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 experimental code generating micro computation tasks
14 #include <simgrid/s4u/VirtualMachine.hpp>
15 #include <simgrid/s4u/host.hpp>
17 #include "msg_private.h"
18 #include "xbt/sysdep.h"
20 #include "simgrid/host.h"
22 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
25 /* **** ******** GENERAL ********* **** */
26 /** \ingroup m_vm_management
27 * \brief Returns the value of a given vm property
30 * \param name a property name
31 * \return value of a property (or nullptr if property not set)
33 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
35 return MSG_host_get_property_value(vm, name);
38 /** \ingroup m_vm_management
39 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
42 * \return a dict containing the properties
44 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
46 xbt_assert((vm != nullptr), "Invalid parameters (vm is nullptr)");
47 return vm->properties();
50 /** \ingroup m_host_management
51 * \brief Change the value of a given host property
54 * \param name a property name
55 * \param value what to change the property to
56 * \param free_ctn the freeing function to use to kill the value on need
58 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
60 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
62 /** \ingroup m_vm_management
63 * \brief Set the parameters of a given host
66 * \param params a parameter object
68 void MSG_vm_set_params(msg_vm_t vm, vm_params_t params)
70 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(params);
73 /** \ingroup m_vm_management
74 * \brief Get the parameters of a given host
77 * \param params a prameter object
79 void MSG_vm_get_params(msg_vm_t vm, vm_params_t params)
81 static_cast<simgrid::s4u::VirtualMachine*>(vm)->parameters(params);
84 /** \ingroup msg_vm_management
85 * \brief Finds a msg_vm_t using its name.
87 * This is a name directory service
88 * \param name the name of a vm.
89 * \return the corresponding vm
91 * Please note that a VM is a specific host. Hence, you should give a different name for each VM/PM.
93 msg_vm_t MSG_vm_get_by_name(const char *name)
95 return MSG_host_by_name(name);
98 /** \ingroup m_vm_management
100 * \brief Return the name of the #msg_host_t.
102 * This functions checks whether \a host is a valid pointer or not and return its name.
104 const char *MSG_vm_get_name(msg_vm_t vm)
106 return MSG_host_get_name(vm);
109 /* **** Check state of a VM **** */
110 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
112 return simcall_vm_get_state(vm) == state;
115 /** @brief Returns whether the given VM has just created, not running.
118 int MSG_vm_is_created(msg_vm_t vm)
120 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
123 /** @brief Returns whether the given VM is currently running
126 int MSG_vm_is_running(msg_vm_t vm)
128 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
131 /** @brief Returns whether the given VM is currently migrating
134 int MSG_vm_is_migrating(msg_vm_t vm)
136 msg_host_priv_t priv = sg_host_msg(vm);
137 return priv->is_migrating;
140 /** @brief Returns whether the given VM is currently suspended, not running.
143 int MSG_vm_is_suspended(msg_vm_t vm)
145 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
148 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
151 int MSG_vm_is_saving(msg_vm_t vm)
153 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
156 /** @brief Returns whether the given VM has been saved, not running.
159 int MSG_vm_is_saved(msg_vm_t vm)
161 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
164 /** @brief Returns whether the given VM is being restored, not running.
167 int MSG_vm_is_restoring(msg_vm_t vm)
169 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
172 /* **** ******** MSG vm actions ********* **** */
173 /** @brief Create a new VM with specified parameters.
175 * @param pm Physical machine that will host the VM
177 * @param ncpus [TODO]
178 * @param ramsize [TODO]
179 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
180 * @param disk_path (unused) Path to the image that boots
181 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
182 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
183 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
185 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name, int ncpus, int ramsize, int net_cap, char *disk_path,
186 int disksize, int mig_netspeed, int dp_intensity)
188 /* For the moment, intensity_rate is the percentage against the migration
190 double host_speed = MSG_host_get_speed(pm);
191 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
193 msg_vm_t vm = MSG_vm_create_core(pm, name);
194 s_vm_params_t params;
195 memset(¶ms, 0, sizeof(params));
196 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
197 //params.overcommit = 0;
199 params.skip_stage2 = 0;
200 params.max_downtime = 0.03;
201 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
202 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
203 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
205 //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate,
206 // params.mig_speed, dp_intensity, update_speed, host_speed);
207 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(¶ms);
212 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
215 * A VM is treated as a host. The name of the VM must be unique among all hosts.
217 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
219 /* make sure the VM of the same name does not exit */
221 simgrid::s4u::Host* ind_host_tmp = sg_host_by_name(name);
222 if (ind_host_tmp != nullptr && sg_host_simix(ind_host_tmp) != nullptr) {
223 XBT_ERROR("host %s already exits", name);
228 /* Note: ind_vm and vm_workstation point to the same elm object. */
229 /* Ask the SIMIX layer to create the surf vm resource */
230 sg_host_t ind_vm_workstation = simcall_vm_create(name, ind_pm);
232 msg_vm_t ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
234 XBT_DEBUG("A new VM (%s) has been created", name);
236 TRACE_msg_vm_create(name, ind_pm);
241 /** @brief Destroy a VM. Destroy the VM object from the simulation.
244 void MSG_vm_destroy(msg_vm_t vm)
246 if (MSG_vm_is_migrating(vm))
247 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
249 /* First, terminate all processes on the VM if necessary */
250 if (MSG_vm_is_running(vm))
251 simcall_vm_shutdown(vm);
253 if (!MSG_vm_is_created(vm)) {
254 XBT_CRITICAL("shutdown the given VM before destroying it");
258 /* Then, destroy the VM object */
259 simcall_vm_destroy(vm);
261 TRACE_msg_vm_end(vm);
264 /** @brief Start a vm (i.e., boot the guest operating system)
267 * If the VM cannot be started, an exception is generated.
269 void MSG_vm_start(msg_vm_t vm)
271 simcall_vm_start(vm);
273 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 use a #MSG_process_sleep() or something.
280 * 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);
290 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
291 * to detect the completion of a 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);
337 struct migration_session {
342 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
343 * know the completion of the migration. */
345 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
349 static int migration_rx_fun(int argc, char *argv[])
351 XBT_DEBUG("mig: rx_start");
353 // The structure has been created in the do_migration function and should only be freed in the same place ;)
354 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
356 s_vm_params_t params;
357 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
361 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
365 msg_task_t task = nullptr;
366 ret = MSG_task_recv(&task, ms->mbox);
369 // An error occurred, clean the code and return
370 // The owner did not change, hence the task should be only destroyed on the other side
371 xbt_free(finalize_task_name);
376 if (strcmp(task->name, finalize_task_name) == 0)
379 MSG_task_destroy(task);
385 // Here Stage 1, 2 and 3 have been performed.
386 // Hence complete the migration
388 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
389 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
390 // structure at the beginning and free it at the end of each function)
391 msg_vm_t vm = ms->vm;
392 msg_host_t src_pm = ms->src_pm;
393 msg_host_t dst_pm = ms-> dst_pm;
395 // TODO: we have an issue, if the DST node is turning off during the three next calls, then the VM is in an inconsistent
396 // state. I should check with Takahiro in order to make this portion of code atomic
398 // /* Update the vm location */
399 // simcall_vm_migrate(vm, dst_pm);
401 // /* Resume the VM */
402 // simcall_vm_resume(vm);
404 simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
407 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
408 msg_host_priv_t priv = sg_host_msg(vm);
409 priv->is_migrating = 0;
410 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
411 sg_host_get_name(ms->dst_pm));
412 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
414 // Inform the SRC that the migration has been correctly performed
416 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
417 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
418 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
419 // xbt_assert(ret == MSG_OK);
420 if(ret == MSG_HOST_FAILURE){
421 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
422 // has been correctly migrated on the DST node
423 // TODO What does it mean ? What should we do ?
424 MSG_task_destroy(task);
425 } else if(ret == MSG_TRANSFER_FAILURE){
426 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
427 MSG_task_destroy(task);
432 xbt_free(finalize_task_name);
434 XBT_DEBUG("mig: rx_done");
438 static void reset_dirty_pages(msg_vm_t vm)
440 msg_host_priv_t priv = sg_host_msg(vm);
443 xbt_dict_cursor_t cursor = nullptr;
444 dirty_page_t dp = nullptr;
445 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
446 double remaining = MSG_task_get_flops_amount(dp->task);
447 dp->prev_clock = MSG_get_clock();
448 dp->prev_remaining = remaining;
450 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
454 static void start_dirty_page_tracking(msg_vm_t vm)
456 msg_host_priv_t priv = sg_host_msg(vm);
457 priv->dp_enabled = 1;
459 reset_dirty_pages(vm);
462 static void stop_dirty_page_tracking(msg_vm_t vm)
464 msg_host_priv_t priv = sg_host_msg(vm);
465 priv->dp_enabled = 0;
468 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
470 double computed = dp->prev_remaining - remaining;
471 double duration = clock - dp->prev_clock;
473 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
474 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
479 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
481 msg_host_priv_t priv = sg_host_msg(vm);
485 xbt_dict_cursor_t cursor = nullptr;
486 dirty_page_t dp = nullptr;
487 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
488 double remaining = MSG_task_get_flops_amount(dp->task);
490 double clock = MSG_get_clock();
492 // total += calc_updated_pages(key, vm, dp, remaining, clock);
493 total += get_computed(key, vm, dp, remaining, clock);
495 dp->prev_remaining = remaining;
496 dp->prev_clock = clock;
499 total += priv->dp_updated_by_deleted_tasks;
501 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
502 stage_for_fancy_debug, stage2_round_for_fancy_debug, total, priv->dp_updated_by_deleted_tasks);
504 priv->dp_updated_by_deleted_tasks = 0;
509 // TODO Is this code redundant with the information provided by
510 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
511 void MSG_host_add_task(msg_host_t host, msg_task_t task)
513 msg_host_priv_t priv = sg_host_msg(host);
514 double remaining = MSG_task_get_flops_amount(task);
515 char *key = bprintf("%s-%p", task->name, task);
517 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
520 /* It should be okay that we add a task onto a migrating VM. */
521 if (priv->dp_enabled) {
522 dp->prev_clock = MSG_get_clock();
523 dp->prev_remaining = remaining;
526 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == nullptr);
527 xbt_dict_set(priv->dp_objs, key, dp, nullptr);
528 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, priv->dp_enabled);
533 void MSG_host_del_task(msg_host_t host, msg_task_t task)
535 msg_host_priv_t priv = sg_host_msg(host);
537 char *key = bprintf("%s-%p", task->name, task);
539 dirty_page_t dp = (dirty_page_t) xbt_dict_get_or_null(priv->dp_objs, key);
540 xbt_assert(dp->task == task);
542 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
543 * the information for the lookup_() function that will called soon. */
544 if (priv->dp_enabled) {
545 double remaining = MSG_task_get_flops_amount(task);
546 double clock = MSG_get_clock();
547 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
548 double updated = get_computed(key, host, dp, remaining, clock);
550 priv->dp_updated_by_deleted_tasks += updated;
553 xbt_dict_remove(priv->dp_objs, key);
556 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
561 static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, sg_size_t size, char *mbox,
562 int stage, int stage2_round, double mig_speed, double timeout)
565 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
566 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
570 double clock_sta = MSG_get_clock();
574 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
576 ret = MSG_task_send(task, mbox);
582 } else if (ret == MSG_TIMEOUT) {
583 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
584 sent = size - remaining;
585 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
588 /* FIXME: why try-and-catch is used here? */
589 if(ret == MSG_HOST_FAILURE){
590 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
591 MSG_task_destroy(task);
592 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
593 }else if(ret == MSG_TRANSFER_FAILURE){
594 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
595 MSG_task_destroy(task);
596 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
599 double clock_end = MSG_get_clock();
600 double duration = clock_end - clock_sta;
601 double actual_speed = size / duration;
604 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)",
605 stage, stage2_round, size, duration, actual_speed, mig_speed);
607 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)",
608 stage, size, duration, actual_speed, mig_speed);
613 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
615 double updated_size = computed * dp_rate;
616 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
617 if (updated_size > dp_cap) {
618 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
620 updated_size = dp_cap;
623 return (sg_size_t) updated_size;
626 static double send_stage1(struct migration_session *ms, sg_size_t ramsize, double mig_speed, double dp_rate,
629 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
630 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
631 sg_size_t remaining = ramsize;
632 double computed_total = 0;
634 while (remaining > 0) {
635 sg_size_t datasize = chunksize;
636 if (remaining < chunksize)
637 datasize = remaining;
639 remaining -= datasize;
640 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
641 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
642 computed_total += computed;
645 return computed_total;
648 static double get_threshold_value(double bandwidth, double max_downtime)
650 return max_downtime * bandwidth;
653 static int migration_tx_fun(int argc, char *argv[])
655 XBT_DEBUG("mig: tx_start");
657 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
658 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
660 s_vm_params_t params;
661 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
662 const sg_size_t ramsize = params.ramsize;
663 const sg_size_t devsize = params.devsize;
664 const int skip_stage1 = params.skip_stage1;
665 int skip_stage2 = params.skip_stage2;
666 const double dp_rate = params.dp_rate;
667 const double dp_cap = params.dp_cap;
668 const double mig_speed = params.mig_speed;
669 double max_downtime = params.max_downtime;
671 /* hard code it temporally. Fix Me */
672 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
673 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
675 double remaining_size = (double) (ramsize + devsize);
676 double threshold = 0.0;
678 /* check parameters */
680 XBT_WARN("migrate a VM, but ramsize is zero");
682 if (max_downtime <= 0) {
683 XBT_WARN("use the default max_downtime value 30ms");
687 /* Stage1: send all memory pages to the destination. */
688 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
689 start_dirty_page_tracking(ms->vm);
691 double computed_during_stage1 = 0;
693 double clock_prev_send = MSG_get_clock();
696 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
697 * transfer exceeds the timeout value. */
698 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
699 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
700 remaining_size -= sent;
701 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
703 if (sent < ramsize) {
704 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
706 } else if (sent > ramsize)
711 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
712 // Stop the dirty page tracking an return (there is no memory space to release)
713 stop_dirty_page_tracking(ms->vm);
717 double clock_post_send = MSG_get_clock();
718 mig_timeout -= (clock_post_send - clock_prev_send);
719 if (mig_timeout < 0) {
720 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
721 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
725 /* estimate bandwidth */
726 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
727 threshold = get_threshold_value(bandwidth, max_downtime);
728 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
732 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
735 int stage2_round = 0;
738 sg_size_t updated_size = 0;
739 if (stage2_round == 0) {
740 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
741 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
743 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
744 updated_size = get_updated_size(computed, dp_rate, dp_cap);
747 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
748 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
750 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
751 remaining_size += updated_size;
752 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
753 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
754 if (remaining_size < threshold)
758 double clock_prev_send = MSG_get_clock();
760 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
761 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
765 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
766 // Stop the dirty page tracking an return (there is no memory space to release)
767 stop_dirty_page_tracking(ms->vm);
770 double clock_post_send = MSG_get_clock();
772 if (sent == updated_size) {
773 /* timeout did not happen */
774 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
775 threshold = get_threshold_value(bandwidth, max_downtime);
776 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
777 remaining_size -= sent;
779 mig_timeout -= (clock_post_send - clock_prev_send);
780 xbt_assert(mig_timeout > 0);
782 } else if (sent < updated_size) {
783 /* When timeout happens, we move to stage 3. The size of memory pages
784 * updated before timeout must be added to the remaining size. */
785 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
786 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
787 remaining_size -= sent;
789 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
790 updated_size = get_updated_size(computed, dp_rate, dp_cap);
791 remaining_size += updated_size;
798 /* Stage3: stop the VM and copy the rest of states. */
799 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
800 simcall_vm_suspend(ms->vm);
801 stop_dirty_page_tracking(ms->vm);
804 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
805 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
808 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
809 // Stop the dirty page tracking an return (there is no memory space to release)
810 simcall_vm_resume(ms->vm);
814 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
815 // effectively the VM on the DST node.
816 XBT_DEBUG("mig: tx_done");
821 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
823 struct migration_session *ms = xbt_new(struct migration_session, 1);
827 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
828 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
830 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
831 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
833 // msg_process_t tx_process, rx_process;
834 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
835 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
838 char **argv = xbt_new(char *, 2);
839 argv[0] = pr_rx_name;
841 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
844 char **argv = xbt_new(char *, 2);
845 argv[0] = pr_tx_name;
847 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
851 /* wait until the migration have finished or on error has occurred */
853 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
854 msg_task_t task = nullptr;
855 msg_error_t ret = MSG_TIMEOUT;
856 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
857 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
859 xbt_free(ms->mbox_ctl);
863 //xbt_assert(ret == MSG_OK);
864 if(ret == MSG_HOST_FAILURE){
865 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
866 // other side. Hence, just throw the execption
867 XBT_ERROR("SRC crashes, throw an exception (m-control)");
868 //MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
872 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){
873 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
874 XBT_ERROR("DST crashes, throw an exception (m-control)");
878 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
879 xbt_assert(strcmp(task->name, expected_task_name) == 0);
880 xbt_free(expected_task_name);
881 MSG_task_destroy(task);
886 /** @brief Migrate the VM to the given host.
889 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
890 * depending on whether you want to do cold or hot migration.
892 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
895 * - One approach is ...
896 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
897 * receive the state of the source
898 * VM over network. We will finally destroy the source VM.
899 * - This behavior is similar to the way of migration in the real world.
900 * Even before a migration is completed, we will see a destination VM, consuming resources.
901 * - We have to relocate all processes. The existing process migration code will work for this?
902 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
904 * - Another one is ...
905 * We update the information of the given VM to place it to the destination physical host.
907 * The second one would be easier.
910 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
912 if(MSG_host_is_off(old_pm))
913 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
915 if(MSG_host_is_off(new_pm))
916 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
918 if (!MSG_vm_is_running(vm))
919 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
921 if (MSG_vm_is_migrating(vm))
922 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
924 msg_host_priv_t priv = sg_host_msg(vm);
925 priv->is_migrating = 1;
928 int ret = do_migration(vm, old_pm, new_pm);
930 priv->is_migrating = 0;
931 THROWF(host_error, 0, "SRC host failed during migration");
934 priv->is_migrating = 0;
935 THROWF(host_error, 0, "DST host failed during migration");
939 // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration
940 // process. In that case, the VM has been already assigned to the DST node.
941 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
942 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
945 /** @brief Immediately suspend the execution of all processes within the given VM.
948 * This function stops the execution of the VM. All the processes on this VM
949 * will pause. The state of the VM is preserved. We can later resume it again.
951 * No suspension cost occurs.
953 void MSG_vm_suspend(msg_vm_t vm)
955 if (MSG_vm_is_migrating(vm))
956 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
958 simcall_vm_suspend(vm);
960 XBT_DEBUG("vm_suspend done");
962 TRACE_msg_vm_suspend(vm);
965 /** @brief Resume the execution of the VM. All processes on the VM run again.
968 * No resume cost occurs.
970 void MSG_vm_resume(msg_vm_t vm)
972 simcall_vm_resume(vm);
974 TRACE_msg_vm_resume(vm);
978 /** @brief Immediately save the execution of all processes within the given VM.
981 * This function stops the execution of the VM. All the processes on this VM
982 * will pause. The state of the VM is preserved. We can later resume it again.
984 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
985 * or after, depending on the exact semantic of VM save to you.
987 void MSG_vm_save(msg_vm_t vm)
989 if (MSG_vm_is_migrating(vm))
990 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
993 TRACE_msg_vm_save(vm);
996 /** @brief Restore the execution of the VM. All processes on the VM run again.
999 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
1000 * after, depending on the exact semantic of VM restore to you.
1002 void MSG_vm_restore(msg_vm_t vm)
1004 simcall_vm_restore(vm);
1006 TRACE_msg_vm_restore(vm);
1009 /** @brief Get the physical host of a given VM.
1012 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1014 return (msg_host_t) simcall_vm_get_pm(vm);
1017 /** @brief Set a CPU bound for a given VM.
1020 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1023 * On PM0, there are Task1 and VM0.
1024 * On VM0, there is Task2.
1025 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1027 * Task1\@PM0 gets 50%.
1028 * Task2\@VM0 gets 25%.
1029 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1031 * This is because a VM has the dummy CPU action in the PM layer. Putting a task on the VM does not affect the bound of
1032 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
1034 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
1035 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
1036 * on the VM in share_resource() or we have to trap both the start and end of task execution.
1038 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
1040 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
1041 * computation task (or a VM) never exceeds the capacity of a CPU core.
1043 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1045 simcall_vm_set_bound(vm, bound);