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)");
49 return vm->getProperties();
52 /** \ingroup m_host_management
53 * \brief Change the value of a given host property
56 * \param name a property name
57 * \param value what to change the property to
58 * \param free_ctn the freeing function to use to kill the value on need
60 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
62 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
65 /** \ingroup msg_vm_management
66 * \brief Finds a msg_vm_t using its name.
68 * This is a name directory service
69 * \param name the name of a vm.
70 * \return the corresponding vm
72 * Please note that a VM is a specific host. Hence, you should give a different name
76 msg_vm_t MSG_vm_get_by_name(const char *name)
78 return MSG_host_by_name(name);
81 /** \ingroup m_vm_management
83 * \brief Return the name of the #msg_host_t.
85 * This functions checks whether \a host is a valid pointer or not and return
88 const char *MSG_vm_get_name(msg_vm_t vm)
90 return MSG_host_get_name(vm);
94 /* **** Check state of a VM **** */
95 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
97 return simcall_vm_get_state(vm) == state;
100 /** @brief Returns whether the given VM has just created, not running.
103 int MSG_vm_is_created(msg_vm_t vm)
105 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
108 /** @brief Returns whether the given VM is currently running
111 int MSG_vm_is_running(msg_vm_t vm)
113 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
116 /** @brief Returns whether the given VM is currently migrating
119 int MSG_vm_is_migrating(msg_vm_t vm)
121 msg_host_priv_t priv = sg_host_msg(vm);
122 return priv->is_migrating;
125 /** @brief Returns whether the given VM is currently suspended, not running.
128 int MSG_vm_is_suspended(msg_vm_t vm)
130 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
133 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
136 int MSG_vm_is_saving(msg_vm_t vm)
138 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
141 /** @brief Returns whether the given VM has been saved, not running.
144 int MSG_vm_is_saved(msg_vm_t vm)
146 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
149 /** @brief Returns whether the given VM is being restored, not running.
152 int MSG_vm_is_restoring(msg_vm_t vm)
154 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
159 /* ------------------------------------------------------------------------- */
160 /* ------------------------------------------------------------------------- */
162 /* **** ******** MSG vm actions ********* **** */
164 /** @brief Create a new VM with specified parameters.
166 * @param pm Physical machine that will host the VM
167 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
168 * @param disk_path (unused) Path to the image that boots
169 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
170 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
171 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
174 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name,
175 int ncpus, int ramsize,
176 int net_cap, char *disk_path, int disksize,
177 int mig_netspeed, int dp_intensity)
179 /* For the moment, intensity_rate is the percentage against the migration
181 double host_speed = MSG_get_host_speed(pm);
182 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
184 msg_vm_t vm = MSG_vm_create_core(pm, name);
185 s_vm_params_t params;
186 memset(¶ms, 0, sizeof(params));
187 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
188 //params.overcommit = 0;
190 params.skip_stage2 = 0;
191 params.max_downtime = 0.03;
192 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
193 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
194 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
196 //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);
197 vm->setParams(¶ms);
203 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
206 * A VM is treated as a host. The name of the VM must be unique among all hosts.
208 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
210 /* make sure the VM of the same name does not exit */
212 simgrid::s4u::Host* ind_host_tmp =
213 (simgrid::s4u::Host*) xbt_dict_get_or_null(host_list, name);
214 if (ind_host_tmp != nullptr && sg_host_simix(ind_host_tmp) != nullptr) {
215 XBT_ERROR("host %s already exits", name);
220 /* Note: ind_vm and vm_workstation point to the same elm object. */
222 /* Ask the SIMIX layer to create the surf vm resource */
223 sg_host_t ind_vm_workstation = (sg_host_t) simcall_vm_create(name, ind_pm);
225 msg_vm_t ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
227 XBT_DEBUG("A new VM (%s) has been created", name);
229 TRACE_msg_vm_create(name, ind_pm);
234 /** @brief Destroy a VM. Destroy the VM object from the simulation.
237 void MSG_vm_destroy(msg_vm_t vm)
239 if (MSG_vm_is_migrating(vm))
240 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
242 /* First, terminate all processes on the VM if necessary */
243 if (MSG_vm_is_running(vm))
244 simcall_vm_shutdown(vm);
246 if (!MSG_vm_is_created(vm)) {
247 XBT_CRITICAL("shutdown the given VM before destroying it");
251 /* Then, destroy the VM object */
252 simcall_vm_destroy(vm);
254 TRACE_msg_vm_end(vm);
258 /** @brief Start a vm (i.e., boot the guest operating system)
261 * If the VM cannot be started, an exception is generated.
264 void MSG_vm_start(msg_vm_t vm)
266 simcall_vm_start(vm);
268 TRACE_msg_vm_start(vm);
273 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
276 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
277 * use a #MSG_process_sleep() or something. I'm not quite sure.
279 void MSG_vm_shutdown(msg_vm_t vm)
281 /* msg_vm_t equals to msg_host_t */
282 simcall_vm_shutdown(vm);
284 // TRACE_msg_vm_(vm);
289 /* We have two mailboxes. mbox is used to transfer migration data between
290 * source and destination PMs. mbox_ctl is used to detect the completion of a
291 * migration. The names of these mailboxes must not conflict with others. */
292 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
294 const char *vm_name = sg_host_get_name(vm);
295 const char *src_pm_name = sg_host_get_name(src_pm);
296 const char *dst_pm_name = sg_host_get_name(dst_pm);
298 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
301 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
303 const char *vm_name = sg_host_get_name(vm);
304 const char *src_pm_name = sg_host_get_name(src_pm);
305 const char *dst_pm_name = sg_host_get_name(dst_pm);
307 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
310 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
312 const char *vm_name = sg_host_get_name(vm);
313 const char *src_pm_name = sg_host_get_name(src_pm);
314 const char *dst_pm_name = sg_host_get_name(dst_pm);
316 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
319 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
321 const char *vm_name = sg_host_get_name(vm);
322 const char *src_pm_name = sg_host_get_name(src_pm);
323 const char *dst_pm_name = sg_host_get_name(dst_pm);
325 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
328 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
330 const char *vm_name = sg_host_get_name(vm);
331 const char *src_pm_name = sg_host_get_name(src_pm);
332 const char *dst_pm_name = sg_host_get_name(dst_pm);
334 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
338 struct migration_session {
343 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
344 * know the completion of the migration. */
346 /* The migration_rx and migration_tx processes use mbox to transfer migration
352 static int migration_rx_fun(int argc, char *argv[])
354 XBT_DEBUG("mig: rx_start");
356 // The structure has been created in the do_migration function and should only be freed in the same place ;)
357 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
359 s_vm_params_t params;
360 ms->vm->getParams(¶ms);
364 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
368 msg_task_t task = NULL;
369 ret = MSG_task_recv(&task, ms->mbox);
372 // An error occured, clean the code and return
373 // The owner did not change, hence the task should be only destroyed on the other side
374 xbt_free(finalize_task_name);
379 if (strcmp(task->name, finalize_task_name) == 0)
382 MSG_task_destroy(task);
388 // Here Stage 1, 2 and 3 have been performed.
389 // Hence complete the migration
391 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
392 // 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)
393 msg_vm_t vm = ms->vm;
394 msg_host_t src_pm = ms->src_pm;
395 msg_host_t dst_pm = ms-> dst_pm;
396 msg_host_priv_t priv = sg_host_msg(vm);
398 // // 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
399 // // I should check with Takahiro in order to make this portion of code atomic
400 // /* deinstall the current affinity setting for the CPU */
401 // simcall_vm_set_affinity(vm, src_pm, 0);
403 // /* Update the vm location */
404 // simcall_vm_migrate(vm, dst_pm);
406 // /* Resume the VM */
407 // simcall_vm_resume(vm);
409 simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
411 /* install the affinity setting of the VM on the destination pm */
414 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));
415 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
416 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
421 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
422 msg_host_priv_t priv = sg_host_msg(vm);
423 priv->is_migrating = 0;
424 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)",
425 sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
426 sg_host_get_name(ms->dst_pm));
427 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
429 // Inform the SRC that the migration has been correctly performed
431 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
432 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
433 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
434 // xbt_assert(ret == MSG_OK);
435 if(ret == MSG_HOST_FAILURE){
436 // 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
437 // TODO What does it mean ? What should we do ?
438 MSG_task_destroy(task);
439 } else if(ret == MSG_TRANSFER_FAILURE){
440 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
441 MSG_task_destroy(task);
448 xbt_free(finalize_task_name);
450 XBT_DEBUG("mig: rx_done");
455 static void reset_dirty_pages(msg_vm_t vm)
457 msg_host_priv_t priv = sg_host_msg(vm);
460 xbt_dict_cursor_t cursor = NULL;
461 dirty_page_t dp = NULL;
462 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
463 double remaining = MSG_task_get_flops_amount(dp->task);
464 dp->prev_clock = MSG_get_clock();
465 dp->prev_remaining = remaining;
467 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
471 static void start_dirty_page_tracking(msg_vm_t vm)
473 msg_host_priv_t priv = sg_host_msg(vm);
474 priv->dp_enabled = 1;
476 reset_dirty_pages(vm);
479 static void stop_dirty_page_tracking(msg_vm_t vm)
481 msg_host_priv_t priv = sg_host_msg(vm);
482 priv->dp_enabled = 0;
485 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
487 double computed = dp->prev_remaining - remaining;
488 double duration = clock - dp->prev_clock;
490 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
491 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
496 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
498 msg_host_priv_t priv = sg_host_msg(vm);
502 xbt_dict_cursor_t cursor = NULL;
503 dirty_page_t dp = NULL;
504 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
505 double remaining = MSG_task_get_flops_amount(dp->task);
507 double clock = MSG_get_clock();
509 // total += calc_updated_pages(key, vm, dp, remaining, clock);
510 total += get_computed(key, vm, dp, remaining, clock);
512 dp->prev_remaining = remaining;
513 dp->prev_clock = clock;
516 total += priv->dp_updated_by_deleted_tasks;
518 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
519 stage_for_fancy_debug,
520 stage2_round_for_fancy_debug,
521 total, priv->dp_updated_by_deleted_tasks);
525 priv->dp_updated_by_deleted_tasks = 0;
531 // TODO Is this code redundant with the information provided by
532 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
533 void MSG_host_add_task(msg_host_t host, msg_task_t task)
535 msg_host_priv_t priv = sg_host_msg(host);
536 double remaining = MSG_task_get_flops_amount(task);
537 char *key = bprintf("%s-%p", task->name, task);
539 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
542 /* It should be okay that we add a task onto a migrating VM. */
543 if (priv->dp_enabled) {
544 dp->prev_clock = MSG_get_clock();
545 dp->prev_remaining = remaining;
548 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
549 xbt_dict_set(priv->dp_objs, key, dp, NULL);
550 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, priv->dp_enabled);
555 void MSG_host_del_task(msg_host_t host, msg_task_t task)
557 msg_host_priv_t priv = sg_host_msg(host);
559 char *key = bprintf("%s-%p", task->name, task);
561 dirty_page_t dp = (dirty_page_t) xbt_dict_get_or_null(priv->dp_objs, key);
562 xbt_assert(dp->task == task);
564 /* If we are in the middle of dirty page tracking, we record how much
565 * computation has been done until now, and keep the information for the
566 * lookup_() function that will called soon. */
567 if (priv->dp_enabled) {
568 double remaining = MSG_task_get_flops_amount(task);
569 double clock = MSG_get_clock();
570 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
571 double updated = get_computed(key, host, dp, remaining, clock);
573 priv->dp_updated_by_deleted_tasks += updated;
576 xbt_dict_remove(priv->dp_objs, key);
579 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
587 static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm,
588 sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double timeout)
591 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
592 msg_task_t task = MSG_task_create(task_name, 0, (double)size, NULL);
596 double clock_sta = MSG_get_clock();
600 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
602 ret = MSG_task_send(task, mbox);
608 } else if (ret == MSG_TIMEOUT) {
609 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
610 sent = size - remaining;
611 XBT_INFO("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu",
612 timeout, remaining, size);
615 /* FIXME: why try-and-catch is used here? */
616 if(ret == MSG_HOST_FAILURE){
617 //XBT_INFO("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
618 MSG_task_destroy(task);
619 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
620 }else if(ret == MSG_TRANSFER_FAILURE){
621 //XBT_INFO("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
622 MSG_task_destroy(task);
623 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
626 double clock_end = MSG_get_clock();
627 double duration = clock_end - clock_sta;
628 double actual_speed = size / duration;
631 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration, actual_speed, mig_speed);
633 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed, mig_speed);
638 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
640 double updated_size = computed * dp_rate;
641 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
642 if (updated_size > dp_cap) {
643 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
644 updated_size = dp_cap;
647 return (sg_size_t) updated_size;
650 static double send_stage1(struct migration_session *ms,
651 sg_size_t ramsize, double mig_speed, double dp_rate, double dp_cap)
654 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
655 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
656 sg_size_t remaining = ramsize;
657 double computed_total = 0;
659 while (remaining > 0) {
660 sg_size_t datasize = chunksize;
661 if (remaining < chunksize)
662 datasize = remaining;
664 remaining -= datasize;
665 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
666 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
667 computed_total += computed;
670 return computed_total;
675 static double get_threshold_value(double bandwidth, double max_downtime)
677 return max_downtime * bandwidth;
680 static int migration_tx_fun(int argc, char *argv[])
682 XBT_DEBUG("mig: tx_start");
684 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
685 migration_session *ms =
686 (migration_session *) MSG_process_get_data(MSG_process_self());
688 s_vm_params_t params;
689 ms->vm->getParams(¶ms);
690 const sg_size_t ramsize = params.ramsize;
691 const sg_size_t devsize = params.devsize;
692 const int skip_stage1 = params.skip_stage1;
693 int skip_stage2 = params.skip_stage2;
694 const double dp_rate = params.dp_rate;
695 const double dp_cap = params.dp_cap;
696 const double mig_speed = params.mig_speed;
697 double max_downtime = params.max_downtime;
699 /* hard code it temporally. Fix Me */
700 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
701 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
703 double remaining_size = (double) (ramsize + devsize);
704 double threshold = 0.0;
706 /* check parameters */
708 XBT_WARN("migrate a VM, but ramsize is zero");
710 if (max_downtime == 0) {
711 XBT_WARN("use the default max_downtime value 30ms");
715 /* Stage1: send all memory pages to the destination. */
716 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
717 start_dirty_page_tracking(ms->vm);
719 double computed_during_stage1 = 0;
721 double clock_prev_send = MSG_get_clock();
724 /* At stage 1, we do not need timeout. We have to send all the memory
725 * pages even though the duration of this tranfer exceeds the timeout
727 XBT_INFO("Stage 1: Gonna send %llu", ramsize);
728 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
729 remaining_size -= sent;
730 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
732 if (sent < ramsize) {
733 XBT_INFO("mig-stage1: timeout, force moving to stage 3");
735 } else if (sent > ramsize)
739 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
740 // Stop the dirty page tracking an return (there is no memory space to release)
741 stop_dirty_page_tracking(ms->vm);
745 double clock_post_send = MSG_get_clock();
746 mig_timeout -= (clock_post_send - clock_prev_send);
747 if (mig_timeout < 0) {
748 XBT_INFO("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
749 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
753 /* estimate bandwidth */
754 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
755 threshold = get_threshold_value(bandwidth, max_downtime);
756 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
760 /* Stage2: send update pages iteratively until the size of remaining states
761 * becomes smaller than the threshold value. */
764 int stage2_round = 0;
767 sg_size_t updated_size = 0;
768 if (stage2_round == 0) {
769 /* just after stage1, nothing has been updated. But, we have to send the
770 * data updated during stage1 */
771 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
773 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
774 updated_size = get_updated_size(computed, dp_rate, dp_cap);
777 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
778 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
781 /* Check whether the remaining size is below the threshold value. If so,
782 * move to stage 3. */
783 remaining_size += updated_size;
784 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
785 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
786 if (remaining_size < threshold)
791 double clock_prev_send = MSG_get_clock();
793 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
794 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed, mig_timeout);
796 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
797 // Stop the dirty page tracking an return (there is no memory space to release)
798 stop_dirty_page_tracking(ms->vm);
801 double clock_post_send = MSG_get_clock();
803 if (sent == updated_size) {
804 /* timeout did not happen */
805 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
806 threshold = get_threshold_value(bandwidth, max_downtime);
807 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
808 remaining_size -= sent;
810 mig_timeout -= (clock_post_send - clock_prev_send);
811 xbt_assert(mig_timeout > 0);
813 } else if (sent < updated_size) {
814 /* When timeout happens, we move to stage 3. The size of memory pages
815 * updated before timeout must be added to the remaining size. */
816 XBT_INFO("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
817 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
818 remaining_size -= sent;
820 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
821 updated_size = get_updated_size(computed, dp_rate, dp_cap);
822 remaining_size += updated_size;
830 /* Stage3: stop the VM and copy the rest of states. */
831 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
832 simcall_vm_suspend(ms->vm);
833 stop_dirty_page_tracking(ms->vm);
836 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
837 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
839 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
840 // Stop the dirty page tracking an return (there is no memory space to release)
841 simcall_vm_resume(ms->vm);
845 // 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.
846 XBT_DEBUG("mig: tx_done");
853 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
855 struct migration_session *ms = xbt_new(struct migration_session, 1);
859 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
860 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
863 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
864 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
866 // msg_process_t tx_process, rx_process;
867 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
868 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
871 char **argv = xbt_new(char *, 2);
872 argv[0] = pr_rx_name;
874 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
877 char **argv = xbt_new(char *, 2);
878 argv[0] = pr_tx_name;
880 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
884 /* wait until the migration have finished or on error has occured */
886 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
887 msg_task_t task = NULL;
888 msg_error_t ret = MSG_TIMEOUT;
889 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
890 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
892 xbt_free(ms->mbox_ctl);
896 //xbt_assert(ret == MSG_OK);
897 if(ret == MSG_HOST_FAILURE){
898 // Note that since the communication failed, the owner did not change and the task should be destroyed on the other side.
899 // Hence, just throw the execption
900 XBT_INFO("SRC crashes, throw an exception (m-control)");
901 //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
904 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){ // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
905 XBT_INFO("DST crashes, throw an exception (m-control)");
910 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
911 xbt_assert(strcmp(task->name, expected_task_name) == 0);
912 xbt_free(expected_task_name);
913 MSG_task_destroy(task);
921 /** @brief Migrate the VM to the given host.
924 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
925 * MSG_task_send() before or after, depending on whether you want to do cold or hot
928 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
931 * - One approach is ...
932 * We first create a new VM (i.e., destination VM) on the destination
933 * physical host. The destination VM will receive the state of the source
934 * VM over network. We will finally destroy the source VM.
935 * - This behavior is similar to the way of migration in the real world.
936 * Even before a migration is completed, we will see a destination VM,
937 * consuming resources.
938 * - We have to relocate all processes. The existing process migraion code
939 * will work for this?
940 * - The name of the VM is a somewhat unique ID in the code. It is tricky
941 * for the destination VM?
943 * - Another one is ...
944 * We update the information of the given VM to place it to the destination
947 * The second one would be easier.
951 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
953 if(MSG_host_is_off(old_pm))
954 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
956 if(MSG_host_is_off(new_pm))
957 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
959 if (!MSG_vm_is_running(vm))
960 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
962 if (MSG_vm_is_migrating(vm))
963 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
965 msg_host_priv_t priv = sg_host_msg(vm);
966 priv->is_migrating = 1;
970 int ret = do_migration(vm, old_pm, new_pm);
972 priv->is_migrating = 0;
973 THROWF(host_error, 0, "SRC host failed during migration");
976 priv->is_migrating = 0;
977 THROWF(host_error, 0, "DST host failed during migration");
981 // 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
982 // In that case, the VM has been already assigned to the DST node.
983 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
984 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
988 /** @brief Immediately suspend the execution of all processes within the given VM.
991 * This function stops the execution of the VM. All the processes on this VM
992 * will pause. The state of the VM is preserved. We can later resume it again.
994 * No suspension cost occurs.
996 void MSG_vm_suspend(msg_vm_t vm)
998 if (MSG_vm_is_migrating(vm))
999 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
1001 simcall_vm_suspend(vm);
1003 XBT_DEBUG("vm_suspend done");
1005 TRACE_msg_vm_suspend(vm);
1009 /** @brief Resume the execution of the VM. All processes on the VM run again.
1012 * No resume cost occurs.
1014 void MSG_vm_resume(msg_vm_t vm)
1016 simcall_vm_resume(vm);
1018 TRACE_msg_vm_resume(vm);
1022 /** @brief Immediately save the execution of all processes within the given VM.
1025 * This function stops the execution of the VM. All the processes on this VM
1026 * will pause. The state of the VM is preserved. We can later resume it again.
1028 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1029 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1030 * of VM save to you.
1032 void MSG_vm_save(msg_vm_t vm)
1034 if (MSG_vm_is_migrating(vm))
1035 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
1037 simcall_vm_save(vm);
1038 TRACE_msg_vm_save(vm);
1041 /** @brief Restore the execution of the VM. All processes on the VM run again.
1044 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1045 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1046 * of VM restore to you.
1048 void MSG_vm_restore(msg_vm_t vm)
1050 simcall_vm_restore(vm);
1052 TRACE_msg_vm_restore(vm);
1056 /** @brief Get the physical host of a given VM.
1059 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1061 return (msg_host_t) simcall_vm_get_pm(vm);
1065 /** @brief Set a CPU bound for a given VM.
1069 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1072 * On PM0, there are Task1 and VM0.
1073 * On VM0, there is Task2.
1074 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1076 * Task1\@PM0 gets 50%.
1077 * Task2\@VM0 gets 25%.
1078 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1080 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1081 * task on the VM does not affect the bound of the dummy CPU action. The bound
1082 * of the dummy CPU action is unlimited.
1084 * There are some solutions for this problem. One option is to update the bound
1085 * of the dummy CPU action automatically. It should be the sum of all tasks on
1086 * the VM. But, this solution might be costly, because we have to scan all tasks
1087 * on the VM in share_resource() or we have to trap both the start and end of
1090 * The current solution is to use MSG_vm_set_bound(), which allows us to
1091 * directly set the bound of the dummy CPU action.
1095 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1096 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1097 * exceeds the capacity of a CPU core.
1099 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1101 simcall_vm_set_bound(vm, bound);
1105 /** @brief Set the CPU affinity of a given VM.
1108 * This function changes the CPU affinity of a given VM. Usage is the same as
1109 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1111 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1113 msg_host_priv_t priv = sg_host_msg(vm);
1116 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1118 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *)(uintptr_t) mask, NULL);
1120 msg_host_t pm_now = MSG_vm_get_pm(vm);
1122 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1123 simcall_vm_set_affinity(vm, pm, mask);
1125 XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));