1 /* Copyright (c) 2012-2015. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
8 * 1. add the support of trace
9 * 2. use parallel tasks to simulate CPU overhead and remove the very
10 * experimental code generating micro computation tasks
15 #include "msg_private.h"
16 #include "xbt/sysdep.h"
18 #include "simgrid/platf.h"
19 #include "simgrid/host.h"
21 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
22 "Cloud-oriented parts of the MSG API");
25 /* **** ******** GENERAL ********* **** */
27 /** \ingroup m_vm_management
28 * \brief Returns the value of a given vm property
31 * \param name a property name
32 * \return value of a property (or NULL if property not set)
35 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
37 return MSG_host_get_property_value(vm, name);
40 /** \ingroup m_vm_management
41 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
44 * \return a dict containing the properties
46 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
48 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
50 return (simcall_host_get_properties(vm));
53 /** \ingroup m_host_management
54 * \brief Change the value of a given host property
57 * \param name a property name
58 * \param value what to change the property to
59 * \param free_ctn the freeing function to use to kill the value on need
61 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
63 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
66 /** \ingroup msg_vm_management
67 * \brief Finds a msg_vm_t using its name.
69 * This is a name directory service
70 * \param name the name of a vm.
71 * \return the corresponding vm
73 * Please note that a VM is a specific host. Hence, you should give a different name
77 msg_vm_t MSG_vm_get_by_name(const char *name)
79 return MSG_host_by_name(name);
82 /** \ingroup m_vm_management
84 * \brief Return the name of the #msg_host_t.
86 * This functions checks whether \a host is a valid pointer or not and return
89 const char *MSG_vm_get_name(msg_vm_t vm)
91 return MSG_host_get_name(vm);
95 /* **** Check state of a VM **** */
96 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
98 return simcall_vm_get_state(vm) == state;
101 /** @brief Returns whether the given VM has just created, not running.
104 int MSG_vm_is_created(msg_vm_t vm)
106 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
109 /** @brief Returns whether the given VM is currently running
112 int MSG_vm_is_running(msg_vm_t vm)
114 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
117 /** @brief Returns whether the given VM is currently migrating
120 int MSG_vm_is_migrating(msg_vm_t vm)
122 msg_host_priv_t priv = sg_host_msg(vm);
123 return priv->is_migrating;
126 /** @brief Returns whether the given VM is currently suspended, not running.
129 int MSG_vm_is_suspended(msg_vm_t vm)
131 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
134 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
137 int MSG_vm_is_saving(msg_vm_t vm)
139 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
142 /** @brief Returns whether the given VM has been saved, not running.
145 int MSG_vm_is_saved(msg_vm_t vm)
147 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
150 /** @brief Returns whether the given VM is being restored, not running.
153 int MSG_vm_is_restoring(msg_vm_t vm)
155 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
160 /* ------------------------------------------------------------------------- */
161 /* ------------------------------------------------------------------------- */
163 /* **** ******** MSG vm actions ********* **** */
165 /** @brief Create a new VM with specified parameters.
167 * @param pm Physical machine that will host the VM
168 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
169 * @param disk_path (unused) Path to the image that boots
170 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
171 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
172 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
175 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name,
176 int ncpus, int ramsize,
177 int net_cap, char *disk_path, int disksize,
178 int mig_netspeed, int dp_intensity)
180 /* For the moment, intensity_rate is the percentage against the migration
182 double host_speed = MSG_get_host_speed(pm);
183 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
185 msg_vm_t vm = MSG_vm_create_core(pm, name);
186 s_vm_params_t params;
187 memset(¶ms, 0, sizeof(params));
188 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
189 //params.overcommit = 0;
191 params.skip_stage2 = 0;
192 params.max_downtime = 0.03;
193 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
194 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
195 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
197 //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);
198 simcall_host_set_params(vm, ¶ms);
204 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
207 * A VM is treated as a host. The name of the VM must be unique among all hosts.
209 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
211 /* make sure the VM of the same name does not exit */
213 simgrid::Host* ind_host_tmp =
214 (simgrid::Host*) xbt_dict_get_or_null(host_list, name);
215 if (ind_host_tmp != nullptr && sg_host_simix(ind_host_tmp) != nullptr) {
216 XBT_ERROR("host %s already exits", name);
221 /* Note: ind_vm and vm_workstation point to the same elm object. */
223 /* Ask the SIMIX layer to create the surf vm resource */
224 sg_host_t ind_vm_workstation = (sg_host_t) simcall_vm_create(name, ind_pm);
226 msg_vm_t ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
228 XBT_DEBUG("A new VM (%s) has been created", name);
230 TRACE_msg_vm_create(name, ind_pm);
235 /** @brief Destroy a VM. Destroy the VM object from the simulation.
238 void MSG_vm_destroy(msg_vm_t vm)
240 if (MSG_vm_is_migrating(vm))
241 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
243 /* First, terminate all processes on the VM if necessary */
244 if (MSG_vm_is_running(vm))
245 simcall_vm_shutdown(vm);
247 if (!MSG_vm_is_created(vm)) {
248 XBT_CRITICAL("shutdown the given VM before destroying it");
252 /* Then, destroy the VM object */
253 simcall_vm_destroy(vm);
255 __MSG_host_destroy(vm);
257 TRACE_msg_vm_end(vm);
261 /** @brief Start a vm (i.e., boot the guest operating system)
264 * If the VM cannot be started, an exception is generated.
267 void MSG_vm_start(msg_vm_t vm)
269 simcall_vm_start(vm);
271 TRACE_msg_vm_start(vm);
276 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
279 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
280 * use a #MSG_process_sleep() or something. I'm not quite sure.
282 void MSG_vm_shutdown(msg_vm_t vm)
284 /* msg_vm_t equals to msg_host_t */
285 simcall_vm_shutdown(vm);
287 // TRACE_msg_vm_(vm);
292 /* We have two mailboxes. mbox is used to transfer migration data between
293 * source and destination PMs. mbox_ctl is used to detect the completion of a
294 * migration. The names of these mailboxes must not conflict with others. */
295 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
297 const char *vm_name = sg_host_get_name(vm);
298 const char *src_pm_name = sg_host_get_name(src_pm);
299 const char *dst_pm_name = sg_host_get_name(dst_pm);
301 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
304 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
306 const char *vm_name = sg_host_get_name(vm);
307 const char *src_pm_name = sg_host_get_name(src_pm);
308 const char *dst_pm_name = sg_host_get_name(dst_pm);
310 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
313 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
315 const char *vm_name = sg_host_get_name(vm);
316 const char *src_pm_name = sg_host_get_name(src_pm);
317 const char *dst_pm_name = sg_host_get_name(dst_pm);
319 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
322 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
324 const char *vm_name = sg_host_get_name(vm);
325 const char *src_pm_name = sg_host_get_name(src_pm);
326 const char *dst_pm_name = sg_host_get_name(dst_pm);
328 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
331 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
333 const char *vm_name = sg_host_get_name(vm);
334 const char *src_pm_name = sg_host_get_name(src_pm);
335 const char *dst_pm_name = sg_host_get_name(dst_pm);
337 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
341 struct migration_session {
346 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
347 * know the completion of the migration. */
349 /* The migration_rx and migration_tx processes use mbox to transfer migration
355 static int migration_rx_fun(int argc, char *argv[])
357 XBT_DEBUG("mig: rx_start");
359 // The structure has been created in the do_migration function and should only be freed in the same place ;)
360 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
362 s_vm_params_t params;
363 simcall_host_get_params(ms->vm, ¶ms);
367 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
371 msg_task_t task = NULL;
372 ret = MSG_task_recv(&task, ms->mbox);
375 // An error occured, clean the code and return
376 // The owner did not change, hence the task should be only destroyed on the other side
377 xbt_free(finalize_task_name);
382 if (strcmp(task->name, finalize_task_name) == 0)
385 MSG_task_destroy(task);
391 // Here Stage 1, 2 and 3 have been performed.
392 // Hence complete the migration
394 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
395 // 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)
396 msg_vm_t vm = ms->vm;
397 msg_host_t src_pm = ms->src_pm;
398 msg_host_t dst_pm = ms-> dst_pm;
399 msg_host_priv_t priv = sg_host_msg(vm);
401 // // 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
402 // // I should check with Takahiro in order to make this portion of code atomic
403 // /* deinstall the current affinity setting for the CPU */
404 // simcall_vm_set_affinity(vm, src_pm, 0);
406 // /* Update the vm location */
407 // simcall_vm_migrate(vm, dst_pm);
409 // /* Resume the VM */
410 // simcall_vm_resume(vm);
412 simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
414 /* install the affinity setting of the VM on the destination pm */
417 unsigned long affinity_mask = (unsigned long)(uintptr_t) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *)dst_pm, sizeof(msg_host_t));
418 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
419 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
424 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
425 msg_host_priv_t priv = sg_host_msg(vm);
426 priv->is_migrating = 0;
427 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)",
428 sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
429 sg_host_get_name(ms->dst_pm));
430 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
432 // Inform the SRC that the migration has been correctly performed
434 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
435 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
436 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
437 // xbt_assert(ret == MSG_OK);
438 if(ret == MSG_HOST_FAILURE){
439 // 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
440 // TODO What does it mean ? What should we do ?
441 MSG_task_destroy(task);
442 } else if(ret == MSG_TRANSFER_FAILURE){
443 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
444 MSG_task_destroy(task);
451 xbt_free(finalize_task_name);
453 XBT_DEBUG("mig: rx_done");
458 static void reset_dirty_pages(msg_vm_t vm)
460 msg_host_priv_t priv = sg_host_msg(vm);
463 xbt_dict_cursor_t cursor = NULL;
464 dirty_page_t dp = NULL;
465 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
466 double remaining = MSG_task_get_flops_amount(dp->task);
467 dp->prev_clock = MSG_get_clock();
468 dp->prev_remaining = remaining;
470 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
474 static void start_dirty_page_tracking(msg_vm_t vm)
476 msg_host_priv_t priv = sg_host_msg(vm);
477 priv->dp_enabled = 1;
479 reset_dirty_pages(vm);
482 static void stop_dirty_page_tracking(msg_vm_t vm)
484 msg_host_priv_t priv = sg_host_msg(vm);
485 priv->dp_enabled = 0;
488 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
490 double computed = dp->prev_remaining - remaining;
491 double duration = clock - dp->prev_clock;
493 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
494 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
499 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
501 msg_host_priv_t priv = sg_host_msg(vm);
505 xbt_dict_cursor_t cursor = NULL;
506 dirty_page_t dp = NULL;
507 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
508 double remaining = MSG_task_get_flops_amount(dp->task);
510 double clock = MSG_get_clock();
512 // total += calc_updated_pages(key, vm, dp, remaining, clock);
513 total += get_computed(key, vm, dp, remaining, clock);
515 dp->prev_remaining = remaining;
516 dp->prev_clock = clock;
519 total += priv->dp_updated_by_deleted_tasks;
521 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
522 stage_for_fancy_debug,
523 stage2_round_for_fancy_debug,
524 total, priv->dp_updated_by_deleted_tasks);
528 priv->dp_updated_by_deleted_tasks = 0;
534 // TODO Is this code redundant with the information provided by
535 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
536 void MSG_host_add_task(msg_host_t host, msg_task_t task)
538 msg_host_priv_t priv = sg_host_msg(host);
539 double remaining = MSG_task_get_flops_amount(task);
540 char *key = bprintf("%s-%p", task->name, task);
542 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
545 /* It should be okay that we add a task onto a migrating VM. */
546 if (priv->dp_enabled) {
547 dp->prev_clock = MSG_get_clock();
548 dp->prev_remaining = remaining;
551 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
552 xbt_dict_set(priv->dp_objs, key, dp, NULL);
553 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, priv->dp_enabled);
558 void MSG_host_del_task(msg_host_t host, msg_task_t task)
560 msg_host_priv_t priv = sg_host_msg(host);
562 char *key = bprintf("%s-%p", task->name, task);
564 dirty_page_t dp = (dirty_page_t) xbt_dict_get_or_null(priv->dp_objs, key);
565 xbt_assert(dp->task == task);
567 /* If we are in the middle of dirty page tracking, we record how much
568 * computation has been done until now, and keep the information for the
569 * lookup_() function that will called soon. */
570 if (priv->dp_enabled) {
571 double remaining = MSG_task_get_flops_amount(task);
572 double clock = MSG_get_clock();
573 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
574 double updated = get_computed(key, host, dp, remaining, clock);
576 priv->dp_updated_by_deleted_tasks += updated;
579 xbt_dict_remove(priv->dp_objs, key);
582 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
590 static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm,
591 sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double timeout)
594 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
595 msg_task_t task = MSG_task_create(task_name, 0, (double)size, NULL);
599 double clock_sta = MSG_get_clock();
603 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
605 ret = MSG_task_send(task, mbox);
611 } else if (ret == MSG_TIMEOUT) {
612 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
613 sent = size - remaining;
614 XBT_INFO("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu",
615 timeout, remaining, size);
618 /* FIXME: why try-and-catch is used here? */
619 if(ret == MSG_HOST_FAILURE){
620 //XBT_INFO("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
621 MSG_task_destroy(task);
622 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
623 }else if(ret == MSG_TRANSFER_FAILURE){
624 //XBT_INFO("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
625 MSG_task_destroy(task);
626 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
629 double clock_end = MSG_get_clock();
630 double duration = clock_end - clock_sta;
631 double actual_speed = size / duration;
634 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration, actual_speed, mig_speed);
636 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed, mig_speed);
641 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
643 double updated_size = computed * dp_rate;
644 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
645 if (updated_size > dp_cap) {
646 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
647 updated_size = dp_cap;
650 return (sg_size_t) updated_size;
653 static double send_stage1(struct migration_session *ms,
654 sg_size_t ramsize, double mig_speed, double dp_rate, double dp_cap)
657 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
658 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
659 sg_size_t remaining = ramsize;
660 double computed_total = 0;
662 while (remaining > 0) {
663 sg_size_t datasize = chunksize;
664 if (remaining < chunksize)
665 datasize = remaining;
667 remaining -= datasize;
668 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
669 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
670 computed_total += computed;
673 return computed_total;
678 static double get_threshold_value(double bandwidth, double max_downtime)
680 return max_downtime * bandwidth;
683 static int migration_tx_fun(int argc, char *argv[])
685 XBT_DEBUG("mig: tx_start");
687 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
688 migration_session *ms =
689 (migration_session *) MSG_process_get_data(MSG_process_self());
691 s_vm_params_t params;
692 simcall_host_get_params(ms->vm, ¶ms);
693 const sg_size_t ramsize = params.ramsize;
694 const sg_size_t devsize = params.devsize;
695 const int skip_stage1 = params.skip_stage1;
696 int skip_stage2 = params.skip_stage2;
697 const double dp_rate = params.dp_rate;
698 const double dp_cap = params.dp_cap;
699 const double mig_speed = params.mig_speed;
700 double max_downtime = params.max_downtime;
702 /* hard code it temporally. Fix Me */
703 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
704 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
706 double remaining_size = (double) (ramsize + devsize);
707 double threshold = 0.0;
709 /* check parameters */
711 XBT_WARN("migrate a VM, but ramsize is zero");
713 if (max_downtime == 0) {
714 XBT_WARN("use the default max_downtime value 30ms");
718 /* Stage1: send all memory pages to the destination. */
719 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
720 start_dirty_page_tracking(ms->vm);
722 double computed_during_stage1 = 0;
724 double clock_prev_send = MSG_get_clock();
727 /* At stage 1, we do not need timeout. We have to send all the memory
728 * pages even though the duration of this tranfer exceeds the timeout
730 XBT_INFO("Stage 1: Gonna send %llu", ramsize);
731 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
732 remaining_size -= sent;
733 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
735 if (sent < ramsize) {
736 XBT_INFO("mig-stage1: timeout, force moving to stage 3");
738 } else if (sent > ramsize)
742 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
743 // Stop the dirty page tracking an return (there is no memory space to release)
744 stop_dirty_page_tracking(ms->vm);
748 double clock_post_send = MSG_get_clock();
749 mig_timeout -= (clock_post_send - clock_prev_send);
750 if (mig_timeout < 0) {
751 XBT_INFO("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
752 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
756 /* estimate bandwidth */
757 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
758 threshold = get_threshold_value(bandwidth, max_downtime);
759 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
763 /* Stage2: send update pages iteratively until the size of remaining states
764 * becomes smaller than the threshold value. */
767 int stage2_round = 0;
770 sg_size_t updated_size = 0;
771 if (stage2_round == 0) {
772 /* just after stage1, nothing has been updated. But, we have to send the
773 * data updated during stage1 */
774 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
776 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
777 updated_size = get_updated_size(computed, dp_rate, dp_cap);
780 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
781 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
784 /* Check whether the remaining size is below the threshold value. If so,
785 * move to stage 3. */
786 remaining_size += updated_size;
787 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
788 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
789 if (remaining_size < threshold)
794 double clock_prev_send = MSG_get_clock();
796 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
797 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed, mig_timeout);
799 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
800 // Stop the dirty page tracking an return (there is no memory space to release)
801 stop_dirty_page_tracking(ms->vm);
804 double clock_post_send = MSG_get_clock();
806 if (sent == updated_size) {
807 /* timeout did not happen */
808 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
809 threshold = get_threshold_value(bandwidth, max_downtime);
810 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
811 remaining_size -= sent;
813 mig_timeout -= (clock_post_send - clock_prev_send);
814 xbt_assert(mig_timeout > 0);
816 } else if (sent < updated_size) {
817 /* When timeout happens, we move to stage 3. The size of memory pages
818 * updated before timeout must be added to the remaining size. */
819 XBT_INFO("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
820 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
821 remaining_size -= sent;
823 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
824 updated_size = get_updated_size(computed, dp_rate, dp_cap);
825 remaining_size += updated_size;
833 /* Stage3: stop the VM and copy the rest of states. */
834 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
835 simcall_vm_suspend(ms->vm);
836 stop_dirty_page_tracking(ms->vm);
839 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
840 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
842 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
843 // Stop the dirty page tracking an return (there is no memory space to release)
844 simcall_vm_resume(ms->vm);
848 // 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.
849 XBT_DEBUG("mig: tx_done");
856 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
858 struct migration_session *ms = xbt_new(struct migration_session, 1);
862 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
863 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
866 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
867 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
869 // msg_process_t tx_process, rx_process;
870 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
871 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
874 char **argv = xbt_new(char *, 2);
875 argv[0] = pr_rx_name;
877 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
880 char **argv = xbt_new(char *, 2);
881 argv[0] = pr_tx_name;
883 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
887 /* wait until the migration have finished or on error has occured */
889 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
890 msg_task_t task = NULL;
891 msg_error_t ret = MSG_TIMEOUT;
892 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
893 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
895 xbt_free(ms->mbox_ctl);
899 //xbt_assert(ret == MSG_OK);
900 if(ret == MSG_HOST_FAILURE){
901 // Note that since the communication failed, the owner did not change and the task should be destroyed on the other side.
902 // Hence, just throw the execption
903 XBT_INFO("SRC crashes, throw an exception (m-control)");
904 //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
907 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){ // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
908 XBT_INFO("DST crashes, throw an exception (m-control)");
913 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
914 xbt_assert(strcmp(task->name, expected_task_name) == 0);
915 xbt_free(expected_task_name);
916 MSG_task_destroy(task);
924 /** @brief Migrate the VM to the given host.
927 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
928 * MSG_task_send() before or after, depending on whether you want to do cold or hot
931 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
934 * - One approach is ...
935 * We first create a new VM (i.e., destination VM) on the destination
936 * physical host. The destination VM will receive the state of the source
937 * VM over network. We will finally destroy the source VM.
938 * - This behavior is similar to the way of migration in the real world.
939 * Even before a migration is completed, we will see a destination VM,
940 * consuming resources.
941 * - We have to relocate all processes. The existing process migraion code
942 * will work for this?
943 * - The name of the VM is a somewhat unique ID in the code. It is tricky
944 * for the destination VM?
946 * - Another one is ...
947 * We update the information of the given VM to place it to the destination
950 * The second one would be easier.
954 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
956 if(MSG_host_is_off(old_pm))
957 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
959 if(MSG_host_is_off(new_pm))
960 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
962 if (!MSG_vm_is_running(vm))
963 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
965 if (MSG_vm_is_migrating(vm))
966 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
968 msg_host_priv_t priv = sg_host_msg(vm);
969 priv->is_migrating = 1;
973 int ret = do_migration(vm, old_pm, new_pm);
975 priv->is_migrating = 0;
976 THROWF(host_error, 0, "SRC host failed during migration");
979 priv->is_migrating = 0;
980 THROWF(host_error, 0, "DST host failed during migration");
984 // 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
985 // In that case, the VM has been already assigned to the DST node.
986 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
987 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
991 /** @brief Immediately suspend the execution of all processes within the given VM.
994 * This function stops the execution of the VM. All the processes on this VM
995 * will pause. The state of the VM is preserved. We can later resume it again.
997 * No suspension cost occurs.
999 void MSG_vm_suspend(msg_vm_t vm)
1001 if (MSG_vm_is_migrating(vm))
1002 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
1004 simcall_vm_suspend(vm);
1006 XBT_DEBUG("vm_suspend done");
1008 TRACE_msg_vm_suspend(vm);
1012 /** @brief Resume the execution of the VM. All processes on the VM run again.
1015 * No resume cost occurs.
1017 void MSG_vm_resume(msg_vm_t vm)
1019 simcall_vm_resume(vm);
1021 TRACE_msg_vm_resume(vm);
1025 /** @brief Immediately save the execution of all processes within the given VM.
1028 * This function stops the execution of the VM. All the processes on this VM
1029 * will pause. The state of the VM is preserved. We can later resume it again.
1031 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1032 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1033 * of VM save to you.
1035 void MSG_vm_save(msg_vm_t vm)
1037 if (MSG_vm_is_migrating(vm))
1038 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
1040 simcall_vm_save(vm);
1041 TRACE_msg_vm_save(vm);
1044 /** @brief Restore the execution of the VM. All processes on the VM run again.
1047 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1048 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1049 * of VM restore to you.
1051 void MSG_vm_restore(msg_vm_t vm)
1053 simcall_vm_restore(vm);
1055 TRACE_msg_vm_restore(vm);
1059 /** @brief Get the physical host of a given VM.
1062 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1064 return (msg_host_t) simcall_vm_get_pm(vm);
1068 /** @brief Set a CPU bound for a given VM.
1072 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1075 * On PM0, there are Task1 and VM0.
1076 * On VM0, there is Task2.
1077 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1079 * Task1\@PM0 gets 50%.
1080 * Task2\@VM0 gets 25%.
1081 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1083 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1084 * task on the VM does not affect the bound of the dummy CPU action. The bound
1085 * of the dummy CPU action is unlimited.
1087 * There are some solutions for this problem. One option is to update the bound
1088 * of the dummy CPU action automatically. It should be the sum of all tasks on
1089 * the VM. But, this solution might be costly, because we have to scan all tasks
1090 * on the VM in share_resource() or we have to trap both the start and end of
1093 * The current solution is to use MSG_vm_set_bound(), which allows us to
1094 * directly set the bound of the dummy CPU action.
1098 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1099 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1100 * exceeds the capacity of a CPU core.
1102 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1104 simcall_vm_set_bound(vm, bound);
1108 /** @brief Set the CPU affinity of a given VM.
1111 * This function changes the CPU affinity of a given VM. Usage is the same as
1112 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1114 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1116 msg_host_priv_t priv = sg_host_msg(vm);
1119 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1121 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *)(uintptr_t) mask, NULL);
1123 msg_host_t pm_now = MSG_vm_get_pm(vm);
1125 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1126 simcall_vm_set_affinity(vm, pm, mask);
1128 XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));