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/host.h"
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
21 "Cloud-oriented parts of the MSG API");
24 /* **** ******** 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 NULL if property not set)
34 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
36 return MSG_host_get_property_value(vm, name);
39 /** \ingroup m_vm_management
40 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
43 * \return a dict containing the properties
45 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
47 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
48 return vm->properties();
51 /** \ingroup m_host_management
52 * \brief Change the value of a given host property
55 * \param name a property name
56 * \param value what to change the property to
57 * \param free_ctn the freeing function to use to kill the value on need
59 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
61 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
64 /** \ingroup msg_vm_management
65 * \brief Finds a msg_vm_t using its name.
67 * This is a name directory service
68 * \param name the name of a vm.
69 * \return the corresponding vm
71 * Please note that a VM is a specific host. Hence, you should give a different name
75 msg_vm_t MSG_vm_get_by_name(const char *name)
77 return MSG_host_by_name(name);
80 /** \ingroup m_vm_management
82 * \brief Return the name of the #msg_host_t.
84 * This functions checks whether \a host is a valid pointer or not and return
87 const char *MSG_vm_get_name(msg_vm_t vm)
89 return MSG_host_get_name(vm);
93 /* **** Check state of a VM **** */
94 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
96 return simcall_vm_get_state(vm) == state;
99 /** @brief Returns whether the given VM has just created, not running.
102 int MSG_vm_is_created(msg_vm_t vm)
104 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
107 /** @brief Returns whether the given VM is currently running
110 int MSG_vm_is_running(msg_vm_t vm)
112 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
115 /** @brief Returns whether the given VM is currently migrating
118 int MSG_vm_is_migrating(msg_vm_t vm)
120 msg_host_priv_t priv = sg_host_msg(vm);
121 return priv->is_migrating;
124 /** @brief Returns whether the given VM is currently suspended, not running.
127 int MSG_vm_is_suspended(msg_vm_t vm)
129 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
132 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
135 int MSG_vm_is_saving(msg_vm_t vm)
137 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
140 /** @brief Returns whether the given VM has been saved, not running.
143 int MSG_vm_is_saved(msg_vm_t vm)
145 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
148 /** @brief Returns whether the given VM is being restored, not running.
151 int MSG_vm_is_restoring(msg_vm_t vm)
153 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
158 /* ------------------------------------------------------------------------- */
159 /* ------------------------------------------------------------------------- */
161 /* **** ******** MSG vm actions ********* **** */
163 /** @brief Create a new VM with specified parameters.
165 * @param pm Physical machine that will host the VM
167 * @param ncpus [TODO]
168 * @param ramsize [TODO]
169 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
170 * @param disk_path (unused) Path to the image that boots
171 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
172 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
173 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
176 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name,
177 int ncpus, int ramsize,
178 int net_cap, char *disk_path, int disksize,
179 int mig_netspeed, int dp_intensity)
181 /* For the moment, intensity_rate is the percentage against the migration
183 double host_speed = MSG_host_get_speed(pm);
184 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
186 msg_vm_t vm = MSG_vm_create_core(pm, name);
187 s_vm_params_t params;
188 memset(¶ms, 0, sizeof(params));
189 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
190 //params.overcommit = 0;
192 params.skip_stage2 = 0;
193 params.max_downtime = 0.03;
194 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
195 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
196 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
198 //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);
199 vm->setParameters(¶ms);
205 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
208 * A VM is treated as a host. The name of the VM must be unique among all hosts.
210 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
212 /* make sure the VM of the same name does not exit */
214 simgrid::s4u::Host* ind_host_tmp =
215 (simgrid::s4u::Host*) xbt_dict_get_or_null(host_list, name);
216 if (ind_host_tmp != nullptr && sg_host_simix(ind_host_tmp) != nullptr) {
217 XBT_ERROR("host %s already exits", name);
222 /* Note: ind_vm and vm_workstation point to the same elm object. */
224 /* Ask the SIMIX layer to create the surf vm resource */
225 sg_host_t ind_vm_workstation = (sg_host_t) simcall_vm_create(name, ind_pm);
227 msg_vm_t ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
229 XBT_DEBUG("A new VM (%s) has been created", name);
231 TRACE_msg_vm_create(name, ind_pm);
236 /** @brief Destroy a VM. Destroy the VM object from the simulation.
239 void MSG_vm_destroy(msg_vm_t vm)
241 if (MSG_vm_is_migrating(vm))
242 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
244 /* First, terminate all processes on the VM if necessary */
245 if (MSG_vm_is_running(vm))
246 simcall_vm_shutdown(vm);
248 if (!MSG_vm_is_created(vm)) {
249 XBT_CRITICAL("shutdown the given VM before destroying it");
253 /* Then, destroy the VM object */
254 simcall_vm_destroy(vm);
256 TRACE_msg_vm_end(vm);
260 /** @brief Start a vm (i.e., boot the guest operating system)
263 * If the VM cannot be started, an exception is generated.
266 void MSG_vm_start(msg_vm_t vm)
268 simcall_vm_start(vm);
270 TRACE_msg_vm_start(vm);
275 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
278 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
279 * use a #MSG_process_sleep() or something. I'm not quite sure.
281 void MSG_vm_shutdown(msg_vm_t vm)
283 /* msg_vm_t equals to msg_host_t */
284 simcall_vm_shutdown(vm);
286 // TRACE_msg_vm_(vm);
291 /* We have two mailboxes. mbox is used to transfer migration data between
292 * source and destination PMs. mbox_ctl is used to detect the completion of a
293 * migration. The names of these mailboxes must not conflict with others. */
294 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
296 const char *vm_name = sg_host_get_name(vm);
297 const char *src_pm_name = sg_host_get_name(src_pm);
298 const char *dst_pm_name = sg_host_get_name(dst_pm);
300 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
303 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
305 const char *vm_name = sg_host_get_name(vm);
306 const char *src_pm_name = sg_host_get_name(src_pm);
307 const char *dst_pm_name = sg_host_get_name(dst_pm);
309 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
312 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
314 const char *vm_name = sg_host_get_name(vm);
315 const char *src_pm_name = sg_host_get_name(src_pm);
316 const char *dst_pm_name = sg_host_get_name(dst_pm);
318 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
321 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
323 const char *vm_name = sg_host_get_name(vm);
324 const char *src_pm_name = sg_host_get_name(src_pm);
325 const char *dst_pm_name = sg_host_get_name(dst_pm);
327 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
330 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
332 const char *vm_name = sg_host_get_name(vm);
333 const char *src_pm_name = sg_host_get_name(src_pm);
334 const char *dst_pm_name = sg_host_get_name(dst_pm);
336 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
340 struct migration_session {
345 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
346 * know the completion of the migration. */
348 /* The migration_rx and migration_tx processes use mbox to transfer migration
354 static int migration_rx_fun(int argc, char *argv[])
356 XBT_DEBUG("mig: rx_start");
358 // The structure has been created in the do_migration function and should only be freed in the same place ;)
359 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
361 s_vm_params_t params;
362 ms->vm->parameters(¶ms);
366 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
370 msg_task_t task = NULL;
371 ret = MSG_task_recv(&task, ms->mbox);
374 // An error occured, clean the code and return
375 // The owner did not change, hence the task should be only destroyed on the other side
376 xbt_free(finalize_task_name);
381 if (strcmp(task->name, finalize_task_name) == 0)
384 MSG_task_destroy(task);
390 // Here Stage 1, 2 and 3 have been performed.
391 // Hence complete the migration
393 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
394 // 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)
395 msg_vm_t vm = ms->vm;
396 msg_host_t src_pm = ms->src_pm;
397 msg_host_t dst_pm = ms-> dst_pm;
398 msg_host_priv_t priv = sg_host_msg(vm);
400 // // 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
401 // // I should check with Takahiro in order to make this portion of code atomic
402 // /* deinstall the current affinity setting for the CPU */
403 // simcall_vm_set_affinity(vm, src_pm, 0);
405 // /* Update the vm location */
406 // simcall_vm_migrate(vm, dst_pm);
408 // /* Resume the VM */
409 // simcall_vm_resume(vm);
411 simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
413 /* install the affinity setting of the VM on the destination pm */
416 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));
417 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
418 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
423 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
424 msg_host_priv_t priv = sg_host_msg(vm);
425 priv->is_migrating = 0;
426 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)",
427 sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
428 sg_host_get_name(ms->dst_pm));
429 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
431 // Inform the SRC that the migration has been correctly performed
433 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
434 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
435 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
436 // xbt_assert(ret == MSG_OK);
437 if(ret == MSG_HOST_FAILURE){
438 // 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
439 // TODO What does it mean ? What should we do ?
440 MSG_task_destroy(task);
441 } else if(ret == MSG_TRANSFER_FAILURE){
442 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
443 MSG_task_destroy(task);
450 xbt_free(finalize_task_name);
452 XBT_DEBUG("mig: rx_done");
457 static void reset_dirty_pages(msg_vm_t vm)
459 msg_host_priv_t priv = sg_host_msg(vm);
462 xbt_dict_cursor_t cursor = NULL;
463 dirty_page_t dp = NULL;
464 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
465 double remaining = MSG_task_get_flops_amount(dp->task);
466 dp->prev_clock = MSG_get_clock();
467 dp->prev_remaining = remaining;
469 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
473 static void start_dirty_page_tracking(msg_vm_t vm)
475 msg_host_priv_t priv = sg_host_msg(vm);
476 priv->dp_enabled = 1;
478 reset_dirty_pages(vm);
481 static void stop_dirty_page_tracking(msg_vm_t vm)
483 msg_host_priv_t priv = sg_host_msg(vm);
484 priv->dp_enabled = 0;
487 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
489 double computed = dp->prev_remaining - remaining;
490 double duration = clock - dp->prev_clock;
492 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
493 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
498 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
500 msg_host_priv_t priv = sg_host_msg(vm);
504 xbt_dict_cursor_t cursor = NULL;
505 dirty_page_t dp = NULL;
506 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
507 double remaining = MSG_task_get_flops_amount(dp->task);
509 double clock = MSG_get_clock();
511 // total += calc_updated_pages(key, vm, dp, remaining, clock);
512 total += get_computed(key, vm, dp, remaining, clock);
514 dp->prev_remaining = remaining;
515 dp->prev_clock = clock;
518 total += priv->dp_updated_by_deleted_tasks;
520 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
521 stage_for_fancy_debug,
522 stage2_round_for_fancy_debug,
523 total, priv->dp_updated_by_deleted_tasks);
527 priv->dp_updated_by_deleted_tasks = 0;
533 // TODO Is this code redundant with the information provided by
534 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
535 void MSG_host_add_task(msg_host_t host, msg_task_t task)
537 msg_host_priv_t priv = sg_host_msg(host);
538 double remaining = MSG_task_get_flops_amount(task);
539 char *key = bprintf("%s-%p", task->name, task);
541 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
544 /* It should be okay that we add a task onto a migrating VM. */
545 if (priv->dp_enabled) {
546 dp->prev_clock = MSG_get_clock();
547 dp->prev_remaining = remaining;
550 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
551 xbt_dict_set(priv->dp_objs, key, dp, NULL);
552 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, priv->dp_enabled);
557 void MSG_host_del_task(msg_host_t host, msg_task_t task)
559 msg_host_priv_t priv = sg_host_msg(host);
561 char *key = bprintf("%s-%p", task->name, task);
563 dirty_page_t dp = (dirty_page_t) xbt_dict_get_or_null(priv->dp_objs, key);
564 xbt_assert(dp->task == task);
566 /* If we are in the middle of dirty page tracking, we record how much
567 * computation has been done until now, and keep the information for the
568 * lookup_() function that will called soon. */
569 if (priv->dp_enabled) {
570 double remaining = MSG_task_get_flops_amount(task);
571 double clock = MSG_get_clock();
572 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
573 double updated = get_computed(key, host, dp, remaining, clock);
575 priv->dp_updated_by_deleted_tasks += updated;
578 xbt_dict_remove(priv->dp_objs, key);
581 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
589 static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm,
590 sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double timeout)
593 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
594 msg_task_t task = MSG_task_create(task_name, 0, (double)size, NULL);
598 double clock_sta = MSG_get_clock();
602 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
604 ret = MSG_task_send(task, mbox);
610 } else if (ret == MSG_TIMEOUT) {
611 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
612 sent = size - remaining;
613 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu",
614 timeout, remaining, size);
617 /* FIXME: why try-and-catch is used here? */
618 if(ret == MSG_HOST_FAILURE){
619 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
620 MSG_task_destroy(task);
621 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
622 }else if(ret == MSG_TRANSFER_FAILURE){
623 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
624 MSG_task_destroy(task);
625 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
628 double clock_end = MSG_get_clock();
629 double duration = clock_end - clock_sta;
630 double actual_speed = size / duration;
633 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration, actual_speed, mig_speed);
635 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed, mig_speed);
640 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
642 double updated_size = computed * dp_rate;
643 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
644 if (updated_size > dp_cap) {
645 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
646 updated_size = dp_cap;
649 return (sg_size_t) updated_size;
652 static double send_stage1(struct migration_session *ms,
653 sg_size_t ramsize, double mig_speed, double dp_rate, double dp_cap)
656 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
657 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
658 sg_size_t remaining = ramsize;
659 double computed_total = 0;
661 while (remaining > 0) {
662 sg_size_t datasize = chunksize;
663 if (remaining < chunksize)
664 datasize = remaining;
666 remaining -= datasize;
667 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
668 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
669 computed_total += computed;
672 return computed_total;
677 static double get_threshold_value(double bandwidth, double max_downtime)
679 return max_downtime * bandwidth;
682 static int migration_tx_fun(int argc, char *argv[])
684 XBT_DEBUG("mig: tx_start");
686 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
687 migration_session *ms =
688 (migration_session *) MSG_process_get_data(MSG_process_self());
690 s_vm_params_t params;
691 ms->vm->parameters(¶ms);
692 const sg_size_t ramsize = params.ramsize;
693 const sg_size_t devsize = params.devsize;
694 const int skip_stage1 = params.skip_stage1;
695 int skip_stage2 = params.skip_stage2;
696 const double dp_rate = params.dp_rate;
697 const double dp_cap = params.dp_cap;
698 const double mig_speed = params.mig_speed;
699 double max_downtime = params.max_downtime;
701 /* hard code it temporally. Fix Me */
702 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
703 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
705 double remaining_size = (double) (ramsize + devsize);
706 double threshold = 0.0;
708 /* check parameters */
710 XBT_WARN("migrate a VM, but ramsize is zero");
712 if (max_downtime == 0) {
713 XBT_WARN("use the default max_downtime value 30ms");
717 /* Stage1: send all memory pages to the destination. */
718 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
719 start_dirty_page_tracking(ms->vm);
721 double computed_during_stage1 = 0;
723 double clock_prev_send = MSG_get_clock();
726 /* At stage 1, we do not need timeout. We have to send all the memory
727 * pages even though the duration of this tranfer exceeds the timeout
729 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
730 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
731 remaining_size -= sent;
732 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
734 if (sent < ramsize) {
735 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
737 } else if (sent > ramsize)
741 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
742 // Stop the dirty page tracking an return (there is no memory space to release)
743 stop_dirty_page_tracking(ms->vm);
747 double clock_post_send = MSG_get_clock();
748 mig_timeout -= (clock_post_send - clock_prev_send);
749 if (mig_timeout < 0) {
750 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
751 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
755 /* estimate bandwidth */
756 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
757 threshold = get_threshold_value(bandwidth, max_downtime);
758 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
762 /* Stage2: send update pages iteratively until the size of remaining states
763 * becomes smaller than the threshold value. */
766 int stage2_round = 0;
769 sg_size_t updated_size = 0;
770 if (stage2_round == 0) {
771 /* just after stage1, nothing has been updated. But, we have to send the
772 * data updated during stage1 */
773 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
775 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
776 updated_size = get_updated_size(computed, dp_rate, dp_cap);
779 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
780 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
783 /* Check whether the remaining size is below the threshold value. If so,
784 * move to stage 3. */
785 remaining_size += updated_size;
786 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
787 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
788 if (remaining_size < threshold)
793 double clock_prev_send = MSG_get_clock();
795 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
796 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed, mig_timeout);
798 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
799 // Stop the dirty page tracking an return (there is no memory space to release)
800 stop_dirty_page_tracking(ms->vm);
803 double clock_post_send = MSG_get_clock();
805 if (sent == updated_size) {
806 /* timeout did not happen */
807 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
808 threshold = get_threshold_value(bandwidth, max_downtime);
809 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
810 remaining_size -= sent;
812 mig_timeout -= (clock_post_send - clock_prev_send);
813 xbt_assert(mig_timeout > 0);
815 } else if (sent < updated_size) {
816 /* When timeout happens, we move to stage 3. The size of memory pages
817 * updated before timeout must be added to the remaining size. */
818 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
819 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
820 remaining_size -= sent;
822 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
823 updated_size = get_updated_size(computed, dp_rate, dp_cap);
824 remaining_size += updated_size;
832 /* Stage3: stop the VM and copy the rest of states. */
833 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
834 simcall_vm_suspend(ms->vm);
835 stop_dirty_page_tracking(ms->vm);
838 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
839 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
841 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
842 // Stop the dirty page tracking an return (there is no memory space to release)
843 simcall_vm_resume(ms->vm);
847 // 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.
848 XBT_DEBUG("mig: tx_done");
855 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
857 struct migration_session *ms = xbt_new(struct migration_session, 1);
861 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
862 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
865 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
866 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
868 // msg_process_t tx_process, rx_process;
869 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
870 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
873 char **argv = xbt_new(char *, 2);
874 argv[0] = pr_rx_name;
876 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
879 char **argv = xbt_new(char *, 2);
880 argv[0] = pr_tx_name;
882 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
886 /* wait until the migration have finished or on error has occured */
888 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
889 msg_task_t task = NULL;
890 msg_error_t ret = MSG_TIMEOUT;
891 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
892 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
894 xbt_free(ms->mbox_ctl);
898 //xbt_assert(ret == MSG_OK);
899 if(ret == MSG_HOST_FAILURE){
900 // Note that since the communication failed, the owner did not change and the task should be destroyed on the other side.
901 // Hence, just throw the execption
902 XBT_ERROR("SRC crashes, throw an exception (m-control)");
903 //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
906 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){ // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
907 XBT_ERROR("DST crashes, throw an exception (m-control)");
912 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
913 xbt_assert(strcmp(task->name, expected_task_name) == 0);
914 xbt_free(expected_task_name);
915 MSG_task_destroy(task);
923 /** @brief Migrate the VM to the given host.
926 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
927 * MSG_task_send() before or after, depending on whether you want to do cold or hot
930 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
933 * - One approach is ...
934 * We first create a new VM (i.e., destination VM) on the destination
935 * physical host. The destination VM will receive the state of the source
936 * VM over network. We will finally destroy the source VM.
937 * - This behavior is similar to the way of migration in the real world.
938 * Even before a migration is completed, we will see a destination VM,
939 * consuming resources.
940 * - We have to relocate all processes. The existing process migraion code
941 * will work for this?
942 * - The name of the VM is a somewhat unique ID in the code. It is tricky
943 * for the destination VM?
945 * - Another one is ...
946 * We update the information of the given VM to place it to the destination
949 * The second one would be easier.
953 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
955 if(MSG_host_is_off(old_pm))
956 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
958 if(MSG_host_is_off(new_pm))
959 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
961 if (!MSG_vm_is_running(vm))
962 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
964 if (MSG_vm_is_migrating(vm))
965 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
967 msg_host_priv_t priv = sg_host_msg(vm);
968 priv->is_migrating = 1;
972 int ret = do_migration(vm, old_pm, new_pm);
974 priv->is_migrating = 0;
975 THROWF(host_error, 0, "SRC host failed during migration");
978 priv->is_migrating = 0;
979 THROWF(host_error, 0, "DST host failed during migration");
983 // 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
984 // In that case, the VM has been already assigned to the DST node.
985 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
986 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
990 /** @brief Immediately suspend the execution of all processes within the given VM.
993 * This function stops the execution of the VM. All the processes on this VM
994 * will pause. The state of the VM is preserved. We can later resume it again.
996 * No suspension cost occurs.
998 void MSG_vm_suspend(msg_vm_t vm)
1000 if (MSG_vm_is_migrating(vm))
1001 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
1003 simcall_vm_suspend(vm);
1005 XBT_DEBUG("vm_suspend done");
1007 TRACE_msg_vm_suspend(vm);
1011 /** @brief Resume the execution of the VM. All processes on the VM run again.
1014 * No resume cost occurs.
1016 void MSG_vm_resume(msg_vm_t vm)
1018 simcall_vm_resume(vm);
1020 TRACE_msg_vm_resume(vm);
1024 /** @brief Immediately save the execution of all processes within the given VM.
1027 * This function stops the execution of the VM. All the processes on this VM
1028 * will pause. The state of the VM is preserved. We can later resume it again.
1030 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1031 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1032 * of VM save to you.
1034 void MSG_vm_save(msg_vm_t vm)
1036 if (MSG_vm_is_migrating(vm))
1037 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
1039 simcall_vm_save(vm);
1040 TRACE_msg_vm_save(vm);
1043 /** @brief Restore the execution of the VM. All processes on the VM run again.
1046 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1047 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1048 * of VM restore to you.
1050 void MSG_vm_restore(msg_vm_t vm)
1052 simcall_vm_restore(vm);
1054 TRACE_msg_vm_restore(vm);
1058 /** @brief Get the physical host of a given VM.
1061 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1063 return (msg_host_t) simcall_vm_get_pm(vm);
1067 /** @brief Set a CPU bound for a given VM.
1071 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1074 * On PM0, there are Task1 and VM0.
1075 * On VM0, there is Task2.
1076 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1078 * Task1\@PM0 gets 50%.
1079 * Task2\@VM0 gets 25%.
1080 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1082 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1083 * task on the VM does not affect the bound of the dummy CPU action. The bound
1084 * of the dummy CPU action is unlimited.
1086 * There are some solutions for this problem. One option is to update the bound
1087 * of the dummy CPU action automatically. It should be the sum of all tasks on
1088 * the VM. But, this solution might be costly, because we have to scan all tasks
1089 * on the VM in share_resource() or we have to trap both the start and end of
1092 * The current solution is to use MSG_vm_set_bound(), which allows us to
1093 * directly set the bound of the dummy CPU action.
1097 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1098 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1099 * exceeds the capacity of a CPU core.
1101 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1103 simcall_vm_set_bound(vm, bound);
1107 /** @brief Set the CPU affinity of a given VM.
1110 * This function changes the CPU affinity of a given VM. Usage is the same as
1111 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1113 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1115 msg_host_priv_t priv = sg_host_msg(vm);
1118 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1120 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *)(uintptr_t) mask, NULL);
1122 msg_host_t pm_now = MSG_vm_get_pm(vm);
1124 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1125 simcall_vm_set_affinity(vm, pm, mask);
1127 XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));