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);
120 /** @brief Returns whether the given VM is currently migrating
123 int MSG_vm_is_migrating(msg_vm_t vm)
125 return __MSG_vm_is_state(vm, SURF_VM_STATE_MIGRATING);
128 /** @brief Returns whether the given VM is currently suspended, not running.
131 int MSG_vm_is_suspended(msg_vm_t vm)
133 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
136 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
139 int MSG_vm_is_saving(msg_vm_t vm)
141 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
144 /** @brief Returns whether the given VM has been saved, not running.
147 int MSG_vm_is_saved(msg_vm_t vm)
149 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
152 /** @brief Returns whether the given VM is being restored, not running.
155 int MSG_vm_is_restoring(msg_vm_t vm)
157 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
162 /* ------------------------------------------------------------------------- */
163 /* ------------------------------------------------------------------------- */
165 /* **** ******** MSG vm actions ********* **** */
167 /** @brief Create a new VM with specified parameters.
171 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name,
172 int ncpus, long ramsize, long net_cap, char *disk_path, long disksize)
174 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
177 s_ws_params_t params;
178 memset(¶ms, 0, sizeof(params));
179 params.ramsize = ramsize;
180 //params.overcommit = 0;
181 simcall_host_set_params(vm, ¶ms);
184 /* TODO: Limit net capability, take into account disk considerations. */
190 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
193 * A VM is treated as a host. The name of the VM must be unique among all hosts.
195 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
197 /* make sure the VM of the same name does not exit */
199 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
201 XBT_ERROR("host %s already exits", name);
206 /* Note: ind_vm and vm_workstation point to the same elm object. */
207 msg_vm_t ind_vm = NULL;
208 void *ind_vm_workstation = NULL;
210 /* Ask the SIMIX layer to create the surf vm resource */
211 ind_vm_workstation = simcall_vm_create(name, ind_pm);
212 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
214 XBT_DEBUG("A new VM (%s) has been created", name);
217 TRACE_msg_vm_create(name, ind_pm);
223 /** @brief Destroy a VM. Destroy the VM object from the simulation.
226 void MSG_vm_destroy(msg_vm_t vm)
228 /* First, terminate all processes on the VM if necessary */
229 if (MSG_vm_is_running(vm))
230 simcall_vm_shutdown(vm);
232 if (!MSG_vm_is_created(vm)) {
233 XBT_CRITICAL("shutdown the given VM before destroying it");
237 /* Then, destroy the VM object */
238 simcall_vm_destroy(vm);
240 __MSG_host_destroy(vm);
243 TRACE_msg_vm_end(vm);
248 /** @brief Start a vm (i.e., boot the guest operating system)
251 * If the VM cannot be started, an exception is generated.
254 void MSG_vm_start(msg_vm_t vm)
256 simcall_vm_start(vm);
259 TRACE_msg_vm_start(vm);
265 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
268 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
269 * use a #MSG_process_sleep() or something. I'm not quite sure.
271 void MSG_vm_shutdown(msg_vm_t vm)
273 /* msg_vm_t equals to msg_host_t */
274 simcall_vm_shutdown(vm);
276 // #ifdef HAVE_TRACING
277 // TRACE_msg_vm_(vm);
283 /* We have two mailboxes. mbox is used to transfer migration data between
284 * source and destiantion PMs. mbox_ctl is used to detect the completion of a
285 * migration. The names of these mailboxes must not conflict with others. */
286 static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
288 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
291 static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
293 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
296 static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
298 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
301 static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
303 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
306 static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
308 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
311 static int migration_rx_fun(int argc, char *argv[])
313 XBT_DEBUG("mig: rx_start");
315 xbt_assert(argc == 4);
316 const char *vm_name = argv[1];
317 const char *src_pm_name = argv[2];
318 const char *dst_pm_name = argv[3];
319 msg_vm_t vm = MSG_get_host_by_name(vm_name);
320 msg_vm_t dst_pm = MSG_get_host_by_name(dst_pm_name);
324 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
325 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
326 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
329 msg_task_t task = NULL;
330 MSG_task_recv(&task, mbox);
332 if (strcmp(task->name, finalize_task_name) == 0)
335 MSG_task_destroy(task);
342 simcall_vm_migrate(vm, dst_pm);
343 simcall_vm_resume(vm);
346 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
348 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
349 msg_error_t ret = MSG_task_send(task, mbox_ctl);
350 xbt_assert(ret == MSG_OK);
358 xbt_free(finalize_task_name);
360 XBT_DEBUG("mig: rx_done");
365 static void reset_dirty_pages(msg_vm_t vm)
367 msg_host_priv_t priv = msg_host_resource_priv(vm);
370 xbt_dict_cursor_t cursor = NULL;
371 dirty_page_t dp = NULL;
372 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
373 double remaining = MSG_task_get_remaining_computation(dp->task);
374 dp->prev_clock = MSG_get_clock();
375 dp->prev_remaining = remaining;
377 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
381 static void start_dirty_page_tracking(msg_vm_t vm)
383 msg_host_priv_t priv = msg_host_resource_priv(vm);
384 priv->dp_enabled = 1;
386 reset_dirty_pages(vm);
389 static void stop_dirty_page_tracking(msg_vm_t vm)
391 msg_host_priv_t priv = msg_host_resource_priv(vm);
392 priv->dp_enabled = 0;
396 /* It might be natural that we define dp_rate for each task. But, we will also
397 * have to care about how each task behavior affects the memory update behavior
398 * at the operating system level. It may not be easy to model it with a simple algorithm. */
399 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
401 double computed = dp->prev_remaining - remaining;
402 double duration = clock - dp->prev_clock;
403 double updated = dp->task->dp_rate * computed;
405 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
406 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
407 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
408 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
414 double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
416 double computed = dp->prev_remaining - remaining;
417 double duration = clock - dp->prev_clock;
419 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
420 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
425 static double lookup_computed_flop_counts(msg_vm_t vm, int stage2_round_for_fancy_debug)
427 msg_host_priv_t priv = msg_host_resource_priv(vm);
431 xbt_dict_cursor_t cursor = NULL;
432 dirty_page_t dp = NULL;
433 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
434 double remaining = MSG_task_get_remaining_computation(dp->task);
435 double clock = MSG_get_clock();
437 // total += calc_updated_pages(key, vm, dp, remaining, clock);
438 total += get_computed(key, vm, dp, remaining, clock);
440 dp->prev_remaining = remaining;
441 dp->prev_clock = clock;
444 total += priv->dp_updated_by_deleted_tasks;
446 XBT_INFO("mig-stage2.%d: computed %f flop_counts (including %f by deleted tasks)",
447 stage2_round_for_fancy_debug,
448 total, priv->dp_updated_by_deleted_tasks);
452 priv->dp_updated_by_deleted_tasks = 0;
458 // TODO Is this code redundant with the information provided by
459 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
460 void MSG_host_add_task(msg_host_t host, msg_task_t task)
462 msg_host_priv_t priv = msg_host_resource_priv(host);
463 double remaining = MSG_task_get_remaining_computation(task);
464 char *key = bprintf("%s-%lld", task->name, task->counter);
466 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
469 /* It should be okay that we add a task onto a migrating VM. */
470 if (priv->dp_enabled) {
471 dp->prev_clock = MSG_get_clock();
472 dp->prev_remaining = remaining;
475 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
476 xbt_dict_set(priv->dp_objs, key, dp, NULL);
477 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
482 void MSG_host_del_task(msg_host_t host, msg_task_t task)
484 msg_host_priv_t priv = msg_host_resource_priv(host);
486 char *key = bprintf("%s-%lld", task->name, task->counter);
488 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
489 xbt_assert(dp->task == task);
491 /* If we are in the middle of dirty page tracking, we record how much
492 * computaion has been done until now, and keep the information for the
493 * lookup_() function that will called soon. */
494 if (priv->dp_enabled) {
495 double remaining = MSG_task_get_remaining_computation(task);
496 double clock = MSG_get_clock();
497 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
498 double updated = get_computed(key, host, dp, remaining, clock);
500 priv->dp_updated_by_deleted_tasks += updated;
503 xbt_dict_remove(priv->dp_objs, key);
506 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
512 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
513 double size, char *mbox, int stage, int stage2_round, double mig_speed)
515 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
516 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
520 ret = MSG_task_send_bounded(task, mbox, mig_speed);
522 ret = MSG_task_send(task, mbox);
523 xbt_assert(ret == MSG_OK);
526 XBT_INFO("mig-stage%d.%d: sent %f", stage, stage2_round, size);
528 XBT_INFO("mig-stage%d: sent %f", stage, size);
534 static int migration_tx_fun(int argc, char *argv[])
536 XBT_DEBUG("mig: tx_start");
538 xbt_assert(argc == 4);
539 const char *vm_name = argv[1];
540 const char *src_pm_name = argv[2];
541 const char *dst_pm_name = argv[3];
542 msg_vm_t vm = MSG_get_host_by_name(vm_name);
545 s_ws_params_t params;
546 simcall_host_get_params(vm, ¶ms);
547 const long ramsize = params.ramsize;
548 const long devsize = params.devsize;
549 const int skip_stage2 = params.skip_stage2;
550 const double dp_rate = params.dp_rate;
551 const double dp_cap = params.dp_cap;
552 const double mig_speed = params.mig_speed;
553 double remaining_size = ramsize + devsize;
555 double max_downtime = params.max_downtime;
556 if (max_downtime == 0) {
557 XBT_WARN("use the default max_downtime value 30ms");
561 double threshold = max_downtime * 125 * 1024 * 1024;
563 /* setting up parameters has done */
567 XBT_WARN("migrate a VM, but ramsize is zero");
569 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
571 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
573 /* Stage1: send all memory pages to the destination. */
574 start_dirty_page_tracking(vm);
576 send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed);
578 remaining_size -= ramsize;
582 /* Stage2: send update pages iteratively until the size of remaining states
583 * becomes smaller than the threshold value. */
586 if (max_downtime == 0) {
587 XBT_WARN("no max_downtime parameter, skip stage2");
592 int stage2_round = 0;
594 // long updated_size = lookup_dirty_pages(vm);
595 double updated_size = lookup_computed_flop_counts(vm, stage2_round) * dp_rate;
596 if (updated_size > dp_cap) {
597 XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f",
598 stage2_round, updated_size, dp_cap);
599 updated_size = dp_cap;
602 remaining_size += updated_size;
604 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
605 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
607 if (remaining_size < threshold)
610 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed);
612 remaining_size -= updated_size;
618 /* Stage3: stop the VM and copy the rest of states. */
619 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
620 simcall_vm_suspend(vm);
621 stop_dirty_page_tracking(vm);
623 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed);
627 XBT_DEBUG("mig: tx_done");
634 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
636 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
639 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
641 char **argv = xbt_new(char *, nargvs);
642 argv[0] = xbt_strdup(pr_name);
643 argv[1] = xbt_strdup(sg_host_name(vm));
644 argv[2] = xbt_strdup(sg_host_name(src_pm));
645 argv[3] = xbt_strdup(sg_host_name(dst_pm));
648 MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
654 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
656 char **argv = xbt_new(char *, nargvs);
657 argv[0] = xbt_strdup(pr_name);
658 argv[1] = xbt_strdup(sg_host_name(vm));
659 argv[2] = xbt_strdup(sg_host_name(src_pm));
660 argv[3] = xbt_strdup(sg_host_name(dst_pm));
662 MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
667 /* wait until the migration have finished */
669 msg_task_t task = NULL;
670 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
671 xbt_assert(ret == MSG_OK);
673 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
674 xbt_assert(strcmp(task->name, expected_task_name) == 0);
675 xbt_free(expected_task_name);
682 /** @brief Migrate the VM to the given host.
685 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
686 * MSG_task_send() before or after, depending on whether you want to do cold or hot
689 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
692 * - One approach is ...
693 * We first create a new VM (i.e., destination VM) on the destination
694 * physical host. The destination VM will receive the state of the source
695 * VM over network. We will finally destroy the source VM.
696 * - This behavior is similar to the way of migration in the real world.
697 * Even before a migration is completed, we will see a destination VM,
698 * consuming resources.
699 * - We have to relocate all processes. The existing process migraion code
700 * will work for this?
701 * - The name of the VM is a somewhat unique ID in the code. It is tricky
702 * for the destination VM?
704 * - Another one is ...
705 * We update the information of the given VM to place it to the destination
708 * The second one would be easier.
712 msg_host_t old_pm = simcall_vm_get_pm(vm);
714 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
715 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
717 do_migration(vm, old_pm, new_pm);
721 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
724 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
729 /** @brief Immediately suspend the execution of all processes within the given VM.
732 * This function stops the exection of the VM. All the processes on this VM
733 * will pause. The state of the VM is perserved. We can later resume it again.
735 * No suspension cost occurs.
737 void MSG_vm_suspend(msg_vm_t vm)
739 simcall_vm_suspend(vm);
741 XBT_DEBUG("vm_suspend done");
744 TRACE_msg_vm_suspend(vm);
749 /** @brief Resume the execution of the VM. All processes on the VM run again.
752 * No resume cost occurs.
754 void MSG_vm_resume(msg_vm_t vm)
756 simcall_vm_resume(vm);
759 TRACE_msg_vm_resume(vm);
764 /** @brief Immediately save the execution of all processes within the given VM.
767 * This function stops the exection of the VM. All the processes on this VM
768 * will pause. The state of the VM is perserved. We can later resume it again.
770 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
771 * use a \ref MSG_file_write() before or after, depending on the exact semantic
774 void MSG_vm_save(msg_vm_t vm)
778 TRACE_msg_vm_save(vm);
782 /** @brief Restore the execution of the VM. All processes on the VM run again.
785 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
786 * use a \ref MSG_file_read() before or after, depending on the exact semantic
787 * of VM restore to you.
789 void MSG_vm_restore(msg_vm_t vm)
791 simcall_vm_restore(vm);
794 TRACE_msg_vm_restore(vm);
801 /** @brief Get the physical host of a given VM.
804 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
806 return simcall_vm_get_pm(vm);