X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/f85b43df272b75c6cf714bde3c3c6f04aaca1b22..f431ee913249e0d6d028d2bfc1db185efa003bc4:/src/msg/msg_vm.c

diff --git a/src/msg/msg_vm.c b/src/msg/msg_vm.c
index 9cce6b7523..7bd7b98d68 100644
--- a/src/msg/msg_vm.c
+++ b/src/msg/msg_vm.c
@@ -61,9 +61,9 @@ xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
  * \param value what to change the property to
  * \param free_ctn the freeing function to use to kill the value on need
  */
-void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value,void_f_pvoid_t free_ctn) {
-
-  xbt_dict_set(MSG_host_get_properties(vm), name, value,free_ctn);
+void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
+{
+  xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
 }
 
 /** \ingroup msg_vm_management
@@ -77,7 +77,8 @@ void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value,void_f
  * for each VM/PM.
  */
 
-msg_vm_t MSG_vm_get_by_name(const char *name){
+msg_vm_t MSG_vm_get_by_name(const char *name)
+{
 	return MSG_get_host_by_name(name);
 }
 
@@ -88,13 +89,15 @@ msg_vm_t MSG_vm_get_by_name(const char *name){
  * This functions checks whether \a host is a valid pointer or not and return
    its name.
  */
-const char *MSG_vm_get_name(msg_vm_t vm) {
+const char *MSG_vm_get_name(msg_vm_t vm)
+{
   return MSG_host_get_name(vm);
 }
 
 
 /* **** Check state of a VM **** */
-static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state) {
+static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
+{
   return simcall_vm_get_state(vm) == state;
 }
 
@@ -178,7 +181,7 @@ msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name,
     simcall_host_set_params(vm, &params);
   }
 
-  /* TODO: We will revisit the disk support later. */
+  /* TODO: Limit net capability, take into account disk considerations. */
 
   return vm;
 }
@@ -204,7 +207,7 @@ msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
   msg_vm_t ind_vm = NULL;
   void *ind_vm_workstation =  NULL;
 
-  // Ask simix to create the surf vm resource
+  /* Ask the SIMIX layer to create the surf vm resource */
   ind_vm_workstation = simcall_vm_create(name, ind_pm);
   ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
 
@@ -217,8 +220,32 @@ 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)
+{
+  /* First, terminate all processes on the VM if necessary */
+  if (MSG_vm_is_running(vm))
+      simcall_vm_shutdown(vm);
 
-/** @brief Start a vm (ie. boot)
+  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
  *
  *  If the VM cannot be started, an exception is generated.
@@ -226,7 +253,6 @@ msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
  */
 void MSG_vm_start(msg_vm_t vm)
 {
-  //Please note that vm start can raise an exception if the VM cannot be started.
   simcall_vm_start(vm);
 
   #ifdef HAVE_TRACING
@@ -253,21 +279,60 @@ 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
+ * 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)
+{
+  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)
+{
+  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)
+{
+  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)
+{
+  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)
+{
+  return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
+}
+
 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);
+
   int need_exit = 0;
 
-  xbt_assert(argc == 3);
-  const char *mbox = argv[1];
-  const char *mbox_ctl = argv[2];
+  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);
 
   for (;;) {
     msg_task_t task = NULL;
     MSG_task_recv(&task, mbox);
 
-    if (strcmp(task->name, "finalize") == 0)
+    if (strcmp(task->name, finalize_task_name) == 0)
       need_exit = 1;
 
     MSG_task_destroy(task);
@@ -277,85 +342,346 @@ static int migration_rx_fun(int argc, char *argv[])
   }
 
 
+  simcall_vm_migrate(vm, dst_pm);
+  simcall_vm_resume(vm);
+
   {
-    msg_task_t task = MSG_task_create("fin", 0, 0, NULL);
+    char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 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);
+
+    xbt_free(task_name);
   }
 
-  XBT_INFO("%s@%s done", pr_name, host_name);
+
+  xbt_free(mbox);
+  xbt_free(mbox_ctl);
+  xbt_free(finalize_task_name);
+
+  XBT_DEBUG("mig: rx_done");
 
   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);
+
+  char *key = NULL;
+  xbt_dict_cursor_t cursor = NULL;
+  dirty_page_t dp = NULL;
+  xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
+    double remaining = MSG_task_get_remaining_computation(dp->task);
+    dp->prev_clock = MSG_get_clock();
+    dp->prev_remaining = remaining;
+
+    // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
+  }
+}
+
+static void start_dirty_page_tracking(msg_vm_t vm)
+{
+  msg_host_priv_t priv = msg_host_resource_priv(vm);
+  priv->dp_enabled = 1;
+
+  reset_dirty_pages(vm);
+}
+
+static void stop_dirty_page_tracking(msg_vm_t vm)
+{
+  msg_host_priv_t priv = msg_host_resource_priv(vm);
+  priv->dp_enabled = 0;
+}
+
+#if 0
+/* It might be natural that we define dp_rate for each task. But, we will also
+ * have to care about how each task behavior affects the memory update behavior
+ * at the operating system level. It may not be easy to model it with a simple algorithm. */
+double calc_updated_pages(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;
+    double updated = dp->task->dp_rate * computed;
+
+    XBT_INFO("%s@%s: computated %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);
+    XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
+        key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
+
+    return updated;
+}
+#endif
+
+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)",
+      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)
+{
+  msg_host_priv_t priv = msg_host_resource_priv(vm);
+  double total = 0;
+
+  char *key = NULL;
+  xbt_dict_cursor_t cursor = NULL;
+  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);
+    total += get_computed(key, vm, dp, remaining, clock);
+
+    dp->prev_remaining = remaining;
+    dp->prev_clock = clock;
+  }
+
+  total += priv->dp_updated_by_deleted_tasks;
+
+  XBT_INFO("mig-stage2.%d: computed %f flop_counts (including %f by deleted tasks)",
+      stage2_round_for_fancy_debug,
+      total, priv->dp_updated_by_deleted_tasks);
+
+
+
+  priv->dp_updated_by_deleted_tasks = 0;
+
+
+  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);
+
+  dirty_page_t dp = xbt_new0(s_dirty_page, 1);
+  dp->task = task;
+
+  /* It should be okay that we add a task onto a migrating VM. */
+  if (priv->dp_enabled) {
+    dp->prev_clock = MSG_get_clock();
+    dp->prev_remaining = remaining;
+  }
+
+  xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
+  xbt_dict_set(priv->dp_objs, key, dp, NULL);
+  XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
+
+  xbt_free(key);
+}
+
+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);
+
+  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
+   * lookup_() function that will called soon. */
+  if (priv->dp_enabled) {
+    double remaining = MSG_task_get_remaining_computation(task);
+    double clock = MSG_get_clock();
+    // double updated = calc_updated_pages(key, host, dp, remaining, clock);
+    double updated = get_computed(key, host, dp, remaining, clock);
+
+    priv->dp_updated_by_deleted_tasks += updated;
+  }
+
+  xbt_dict_remove(priv->dp_objs, key);
+  xbt_free(dp);
+
+  XBT_DEBUG("del %s on %s", key, sg_host_name(host));
+
+  xbt_free(key);
+}
+
+
+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, double mig_speed)
+{
+  char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
+  msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
+
+  msg_error_t ret;
+  if (mig_speed > 0)
+    ret = MSG_task_send_bounded(task, mbox, mig_speed);
+  else
+    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);
+  else
+    XBT_INFO("mig-stage%d: sent %f", stage, size);
+
+  xbt_free(task_name);
+}
+
+
 static int migration_tx_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_assert(argc == 3);
-  const char *mbox = argv[1];
-  long ramsize = atol(argv[2]);
+  XBT_DEBUG("mig: tx_start");
 
-  char *task_name = bprintf("task-%s", mbox);
-  msg_task_t task = MSG_task_create(task_name, 0, ramsize, NULL);
-  msg_error_t ret = MSG_task_send(task, mbox);
-  xbt_assert(ret == MSG_OK);
+  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);
 
-  xbt_free(task_name);
 
-  {
-    msg_task_t task = MSG_task_create("finalize", 0, 0, NULL);
-    msg_error_t ret = MSG_task_send(task, mbox);
-    xbt_assert(ret == MSG_OK);
+  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;
+  const double dp_rate      = params.dp_rate;
+  const double dp_cap       = params.dp_cap;
+  const double mig_speed    = params.mig_speed;
+  double remaining_size = ramsize + devsize;
+  double threshold = max_downtime * 125 * 1000 * 1000;
+
+
+  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);
+
+  /* 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, mig_speed);
+
+  remaining_size -= ramsize;
+
+
+
+  /* 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;
   }
 
-  XBT_INFO("%s@%s done", pr_name, host_name);
+
+  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;
+    }
+
+    remaining_size += updated_size;
+
+    XBT_INFO("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, mig_speed);
+
+    remaining_size -= updated_size;
+    stage2_round += 1;
+  }
+
+
+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, mig_speed);
+
+  xbt_free(mbox);
+
+  XBT_DEBUG("mig: tx_done");
 
   return 0;
 }
 
-static void create_dummy_task(msg_vm_t vm, msg_host_t old_pm, msg_host_t new_pm, long ramsize)
-{
-  if (ramsize == 0)
-    XBT_WARN("migrate a VM, but ramsize is zero");
 
-  char *suffix = bprintf("mig-%s(%s-%s)", vm->key, old_pm->key, new_pm->key);
-  char *mbox = bprintf("MBOX:%s", suffix);
-  char *mbox_ctl = bprintf("MBOX:%s:CTL", suffix);
+
+static void 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));
 
   {
-    const char *pr_name = "mig_tx";
-    int nargvs = 4;
+    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(mbox);
-    argv[2] = bprintf("%ld", ramsize);
-    argv[3] = NULL;
+    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, old_pm, nargvs - 1, argv);
+    msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
+
+    xbt_free(pr_name);
   }
 
   {
-    const char *pr_name = "mig_rx";
-    int nargvs = 4;
+    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(mbox);
-    argv[2] = xbt_strdup(mbox_ctl);
-    argv[3] = NULL;
+    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);
 
-    msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, new_pm, nargvs - 1, argv);
+    xbt_free(pr_name);
   }
 
+  /* wait until the migration have finished */
   {
     msg_task_t task = NULL;
     msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
     xbt_assert(ret == MSG_OK);
+
+    char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
+    xbt_assert(strcmp(task->name, expected_task_name) == 0);
+    xbt_free(expected_task_name);
   }
 
-  xbt_free(suffix);
-  xbt_free(mbox);
   xbt_free(mbox_ctl);
 }
 
@@ -392,13 +718,12 @@ void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
 
   msg_host_t old_pm = simcall_vm_get_pm(vm);
 
-  s_ws_params_t params;
-  simcall_host_get_params(vm, &params);
-  long ramsize = params.ramsize;
+  if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
+    THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
+
+  do_migration(vm, old_pm, new_pm);
 
-  create_dummy_task(vm, old_pm, new_pm, ramsize);
 
-  simcall_vm_migrate(vm, new_pm);
 
   XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
 
@@ -478,32 +803,9 @@ void MSG_vm_restore(msg_vm_t vm)
 }
 
 
-/** @brief Destroy a VM. Destroy the VM object from the simulation.
- *  @ingroup msg_VMs
- */
-void MSG_vm_destroy(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);
-
-  __MSG_host_destroy(vm);
-
-  #ifdef HAVE_TRACING
-  TRACE_msg_vm_end(vm);
-  #endif
-}
 
 
-/** @brief Get the physical host of a givne VM.
+/** @brief Get the physical host of a given VM.
  *  @ingroup msg_VMs
  */
 msg_host_t MSG_vm_get_pm(msg_vm_t vm)