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
12 #include "msg_private.h"
13 #include "xbt/sysdep.h"
15 #include "simgrid/host.h"
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
20 /* **** ******** GENERAL ********* **** */
21 /** \ingroup m_vm_management
22 * \brief Returns the value of a given vm property
25 * \param name a property name
26 * \return value of a property (or NULL if property not set)
28 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
30 return MSG_host_get_property_value(vm, name);
33 /** \ingroup m_vm_management
34 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
37 * \return a dict containing the properties
39 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
41 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
42 return vm->properties();
45 /** \ingroup m_host_management
46 * \brief Change the value of a given host property
49 * \param name a property name
50 * \param value what to change the property to
51 * \param free_ctn the freeing function to use to kill the value on need
53 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
55 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
58 /** \ingroup msg_vm_management
59 * \brief Finds a msg_vm_t using its name.
61 * This is a name directory service
62 * \param name the name of a vm.
63 * \return the corresponding vm
65 * Please note that a VM is a specific host. Hence, you should give a different name for each VM/PM.
67 msg_vm_t MSG_vm_get_by_name(const char *name)
69 return MSG_host_by_name(name);
72 /** \ingroup m_vm_management
74 * \brief Return the name of the #msg_host_t.
76 * This functions checks whether \a host is a valid pointer or not and return its name.
78 const char *MSG_vm_get_name(msg_vm_t vm)
80 return MSG_host_get_name(vm);
83 /* **** Check state of a VM **** */
84 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
86 return simcall_vm_get_state(vm) == state;
89 /** @brief Returns whether the given VM has just created, not running.
92 int MSG_vm_is_created(msg_vm_t vm)
94 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
97 /** @brief Returns whether the given VM is currently running
100 int MSG_vm_is_running(msg_vm_t vm)
102 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
105 /** @brief Returns whether the given VM is currently migrating
108 int MSG_vm_is_migrating(msg_vm_t vm)
110 msg_host_priv_t priv = sg_host_msg(vm);
111 return priv->is_migrating;
114 /** @brief Returns whether the given VM is currently suspended, not running.
117 int MSG_vm_is_suspended(msg_vm_t vm)
119 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
122 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
125 int MSG_vm_is_saving(msg_vm_t vm)
127 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
130 /** @brief Returns whether the given VM has been saved, not running.
133 int MSG_vm_is_saved(msg_vm_t vm)
135 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
138 /** @brief Returns whether the given VM is being restored, not running.
141 int MSG_vm_is_restoring(msg_vm_t vm)
143 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
146 /* **** ******** MSG vm actions ********* **** */
147 /** @brief Create a new VM with specified parameters.
149 * @param pm Physical machine that will host the VM
151 * @param ncpus [TODO]
152 * @param ramsize [TODO]
153 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
154 * @param disk_path (unused) Path to the image that boots
155 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
156 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
157 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
159 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name, int ncpus, int ramsize, int net_cap, char *disk_path,
160 int disksize, int mig_netspeed, int dp_intensity)
162 /* For the moment, intensity_rate is the percentage against the migration
164 double host_speed = MSG_host_get_speed(pm);
165 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
167 msg_vm_t vm = MSG_vm_create_core(pm, name);
168 s_vm_params_t params;
169 memset(¶ms, 0, sizeof(params));
170 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
171 //params.overcommit = 0;
173 params.skip_stage2 = 0;
174 params.max_downtime = 0.03;
175 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
176 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
177 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
179 //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate,
180 // params.mig_speed, dp_intensity, update_speed, host_speed);
181 vm->setParameters(¶ms);
186 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
189 * A VM is treated as a host. The name of the VM must be unique among all hosts.
191 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
193 /* make sure the VM of the same name does not exit */
195 simgrid::s4u::Host* ind_host_tmp = (simgrid::s4u::Host*) xbt_dict_get_or_null(host_list, name);
196 if (ind_host_tmp != nullptr && sg_host_simix(ind_host_tmp) != nullptr) {
197 XBT_ERROR("host %s already exits", name);
202 /* Note: ind_vm and vm_workstation point to the same elm object. */
203 /* Ask the SIMIX layer to create the surf vm resource */
204 sg_host_t ind_vm_workstation = (sg_host_t) simcall_vm_create(name, ind_pm);
206 msg_vm_t ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
208 XBT_DEBUG("A new VM (%s) has been created", name);
210 TRACE_msg_vm_create(name, ind_pm);
215 /** @brief Destroy a VM. Destroy the VM object from the simulation.
218 void MSG_vm_destroy(msg_vm_t vm)
220 if (MSG_vm_is_migrating(vm))
221 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
223 /* First, terminate all processes on the VM if necessary */
224 if (MSG_vm_is_running(vm))
225 simcall_vm_shutdown(vm);
227 if (!MSG_vm_is_created(vm)) {
228 XBT_CRITICAL("shutdown the given VM before destroying it");
232 /* Then, destroy the VM object */
233 simcall_vm_destroy(vm);
235 TRACE_msg_vm_end(vm);
238 /** @brief Start a vm (i.e., boot the guest operating system)
241 * If the VM cannot be started, an exception is generated.
243 void MSG_vm_start(msg_vm_t vm)
245 simcall_vm_start(vm);
247 TRACE_msg_vm_start(vm);
250 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
253 * FIXME: No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep() or something.
254 * I'm not quite sure.
256 void MSG_vm_shutdown(msg_vm_t vm)
258 /* msg_vm_t equals to msg_host_t */
259 simcall_vm_shutdown(vm);
261 // TRACE_msg_vm_(vm);
264 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
265 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
266 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
268 const char *vm_name = sg_host_get_name(vm);
269 const char *src_pm_name = sg_host_get_name(src_pm);
270 const char *dst_pm_name = sg_host_get_name(dst_pm);
272 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
275 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
277 const char *vm_name = sg_host_get_name(vm);
278 const char *src_pm_name = sg_host_get_name(src_pm);
279 const char *dst_pm_name = sg_host_get_name(dst_pm);
281 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
284 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
286 const char *vm_name = sg_host_get_name(vm);
287 const char *src_pm_name = sg_host_get_name(src_pm);
288 const char *dst_pm_name = sg_host_get_name(dst_pm);
290 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
293 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
295 const char *vm_name = sg_host_get_name(vm);
296 const char *src_pm_name = sg_host_get_name(src_pm);
297 const char *dst_pm_name = sg_host_get_name(dst_pm);
299 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
302 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
304 const char *vm_name = sg_host_get_name(vm);
305 const char *src_pm_name = sg_host_get_name(src_pm);
306 const char *dst_pm_name = sg_host_get_name(dst_pm);
308 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
311 struct migration_session {
316 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
317 * know the completion of the migration. */
319 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
323 static int migration_rx_fun(int argc, char *argv[])
325 XBT_DEBUG("mig: rx_start");
327 // The structure has been created in the do_migration function and should only be freed in the same place ;)
328 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
330 s_vm_params_t params;
331 ms->vm->parameters(¶ms);
335 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
339 msg_task_t task = NULL;
340 ret = MSG_task_recv(&task, ms->mbox);
343 // An error occured, clean the code and return
344 // The owner did not change, hence the task should be only destroyed on the other side
345 xbt_free(finalize_task_name);
350 if (strcmp(task->name, finalize_task_name) == 0)
353 MSG_task_destroy(task);
359 // Here Stage 1, 2 and 3 have been performed.
360 // Hence complete the migration
362 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
363 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
364 // structure at the beginning and free it at the end of each function)
365 msg_vm_t vm = ms->vm;
366 msg_host_t src_pm = ms->src_pm;
367 msg_host_t dst_pm = ms-> dst_pm;
368 msg_host_priv_t priv = sg_host_msg(vm);
370 // TODO: we have an issue, if the DST node is turning off during the three next calls, then the VM is in an inconsistent
371 // state. I should check with Takahiro in order to make this portion of code atomic
372 // /* deinstall the current affinity setting for the CPU */
373 // simcall_vm_set_affinity(vm, src_pm, 0);
375 // /* Update the vm location */
376 // simcall_vm_migrate(vm, dst_pm);
378 // /* Resume the VM */
379 // simcall_vm_resume(vm);
381 simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
383 /* install the affinity setting of the VM on the destination pm */
386 unsigned long affinity_mask =
387 (unsigned long)(uintptr_t) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *)dst_pm, sizeof(msg_host_t));
388 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
389 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
393 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
394 msg_host_priv_t priv = sg_host_msg(vm);
395 priv->is_migrating = 0;
396 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
397 sg_host_get_name(ms->dst_pm));
398 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
400 // Inform the SRC that the migration has been correctly performed
402 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
403 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
404 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
405 // xbt_assert(ret == MSG_OK);
406 if(ret == MSG_HOST_FAILURE){
407 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
408 // has been correctly migrated on the DST node
409 // TODO What does it mean ? What should we do ?
410 MSG_task_destroy(task);
411 } else if(ret == MSG_TRANSFER_FAILURE){
412 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
413 MSG_task_destroy(task);
418 xbt_free(finalize_task_name);
420 XBT_DEBUG("mig: rx_done");
424 static void reset_dirty_pages(msg_vm_t vm)
426 msg_host_priv_t priv = sg_host_msg(vm);
429 xbt_dict_cursor_t cursor = NULL;
430 dirty_page_t dp = NULL;
431 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
432 double remaining = MSG_task_get_flops_amount(dp->task);
433 dp->prev_clock = MSG_get_clock();
434 dp->prev_remaining = remaining;
436 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
440 static void start_dirty_page_tracking(msg_vm_t vm)
442 msg_host_priv_t priv = sg_host_msg(vm);
443 priv->dp_enabled = 1;
445 reset_dirty_pages(vm);
448 static void stop_dirty_page_tracking(msg_vm_t vm)
450 msg_host_priv_t priv = sg_host_msg(vm);
451 priv->dp_enabled = 0;
454 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
456 double computed = dp->prev_remaining - remaining;
457 double duration = clock - dp->prev_clock;
459 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
460 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
465 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
467 msg_host_priv_t priv = sg_host_msg(vm);
471 xbt_dict_cursor_t cursor = NULL;
472 dirty_page_t dp = NULL;
473 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
474 double remaining = MSG_task_get_flops_amount(dp->task);
476 double clock = MSG_get_clock();
478 // total += calc_updated_pages(key, vm, dp, remaining, clock);
479 total += get_computed(key, vm, dp, remaining, clock);
481 dp->prev_remaining = remaining;
482 dp->prev_clock = clock;
485 total += priv->dp_updated_by_deleted_tasks;
487 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
488 stage_for_fancy_debug, stage2_round_for_fancy_debug, total, priv->dp_updated_by_deleted_tasks);
490 priv->dp_updated_by_deleted_tasks = 0;
495 // TODO Is this code redundant with the information provided by
496 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
497 void MSG_host_add_task(msg_host_t host, msg_task_t task)
499 msg_host_priv_t priv = sg_host_msg(host);
500 double remaining = MSG_task_get_flops_amount(task);
501 char *key = bprintf("%s-%p", task->name, task);
503 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
506 /* It should be okay that we add a task onto a migrating VM. */
507 if (priv->dp_enabled) {
508 dp->prev_clock = MSG_get_clock();
509 dp->prev_remaining = remaining;
512 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
513 xbt_dict_set(priv->dp_objs, key, dp, NULL);
514 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, priv->dp_enabled);
519 void MSG_host_del_task(msg_host_t host, msg_task_t task)
521 msg_host_priv_t priv = sg_host_msg(host);
523 char *key = bprintf("%s-%p", task->name, task);
525 dirty_page_t dp = (dirty_page_t) xbt_dict_get_or_null(priv->dp_objs, key);
526 xbt_assert(dp->task == task);
528 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
529 * the information for the lookup_() function that will called soon. */
530 if (priv->dp_enabled) {
531 double remaining = MSG_task_get_flops_amount(task);
532 double clock = MSG_get_clock();
533 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
534 double updated = get_computed(key, host, dp, remaining, clock);
536 priv->dp_updated_by_deleted_tasks += updated;
539 xbt_dict_remove(priv->dp_objs, key);
542 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
547 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,
548 int stage, int stage2_round, double mig_speed, double timeout)
551 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
552 msg_task_t task = MSG_task_create(task_name, 0, (double)size, NULL);
556 double clock_sta = MSG_get_clock();
560 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
562 ret = MSG_task_send(task, mbox);
568 } else if (ret == MSG_TIMEOUT) {
569 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
570 sent = size - remaining;
571 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
574 /* FIXME: why try-and-catch is used here? */
575 if(ret == MSG_HOST_FAILURE){
576 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
577 MSG_task_destroy(task);
578 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
579 }else if(ret == MSG_TRANSFER_FAILURE){
580 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
581 MSG_task_destroy(task);
582 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
585 double clock_end = MSG_get_clock();
586 double duration = clock_end - clock_sta;
587 double actual_speed = size / duration;
590 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)",
591 stage, stage2_round, size, duration, actual_speed, mig_speed);
593 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)",
594 stage, size, duration, actual_speed, mig_speed);
599 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
601 double updated_size = computed * dp_rate;
602 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
603 if (updated_size > dp_cap) {
604 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
606 updated_size = dp_cap;
609 return (sg_size_t) updated_size;
612 static double send_stage1(struct migration_session *ms, sg_size_t ramsize, double mig_speed, double dp_rate,
615 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
616 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
617 sg_size_t remaining = ramsize;
618 double computed_total = 0;
620 while (remaining > 0) {
621 sg_size_t datasize = chunksize;
622 if (remaining < chunksize)
623 datasize = remaining;
625 remaining -= datasize;
626 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
627 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
628 computed_total += computed;
631 return computed_total;
634 static double get_threshold_value(double bandwidth, double max_downtime)
636 return max_downtime * bandwidth;
639 static int migration_tx_fun(int argc, char *argv[])
641 XBT_DEBUG("mig: tx_start");
643 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
644 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
646 s_vm_params_t params;
647 ms->vm->parameters(¶ms);
648 const sg_size_t ramsize = params.ramsize;
649 const sg_size_t devsize = params.devsize;
650 const int skip_stage1 = params.skip_stage1;
651 int skip_stage2 = params.skip_stage2;
652 const double dp_rate = params.dp_rate;
653 const double dp_cap = params.dp_cap;
654 const double mig_speed = params.mig_speed;
655 double max_downtime = params.max_downtime;
657 /* hard code it temporally. Fix Me */
658 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
659 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
661 double remaining_size = (double) (ramsize + devsize);
662 double threshold = 0.0;
664 /* check parameters */
666 XBT_WARN("migrate a VM, but ramsize is zero");
668 if (max_downtime == 0) {
669 XBT_WARN("use the default max_downtime value 30ms");
673 /* Stage1: send all memory pages to the destination. */
674 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
675 start_dirty_page_tracking(ms->vm);
677 double computed_during_stage1 = 0;
679 double clock_prev_send = MSG_get_clock();
682 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
683 * transfer exceeds the timeout value. */
684 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
685 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
686 remaining_size -= sent;
687 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
689 if (sent < ramsize) {
690 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
692 } else if (sent > ramsize)
696 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
697 // Stop the dirty page tracking an return (there is no memory space to release)
698 stop_dirty_page_tracking(ms->vm);
702 double clock_post_send = MSG_get_clock();
703 mig_timeout -= (clock_post_send - clock_prev_send);
704 if (mig_timeout < 0) {
705 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
706 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
710 /* estimate bandwidth */
711 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
712 threshold = get_threshold_value(bandwidth, max_downtime);
713 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
717 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
720 int stage2_round = 0;
723 sg_size_t updated_size = 0;
724 if (stage2_round == 0) {
725 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
726 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
728 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
729 updated_size = get_updated_size(computed, dp_rate, dp_cap);
732 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
733 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
735 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
736 remaining_size += updated_size;
737 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
738 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
739 if (remaining_size < threshold)
743 double clock_prev_send = MSG_get_clock();
745 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
746 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
749 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
750 // Stop the dirty page tracking an return (there is no memory space to release)
751 stop_dirty_page_tracking(ms->vm);
754 double clock_post_send = MSG_get_clock();
756 if (sent == updated_size) {
757 /* timeout did not happen */
758 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
759 threshold = get_threshold_value(bandwidth, max_downtime);
760 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
761 remaining_size -= sent;
763 mig_timeout -= (clock_post_send - clock_prev_send);
764 xbt_assert(mig_timeout > 0);
766 } else if (sent < updated_size) {
767 /* When timeout happens, we move to stage 3. The size of memory pages
768 * updated before timeout must be added to the remaining size. */
769 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
770 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
771 remaining_size -= sent;
773 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
774 updated_size = get_updated_size(computed, dp_rate, dp_cap);
775 remaining_size += updated_size;
782 /* Stage3: stop the VM and copy the rest of states. */
783 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
784 simcall_vm_suspend(ms->vm);
785 stop_dirty_page_tracking(ms->vm);
788 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
789 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
791 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
792 // Stop the dirty page tracking an return (there is no memory space to release)
793 simcall_vm_resume(ms->vm);
797 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
798 // effectively the VM on the DST node.
799 XBT_DEBUG("mig: tx_done");
804 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
806 struct migration_session *ms = xbt_new(struct migration_session, 1);
810 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
811 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
813 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
814 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
816 // msg_process_t tx_process, rx_process;
817 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
818 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
821 char **argv = xbt_new(char *, 2);
822 argv[0] = pr_rx_name;
824 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
827 char **argv = xbt_new(char *, 2);
828 argv[0] = pr_tx_name;
830 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
834 /* wait until the migration have finished or on error has occured */
836 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
837 msg_task_t task = NULL;
838 msg_error_t ret = MSG_TIMEOUT;
839 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
840 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
842 xbt_free(ms->mbox_ctl);
846 //xbt_assert(ret == MSG_OK);
847 if(ret == MSG_HOST_FAILURE){
848 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
849 // other side. Hence, just throw the execption
850 XBT_ERROR("SRC crashes, throw an exception (m-control)");
851 //MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
855 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){
856 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
857 XBT_ERROR("DST crashes, throw an exception (m-control)");
861 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
862 xbt_assert(strcmp(task->name, expected_task_name) == 0);
863 xbt_free(expected_task_name);
864 MSG_task_destroy(task);
869 /** @brief Migrate the VM to the given host.
872 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
873 * depending on whether you want to do cold or hot migration.
875 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
878 * - One approach is ...
879 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
880 * receive the state of the source
881 * VM over network. We will finally destroy the source VM.
882 * - This behavior is similar to the way of migration in the real world.
883 * Even before a migration is completed, we will see a destination VM, consuming resources.
884 * - We have to relocate all processes. The existing process migration code will work for this?
885 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
887 * - Another one is ...
888 * We update the information of the given VM to place it to the destination physical host.
890 * The second one would be easier.
893 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
895 if(MSG_host_is_off(old_pm))
896 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
898 if(MSG_host_is_off(new_pm))
899 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
901 if (!MSG_vm_is_running(vm))
902 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
904 if (MSG_vm_is_migrating(vm))
905 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
907 msg_host_priv_t priv = sg_host_msg(vm);
908 priv->is_migrating = 1;
911 int ret = do_migration(vm, old_pm, new_pm);
913 priv->is_migrating = 0;
914 THROWF(host_error, 0, "SRC host failed during migration");
917 priv->is_migrating = 0;
918 THROWF(host_error, 0, "DST host failed during migration");
922 // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration
923 // process. In that case, the VM has been already assigned to the DST node.
924 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
925 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
928 /** @brief Immediately suspend the execution of all processes within the given VM.
931 * This function stops the execution of the VM. All the processes on this VM
932 * will pause. The state of the VM is preserved. We can later resume it again.
934 * No suspension cost occurs.
936 void MSG_vm_suspend(msg_vm_t vm)
938 if (MSG_vm_is_migrating(vm))
939 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
941 simcall_vm_suspend(vm);
943 XBT_DEBUG("vm_suspend done");
945 TRACE_msg_vm_suspend(vm);
948 /** @brief Resume the execution of the VM. All processes on the VM run again.
951 * No resume cost occurs.
953 void MSG_vm_resume(msg_vm_t vm)
955 simcall_vm_resume(vm);
957 TRACE_msg_vm_resume(vm);
961 /** @brief Immediately save the execution of all processes within the given VM.
964 * This function stops the execution of the VM. All the processes on this VM
965 * will pause. The state of the VM is preserved. We can later resume it again.
967 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
968 * or after, depending on the exact semantic of VM save to you.
970 void MSG_vm_save(msg_vm_t vm)
972 if (MSG_vm_is_migrating(vm))
973 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
976 TRACE_msg_vm_save(vm);
979 /** @brief Restore the execution of the VM. All processes on the VM run again.
982 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
983 * after, depending on the exact semantic of VM restore to you.
985 void MSG_vm_restore(msg_vm_t vm)
987 simcall_vm_restore(vm);
989 TRACE_msg_vm_restore(vm);
992 /** @brief Get the physical host of a given VM.
995 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
997 return (msg_host_t) simcall_vm_get_pm(vm);
1000 /** @brief Set a CPU bound for a given VM.
1003 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1006 * On PM0, there are Task1 and VM0.
1007 * On VM0, there is Task2.
1008 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1010 * Task1\@PM0 gets 50%.
1011 * Task2\@VM0 gets 25%.
1012 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1014 * 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
1015 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
1017 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
1018 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
1019 * on the VM in share_resource() or we have to trap both the start and end of task execution.
1021 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
1023 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
1024 * computation task (or a VM) never exceeds the capacity of a CPU core.
1026 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1028 simcall_vm_set_bound(vm, bound);
1031 /** @brief Set the CPU affinity of a given VM.
1034 * This function changes the CPU affinity of a given VM. Usage is the same as
1035 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1037 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1039 msg_host_priv_t priv = sg_host_msg(vm);
1042 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1044 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *)(uintptr_t) mask, NULL);
1046 msg_host_t pm_now = MSG_vm_get_pm(vm);
1048 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1049 simcall_vm_set_affinity(vm, pm, mask);
1051 XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));