1 /* Copyright (c) 2012-2014. 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"
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)");
49 return (simcall_host_get_properties(vm));
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_get_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 = msg_host_resource_priv(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 * All parameters are in MBytes
169 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name,
170 int ncpus, int ramsize,
171 int net_cap, char *disk_path, int disksize,
172 int mig_netspeed, int dp_intensity)
174 /* For the moment, intensity_rate is the percentage against the migration
176 double host_speed = MSG_get_host_speed(ind_pm);
177 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
179 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
180 s_ws_params_t params;
181 memset(¶ms, 0, sizeof(params));
182 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
183 //params.overcommit = 0;
185 params.skip_stage2 = 0;
186 params.max_downtime = 0.03;
187 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
188 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
189 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
191 //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);
192 simcall_host_set_params(vm, ¶ms);
198 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
201 * A VM is treated as a host. The name of the VM must be unique among all hosts.
203 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
205 /* make sure the VM of the same name does not exit */
207 xbt_dictelm_t ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
208 if (ind_host_tmp && xbt_lib_get_level(ind_host_tmp, SIMIX_HOST_LEVEL) != NULL) {
209 XBT_ERROR("host %s already exits", name);
214 /* Note: ind_vm and vm_workstation point to the same elm object. */
215 msg_vm_t ind_vm = NULL;
216 void *ind_vm_workstation = NULL;
218 /* Ask the SIMIX layer to create the surf vm resource */
219 ind_vm_workstation = simcall_vm_create(name, ind_pm);
220 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
222 XBT_DEBUG("A new VM (%s) has been created", name);
224 TRACE_msg_vm_create(name, ind_pm);
229 /** @brief Destroy a VM. Destroy the VM object from the simulation.
232 void MSG_vm_destroy(msg_vm_t vm)
234 if (MSG_vm_is_migrating(vm))
235 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
237 /* First, terminate all processes on the VM if necessary */
238 if (MSG_vm_is_running(vm))
239 simcall_vm_shutdown(vm);
241 if (!MSG_vm_is_created(vm)) {
242 XBT_CRITICAL("shutdown the given VM before destroying it");
246 /* Then, destroy the VM object */
247 simcall_vm_destroy(vm);
249 __MSG_host_destroy(vm);
251 TRACE_msg_vm_end(vm);
255 /** @brief Start a vm (i.e., boot the guest operating system)
258 * If the VM cannot be started, an exception is generated.
261 void MSG_vm_start(msg_vm_t vm)
263 simcall_vm_start(vm);
265 TRACE_msg_vm_start(vm);
270 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
273 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
274 * use a #MSG_process_sleep() or something. I'm not quite sure.
276 void MSG_vm_shutdown(msg_vm_t vm)
278 /* msg_vm_t equals to msg_host_t */
279 simcall_vm_shutdown(vm);
281 // TRACE_msg_vm_(vm);
286 /* We have two mailboxes. mbox is used to transfer migration data between
287 * source and destination PMs. mbox_ctl is used to detect the completion of a
288 * migration. The names of these mailboxes must not conflict with others. */
289 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
291 char *vm_name = sg_host_name(vm);
292 char *src_pm_name = sg_host_name(src_pm);
293 char *dst_pm_name = sg_host_name(dst_pm);
295 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
298 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
300 char *vm_name = sg_host_name(vm);
301 char *src_pm_name = sg_host_name(src_pm);
302 char *dst_pm_name = sg_host_name(dst_pm);
304 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
307 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
309 char *vm_name = sg_host_name(vm);
310 char *src_pm_name = sg_host_name(src_pm);
311 char *dst_pm_name = sg_host_name(dst_pm);
313 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
316 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
318 char *vm_name = sg_host_name(vm);
319 char *src_pm_name = sg_host_name(src_pm);
320 char *dst_pm_name = sg_host_name(dst_pm);
322 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
325 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
327 char *vm_name = sg_host_name(vm);
328 char *src_pm_name = sg_host_name(src_pm);
329 char *dst_pm_name = sg_host_name(dst_pm);
331 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
335 struct migration_session {
340 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
341 * know the completion of the migration. */
343 /* The migration_rx and migration_tx processes use mbox to transfer migration
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 = MSG_process_get_data(MSG_process_self());
356 s_ws_params_t params;
357 simcall_host_get_params(ms->vm, ¶ms);
361 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
365 msg_task_t task = NULL;
366 ret = MSG_task_recv(&task, ms->mbox);
369 // An error occured, 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 structure at the begining and free it at the end of each function)
390 msg_vm_t vm = ms->vm;
391 msg_host_t src_pm = ms->src_pm;
392 msg_host_t dst_pm = ms-> dst_pm;
393 msg_host_priv_t priv = msg_host_resource_priv(vm);
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 state
396 // I should check with Takahiro in order to make this portion of code atomic
397 /* deinstall the current affinity setting for the CPU */
398 simcall_vm_set_affinity(vm, src_pm, 0);
400 /* Update the vm location */
401 simcall_vm_migrate(vm, dst_pm);
404 simcall_vm_resume(vm);
406 /* install the affinity setting of the VM on the destination pm */
409 unsigned long affinity_mask = (unsigned long) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *)dst_pm, sizeof(msg_host_t));
410 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
411 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
416 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
417 msg_host_priv_t priv = msg_host_resource_priv(vm);
418 priv->is_migrating = 0;
419 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->key, ms->src_pm->key, ms->dst_pm->key);
420 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
422 // Inform the SRC that the migration has been correctly performed
424 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
425 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
426 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
427 // xbt_assert(ret == MSG_OK);
428 if(ret == MSG_HOST_FAILURE){
429 // 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
430 // TODO What does it mean ? What should we do ?
431 MSG_task_destroy(task);
432 } else if(ret == MSG_TRANSFER_FAILURE){
433 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
434 MSG_task_destroy(task);
441 xbt_free(finalize_task_name);
443 XBT_DEBUG("mig: rx_done");
448 static void reset_dirty_pages(msg_vm_t vm)
450 msg_host_priv_t priv = msg_host_resource_priv(vm);
453 xbt_dict_cursor_t cursor = NULL;
454 dirty_page_t dp = NULL;
455 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
456 double remaining = MSG_task_get_remaining_computation(dp->task);
457 dp->prev_clock = MSG_get_clock();
458 dp->prev_remaining = remaining;
460 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
464 static void start_dirty_page_tracking(msg_vm_t vm)
466 msg_host_priv_t priv = msg_host_resource_priv(vm);
467 priv->dp_enabled = 1;
469 reset_dirty_pages(vm);
472 static void stop_dirty_page_tracking(msg_vm_t vm)
474 msg_host_priv_t priv = msg_host_resource_priv(vm);
475 priv->dp_enabled = 0;
479 /* It might be natural that we define dp_rate for each task. But, we will also
480 * have to care about how each task behavior affects the memory update behavior
481 * at the operating system level. It may not be easy to model it with a simple algorithm. */
482 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
484 double computed = dp->prev_remaining - remaining;
485 double duration = clock - dp->prev_clock;
486 double updated = dp->task->dp_rate * computed;
488 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
489 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
490 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
491 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
497 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
499 double computed = dp->prev_remaining - remaining;
500 double duration = clock - dp->prev_clock;
502 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
503 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
508 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
510 msg_host_priv_t priv = msg_host_resource_priv(vm);
514 xbt_dict_cursor_t cursor = NULL;
515 dirty_page_t dp = NULL;
516 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
517 double remaining = MSG_task_get_remaining_computation(dp->task);
519 double clock = MSG_get_clock();
521 // total += calc_updated_pages(key, vm, dp, remaining, clock);
522 total += get_computed(key, vm, dp, remaining, clock);
524 dp->prev_remaining = remaining;
525 dp->prev_clock = clock;
528 total += priv->dp_updated_by_deleted_tasks;
530 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
531 stage_for_fancy_debug,
532 stage2_round_for_fancy_debug,
533 total, priv->dp_updated_by_deleted_tasks);
537 priv->dp_updated_by_deleted_tasks = 0;
543 // TODO Is this code redundant with the information provided by
544 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
545 void MSG_host_add_task(msg_host_t host, msg_task_t task)
547 msg_host_priv_t priv = msg_host_resource_priv(host);
548 double remaining = MSG_task_get_remaining_computation(task);
549 char *key = bprintf("%s-%p", task->name, task);
551 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
554 /* It should be okay that we add a task onto a migrating VM. */
555 if (priv->dp_enabled) {
556 dp->prev_clock = MSG_get_clock();
557 dp->prev_remaining = remaining;
560 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
561 xbt_dict_set(priv->dp_objs, key, dp, NULL);
562 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
567 void MSG_host_del_task(msg_host_t host, msg_task_t task)
569 msg_host_priv_t priv = msg_host_resource_priv(host);
571 char *key = bprintf("%s-%p", task->name, task);
573 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
574 xbt_assert(dp->task == task);
576 /* If we are in the middle of dirty page tracking, we record how much
577 * computation has been done until now, and keep the information for the
578 * lookup_() function that will called soon. */
579 if (priv->dp_enabled) {
580 double remaining = MSG_task_get_remaining_computation(task);
581 double clock = MSG_get_clock();
582 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
583 double updated = get_computed(key, host, dp, remaining, clock);
585 priv->dp_updated_by_deleted_tasks += updated;
588 xbt_dict_remove(priv->dp_objs, key);
591 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
599 static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm,
600 sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double timeout)
603 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
604 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
608 double clock_sta = MSG_get_clock();
612 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
614 ret = MSG_task_send(task, mbox);
620 } else if (ret == MSG_TIMEOUT) {
621 sg_size_t remaining = MSG_task_get_remaining_communication(task);
622 sent = size - remaining;
623 XBT_INFO("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu",
624 timeout, remaining, size);
627 /* FIXME: why try-and-catch is used here? */
628 if(ret == MSG_HOST_FAILURE){
629 //XBT_INFO("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
630 MSG_task_destroy(task);
631 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
632 }else if(ret == MSG_TRANSFER_FAILURE){
633 //XBT_INFO("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
634 MSG_task_destroy(task);
635 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
638 double clock_end = MSG_get_clock();
639 double duration = clock_end - clock_sta;
640 double actual_speed = size / duration;
643 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration, actual_speed, mig_speed);
645 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed, mig_speed);
650 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
652 double updated_size = computed * dp_rate;
653 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
654 if (updated_size > dp_cap) {
655 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
656 updated_size = dp_cap;
659 return (sg_size_t) updated_size;
662 static double send_stage1(struct migration_session *ms,
663 sg_size_t ramsize, double mig_speed, double dp_rate, double dp_cap)
666 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
667 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
668 sg_size_t remaining = ramsize;
669 double computed_total = 0;
671 while (remaining > 0) {
672 sg_size_t datasize = chunksize;
673 if (remaining < chunksize)
674 datasize = remaining;
676 remaining -= datasize;
677 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
678 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
679 computed_total += computed;
682 return computed_total;
687 static double get_threshold_value(double bandwidth, double max_downtime)
689 return max_downtime * bandwidth;
692 static int migration_tx_fun(int argc, char *argv[])
694 XBT_DEBUG("mig: tx_start");
696 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
697 struct migration_session *ms = MSG_process_get_data(MSG_process_self());
699 s_ws_params_t params;
700 simcall_host_get_params(ms->vm, ¶ms);
701 const sg_size_t ramsize = params.ramsize;
702 const sg_size_t devsize = params.devsize;
703 const int skip_stage1 = params.skip_stage1;
704 int skip_stage2 = params.skip_stage2;
705 const double dp_rate = params.dp_rate;
706 const double dp_cap = params.dp_cap;
707 const double mig_speed = params.mig_speed;
708 double max_downtime = params.max_downtime;
710 /* hard code it temporally. Fix Me */
711 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
712 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
714 double remaining_size = ramsize + devsize;
715 double threshold = 0.0;
717 /* check parameters */
719 XBT_WARN("migrate a VM, but ramsize is zero");
721 if (max_downtime == 0) {
722 XBT_WARN("use the default max_downtime value 30ms");
726 /* Stage1: send all memory pages to the destination. */
727 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
728 start_dirty_page_tracking(ms->vm);
730 double computed_during_stage1 = 0;
732 double clock_prev_send = MSG_get_clock();
735 /* At stage 1, we do not need timeout. We have to send all the memory
736 * pages even though the duration of this tranfer exceeds the timeout
738 XBT_INFO("Stage 1: Gonna send %llu", ramsize);
739 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
740 remaining_size -= sent;
741 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
743 if (sent < ramsize) {
744 XBT_INFO("mig-stage1: timeout, force moving to stage 3");
746 } else if (sent > ramsize)
750 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
751 // Stop the dirty page tracking an return (there is no memory space to release)
752 stop_dirty_page_tracking(ms->vm);
756 double clock_post_send = MSG_get_clock();
757 mig_timeout -= (clock_post_send - clock_prev_send);
758 if (mig_timeout < 0) {
759 XBT_INFO("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
760 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
764 /* estimate bandwidth */
765 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
766 threshold = get_threshold_value(bandwidth, max_downtime);
767 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
771 /* Stage2: send update pages iteratively until the size of remaining states
772 * becomes smaller than the threshold value. */
777 int stage2_round = 0;
780 sg_size_t updated_size = 0;
781 if (stage2_round == 0) {
782 /* just after stage1, nothing has been updated. But, we have to send the
783 * data updated during stage1 */
784 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
786 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
787 updated_size = get_updated_size(computed, dp_rate, dp_cap);
790 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
791 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
794 /* Check whether the remaining size is below the threshold value. If so,
795 * move to stage 3. */
796 remaining_size += updated_size;
797 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
798 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
799 if (remaining_size < threshold)
804 double clock_prev_send = MSG_get_clock();
806 XBT_INFO("Stage 2, gonna send %llu", updated_size);
807 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed, mig_timeout);
809 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
810 // Stop the dirty page tracking an return (there is no memory space to release)
811 stop_dirty_page_tracking(ms->vm);
814 double clock_post_send = MSG_get_clock();
816 if (sent == updated_size) {
817 /* timeout did not happen */
818 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
819 threshold = get_threshold_value(bandwidth, max_downtime);
820 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
821 remaining_size -= sent;
823 mig_timeout -= (clock_post_send - clock_prev_send);
824 xbt_assert(mig_timeout > 0);
826 } else if (sent < updated_size) {
827 /* When timeout happens, we move to stage 3. The size of memory pages
828 * updated before timeout must be added to the remaining size. */
829 XBT_INFO("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
830 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
831 remaining_size -= sent;
833 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
834 updated_size = get_updated_size(computed, dp_rate, dp_cap);
835 remaining_size += updated_size;
844 /* Stage3: stop the VM and copy the rest of states. */
845 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
846 simcall_vm_suspend(ms->vm);
847 stop_dirty_page_tracking(ms->vm);
850 XBT_INFO("Stage 3: Gonna send %f", remaining_size);
851 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, remaining_size, ms->mbox, 3, 0, mig_speed, -1);
853 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
854 // Stop the dirty page tracking an return (there is no memory space to release)
855 simcall_vm_resume(ms->vm);
859 // 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.
860 XBT_INFO("mig: tx_done");
867 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
869 struct migration_session *ms = xbt_new(struct migration_session, 1);
873 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
874 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
877 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
878 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
880 // msg_process_t tx_process, rx_process;
881 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
882 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
885 char **argv = xbt_new(char *, 2);
886 argv[0] = pr_rx_name;
888 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
891 char **argv = xbt_new(char *, 2);
892 argv[0] = pr_tx_name;
894 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
898 /* wait until the migration have finished or on error has occured */
900 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
901 msg_task_t task = NULL;
902 msg_error_t ret = MSG_TIMEOUT;
903 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
904 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
906 xbt_free(ms->mbox_ctl);
910 //xbt_assert(ret == MSG_OK);
911 if(ret == MSG_HOST_FAILURE){
912 // Note that since the communication failed, the owner did not change and the task should be destroyed on the other side.
913 // Hence, just throw the execption
914 XBT_INFO("SRC crashes, throw an exception (m-control)");
915 //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
918 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){ // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
919 XBT_INFO("DST crashes, throw an exception (m-control)");
924 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
925 xbt_assert(strcmp(task->name, expected_task_name) == 0);
926 xbt_free(expected_task_name);
927 MSG_task_destroy(task);
935 /** @brief Migrate the VM to the given host.
938 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
939 * MSG_task_send() before or after, depending on whether you want to do cold or hot
942 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
945 * - One approach is ...
946 * We first create a new VM (i.e., destination VM) on the destination
947 * physical host. The destination VM will receive the state of the source
948 * VM over network. We will finally destroy the source VM.
949 * - This behavior is similar to the way of migration in the real world.
950 * Even before a migration is completed, we will see a destination VM,
951 * consuming resources.
952 * - We have to relocate all processes. The existing process migraion code
953 * will work for this?
954 * - The name of the VM is a somewhat unique ID in the code. It is tricky
955 * for the destination VM?
957 * - Another one is ...
958 * We update the information of the given VM to place it to the destination
961 * The second one would be easier.
965 msg_host_t old_pm = simcall_vm_get_pm(vm);
967 if (!MSG_vm_is_running(vm))
968 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
970 if (MSG_vm_is_migrating(vm))
971 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_name(vm));
973 msg_host_priv_t priv = msg_host_resource_priv(vm);
974 priv->is_migrating = 1;
978 int ret = do_migration(vm, old_pm, new_pm);
980 priv->is_migrating = 0;
981 THROWF(host_error, 0, "SRC host failed during migration");
984 priv->is_migrating = 0;
985 THROWF(host_error, 0, "DST host failed during migration");
989 // 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
990 // In that case, the VM has been already assigned to the DST node.
991 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
992 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
996 /** @brief Immediately suspend the execution of all processes within the given VM.
999 * This function stops the execution of the VM. All the processes on this VM
1000 * will pause. The state of the VM is preserved. We can later resume it again.
1002 * No suspension cost occurs.
1004 void MSG_vm_suspend(msg_vm_t vm)
1006 if (MSG_vm_is_migrating(vm))
1007 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
1009 simcall_vm_suspend(vm);
1011 XBT_DEBUG("vm_suspend done");
1013 TRACE_msg_vm_suspend(vm);
1017 /** @brief Resume the execution of the VM. All processes on the VM run again.
1020 * No resume cost occurs.
1022 void MSG_vm_resume(msg_vm_t vm)
1024 simcall_vm_resume(vm);
1026 TRACE_msg_vm_resume(vm);
1030 /** @brief Immediately save the execution of all processes within the given VM.
1033 * This function stops the execution of the VM. All the processes on this VM
1034 * will pause. The state of the VM is preserved. We can later resume it again.
1036 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1037 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1038 * of VM save to you.
1040 void MSG_vm_save(msg_vm_t vm)
1042 if (MSG_vm_is_migrating(vm))
1043 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
1045 simcall_vm_save(vm);
1046 TRACE_msg_vm_save(vm);
1049 /** @brief Restore the execution of the VM. All processes on the VM run again.
1052 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1053 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1054 * of VM restore to you.
1056 void MSG_vm_restore(msg_vm_t vm)
1058 simcall_vm_restore(vm);
1060 TRACE_msg_vm_restore(vm);
1064 /** @brief Get the physical host of a given VM.
1067 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1069 return simcall_vm_get_pm(vm);
1073 /** @brief Set a CPU bound for a given VM.
1077 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1080 * On PM0, there are Task1 and VM0.
1081 * On VM0, there is Task2.
1082 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1084 * Task1\@PM0 gets 50%.
1085 * Task2\@VM0 gets 25%.
1086 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1088 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1089 * task on the VM does not affect the bound of the dummy CPU action. The bound
1090 * of the dummy CPU action is unlimited.
1092 * There are some solutions for this problem. One option is to update the bound
1093 * of the dummy CPU action automatically. It should be the sum of all tasks on
1094 * the VM. But, this solution might be costly, because we have to scan all tasks
1095 * on the VM in share_resource() or we have to trap both the start and end of
1098 * The current solution is to use MSG_vm_set_bound(), which allows us to
1099 * directly set the bound of the dummy CPU action.
1103 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1104 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1105 * exceeds the capacity of a CPU core.
1107 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1109 return simcall_vm_set_bound(vm, bound);
1113 /** @brief Set the CPU affinity of a given VM.
1116 * This function changes the CPU affinity of a given VM. Usage is the same as
1117 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1119 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1121 msg_host_priv_t priv = msg_host_resource_priv(vm);
1124 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1126 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *) mask, NULL);
1128 msg_host_t pm_now = MSG_vm_get_pm(vm);
1130 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1131 simcall_vm_set_affinity(vm, pm, mask);
1133 XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));