1 /* Copyright (c) 2012. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
7 // 1./ check how and where a new VM is added to the list of the hosts
8 // 2./ Diff between SIMIX_Actions and SURF_Actions
9 // => SIMIX_actions : point synchro entre processus de niveau (theoretically speaking I do not have to create such SIMIX_ACTION
13 // MSG_TRACE can be revisited in order to use the host
14 // To implement a mixed model between workstation and vm_workstation,
15 // please give a look at surf_model_private_t model_private at SURF Level and to the share resource functions
16 // double (*share_resources) (double now);
17 // For the action into the vm workstation model, we should be able to leverage the usual one (and if needed, look at
18 // the workstation model.
20 #include "msg_private.h"
21 #include "xbt/sysdep.h"
24 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
25 "Cloud-oriented parts of the MSG API");
28 /* **** ******** GENERAL ********* **** */
30 /** \ingroup m_vm_management
31 * \brief Returns the value of a given vm property
34 * \param name a property name
35 * \return value of a property (or NULL if property not set)
38 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
40 return MSG_host_get_property_value(vm, name);
43 /** \ingroup m_vm_management
44 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
47 * \return a dict containing the properties
49 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
51 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
53 return (simcall_host_get_properties(vm));
56 /** \ingroup m_host_management
57 * \brief Change the value of a given host property
60 * \param name a property name
61 * \param value what to change the property to
62 * \param free_ctn the freeing function to use to kill the value on need
64 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
66 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
69 /** \ingroup msg_vm_management
70 * \brief Finds a msg_vm_t using its name.
72 * This is a name directory service
73 * \param name the name of a vm.
74 * \return the corresponding vm
76 * Please note that a VM is a specific host. Hence, you should give a different name
80 msg_vm_t MSG_vm_get_by_name(const char *name)
82 return MSG_get_host_by_name(name);
85 /** \ingroup m_vm_management
87 * \brief Return the name of the #msg_host_t.
89 * This functions checks whether \a host is a valid pointer or not and return
92 const char *MSG_vm_get_name(msg_vm_t vm)
94 return MSG_host_get_name(vm);
98 /* **** Check state of a VM **** */
99 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
101 return simcall_vm_get_state(vm) == state;
104 /** @brief Returns whether the given VM has just reated, not running.
107 int MSG_vm_is_created(msg_vm_t vm)
109 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
112 /** @brief Returns whether the given VM is currently running
115 int MSG_vm_is_running(msg_vm_t vm)
117 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
121 /** @brief Returns whether the given VM is currently migrating
124 int MSG_vm_is_migrating(msg_vm_t vm)
126 return __MSG_vm_is_state(vm, SURF_VM_STATE_MIGRATING);
130 /** @brief Returns whether the given VM is currently suspended, not running.
133 int MSG_vm_is_suspended(msg_vm_t vm)
135 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
138 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
141 int MSG_vm_is_saving(msg_vm_t vm)
143 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
146 /** @brief Returns whether the given VM has been saved, not running.
149 int MSG_vm_is_saved(msg_vm_t vm)
151 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
154 /** @brief Returns whether the given VM is being restored, not running.
157 int MSG_vm_is_restoring(msg_vm_t vm)
159 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
164 /* ------------------------------------------------------------------------- */
165 /* ------------------------------------------------------------------------- */
167 /* **** ******** MSG vm actions ********* **** */
169 /** @brief Create a new VM with specified parameters.
173 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name,
174 int ncpus, int ramsize, int net_cap, char *disk_path, int disksize)
176 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
179 s_ws_params_t params;
180 memset(¶ms, 0, sizeof(params));
181 params.ramsize = ramsize;
182 params.overcommit = 0;
183 simcall_host_set_params(vm, ¶ms);
186 /* TODO: We will revisit the disk support later. */
192 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
195 * A VM is treated as a host. The name of the VM must be unique among all hosts.
197 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
199 /* make sure the VM of the same name does not exit */
201 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
203 XBT_ERROR("host %s already exits", name);
208 /* Note: ind_vm and vm_workstation point to the same elm object. */
209 msg_vm_t ind_vm = NULL;
210 void *ind_vm_workstation = NULL;
212 /* Ask the SIMIX layer to create the surf vm resource */
213 ind_vm_workstation = simcall_vm_create(name, ind_pm);
214 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
216 XBT_DEBUG("A new VM (%s) has been created", name);
219 TRACE_msg_vm_create(name, ind_pm);
226 /** @brief Start a vm (i.e., boot the guest operating system)
229 * If the VM cannot be started, an exception is generated.
232 void MSG_vm_start(msg_vm_t vm)
234 simcall_vm_start(vm);
237 TRACE_msg_vm_start(vm);
243 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
246 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
247 * use a #MSG_process_sleep() or something. I'm not quite sure.
249 void MSG_vm_shutdown(msg_vm_t vm)
251 /* msg_vm_t equals to msg_host_t */
252 simcall_vm_shutdown(vm);
254 // #ifdef HAVE_TRACING
255 // TRACE_msg_vm_(vm);
261 /* We have two mailboxes. mbox is used to transfer migration data between
262 * source and destiantion PMs. mbox_ctl is used to detect the completion of a
263 * migration. The names of these mailboxes must not conflict with others. */
264 static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
266 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
269 static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
271 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
274 static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
276 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
279 static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
281 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
284 static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
286 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
289 static int migration_rx_fun(int argc, char *argv[])
291 const char *pr_name = MSG_process_get_name(MSG_process_self());
292 const char *host_name = MSG_host_get_name(MSG_host_self());
294 XBT_DEBUG("mig: rx_start");
296 xbt_assert(argc == 4);
297 const char *vm_name = argv[1];
298 const char *src_pm_name = argv[2];
299 const char *dst_pm_name = argv[3];
300 msg_vm_t vm = MSG_get_host_by_name(vm_name);
301 msg_vm_t dst_pm = MSG_get_host_by_name(dst_pm_name);
305 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
306 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
307 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
310 msg_task_t task = NULL;
311 MSG_task_recv(&task, mbox);
313 if (strcmp(task->name, finalize_task_name) == 0)
316 MSG_task_destroy(task);
323 simcall_vm_migrate(vm, dst_pm);
324 simcall_vm_resume(vm);
327 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
329 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
330 msg_error_t ret = MSG_task_send(task, mbox_ctl);
331 xbt_assert(ret == MSG_OK);
339 xbt_free(finalize_task_name);
341 XBT_DEBUG("mig: rx_done");
347 typedef struct dirty_page {
349 double prev_remaining;
351 } s_dirty_page, *dirty_page_t;
354 static void reset_dirty_pages(msg_vm_t vm)
356 msg_host_priv_t priv = msg_host_resource_priv(vm);
359 xbt_dict_cursor_t cursor = NULL;
360 dirty_page_t dp = NULL;
361 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
362 double remaining = MSG_task_get_remaining_computation(dp->task);
363 dp->prev_clock = MSG_get_clock();
364 dp->prev_remaining = remaining;
366 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
370 static void start_dirty_page_tracking(msg_vm_t vm)
372 msg_host_priv_t priv = msg_host_resource_priv(vm);
373 priv->dp_enabled = 1;
375 reset_dirty_pages(vm);
378 static void stop_dirty_page_tracking(msg_vm_t vm)
380 msg_host_priv_t priv = msg_host_resource_priv(vm);
381 priv->dp_enabled = 0;
385 /* It might be natural that we define dp_rate for each task. But, we will also
386 * have to care about how each task behavior affects the memory update behavior
387 * at the operating system level. It may not be easy to model it with a simple algorithm. */
388 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
390 double computed = dp->prev_remaining - remaining;
391 double duration = clock - dp->prev_clock;
392 double updated = dp->task->dp_rate * computed;
394 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
395 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
396 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
397 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
403 double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
405 double computed = dp->prev_remaining - remaining;
406 double duration = clock - dp->prev_clock;
408 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
409 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
414 static double lookup_computed_flop_counts(msg_vm_t vm, int stage2_round_for_fancy_debug)
416 msg_host_priv_t priv = msg_host_resource_priv(vm);
420 xbt_dict_cursor_t cursor = NULL;
421 dirty_page_t dp = NULL;
422 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
423 double remaining = MSG_task_get_remaining_computation(dp->task);
424 double clock = MSG_get_clock();
426 // total += calc_updated_pages(key, vm, dp, remaining, clock);
427 total += get_computed(key, vm, dp, remaining, clock);
429 dp->prev_remaining = remaining;
430 dp->prev_clock = clock;
433 total += priv->dp_updated_by_deleted_tasks;
435 XBT_INFO("mig-stage2.%d: computed %f flop_counts (including %f by deleted tasks)",
436 stage2_round_for_fancy_debug,
437 total, priv->dp_updated_by_deleted_tasks);
441 priv->dp_updated_by_deleted_tasks = 0;
447 void MSG_host_add_task(msg_host_t host, msg_task_t task)
449 msg_host_priv_t priv = msg_host_resource_priv(host);
450 double remaining = MSG_task_get_remaining_computation(task);
451 char *key = bprintf("%s-%lld", task->name, task->counter);
453 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
456 /* It should be okay that we add a task onto a migrating VM. */
457 if (priv->dp_enabled) {
458 dp->prev_clock = MSG_get_clock();
459 dp->prev_remaining = remaining;
462 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
463 xbt_dict_set(priv->dp_objs, key, dp, NULL);
464 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
469 void MSG_host_del_task(msg_host_t host, msg_task_t task)
471 msg_host_priv_t priv = msg_host_resource_priv(host);
473 char *key = bprintf("%s-%lld", task->name, task->counter);
475 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
476 xbt_assert(dp->task == task);
478 /* If we are in the middle of dirty page tracking, we record how much
479 * computaion has been done until now, and keep the information for the
480 * lookup_() function that will called soon. */
481 if (priv->dp_enabled) {
482 double remaining = MSG_task_get_remaining_computation(task);
483 double clock = MSG_get_clock();
484 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
485 double updated = get_computed(key, host, dp, remaining, clock);
487 priv->dp_updated_by_deleted_tasks += updated;
490 xbt_dict_remove(priv->dp_objs, key);
493 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
499 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
500 double size, char *mbox, int stage, int stage2_round)
502 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
503 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
504 msg_error_t ret = MSG_task_send(task, mbox);
505 xbt_assert(ret == MSG_OK);
508 XBT_INFO("mig-stage%d.%d: sent %f", stage, stage2_round, size);
510 XBT_INFO("mig-stage%d: sent %f", stage, size);
516 static int migration_tx_fun(int argc, char *argv[])
518 const char *pr_name = MSG_process_get_name(MSG_process_self());
519 const char *host_name = MSG_host_get_name(MSG_host_self());
521 XBT_DEBUG("mig: tx_start");
523 xbt_assert(argc == 4);
524 const char *vm_name = argv[1];
525 const char *src_pm_name = argv[2];
526 const char *dst_pm_name = argv[3];
527 msg_vm_t vm = MSG_get_host_by_name(vm_name);
530 s_ws_params_t params;
531 simcall_host_get_params(vm, ¶ms);
532 const long ramsize = params.ramsize;
533 const long devsize = params.devsize;
534 const int skip_stage2 = params.skip_stage2;
535 const double max_downtime = params.max_downtime;
536 const double dp_rate = params.dp_rate;
537 const double dp_cap = params.dp_cap;
538 double remaining_size = ramsize + devsize;
539 double threshold = max_downtime * 125 * 1000 * 1000;
543 XBT_WARN("migrate a VM, but ramsize is zero");
545 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
547 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
549 /* Stage1: send all memory pages to the destination. */
550 start_dirty_page_tracking(vm);
552 send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0);
554 remaining_size -= ramsize;
558 /* Stage2: send update pages iteratively until the size of remaining states
559 * becomes smaller than the threshold value. */
562 if (max_downtime == 0) {
563 XBT_WARN("no max_downtime parameter, skip stage2");
568 int stage2_round = 0;
570 // long updated_size = lookup_dirty_pages(vm);
571 double updated_size = lookup_computed_flop_counts(vm, stage2_round) * dp_rate;
572 if (updated_size > dp_cap) {
573 XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f",
574 stage2_round, updated_size, dp_cap);
575 updated_size = dp_cap;
578 remaining_size += updated_size;
580 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
581 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
583 if (remaining_size < threshold)
586 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round);
588 remaining_size -= updated_size;
594 /* Stage3: stop the VM and copy the rest of states. */
595 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
596 simcall_vm_suspend(vm);
597 stop_dirty_page_tracking(vm);
599 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0);
603 XBT_DEBUG("mig: tx_done");
610 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
612 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
615 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
617 char **argv = xbt_new(char *, nargvs);
618 argv[0] = xbt_strdup(pr_name);
619 argv[1] = xbt_strdup(sg_host_name(vm));
620 argv[2] = xbt_strdup(sg_host_name(src_pm));
621 argv[3] = xbt_strdup(sg_host_name(dst_pm));
624 msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
630 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
632 char **argv = xbt_new(char *, nargvs);
633 argv[0] = xbt_strdup(pr_name);
634 argv[1] = xbt_strdup(sg_host_name(vm));
635 argv[2] = xbt_strdup(sg_host_name(src_pm));
636 argv[3] = xbt_strdup(sg_host_name(dst_pm));
639 msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
644 /* wait until the migration have finished */
646 msg_task_t task = NULL;
647 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
648 xbt_assert(ret == MSG_OK);
650 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
651 xbt_assert(strcmp(task->name, expected_task_name) == 0);
652 xbt_free(expected_task_name);
659 /** @brief Migrate the VM to the given host.
662 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
663 * MSG_task_send() before or after, depending on whether you want to do cold or hot
666 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
669 * - One approach is ...
670 * We first create a new VM (i.e., destination VM) on the destination
671 * physical host. The destination VM will receive the state of the source
672 * VM over network. We will finally destroy the source VM.
673 * - This behavior is similar to the way of migration in the real world.
674 * Even before a migration is completed, we will see a destination VM,
675 * consuming resources.
676 * - We have to relocate all processes. The existing process migraion code
677 * will work for this?
678 * - The name of the VM is a somewhat unique ID in the code. It is tricky
679 * for the destination VM?
681 * - Another one is ...
682 * We update the information of the given VM to place it to the destination
685 * The second one would be easier.
689 msg_host_t old_pm = simcall_vm_get_pm(vm);
691 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
692 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
694 do_migration(vm, old_pm, new_pm);
698 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
701 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
706 /** @brief Immediately suspend the execution of all processes within the given VM.
709 * This function stops the exection of the VM. All the processes on this VM
710 * will pause. The state of the VM is perserved. We can later resume it again.
712 * No suspension cost occurs.
714 void MSG_vm_suspend(msg_vm_t vm)
716 simcall_vm_suspend(vm);
718 XBT_DEBUG("vm_suspend done");
721 TRACE_msg_vm_suspend(vm);
726 /** @brief Resume the execution of the VM. All processes on the VM run again.
729 * No resume cost occurs.
731 void MSG_vm_resume(msg_vm_t vm)
733 simcall_vm_resume(vm);
736 TRACE_msg_vm_resume(vm);
741 /** @brief Immediately save the execution of all processes within the given VM.
744 * This function stops the exection of the VM. All the processes on this VM
745 * will pause. The state of the VM is perserved. We can later resume it again.
747 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
748 * use a \ref MSG_file_write() before or after, depending on the exact semantic
751 void MSG_vm_save(msg_vm_t vm)
755 TRACE_msg_vm_save(vm);
759 /** @brief Restore the execution of the VM. All processes on the VM run again.
762 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
763 * use a \ref MSG_file_read() before or after, depending on the exact semantic
764 * of VM restore to you.
766 void MSG_vm_restore(msg_vm_t vm)
768 simcall_vm_restore(vm);
771 TRACE_msg_vm_restore(vm);
776 /** @brief Destroy a VM. Destroy the VM object from the simulation.
779 void MSG_vm_destroy(msg_vm_t vm)
781 /* First, terminate all processes on the VM if necessary */
782 if (MSG_vm_is_running(vm))
783 simcall_vm_shutdown(vm);
785 if (!MSG_vm_is_created(vm)) {
786 XBT_CRITICAL("shutdown the given VM before destroying it");
790 /* Then, destroy the VM object */
791 simcall_vm_destroy(vm);
793 __MSG_host_destroy(vm);
796 TRACE_msg_vm_end(vm);
801 /** @brief Get the physical host of a givne VM.
804 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
806 return simcall_vm_get_pm(vm);