/* 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 <unistd.h>
+#include <sys/syscall.h>
+#include <linux/futex.h>
+#include <errno.h>
#include "parmap_private.h"
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(xbt_parmap, xbt, "parmap: parallel map");
static void *_xbt_parmap_worker_main(void *parmap);
+static void futex_wait(int *uaddr, int val);
+static void futex_wake(int *uaddr, int val);
+
xbt_parmap_t xbt_parmap_new(unsigned int num_workers)
{
unsigned int i;
DEBUG1("Create new parmap (%u workers)", num_workers);
- /* Initialize thread pool data structure */
+ /* Initialize the thread pool data structure */
xbt_parmap_t parmap = xbt_new0(s_xbt_parmap_t, 1);
- parmap->mutex = xbt_os_mutex_init();
- parmap->job_posted = xbt_os_cond_init();
- parmap->all_done = xbt_os_cond_init();
- parmap->flags = xbt_new0(e_xbt_parmap_flag_t, num_workers + 1);
parmap->num_workers = num_workers;
- parmap->num_idle_workers = 0;
- parmap->workers_max_id = 0;
+ parmap->status = PARMAP_WORK;
- /* Init our flag to wait (for workers' initialization) */
- parmap->flags[num_workers] = PARMAP_WAIT;
+ parmap->workers_ready = xbt_new0(s_xbt_barrier_t, 1);
+ xbt_barrier_init(parmap->workers_ready, num_workers + 1);
+ parmap->workers_done = xbt_new0(s_xbt_barrier_t, 1);
+ xbt_barrier_init(parmap->workers_done, num_workers + 1);
/* Create the pool of worker threads */
for(i=0; i < num_workers; i++){
xbt_os_thread_detach(worker);
}
- /* wait for the workers to initialize */
- xbt_os_mutex_acquire(parmap->mutex);
- while(parmap->flags[num_workers] == PARMAP_WAIT)
- xbt_os_cond_wait(parmap->all_done, parmap->mutex);
- xbt_os_mutex_release(parmap->mutex);
-
return parmap;
}
{
DEBUG1("Destroy parmap %p", parmap);
- unsigned int i;
-
- /* Lock the parmap, then signal every worker an wait for each to finish */
- xbt_os_mutex_acquire(parmap->mutex);
- for(i=0; i < parmap->num_workers; i++){
- parmap->flags[i] = PARMAP_DESTROY;
- }
+ parmap->status = PARMAP_DESTROY;
- xbt_os_cond_broadcast(parmap->job_posted);
- while(parmap->num_workers){
- DEBUG1("Still %u workers, waiting...", parmap->num_workers);
- xbt_os_cond_wait(parmap->all_done, parmap->mutex);
- }
+ xbt_barrier_wait(parmap->workers_ready);
+ DEBUG0("Kill job sent");
+ xbt_barrier_wait(parmap->workers_done);
- /* Destroy pool's data structures */
- xbt_os_cond_destroy(parmap->job_posted);
- xbt_os_cond_destroy(parmap->all_done);
- xbt_free(parmap->flags);
- xbt_os_mutex_release(parmap->mutex);
- xbt_os_mutex_destroy(parmap->mutex);
+ xbt_free(parmap->workers_ready);
+ xbt_free(parmap->workers_done);
xbt_free(parmap);
}
-void xbt_parmap_apply(xbt_parmap_t parmap, void_f_pvoid_t fun, xbt_dynar_t data)
+ void xbt_parmap_apply(xbt_parmap_t parmap, void_f_pvoid_t fun, xbt_dynar_t data)
{
- unsigned int i;
- unsigned int myflag_idx = parmap->num_workers;
-
- /* Assign resources to worker threads */
- xbt_os_mutex_acquire(parmap->mutex);
+ /* Assign resources to worker threads*/
parmap->fun = fun;
parmap->data = data;
- parmap->num_idle_workers = 0;
-
- /* Set worker flags to work */
- for(i=0; i < parmap->num_workers; i++){
- parmap->flags[i] = PARMAP_WORK;
- }
-
- /* Set our flag to wait (for the job to be completed)*/
- parmap->flags[myflag_idx] = PARMAP_WAIT;
/* Notify workers that there is a job */
- xbt_os_cond_broadcast(parmap->job_posted);
+ xbt_barrier_wait(parmap->workers_ready);
DEBUG0("Job dispatched, lets wait...");
-
- /* wait for the workers to finish */
- while(parmap->flags[myflag_idx] == PARMAP_WAIT)
- xbt_os_cond_wait(parmap->all_done, parmap->mutex);
+ xbt_barrier_wait(parmap->workers_done);
DEBUG0("Job done");
parmap->fun = NULL;
parmap->data = NULL;
-
- xbt_os_mutex_release(parmap->mutex);
- return;
}
static void *_xbt_parmap_worker_main(void *arg)
xbt_parmap_t parmap = (xbt_parmap_t)arg;
/* Fetch a worker id */
- xbt_os_mutex_acquire(parmap->mutex);
- worker_id = parmap->workers_max_id++;
- xbt_os_mutex_release(parmap->mutex);
+ worker_id = __sync_fetch_and_add(&parmap->workers_max_id, 1);
DEBUG1("New worker thread created (%u)", worker_id);
/* Worker's main loop */
while(1){
- xbt_os_mutex_acquire(parmap->mutex);
- parmap->flags[worker_id] = PARMAP_WAIT;
- parmap->num_idle_workers++;
-
- /* If everybody is done set the parmap work flag and signal it */
- if(parmap->num_idle_workers == parmap->num_workers){
- DEBUG1("Worker %u: All done, signal the parmap", worker_id);
- parmap->flags[parmap->num_workers] = PARMAP_WORK;
- xbt_os_cond_signal(parmap->all_done);
- }
+ xbt_barrier_wait(parmap->workers_ready);
+
+ if(parmap->status == PARMAP_WORK){
+ DEBUG1("Worker %u got a job", worker_id);
- /* If the wait flag is set then ... wait. */
- while(parmap->flags[worker_id] == PARMAP_WAIT)
- xbt_os_cond_wait(parmap->job_posted, parmap->mutex);
+ /* Compute how much data does every worker gets */
+ data_size = (xbt_dynar_length(parmap->data) / parmap->num_workers)
+ + ((xbt_dynar_length(parmap->data) % parmap->num_workers) ? 1 : 0);
- DEBUG1("Worker %u got a job", worker_id);
+ /* Each worker data segment starts in a position associated with its id*/
+ data_start = data_size * worker_id;
- /* If we are shutting down, the last worker is going to signal the
- * parmap so it can finish destroying the data structure */
- if(parmap->flags[worker_id] == PARMAP_DESTROY){
+ /* The end of the worker data segment must be bounded by the end of the data vector */
+ data_end = MIN(data_start + data_size, xbt_dynar_length(parmap->data));
+
+ DEBUG4("Worker %u: data_start=%u data_end=%u (data_size=%u)",
+ worker_id, data_start, data_end, data_size);
+
+ /* While the worker don't pass the end of it data segment apply the function */
+ while(data_start < data_end){
+ parmap->fun(*(void **)xbt_dynar_get_ptr(parmap->data, data_start));
+ data_start++;
+ }
+
+ xbt_barrier_wait(parmap->workers_done);
+
+ /* We are destroying the parmap */
+ }else{
+ xbt_barrier_wait(parmap->workers_done);
DEBUG1("Shutting down worker %u", worker_id);
- parmap->num_workers--;
- if(parmap->num_workers == 0)
- xbt_os_cond_signal(parmap->all_done);
- xbt_os_mutex_release(parmap->mutex);
return NULL;
}
- xbt_os_mutex_release(parmap->mutex);
-
- /* Compute how much data does every worker gets */
- data_size = (xbt_dynar_length(parmap->data) / parmap->num_workers)
- + ((xbt_dynar_length(parmap->data) % parmap->num_workers) ? 1 : 0);
+ }
+}
- /* Each worker data segment starts in a position associated with its id*/
- data_start = data_size * worker_id;
+static void futex_wait(int *uaddr, int val)
+{
+ DEBUG1("Waiting on futex %d", *uaddr);
+ syscall(SYS_futex, uaddr, FUTEX_WAIT_PRIVATE, val, NULL, NULL, 0);
+}
- /* The end of the worker data segment must be bounded by the end of the data vector */
- data_end = MIN(data_start + data_size, xbt_dynar_length(parmap->data));
+static void futex_wake(int *uaddr, int val)
+{
+ DEBUG1("Waking futex %d", *uaddr);
+ syscall(SYS_futex, uaddr, FUTEX_WAKE_PRIVATE, val, NULL, NULL, 0);
+}
- DEBUG4("Worker %u: data_start=%u data_end=%u (data_size=%u)", worker_id, data_start, data_end, data_size);
+/* Futex based implementation of the barrier */
+void xbt_barrier_init(xbt_barrier_t barrier, unsigned int threads_to_wait)
+{
+ barrier->threads_to_wait = threads_to_wait;
+ barrier->thread_count = 0;
+}
- /* While the worker don't pass the end of it data segment apply the function */
- while(data_start < data_end){
- parmap->fun(*(void **)xbt_dynar_get_ptr(parmap->data, data_start));
- data_start++;
- }
+void xbt_barrier_wait(xbt_barrier_t barrier)
+{
+ int myflag = 0;
+ unsigned int mycount = 0;
+
+ myflag = barrier->futex;
+ mycount = __sync_add_and_fetch(&barrier->thread_count, 1);
+ if(mycount < barrier->threads_to_wait){
+ futex_wait(&barrier->futex, myflag);
+ }else{
+ barrier->futex = __sync_add_and_fetch(&barrier->futex, 1);
+ barrier->thread_count = 0;
+ futex_wake(&barrier->futex, barrier->threads_to_wait);
}
}
void fun(void *arg)
{
- INFO1("I'm job %lu", (unsigned long)arg);
+ //INFO1("I'm job %lu", (unsigned long)arg);
}
XBT_TEST_UNIT("basic", test_parmap_basic, "Basic usage")
{
xbt_test_add0("Create the parmap");
- unsigned long j;
+ unsigned long i,j;
xbt_dynar_t data = xbt_dynar_new(sizeof(void *), NULL);
/* Create the parallel map */
- parmap = xbt_parmap_new(5);
+ parmap = xbt_parmap_new(10);
- for(j=0; j < 200; j++){
+ for(j=0; j < 100; j++){
xbt_dynar_push_as(data, void *, (void *)j);
}
- xbt_parmap_apply(parmap, fun, data);
+ for(i=0; i < 5; i++)
+ xbt_parmap_apply(parmap, fun, data);
/* Destroy the parmap */
xbt_parmap_destroy(parmap);
