Parmap review.

* synchronizations are made a bit simpler;
* helper functions are renamed for more clarity;
* synchronization at the end of xbt_parmap_new() is removed;
* the working loop is made faster by manually inlining xbt_parmap_next();
* finally, the posix mode is fixed.

diff --git a/src/xbt/parmap.c b/src/xbt/parmap.c
index 13c7baf..fa073e8 100644 (file)
--- a/src/xbt/parmap.c
+++ b/src/xbt/parmap.c
@@ -25,40 +25,41 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(xbt_parmap, xbt, "parmap: parallel map");
 XBT_LOG_NEW_SUBCATEGORY(xbt_parmap_unit, xbt_parmap, "parmap unit testing");
 
 typedef enum {
-  XBT_PARMAP_WORK = 0,
+  XBT_PARMAP_WORK,
   XBT_PARMAP_DESTROY
 } e_xbt_parmap_flag_t;
 
 static void xbt_parmap_set_mode(xbt_parmap_t parmap, e_xbt_parmap_mode_t mode);
 static void *xbt_parmap_worker_main(void *parmap);
+static void xbt_parmap_work(xbt_parmap_t parmap);
 
-static void xbt_parmap_posix_start(xbt_parmap_t parmap);
-static void xbt_parmap_posix_end(xbt_parmap_t parmap);
-static void xbt_parmap_posix_signal(xbt_parmap_t parmap);
-static void xbt_parmap_posix_wait(xbt_parmap_t parmap);
+static void xbt_parmap_posix_master_wait(xbt_parmap_t parmap);
+static void xbt_parmap_posix_worker_signal(xbt_parmap_t parmap);
+static void xbt_parmap_posix_master_signal(xbt_parmap_t parmap);
+static void xbt_parmap_posix_worker_wait(xbt_parmap_t parmap, unsigned round);
 
 #ifdef HAVE_FUTEX_H
-static void xbt_parmap_futex_start(xbt_parmap_t parmap);
-static void xbt_parmap_futex_end(xbt_parmap_t parmap);
-static void xbt_parmap_futex_signal(xbt_parmap_t parmap);
-static void xbt_parmap_futex_wait(xbt_parmap_t parmap);
-static void futex_wait(int *uaddr, int val);
-static void futex_wake(int *uaddr, int val);
+static void xbt_parmap_futex_master_wait(xbt_parmap_t parmap);
+static void xbt_parmap_futex_worker_signal(xbt_parmap_t parmap);
+static void xbt_parmap_futex_master_signal(xbt_parmap_t parmap);
+static void xbt_parmap_futex_worker_wait(xbt_parmap_t parmap, unsigned round);
+static void futex_wait(unsigned *uaddr, unsigned val);
+static void futex_wake(unsigned *uaddr, unsigned val);
 #endif
 
-static void xbt_parmap_busy_start(xbt_parmap_t parmap);
-static void xbt_parmap_busy_end(xbt_parmap_t parmap);
-static void xbt_parmap_busy_signal(xbt_parmap_t parmap);
-static void xbt_parmap_busy_wait(xbt_parmap_t parmap);
+static void xbt_parmap_busy_master_wait(xbt_parmap_t parmap);
+static void xbt_parmap_busy_worker_signal(xbt_parmap_t parmap);
+static void xbt_parmap_busy_master_signal(xbt_parmap_t parmap);
+static void xbt_parmap_busy_worker_wait(xbt_parmap_t parmap, unsigned round);
 
 /**
  * \brief Parallel map structure
  */
 typedef struct s_xbt_parmap {
   e_xbt_parmap_flag_t status;      /**< is the parmap active or being destroyed? */
-  int work;                        /**< index of the current round (1 is the first) */
-  int done;                        /**< number of rounds already done (futexes only) */
-  unsigned int thread_counter;     /**< number of threads currently working */
+  unsigned work;                   /**< index of the current round */
+  unsigned thread_counter;         /**< number of workers that have done the work */
+
   unsigned int num_workers;        /**< total number of worker threads including the controller */
   void_f_pvoid_t fun;              /**< function to run in parallel on each element of data */
   xbt_dynar_t data;                /**< parameters to pass to fun in parallel */
@@ -72,10 +73,10 @@ typedef struct s_xbt_parmap {
 
   /* fields that depend on the synchronization mode */
   e_xbt_parmap_mode_t mode;        /**< synchronization mode */
-  void (*start_f)(xbt_parmap_t);   /**< initializes the worker threads */
-  void (*end_f)(xbt_parmap_t);     /**< finalizes the worker threads */
-  void (*signal_f)(xbt_parmap_t);  /**< wakes the workers threads to process tasks */
-  void (*wait_f)(xbt_parmap_t);    /**< waits for more work */
+  void (*master_wait_f)(xbt_parmap_t);    /**< wait for the workers to have done the work */
+  void (*worker_signal_f)(xbt_parmap_t);  /**< signal the master that a worker has done the work */
+  void (*master_signal_f)(xbt_parmap_t);  /**< wakes the workers threads to process tasks */
+  void (*worker_wait_f)(xbt_parmap_t, unsigned); /**< waits for more work */
 } s_xbt_parmap_t;
 
 /**
@@ -99,11 +100,10 @@ xbt_parmap_t xbt_parmap_new(unsigned int num_workers, e_xbt_parmap_mode_t mode)
   xbt_parmap_set_mode(parmap, mode);
 
   /* Create the pool of worker threads */
-  for (i = 0; i < num_workers - 1; i++) {
+  for (i = 1; i < num_workers; i++) {
     worker = xbt_os_thread_create(NULL, xbt_parmap_worker_main, parmap, NULL);
     xbt_os_thread_detach(worker);
   }
-  parmap->start_f(parmap);
   return parmap;
 }
 
@@ -118,7 +118,8 @@ void xbt_parmap_destroy(xbt_parmap_t parmap)
   }
 
   parmap->status = XBT_PARMAP_DESTROY;
-  parmap->signal_f(parmap);
+  parmap->master_signal_f(parmap);
+  parmap->master_wait_f(parmap);
 
   xbt_os_cond_destroy(parmap->ready_cond);
   xbt_os_mutex_destroy(parmap->ready_mutex);
@@ -147,10 +148,10 @@ static void xbt_parmap_set_mode(xbt_parmap_t parmap, e_xbt_parmap_mode_t mode)
   switch (mode) {
 
     case XBT_PARMAP_POSIX:
-      parmap->start_f = xbt_parmap_posix_start;
-      parmap->end_f = xbt_parmap_posix_end;
-      parmap->signal_f = xbt_parmap_posix_signal;
-      parmap->wait_f = xbt_parmap_posix_wait;
+      parmap->master_wait_f = xbt_parmap_posix_master_wait;
+      parmap->worker_signal_f = xbt_parmap_posix_worker_signal;
+      parmap->master_signal_f = xbt_parmap_posix_master_signal;
+      parmap->worker_wait_f = xbt_parmap_posix_worker_wait;
 
       parmap->ready_cond = xbt_os_cond_init();
       parmap->ready_mutex = xbt_os_mutex_init();
@@ -161,10 +162,10 @@ static void xbt_parmap_set_mode(xbt_parmap_t parmap, e_xbt_parmap_mode_t mode)
 
     case XBT_PARMAP_FUTEX:
 #ifdef HAVE_FUTEX_H
-      parmap->start_f = xbt_parmap_futex_start;
-      parmap->end_f = xbt_parmap_futex_end;
-      parmap->signal_f = xbt_parmap_futex_signal;
-      parmap->wait_f = xbt_parmap_futex_wait;
+      parmap->master_wait_f = xbt_parmap_futex_master_wait;
+      parmap->worker_signal_f = xbt_parmap_futex_worker_signal;
+      parmap->master_signal_f = xbt_parmap_futex_master_signal;
+      parmap->worker_wait_f = xbt_parmap_futex_worker_wait;
 
       xbt_os_cond_destroy(parmap->ready_cond);
       xbt_os_mutex_destroy(parmap->ready_mutex);
@@ -176,10 +177,10 @@ static void xbt_parmap_set_mode(xbt_parmap_t parmap, e_xbt_parmap_mode_t mode)
 #endif
 
     case XBT_PARMAP_BUSY_WAIT:
-      parmap->start_f = xbt_parmap_busy_start;
-      parmap->end_f = xbt_parmap_busy_end;
-      parmap->signal_f = xbt_parmap_busy_signal;
-      parmap->wait_f = xbt_parmap_busy_wait;
+      parmap->master_wait_f = xbt_parmap_busy_master_wait;
+      parmap->worker_signal_f = xbt_parmap_busy_worker_signal;
+      parmap->master_signal_f = xbt_parmap_busy_master_signal;
+      parmap->worker_wait_f = xbt_parmap_busy_worker_wait;
 
       xbt_os_cond_destroy(parmap->ready_cond);
       xbt_os_mutex_destroy(parmap->ready_mutex);
@@ -205,7 +206,9 @@ void xbt_parmap_apply(xbt_parmap_t parmap, void_f_pvoid_t fun, xbt_dynar_t data)
   parmap->fun = fun;
   parmap->data = data;
   parmap->index = 0;
-  parmap->signal_f(parmap);
+  parmap->master_signal_f(parmap);
+  xbt_parmap_work(parmap);
+  parmap->master_wait_f(parmap);
   XBT_DEBUG("Job done");
 }
 
@@ -225,6 +228,14 @@ void* xbt_parmap_next(xbt_parmap_t parmap)
   return NULL;
 }
 
+static void xbt_parmap_work(xbt_parmap_t parmap)
+{
+  unsigned index;
+  while ((index = __sync_fetch_and_add(&parmap->index, 1))
+         < xbt_dynar_length(parmap->data))
+    parmap->fun(xbt_dynar_get_as(parmap->data, index, void*));
+}
+
 /**
  * \brief Main function of a worker thread.
  * \param arg the parmap
@@ -232,40 +243,38 @@ void* xbt_parmap_next(xbt_parmap_t parmap)
 static void *xbt_parmap_worker_main(void *arg)
 {
   xbt_parmap_t parmap = (xbt_parmap_t) arg;
+  unsigned round = 0;
 
   XBT_DEBUG("New worker thread created");
 
   /* Worker's main loop */
   while (1) {
-    parmap->wait_f(parmap);
+    parmap->worker_wait_f(parmap, ++round);
     if (parmap->status == XBT_PARMAP_WORK) {
 
       XBT_DEBUG("Worker got a job");
 
-      void* work = xbt_parmap_next(parmap);
-      while (work != NULL) {
-        parmap->fun(work);
-        work = xbt_parmap_next(parmap);
-      }
+      xbt_parmap_work(parmap);
+      parmap->worker_signal_f(parmap);
 
       XBT_DEBUG("Worker has finished");
 
     /* We are destroying the parmap */
     } else {
-      parmap->end_f(parmap);
+      parmap->worker_signal_f(parmap);
       return NULL;
     }
   }
 }
 
 #ifdef HAVE_FUTEX_H
-static void futex_wait(int *uaddr, int val)
+static void futex_wait(unsigned *uaddr, unsigned val)
 {
   XBT_VERB("Waiting on futex %p", uaddr);
   syscall(SYS_futex, uaddr, FUTEX_WAIT_PRIVATE, val, NULL, NULL, 0);
 }
 
-static void futex_wake(int *uaddr, int val)
+static void futex_wake(unsigned *uaddr, unsigned val)
 {
   XBT_VERB("Waking futex %p", uaddr);
   syscall(SYS_futex, uaddr, FUTEX_WAKE_PRIVATE, val, NULL, NULL, 0);
@@ -279,13 +288,14 @@ static void futex_wake(int *uaddr, int val)
  *
  * \param parmap a parmap
  */
-static void xbt_parmap_posix_start(xbt_parmap_t parmap)
+static void xbt_parmap_posix_master_wait(xbt_parmap_t parmap)
 {
-  unsigned int counter = __sync_fetch_and_add(&parmap->thread_counter, 1);
-  if (counter < parmap->num_workers) {
-    /* wait for all workers to be initialized */
+  xbt_os_mutex_acquire(parmap->done_mutex);
+  if (parmap->thread_counter < parmap->num_workers) {
+    /* wait for all workers to be ready */
     xbt_os_cond_wait(parmap->done_cond, parmap->done_mutex);
   }
+  xbt_os_mutex_release(parmap->done_mutex);
 }
 
 /**
@@ -296,14 +306,14 @@ static void xbt_parmap_posix_start(xbt_parmap_t parmap)
  *
  * \param parmap a parmap
  */
-static void xbt_parmap_posix_end(xbt_parmap_t parmap)
+static void xbt_parmap_posix_worker_signal(xbt_parmap_t parmap)
 {
-  unsigned int counter = __sync_add_and_fetch(&parmap->thread_counter, 1);
-  XBT_DEBUG("Shutting down worker %d", counter);
-  if (counter == parmap->num_workers) {
+  xbt_os_mutex_acquire(parmap->done_mutex);
+  if (++parmap->thread_counter == parmap->num_workers) {
     /* all workers have finished, wake the controller */
     xbt_os_cond_signal(parmap->done_cond);
   }
+  xbt_os_mutex_release(parmap->done_mutex);
 }
 
 /**
@@ -313,30 +323,14 @@ static void xbt_parmap_posix_end(xbt_parmap_t parmap)
  *
  * \param parmap a parmap
  */
-static void xbt_parmap_posix_signal(xbt_parmap_t parmap)
+static void xbt_parmap_posix_master_signal(xbt_parmap_t parmap)
 {
-  parmap->thread_counter = 0;
+  xbt_os_mutex_acquire(parmap->ready_mutex);
+  parmap->thread_counter = 1;
   parmap->work++;
-  XBT_DEBUG("Starting work %d", parmap->work);
-
   /* wake all workers */
   xbt_os_cond_broadcast(parmap->ready_cond);
-
-  if (parmap->status == XBT_PARMAP_WORK) {
-    /* also work myself */
-    void* work = xbt_parmap_next(parmap);
-    while (work != NULL) {
-      parmap->fun(work);
-      work = xbt_parmap_next(parmap);
-    }
-  }
-
-  unsigned int counter = __sync_add_and_fetch(&parmap->thread_counter, 1);
-  if (counter < parmap->num_workers) {
-    /* some workers have not finished yet */
-    XBT_DEBUG("Some workers have not finished yet, waiting for them");
-    xbt_os_cond_wait(parmap->done_cond, parmap->done_mutex);
-  }
+  xbt_os_mutex_release(parmap->ready_mutex);
 }
 
 /**
@@ -346,23 +340,16 @@ static void xbt_parmap_posix_signal(xbt_parmap_t parmap)
  * when it has no more work to do.
  *
  * \param parmap a parmap
+ * \param round  the expected round number
  */
-static void xbt_parmap_posix_wait(xbt_parmap_t parmap)
+static void xbt_parmap_posix_worker_wait(xbt_parmap_t parmap, unsigned round)
 {
-  int work = parmap->work;
-  unsigned int counter = __sync_add_and_fetch(&parmap->thread_counter, 1);
-  if (counter == parmap->num_workers) {
-    /* all workers have finished, wake the controller */
-    parmap->done++;
-    XBT_DEBUG("Last worker has finished, waking the controller");
-    xbt_os_cond_signal(parmap->done_cond);
-  }
-
+  xbt_os_mutex_acquire(parmap->ready_mutex);
   /* wait for more work */
-  XBT_DEBUG("Worker %d waiting for more work", counter);
-  if (parmap->work == work) {
+  if (parmap->work < round) {
     xbt_os_cond_wait(parmap->ready_cond, parmap->ready_mutex);
   }
+  xbt_os_mutex_release(parmap->ready_mutex);
 }
 
 #ifdef HAVE_FUTEX_H
@@ -373,13 +360,13 @@ static void xbt_parmap_posix_wait(xbt_parmap_t parmap)
  *
  * \param parmap a parmap
  */
-static void xbt_parmap_futex_start(xbt_parmap_t parmap)
+static void xbt_parmap_futex_master_wait(xbt_parmap_t parmap)
 {
-  int myflag = parmap->done;
-  __sync_fetch_and_add(&parmap->thread_counter, 1);
-  if (parmap->thread_counter < parmap->num_workers) {
+  unsigned count = parmap->thread_counter;
+  while (count < parmap->num_workers) {
     /* wait for all workers to be ready */
-    futex_wait(&parmap->done, myflag);
+    futex_wait(&parmap->thread_counter, count);
+    count = parmap->thread_counter;
   }
 }
 
@@ -391,15 +378,12 @@ static void xbt_parmap_futex_start(xbt_parmap_t parmap)
  *
  * \param parmap a parmap
  */
-static void xbt_parmap_futex_end(xbt_parmap_t parmap)
+static void xbt_parmap_futex_worker_signal(xbt_parmap_t parmap)
 {
-  unsigned int mycount;
-
-  mycount = __sync_add_and_fetch(&parmap->thread_counter, 1);
-  if (mycount == parmap->num_workers) {
+  unsigned count = __sync_add_and_fetch(&parmap->thread_counter, 1);
+  if (count == parmap->num_workers) {
     /* all workers have finished, wake the controller */
-    parmap->done++;
-    futex_wake(&parmap->done, 1);
+    futex_wake(&parmap->thread_counter, 1);
   }
 }
 
@@ -410,29 +394,12 @@ static void xbt_parmap_futex_end(xbt_parmap_t parmap)
  *
  * \param parmap a parmap
  */
-static void xbt_parmap_futex_signal(xbt_parmap_t parmap)
+static void xbt_parmap_futex_master_signal(xbt_parmap_t parmap)
 {
-  int myflag = parmap->done;
-  parmap->thread_counter = 0;
-  parmap->work++;
-
+  parmap->thread_counter = 1;
+  __sync_add_and_fetch(&parmap->work, 1);
   /* wake all workers */
-  futex_wake(&parmap->work, parmap->num_workers);
-
-  if (parmap->status == XBT_PARMAP_WORK) {
-    /* also work myself */
-    void* work = xbt_parmap_next(parmap);
-    while (work != NULL) {
-      parmap->fun(work);
-      work = xbt_parmap_next(parmap);
-    }
-  }
-
-  unsigned int mycount = __sync_add_and_fetch(&parmap->thread_counter, 1);
-  if (mycount < parmap->num_workers) {
-    /* some workers have not finished yet */
-    futex_wait(&parmap->done, myflag);
-  }
+  futex_wake(&parmap->work, parmap->num_workers - 1);
 }
 
 /**
@@ -442,22 +409,14 @@ static void xbt_parmap_futex_signal(xbt_parmap_t parmap)
  * when it has no more work to do.
  *
  * \param parmap a parmap
+ * \param round  the expected round number
  */
-static void xbt_parmap_futex_wait(xbt_parmap_t parmap)
+static void xbt_parmap_futex_worker_wait(xbt_parmap_t parmap, unsigned round)
 {
-  int myflag;
-  unsigned int mycount;
-
-  myflag = parmap->work;
-  mycount = __sync_add_and_fetch(&parmap->thread_counter, 1);
-  if (mycount == parmap->num_workers) {
-    /* all workers have finished, wake the controller */
-    parmap->done++;
-    futex_wake(&parmap->done, 1);
-  }
-
+  unsigned work = parmap->work;
   /* wait for more work */
-  futex_wait(&parmap->work, myflag);
+  if (work < round)
+    futex_wait(&parmap->work, work);
 }
 #endif
 
@@ -468,9 +427,8 @@ static void xbt_parmap_futex_wait(xbt_parmap_t parmap)
  *
  * \param parmap a parmap
  */
-static void xbt_parmap_busy_start(xbt_parmap_t parmap)
+static void xbt_parmap_busy_master_wait(xbt_parmap_t parmap)
 {
-  __sync_fetch_and_add(&parmap->thread_counter, 1);
   while (parmap->thread_counter < parmap->num_workers) {
     xbt_os_thread_yield();
   }
@@ -483,7 +441,7 @@ static void xbt_parmap_busy_start(xbt_parmap_t parmap)
  *
  * \param parmap a parmap
  */
-static void xbt_parmap_busy_end(xbt_parmap_t parmap)
+static void xbt_parmap_busy_worker_signal(xbt_parmap_t parmap)
 {
   __sync_add_and_fetch(&parmap->thread_counter, 1);
 }
@@ -495,25 +453,10 @@ static void xbt_parmap_busy_end(xbt_parmap_t parmap)
  *
  * \param parmap a parmap
  */
-static void xbt_parmap_busy_signal(xbt_parmap_t parmap)
+static void xbt_parmap_busy_master_signal(xbt_parmap_t parmap)
 {
-  parmap->thread_counter = 0;
-  parmap->work++;
-
-  if (parmap->status == XBT_PARMAP_WORK) {
-    /* also work myself */
-    void* work = xbt_parmap_next(parmap);
-    while (work != NULL) {
-      parmap->fun(work);
-      work = xbt_parmap_next(parmap);
-    }
-  }
-
-  /* I have finished, wait for the others */
-  __sync_add_and_fetch(&parmap->thread_counter, 1);
-  while (parmap->thread_counter < parmap->num_workers) {
-    xbt_os_thread_yield();
-  }
+  parmap->thread_counter = 1;
+  __sync_add_and_fetch(&parmap->work, 1);
 }
 
 /**
@@ -523,14 +466,12 @@ static void xbt_parmap_busy_signal(xbt_parmap_t parmap)
  * when it has no more work to do.
  *
  * \param parmap a parmap
+ * \param round  the expected round number
  */
-static void xbt_parmap_busy_wait(xbt_parmap_t parmap)
+static void xbt_parmap_busy_worker_wait(xbt_parmap_t parmap, unsigned round)
 {
-  int work = parmap->work;
-  __sync_add_and_fetch(&parmap->thread_counter, 1);
-
   /* wait for more work */
-  while (parmap->work == work) {
+  while (parmap->work < round) {
     xbt_os_thread_yield();
   }
 }