X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/f185f7337939bd2116d71563cedda3c33a39a54c..7018d53f4b991df9ddfa5b0c53293118d200efd6:/src/msg/msg_vm.c

diff --git a/src/msg/msg_vm.c b/src/msg/msg_vm.c
index 9ab63aef4e..9bc7ab0c84 100644
--- a/src/msg/msg_vm.c
+++ b/src/msg/msg_vm.c
@@ -1,25 +1,21 @@
-/* Copyright (c) 2012. The SimGrid Team. All rights reserved.               */
+/* Copyright (c) 2012-2014. 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. */
 
-// QUESTIONS:
-// 1./ check how and where a new VM is added to the list of the hosts
-// 2./ Diff between SIMIX_Actions and SURF_Actions
-// => SIMIX_actions : point synchro entre processus de niveau (theoretically speaking I do not have to create such SIMIX_ACTION
-// =>  Surf_Actions
+/* TODO:
+ * 1. add the support of trace
+ * 2. use parallel tasks to simulate CPU overhead and remove the very
+ *    experimental code generating micro computation tasks
+ */
+
 
-// TODO
-//	MSG_TRACE can be revisited in order to use  the host
-//	To implement a mixed model between workstation and vm_workstation,
-//     please give a look at surf_model_private_t model_private at SURF Level and to the share resource functions
-//     double (*share_resources) (double now);
-//	For the action into the vm workstation model, we should be able to leverage the usual one (and if needed, look at
-// 		the workstation model.
 
 #include "msg_private.h"
 #include "xbt/sysdep.h"
 #include "xbt/log.h"
+#include "simgrid/platf.h"
 
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
                                 "Cloud-oriented parts of the MSG API");
@@ -56,7 +52,7 @@ xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
 /** \ingroup m_host_management
  * \brief Change the value of a given host property
  *
- * \param host a host
+ * \param vm a vm
  * \param name a property name
  * \param value what to change the property to
  * \param free_ctn the freeing function to use to kill the value on need
@@ -79,7 +75,7 @@ void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_
 
 msg_vm_t MSG_vm_get_by_name(const char *name)
 {
-	return MSG_get_host_by_name(name);
+  return MSG_get_host_by_name(name);
 }
 
 /** \ingroup m_vm_management
@@ -101,7 +97,7 @@ static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
   return simcall_vm_get_state(vm) == state;
 }
 
-/** @brief Returns whether the given VM has just reated, not running.
+/** @brief Returns whether the given VM has just created, not running.
  *  @ingroup msg_VMs
  */
 int MSG_vm_is_created(msg_vm_t vm)
@@ -117,15 +113,14 @@ int MSG_vm_is_running(msg_vm_t vm)
   return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
 }
 
-#if 0
 /** @brief Returns whether the given VM is currently migrating
  *  @ingroup msg_VMs
  */
 int MSG_vm_is_migrating(msg_vm_t vm)
 {
-  return __MSG_vm_is_state(vm, SURF_VM_STATE_MIGRATING);
+  msg_host_priv_t priv = msg_host_resource_priv(vm);
+  return priv->is_migrating;
 }
-#endif
 
 /** @brief Returns whether the given VM is currently suspended, not running.
  *  @ingroup msg_VMs
@@ -168,22 +163,33 @@ int MSG_vm_is_restoring(msg_vm_t vm)
 
 /** @brief Create a new VM with specified parameters.
  *  @ingroup msg_VMs*
+ *  All parameters are in MBytes
  *
  */
 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name,
-	                                     int ncpus, int ramsize, int net_cap, char *disk_path, int disksize)
+                       int ncpus, int ramsize,
+                       int net_cap, char *disk_path, int disksize,
+                       int mig_netspeed, int dp_intensity)
 {
-  msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
-
-  {
-    s_ws_params_t params;
-    memset(&params, 0, sizeof(params));
-    params.ramsize = ramsize;
-    params.overcommit = 0;
-    simcall_host_set_params(vm, &params);
-  }
+  /* For the moment, intensity_rate is the percentage against the migration
+   * bandwidth */
+  double host_speed = MSG_get_host_speed(ind_pm);
+  double update_speed = ((double)dp_intensity/100) * mig_netspeed;
 
-  /* TODO: Limit net capability, take into account disk considerations. */
+  msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
+  s_ws_params_t params;
+  memset(&params, 0, sizeof(params));
+  params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
+  //params.overcommit = 0;
+  params.devsize = 0;
+  params.skip_stage2 = 0;
+  params.max_downtime = 0.03;
+  params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
+  params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
+  params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
+
+  //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate, params.mig_speed, dp_intensity, update_speed, host_speed);
+  simcall_host_set_params(vm, &params);
 
   return vm;
 }
@@ -198,8 +204,8 @@ msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
 {
   /* make sure the VM of the same name does not exit */
   {
-    void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
-    if (ind_host_tmp) {
+    xbt_dictelm_t ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
+    if (ind_host_tmp && xbt_lib_get_level(ind_host_tmp, SIMIX_HOST_LEVEL) != NULL) {
       XBT_ERROR("host %s already exits", name);
       return NULL;
     }
@@ -222,6 +228,33 @@ msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
   return ind_vm;
 }
 
+/** @brief Destroy a VM. Destroy the VM object from the simulation.
+ *  @ingroup msg_VMs
+ */
+void MSG_vm_destroy(msg_vm_t vm)
+{
+  if (MSG_vm_is_migrating(vm))
+    THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
+
+  /* First, terminate all processes on the VM if necessary */
+  if (MSG_vm_is_running(vm))
+      simcall_vm_shutdown(vm);
+
+  if (!MSG_vm_is_created(vm)) {
+    XBT_CRITICAL("shutdown the given VM before destroying it");
+    DIE_IMPOSSIBLE;
+  }
+
+  /* Then, destroy the VM object */
+  simcall_vm_destroy(vm);
+
+  __MSG_host_destroy(vm);
+
+  #ifdef HAVE_TRACING
+  TRACE_msg_vm_end(vm);
+  #endif
+}
+
 
 /** @brief Start a vm (i.e., boot the guest operating system)
  *  @ingroup msg_VMs
@@ -259,56 +292,94 @@ void MSG_vm_shutdown(msg_vm_t vm)
 
 
 /* We have two mailboxes. mbox is used to transfer migration data between
- * source and destiantion PMs. mbox_ctl is used to detect the completion of a
+ * source and destination PMs. mbox_ctl is used to detect the completion of a
  * migration. The names of these mailboxes must not conflict with others. */
-static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
+static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
 {
+  char *vm_name = sg_host_name(vm);
+  char *src_pm_name = sg_host_name(src_pm);
+  char *dst_pm_name = sg_host_name(dst_pm);
+
   return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
 }
 
-static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
+static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
 {
+  char *vm_name = sg_host_name(vm);
+  char *src_pm_name = sg_host_name(src_pm);
+  char *dst_pm_name = sg_host_name(dst_pm);
+
   return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
 }
 
-static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
+static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
 {
+  char *vm_name = sg_host_name(vm);
+  char *src_pm_name = sg_host_name(src_pm);
+  char *dst_pm_name = sg_host_name(dst_pm);
+
   return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
 }
 
-static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
+static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
 {
+  char *vm_name = sg_host_name(vm);
+  char *src_pm_name = sg_host_name(src_pm);
+  char *dst_pm_name = sg_host_name(dst_pm);
+
   return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
 }
 
-static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
+static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
 {
+  char *vm_name = sg_host_name(vm);
+  char *src_pm_name = sg_host_name(src_pm);
+  char *dst_pm_name = sg_host_name(dst_pm);
+
   return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
 }
 
+
+struct migration_session {
+  msg_vm_t vm;
+  msg_host_t src_pm;
+  msg_host_t dst_pm;
+
+  /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
+   * know the completion of the migration. */
+  char *mbox_ctl;
+  /* The migration_rx and migration_tx processes use mbox to transfer migration
+   * data. */
+  char *mbox;
+};
+
+
 static int migration_rx_fun(int argc, char *argv[])
 {
-  const char *pr_name = MSG_process_get_name(MSG_process_self());
-  const char *host_name = MSG_host_get_name(MSG_host_self());
-
   XBT_DEBUG("mig: rx_start");
 
-  xbt_assert(argc == 4);
-  const char *vm_name = argv[1];
-  const char *src_pm_name  = argv[2];
-  const char *dst_pm_name  = argv[3];
-  msg_vm_t vm = MSG_get_host_by_name(vm_name);
-  msg_vm_t dst_pm = MSG_get_host_by_name(dst_pm_name);
+  // The structure has been created in the do_migration function and should only be freed in the same place ;)
+  struct migration_session *ms = MSG_process_get_data(MSG_process_self());
+
+  s_ws_params_t params;
+  simcall_host_get_params(ms->vm, &params);
 
   int need_exit = 0;
 
-  char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
-  char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
-  char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
+  char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
 
+  int ret = 0;
   for (;;) {
     msg_task_t task = NULL;
-    MSG_task_recv(&task, mbox);
+    ret = MSG_task_recv(&task, ms->mbox);
+    {
+      if (ret != MSG_OK) {
+        // An error occured, clean the code and return
+        // The owner did not change, hence the task should be only destroyed on the other side
+        xbt_free(finalize_task_name);
+        return 0;
+      }
+    }
 
     if (strcmp(task->name, finalize_task_name) == 0)
       need_exit = 1;
@@ -319,23 +390,65 @@ static int migration_rx_fun(int argc, char *argv[])
       break;
   }
 
+  // Here Stage 1, 2  and 3 have been performed.
+  // Hence complete the migration
+
+  // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
+  // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms structure at the begining and free it at the end of each function)
+   msg_vm_t vm = ms->vm;
+   msg_host_t src_pm = ms->src_pm;
+   msg_host_t dst_pm = ms-> dst_pm;
+   msg_host_priv_t priv = msg_host_resource_priv(vm);
+
+// TODO: we have an issue, if the DST node is turning off during the three next calls, then the VM is in an inconsistent state
+// I should check with Takahiro in order to make this portion of code atomic
+  /* deinstall the current affinity setting for the CPU */
+  simcall_vm_set_affinity(vm, src_pm, 0);
 
+  /* Update the vm location */
   simcall_vm_migrate(vm, dst_pm);
+ 
+  /* Resume the VM */
   simcall_vm_resume(vm);
 
+  /* install the affinity setting of the VM on the destination pm */
   {
-    char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
 
+    unsigned long affinity_mask = (unsigned long) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *)dst_pm, sizeof(msg_host_t));
+    simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
+    XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
+  }
+
+  {
+
+   // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
+   msg_host_priv_t priv = msg_host_resource_priv(vm);
+   priv->is_migrating = 0;
+   XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->key, ms->src_pm->key, ms->dst_pm->key);
+   #ifdef HAVE_TRACING
+    TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
+   #endif
+
+  }
+  // Inform the SRC that the migration has been correctly performed
+  {
+    char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
     msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
-    msg_error_t ret = MSG_task_send(task, mbox_ctl);
-    xbt_assert(ret == MSG_OK);
+    msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
+    // xbt_assert(ret == MSG_OK);
+    if(ret == MSG_HOST_FAILURE){
+      // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM has been correctly migrated on the DST node
+      // TODO What does it mean ? What should we do ?
+      MSG_task_destroy(task);
+    } else if(ret == MSG_TRANSFER_FAILURE){
+      // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
+      MSG_task_destroy(task);
+    }
 
     xbt_free(task_name);
   }
 
 
-  xbt_free(mbox);
-  xbt_free(mbox_ctl);
   xbt_free(finalize_task_name);
 
   XBT_DEBUG("mig: rx_done");
@@ -343,14 +456,6 @@ static int migration_rx_fun(int argc, char *argv[])
   return 0;
 }
 
-
-typedef struct dirty_page {
-  double prev_clock;
-  double prev_remaining;
-  msg_task_t task;
-} s_dirty_page, *dirty_page_t;
-
-
 static void reset_dirty_pages(msg_vm_t vm)
 {
   msg_host_priv_t priv = msg_host_resource_priv(vm);
@@ -400,18 +505,18 @@ double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remain
 }
 #endif
 
-double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
+static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
 {
   double computed = dp->prev_remaining - remaining;
   double duration = clock - dp->prev_clock;
 
-  XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
+  XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
       key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
 
   return computed;
 }
 
-static double lookup_computed_flop_counts(msg_vm_t vm, int stage2_round_for_fancy_debug)
+static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
 {
   msg_host_priv_t priv = msg_host_resource_priv(vm);
   double total = 0;
@@ -421,6 +526,7 @@ static double lookup_computed_flop_counts(msg_vm_t vm, int stage2_round_for_fanc
   dirty_page_t dp = NULL;
   xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
     double remaining = MSG_task_get_remaining_computation(dp->task);
+
     double clock = MSG_get_clock();
 
     // total += calc_updated_pages(key, vm, dp, remaining, clock);
@@ -432,7 +538,8 @@ static double lookup_computed_flop_counts(msg_vm_t vm, int stage2_round_for_fanc
 
   total += priv->dp_updated_by_deleted_tasks;
 
-  XBT_INFO("mig-stage2.%d: computed %f flop_counts (including %f by deleted tasks)",
+  XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
+      stage_for_fancy_debug,
       stage2_round_for_fancy_debug,
       total, priv->dp_updated_by_deleted_tasks);
 
@@ -444,11 +551,13 @@ static double lookup_computed_flop_counts(msg_vm_t vm, int stage2_round_for_fanc
   return total;
 }
 
+// TODO Is this code redundant with the information provided by
+// msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
 void MSG_host_add_task(msg_host_t host, msg_task_t task)
 {
   msg_host_priv_t priv = msg_host_resource_priv(host);
   double remaining = MSG_task_get_remaining_computation(task);
-  char *key = bprintf("%s-%lld", task->name, task->counter);
+  char *key = bprintf("%s-%p", task->name, task);
 
   dirty_page_t dp = xbt_new0(s_dirty_page, 1);
   dp->task = task;
@@ -470,13 +579,13 @@ void MSG_host_del_task(msg_host_t host, msg_task_t task)
 {
   msg_host_priv_t priv = msg_host_resource_priv(host);
 
-  char *key = bprintf("%s-%lld", task->name, task->counter);
+  char *key = bprintf("%s-%p", task->name, task);
 
   dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
   xbt_assert(dp->task == task);
 
   /* If we are in the middle of dirty page tracking, we record how much
-   * computaion has been done until now, and keep the information for the
+   * computation has been done until now, and keep the information for the
    * lookup_() function that will called soon. */
   if (priv->dp_enabled) {
     double remaining = MSG_task_get_remaining_computation(task);
@@ -496,166 +605,344 @@ void MSG_host_del_task(msg_host_t host, msg_task_t task)
 }
 
 
-static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
-    double size, char *mbox, int stage, int stage2_round)
+
+
+static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm,
+    sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double timeout)
 {
-  char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
+  sg_size_t sent = 0;
+  char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
   msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
-  msg_error_t ret = MSG_task_send(task, mbox);
-  xbt_assert(ret == MSG_OK);
 
-  if (stage == 2)
-    XBT_INFO("mig-stage%d.%d: sent %f", stage, stage2_round, size);
+  /* TODO: clean up */
+
+  double clock_sta = MSG_get_clock();
+
+  msg_error_t ret;
+  if (mig_speed > 0)
+    ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
   else
-    XBT_INFO("mig-stage%d: sent %f", stage, size);
+    ret = MSG_task_send(task, mbox);
 
   xbt_free(task_name);
+
+  if (ret == MSG_OK) {
+    sent = size;
+  } else if (ret == MSG_TIMEOUT) {
+    sg_size_t remaining = MSG_task_get_remaining_communication(task);
+    sent = size - remaining;
+    XBT_INFO("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu",
+        timeout, remaining, size);
+  }
+
+  /* FIXME: why try-and-catch is used here? */
+  if(ret == MSG_HOST_FAILURE){
+    //XBT_INFO("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
+    MSG_task_destroy(task);
+    THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
+  }else if(ret == MSG_TRANSFER_FAILURE){
+    //XBT_INFO("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
+    MSG_task_destroy(task);
+    THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
+  }
+
+  double clock_end = MSG_get_clock();
+  double duration = clock_end - clock_sta;
+  double actual_speed = size / duration;
+
+  if (stage == 2)
+    XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration, actual_speed, mig_speed);
+  else
+    XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed, mig_speed);
+
+  return sent;
 }
 
+static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
+{
+  double updated_size = computed * dp_rate;
+  XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
+  if (updated_size > dp_cap) {
+    // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
+    updated_size = dp_cap;
+  }
 
-static int migration_tx_fun(int argc, char *argv[])
+  return (sg_size_t) updated_size;
+}
+
+static double send_stage1(struct migration_session *ms,
+    sg_size_t ramsize, double mig_speed, double dp_rate, double dp_cap)
 {
-  const char *pr_name = MSG_process_get_name(MSG_process_self());
-  const char *host_name = MSG_host_get_name(MSG_host_self());
 
-  XBT_DEBUG("mig: tx_start");
+  // const long chunksize = (sg_size_t)1024 * 1024 * 100;
+  const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
+  sg_size_t remaining = ramsize;
+  double computed_total = 0;
+
+  while (remaining > 0) {
+    sg_size_t datasize = chunksize;
+    if (remaining < chunksize)
+      datasize = remaining;
+
+    remaining -= datasize;
+    send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
+    double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
+    computed_total += computed;
+  }
+
+  return computed_total;
+}
 
-  xbt_assert(argc == 4);
-  const char *vm_name = argv[1];
-  const char *src_pm_name  = argv[2];
-  const char *dst_pm_name  = argv[3];
-  msg_vm_t vm = MSG_get_host_by_name(vm_name);
 
 
+static double get_threshold_value(double bandwidth, double max_downtime)
+{
+  return max_downtime * bandwidth;
+}
+
+static int migration_tx_fun(int argc, char *argv[])
+{
+  XBT_DEBUG("mig: tx_start");
+
+  // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
+  struct migration_session *ms = MSG_process_get_data(MSG_process_self());
+
   s_ws_params_t params;
-  simcall_host_get_params(vm, &params);
-  const long ramsize        = params.ramsize;
-  const long devsize        = params.devsize;
-  const int skip_stage2     = params.skip_stage2;
-  const double max_downtime = params.max_downtime;
+  simcall_host_get_params(ms->vm, &params);
+  const sg_size_t ramsize   = params.ramsize;
+  const sg_size_t devsize   = params.devsize;
+  const int skip_stage1     = params.skip_stage1;
+  int skip_stage2           = params.skip_stage2;
   const double dp_rate      = params.dp_rate;
   const double dp_cap       = params.dp_cap;
-  double remaining_size = ramsize + devsize;
-  double threshold = max_downtime * 125 * 1000 * 1000;
+  const double mig_speed    = params.mig_speed;
+  double max_downtime       = params.max_downtime;
 
+  /* hard code it temporally. Fix Me */
+#define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
+  double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
 
+  double remaining_size = ramsize + devsize;
+  double threshold = 0.0;
+
+  /* check parameters */
   if (ramsize == 0)
     XBT_WARN("migrate a VM, but ramsize is zero");
 
-  char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
-
-  XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
+  if (max_downtime == 0) {
+    XBT_WARN("use the default max_downtime value 30ms");
+    max_downtime = 0.03;
+  }
 
   /* Stage1: send all memory pages to the destination. */
-  start_dirty_page_tracking(vm);
-
-  send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0);
+  XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
+  start_dirty_page_tracking(ms->vm);
+
+  double computed_during_stage1 = 0;
+  if (!skip_stage1) {
+    double clock_prev_send = MSG_get_clock();
+
+    TRY {
+      /* At stage 1, we do not need timeout. We have to send all the memory
+       * pages even though the duration of this tranfer exceeds the timeout
+       * value. */
+      XBT_INFO("Stage 1: Gonna send %llu", ramsize);
+      sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
+      remaining_size -= sent;
+      computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
+
+      if (sent < ramsize) {
+        XBT_INFO("mig-stage1: timeout, force moving to stage 3");
+        skip_stage2 = 1;
+      } else if (sent > ramsize)
+        XBT_CRITICAL("bug");
+
+    } CATCH_ANONYMOUS {
+      //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
+      // Stop the dirty page tracking an return (there is no memory space to release)
+      stop_dirty_page_tracking(ms->vm);
+      return 0;
+    }
 
-  remaining_size -= ramsize;
+    double clock_post_send = MSG_get_clock();
+    mig_timeout -= (clock_post_send - clock_prev_send);
+    if (mig_timeout < 0) {
+      XBT_INFO("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
+          (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
+      skip_stage2 = 1;
+    }
 
+    /* estimate bandwidth */
+    double bandwidth = ramsize / (clock_post_send - clock_prev_send);
+    threshold = get_threshold_value(bandwidth, max_downtime);
+    XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
+  }
 
 
   /* Stage2: send update pages iteratively until the size of remaining states
    * becomes smaller than the threshold value. */
   if (skip_stage2)
     goto stage3;
-  if (max_downtime == 0) {
-    XBT_WARN("no max_downtime parameter, skip stage2");
-    goto stage3;
-  }
 
 
   int stage2_round = 0;
   for (;;) {
-    // long updated_size = lookup_dirty_pages(vm);
-    double updated_size = lookup_computed_flop_counts(vm, stage2_round) * dp_rate;
-    if (updated_size > dp_cap) {
-      XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f",
-          stage2_round, updated_size, dp_cap);
-      updated_size = dp_cap;
+
+    sg_size_t updated_size = 0;
+    if (stage2_round == 0) {
+      /* just after stage1, nothing has been updated. But, we have to send the
+       * data updated during stage1 */
+      updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
+    } else {
+      double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
+      updated_size = get_updated_size(computed, dp_rate, dp_cap);
     }
 
-    remaining_size += updated_size;
+    XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
+        stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
 
-    XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
-        remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
 
+    /* Check whether the remaining size is below the threshold value. If so,
+     * move to stage 3. */
+    remaining_size += updated_size;
+    XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
+        remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
     if (remaining_size < threshold)
       break;
 
-    send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round);
 
-    remaining_size -= updated_size;
-    stage2_round += 1;
+    sg_size_t sent = 0;
+    double clock_prev_send = MSG_get_clock();
+    TRY {
+      XBT_INFO("Stage 2, gonna send %llu", updated_size);
+      sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed, mig_timeout);
+    } CATCH_ANONYMOUS {
+      //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
+      // Stop the dirty page tracking an return (there is no memory space to release)
+      stop_dirty_page_tracking(ms->vm);
+      return 0;
+    }
+    double clock_post_send = MSG_get_clock();
+
+    if (sent == updated_size) {
+      /* timeout did not happen */
+      double bandwidth = updated_size / (clock_post_send - clock_prev_send);
+      threshold = get_threshold_value(bandwidth, max_downtime);
+      XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
+      remaining_size -= sent;
+      stage2_round += 1;
+      mig_timeout -= (clock_post_send - clock_prev_send);
+      xbt_assert(mig_timeout > 0);
+
+    } else if (sent < updated_size) {
+      /* When timeout happens, we move to stage 3. The size of memory pages
+       * updated before timeout must be added to the remaining size. */
+      XBT_INFO("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
+          stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
+      remaining_size -= sent;
+
+      double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
+      updated_size = get_updated_size(computed, dp_rate, dp_cap);
+      remaining_size += updated_size;
+      break;
+
+    } else
+      XBT_CRITICAL("bug");
   }
 
 
 stage3:
   /* Stage3: stop the VM and copy the rest of states. */
-  XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
-  simcall_vm_suspend(vm);
-  stop_dirty_page_tracking(vm);
-
-  send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0);
-
-  xbt_free(mbox);
+  XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
+  simcall_vm_suspend(ms->vm);
+  stop_dirty_page_tracking(ms->vm);
+
+  TRY {
+    XBT_INFO("Stage 3: Gonna send %f", remaining_size);
+    send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, remaining_size, ms->mbox, 3, 0, mig_speed, -1);
+  } CATCH_ANONYMOUS {
+    //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
+    // Stop the dirty page tracking an return (there is no memory space to release)
+    simcall_vm_resume(ms->vm);
+    return 0;
+  }
 
-  XBT_DEBUG("mig: tx_done");
+  // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate effectively the VM on the DST node.
+  XBT_INFO("mig: tx_done");
 
   return 0;
 }
 
 
 
-static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
+static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
 {
-  char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
-
-  {
-    char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
-    int nargvs = 5;
-    char **argv = xbt_new(char *, nargvs);
-    argv[0] = xbt_strdup(pr_name);
-    argv[1] = xbt_strdup(sg_host_name(vm));
-    argv[2] = xbt_strdup(sg_host_name(src_pm));
-    argv[3] = xbt_strdup(sg_host_name(dst_pm));
-    argv[4] = NULL;
-
-    msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
-
-    xbt_free(pr_name);
-  }
-
-  {
-    char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
-    int nargvs = 5;
-    char **argv = xbt_new(char *, nargvs);
-    argv[0] = xbt_strdup(pr_name);
-    argv[1] = xbt_strdup(sg_host_name(vm));
-    argv[2] = xbt_strdup(sg_host_name(src_pm));
-    argv[3] = xbt_strdup(sg_host_name(dst_pm));
-    argv[4] = NULL;
-
-    msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
-
-    xbt_free(pr_name);
-  }
+  struct migration_session *ms = xbt_new(struct migration_session, 1);
+  ms->vm = vm;
+  ms->src_pm = src_pm;
+  ms->dst_pm = dst_pm;
+  ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
+  ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
+ 
+
+  char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
+  char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
+
+//  msg_process_t tx_process, rx_process; 
+//  MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
+//  MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
+#if 1
+ {
+ char **argv = xbt_new(char *, 2);
+ argv[0] = pr_rx_name;
+ argv[1] = NULL;
+/*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
+ }
+ {
+ char **argv = xbt_new(char *, 2);
+ argv[0] = pr_tx_name;
+ argv[1] = NULL;
+/* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
+ }
+#endif
 
-  /* wait until the migration have finished */
+  /* wait until the migration have finished or on error has occured */
   {
+    XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
     msg_task_t task = NULL;
-    msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
-    xbt_assert(ret == MSG_OK);
+    msg_error_t ret = MSG_TIMEOUT;
+    while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
+     ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
+
+    xbt_free(ms->mbox_ctl);
+    xbt_free(ms->mbox);
+    xbt_free(ms);
+   
+    //xbt_assert(ret == MSG_OK);
+    if(ret == MSG_HOST_FAILURE){
+        // Note that since the communication failed, the owner did not change and the task should be destroyed on the other side.
+        // Hence, just throw the execption
+        XBT_INFO("SRC crashes, throw an exception (m-control)");
+        //MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is required or not 
+        return -1; 
+    } 
+    else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){ // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
+        XBT_INFO("DST crashes, throw an exception (m-control)");
+        return -2;  
+    }
 
-    char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
+   
+    char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
     xbt_assert(strcmp(task->name, expected_task_name) == 0);
     xbt_free(expected_task_name);
+    MSG_task_destroy(task);
+    return 0;
   }
-
-  xbt_free(mbox_ctl);
 }
 
 
+
+
 /** @brief Migrate the VM to the given host.
  *  @ingroup msg_VMs
  *
@@ -683,36 +970,61 @@ void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
    *   physical host.
    *
    * The second one would be easier.
-   *   
+   *  
    */
 
   msg_host_t old_pm = simcall_vm_get_pm(vm);
 
-  if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
+  if(!MSG_host_is_off(old_pm))
+    THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_name(old_pm));
+ 
+  if(!MSG_host_is_off(new_pm))
+    THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_name(new_pm));
+  
+  if (!MSG_vm_is_running(vm))
     THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
 
-  do_migration(vm, old_pm, new_pm);
-
+  if (MSG_vm_is_migrating(vm))
+    THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_name(vm));
 
+  msg_host_priv_t priv = msg_host_resource_priv(vm);
+  priv->is_migrating = 1;
 
-  XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
+  {
+  
+    int ret = do_migration(vm, old_pm, new_pm);
+    if (ret == -1){
+     priv->is_migrating = 0;
+     THROWF(host_error, 0, "SRC host failed during migration");
+    }
+    else if(ret == -2){
+     priv->is_migrating = 0;
+     THROWF(host_error, 0, "DST host failed during migration");
+    }
+  }
 
-  #ifdef HAVE_TRACING
-  TRACE_msg_vm_change_host(vm, old_pm, new_pm);
-  #endif
+  // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration process
+  // In that case, the VM has been already assigned to the DST node.
+  //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
+  //#ifdef HAVE_TRACING
+  //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
+  //#endif
 }
 
 
 /** @brief Immediately suspend the execution of all processes within the given VM.
  *  @ingroup msg_VMs
  *
- * This function stops the exection of the VM. All the processes on this VM
- * will pause. The state of the VM is perserved. We can later resume it again.
+ * This function stops the execution of the VM. All the processes on this VM
+ * will pause. The state of the VM is preserved. We can later resume it again.
  *
  * No suspension cost occurs.
  */
 void MSG_vm_suspend(msg_vm_t vm)
 {
+  if (MSG_vm_is_migrating(vm))
+    THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
+
   simcall_vm_suspend(vm);
 
   XBT_DEBUG("vm_suspend done");
@@ -741,8 +1053,8 @@ void MSG_vm_resume(msg_vm_t vm)
 /** @brief Immediately save the execution of all processes within the given VM.
  *  @ingroup msg_VMs
  *
- * This function stops the exection of the VM. All the processes on this VM
- * will pause. The state of the VM is perserved. We can later resume it again.
+ * This function stops the execution of the VM. All the processes on this VM
+ * will pause. The state of the VM is preserved. We can later resume it again.
  *
  * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
  * use a \ref MSG_file_write() before or after, depending on the exact semantic
@@ -750,6 +1062,9 @@ void MSG_vm_resume(msg_vm_t vm)
  */
 void MSG_vm_save(msg_vm_t vm)
 {
+  if (MSG_vm_is_migrating(vm))
+    THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
+
   simcall_vm_save(vm);
   #ifdef HAVE_TRACING
   TRACE_msg_vm_save(vm);
@@ -773,35 +1088,74 @@ void MSG_vm_restore(msg_vm_t vm)
 }
 
 
-/** @brief Destroy a VM. Destroy the VM object from the simulation.
+/** @brief Get the physical host of a given VM.
  *  @ingroup msg_VMs
  */
-void MSG_vm_destroy(msg_vm_t vm)
+msg_host_t MSG_vm_get_pm(msg_vm_t vm)
 {
-  /* First, terminate all processes on the VM if necessary */
-  if (MSG_vm_is_running(vm))
-      simcall_vm_shutdown(vm);
-
-  if (!MSG_vm_is_created(vm)) {
-    XBT_CRITICAL("shutdown the given VM before destroying it");
-    DIE_IMPOSSIBLE;
-  }
-
-  /* Then, destroy the VM object */
-  simcall_vm_destroy(vm);
+  return simcall_vm_get_pm(vm);
+}
 
-  __MSG_host_destroy(vm);
 
-  #ifdef HAVE_TRACING
-  TRACE_msg_vm_end(vm);
-  #endif
+/** @brief Set a CPU bound for a given VM.
+ *  @ingroup msg_VMs
+ *
+ * 1.
+ * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
+ *
+ * For example,
+ *  On PM0, there are Task1 and VM0.
+ *  On VM0, there is Task2.
+ * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
+ * Then,
+ *  Task1\@PM0 gets 50%.
+ *  Task2\@VM0 gets 25%.
+ * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
+ *
+ * This is because a VM has the dummy CPU action in the PM layer. Putting a
+ * task on the VM does not affect the bound of the dummy CPU action. The bound
+ * of the dummy CPU action is unlimited.
+ *
+ * There are some solutions for this problem. One option is to update the bound
+ * of the dummy CPU action automatically. It should be the sum of all tasks on
+ * the VM. But, this solution might be costly, because we have to scan all tasks
+ * on the VM in share_resource() or we have to trap both the start and end of
+ * task execution.
+ *
+ * The current solution is to use MSG_vm_set_bound(), which allows us to
+ * directly set the bound of the dummy CPU action.
+ *
+ *
+ * 2.
+ * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
+ * multiple CPU cores, the CPU share of a computation task (or a VM) never
+ * exceeds the capacity of a CPU core.
+ */
+void MSG_vm_set_bound(msg_vm_t vm, double bound)
+{
+  return simcall_vm_set_bound(vm, bound);
 }
 
 
-/** @brief Get the physical host of a givne VM.
+/** @brief Set the CPU affinity of a given VM.
  *  @ingroup msg_VMs
+ *
+ * This function changes the CPU affinity of a given VM. Usage is the same as
+ * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
  */
-msg_host_t MSG_vm_get_pm(msg_vm_t vm)
+void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
 {
-  return simcall_vm_get_pm(vm);
+  msg_host_priv_t priv = msg_host_resource_priv(vm);
+
+  if (mask == 0)
+    xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
+  else
+    xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *) mask, NULL);
+
+  msg_host_t pm_now = MSG_vm_get_pm(vm);
+  if (pm_now == pm) {
+    XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
+    simcall_vm_set_affinity(vm, pm, mask);
+  } else
+    XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
 }