)->suspend = NULL;
+ timer = xbt_os_timer_new();
+#endif
+
+#ifdef ADAPTIVE_THRESHOLD
+ round_time = xbt_os_timer_new();
+ reached_seq_limit = 0;
+ reached_par_limit = 0;
+#endif
+
#ifdef TIME_BENCH_PER_SR
timer = xbt_os_timer_new();
#endif
static int smx_ctx_raw_factory_finalize(smx_context_factory_t *factory)
{
#ifdef TIME_BENCH_PER_SR
- 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);
+ XBT_VERB("Total wasted time in %u SR: %f", sr_count, time_wasted_sr);
+ XBT_VERB("Total wasted time in %u SSR: %f", ssr_count, time_wasted_ssr);
#endif
#ifdef CONTEXT_THREADS
* \param argv arguments to pass to the main function
* \param cleanup_func a function to call to free the user data when the
* context finished
- * \param data user data
+ * \param process SIMIX process
*/
static smx_context_t
smx_ctx_raw_create_context(xbt_main_func_t code, int argc, char **argv,
- void_pfn_smxprocess_t cleanup_func,
- void *data)
+ void_pfn_smxprocess_t cleanup_func,
+ smx_process_t process)
{
smx_ctx_raw_t context =
argc,
argv,
cleanup_func,
- data);
+ process);
/* if the user provided a function for the process then use it,
otherwise it is the context for maestro */
if (code) {
- context->malloced_stack = xbt_malloc0(smx_context_stack_size);
+ context->malloced_stack = SIMIX_context_stack_new();
context->stack_top =
- raw_makecontext(context->malloced_stack, smx_context_stack_size,
- (void_f_pvoid_t) smx_ctx_raw_wrapper, context);
-
-#ifdef HAVE_VALGRIND_VALGRIND_H
- context->valgrind_stack_id =
- VALGRIND_STACK_REGISTER(context->malloced_stack,
- context->malloced_stack + smx_context_stack_size);
-#endif /* HAVE_VALGRIND_VALGRIND_H */
+ raw_makecontext(context->malloced_stack,
+ smx_context_usable_stack_size,
+ (void_f_pvoid_t)smx_ctx_raw_wrapper, context);
} else {
- if(data != NULL && raw_maestro_context==NULL)
+ if(process != NULL && raw_maestro_context==NULL)
raw_maestro_context = context;
if(MC_is_active())
static void smx_ctx_raw_free(smx_context_t context)
{
if (context) {
-
-#ifdef HAVE_VALGRIND_VALGRIND_H
- VALGRIND_STACK_DEREGISTER(((smx_ctx_raw_t)
- context)->valgrind_stack_id);
-#endif /* HAVE_VALGRIND_VALGRIND_H */
-
- free(((smx_ctx_raw_t) context)->malloced_stack);
+ SIMIX_context_stack_delete(((smx_ctx_raw_t) context)->malloced_stack);
}
smx_ctx_base_free(context);
}
{
/* determine the next context */
smx_context_t next_context;
- unsigned long int i = raw_process_index++;
-
+ unsigned long int i;
+#ifdef TIME_BENCH_PER_SR
+ i = ++raw_process_index;
+#else
+ i = raw_process_index++;
+#endif
if (i < xbt_dynar_length(simix_global->process_to_run)) {
/* execute the next process */
XBT_DEBUG("Run next process");
{
smx_process_t process;
unsigned int cursor;
-
double elapsed = 0;
double tmax = 0;
- unsigned long num_proc = xbt_dynar_length(processes);
+ unsigned long num_proc = xbt_dynar_length(simix_global->process_to_run);
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) {
- 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_cputimer_start(timer);
- smx_ctx_raw_resume(process);
- xbt_os_cputimer_stop(timer);
- elapsed = xbt_os_timer_elapsed(timer);
- time_thread_ssr[t] += elapsed;
- time_thread_sr[((smx_ctx_raw_t)process->context)->thread] += elapsed;
+ xbt_dynar_foreach(processes, cursor, process){
+ XBT_VERB("Schedule item %u of %lu",cursor,num_proc);
+ 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_cputimer_start(timer);
+ smx_ctx_raw_resume_serial(process);
+ xbt_os_cputimer_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)
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);
+ XBT_VERB("Time SSR thread %u = %f (max %f)", cursor, time_thread_ssr[cursor], tmax);
time_wasted_ssr += tmax - time_thread_ssr[cursor];
}
}
-void smx_ctx_raw_new_sr(void);
void smx_ctx_raw_new_sr(void)
{
int i;
}
for(i=0; i < NUM_THREADS; i++){
- XBT_VERB("Time SR thread %u = %lf (max %lf)", i, time_thread_sr[i], tmax);
+ XBT_CRITICAL("Time SR thread %u = %f (max %f)", i, time_thread_sr[i], tmax);
time_wasted_sr += tmax - time_thread_sr[i];
}
- XBT_VERB("New scheduling round");
+ XBT_CRITICAL("Total time SR %u = %f, %d", sr_count, tmax, xbt_dynar_length(simix_global->process_that_ran));
+ XBT_CRITICAL("New scheduling round");
}
#else
-
/**
* \brief Resumes sequentially all processes ready to run.
*/
xbt_parmap_apply(raw_parmap, (void_f_pvoid_t) smx_ctx_raw_resume_parallel,
simix_global->process_to_run);
#else
- xbt_die("You asked for a parallel execution, but you don't have any threads.")
+ xbt_die("You asked for a parallel execution, but you don't have any threads.");
#endif
}
/**
* \brief Resumes all processes ready to run.
*/
+#ifdef ADAPTIVE_THRESHOLD
static void smx_ctx_raw_runall(void)
{
unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
+ unsigned long threshold = SIMIX_context_get_parallel_threshold();
+ reached_seq_limit = (seq_sched_round % SCHED_ROUND_LIMIT == 0);
+ reached_par_limit = (par_sched_round % SCHED_ROUND_LIMIT == 0);
+
+ if(reached_seq_limit && reached_par_limit){
+ par_ratio = (par_proc_that_ran != 0) ? (par_time / (double)par_proc_that_ran) : 0;
+ seq_ratio = (seq_proc_that_ran != 0) ? (seq_time / (double)seq_proc_that_ran) : 0;
+ if(seq_ratio > par_ratio){
+ if(nb_processes < avg_par_proc) {
+ threshold = (threshold>2) ? threshold - 1 : threshold ;
+ SIMIX_context_set_parallel_threshold(threshold);
+ }
+ } else {
+ if(nb_processes > avg_seq_proc){
+ SIMIX_context_set_parallel_threshold(threshold+1);
+ }
+ }
+ }
+
+ //XBT_CRITICAL("Thresh: %d", SIMIX_context_get_parallel_threshold());
if (nb_processes >= SIMIX_context_get_parallel_threshold()) {
- XBT_DEBUG("Runall // %lu", nb_processes);
simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel;
- smx_ctx_raw_runall_parallel();
+ if(nb_processes < par_window){
+ par_sched_round++;
+ xbt_os_walltimer_start(round_time);
+ smx_ctx_raw_runall_parallel();
+ xbt_os_walltimer_stop(round_time);
+ par_time += xbt_os_timer_elapsed(round_time);
+
+ prev_avg_par_proc = avg_par_proc;
+ delta = nb_processes - avg_par_proc;
+ avg_par_proc = (par_sched_round==1) ? nb_processes : avg_par_proc + delta / (double) par_sched_round;
+
+ if(par_sched_round>=2){
+ s_par_proc = s_par_proc + (nb_processes - prev_avg_par_proc) * delta;
+ sd_par_proc = sqrt(s_par_proc / (par_sched_round-1));
+ par_window = (int) (avg_par_proc + sd_par_proc);
+ }else{
+ sd_par_proc = 0;
+ }
+
+ par_proc_that_ran += nb_processes;
+ } else{
+ smx_ctx_raw_runall_parallel();
+ }
} else {
- XBT_DEBUG("Runall serial %lu", nb_processes);
simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial;
- smx_ctx_raw_runall_serial();
+ if(nb_processes > seq_window){
+ seq_sched_round++;
+ xbt_os_walltimer_start(round_time);
+ smx_ctx_raw_runall_serial();
+ xbt_os_walltimer_stop(round_time);
+ seq_time += xbt_os_timer_elapsed(round_time);
+
+ prev_avg_seq_proc = avg_seq_proc;
+ delta = (nb_processes-avg_seq_proc);
+ avg_seq_proc = (seq_sched_round==1) ? nb_processes : avg_seq_proc + delta / (double) seq_sched_round;
+
+ if(seq_sched_round>=2){
+ s_seq_proc = s_seq_proc + (nb_processes - prev_avg_seq_proc)*delta;
+ sd_seq_proc = sqrt(s_seq_proc / (seq_sched_round-1));
+ seq_window = (int) (avg_seq_proc - sd_seq_proc);
+ } else {
+ sd_seq_proc = 0;
+ }
+
+ seq_proc_that_ran += nb_processes;
+ } else {
+ smx_ctx_raw_runall_serial();
+ }
}
}
+
+#else
+
+static void smx_ctx_raw_runall(void)
+{
+#ifdef TIME_BENCH_ENTIRE_SRS
+ sr_count++;
+ timer = xbt_os_timer_new();
+ double elapsed = 0;
+#endif
+ unsigned long nb_processes = xbt_dynar_length(simix_global->process_to_run);
+ if (SIMIX_context_is_parallel() && SIMIX_context_get_parallel_threshold()<nb_processes) {
+ XBT_DEBUG("Runall // %lu", nb_processes);
+ simix_global->context_factory->suspend = smx_ctx_raw_suspend_parallel;
+
+ #ifdef TIME_BENCH_ENTIRE_SRS
+ xbt_os_walltimer_start(timer);
+ #endif
+
+ smx_ctx_raw_runall_parallel();
+
+ #ifdef TIME_BENCH_ENTIRE_SRS
+ xbt_os_walltimer_stop(timer);
+ elapsed = xbt_os_timer_elapsed(timer);
+ #endif
+ } else {
+ XBT_DEBUG("Runall serial %lu", nb_processes);
+ simix_global->context_factory->suspend = smx_ctx_raw_suspend_serial;
+
+ #ifdef TIME_BENCH_PER_SR
+ smx_ctx_raw_runall_serial(simix_global->process_to_run);
+ #else
+
+ #ifdef TIME_BENCH_ENTIRE_SRS
+ xbt_os_walltimer_start(timer);
+ #endif
+
+ smx_ctx_raw_runall_serial();
+
+ #ifdef TIME_BENCH_ENTIRE_SRS
+ xbt_os_walltimer_stop(timer);
+ elapsed = xbt_os_timer_elapsed(timer);
+ #endif
+ #endif
+ }
+
+#ifdef TIME_BENCH_ENTIRE_SRS
+ XBT_CRITICAL("Total time SR %u = %f, %d", sr_count, elapsed, nb_processes);
+#endif
+}
+#endif
