1 /* Copyright (c) 2012-2015. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
8 * 1. add the support of trace
9 * 2. use parallel tasks to simulate CPU overhead and remove the very
10 * experimental code generating micro computation tasks
15 #include "msg_private.h"
16 #include "xbt/sysdep.h"
18 #include "simgrid/platf.h"
19 #include "simgrid/host.h"
21 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
22 "Cloud-oriented parts of the MSG API");
25 /* **** ******** GENERAL ********* **** */
27 /** \ingroup m_vm_management
28 * \brief Returns the value of a given vm property
31 * \param name a property name
32 * \return value of a property (or NULL if property not set)
35 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
37 return MSG_host_get_property_value(vm, name);
40 /** \ingroup m_vm_management
41 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
44 * \return a dict containing the properties
46 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
48 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
50 return (simcall_host_get_properties(vm));
53 /** \ingroup m_host_management
54 * \brief Change the value of a given host property
57 * \param name a property name
58 * \param value what to change the property to
59 * \param free_ctn the freeing function to use to kill the value on need
61 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
63 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
66 /** \ingroup msg_vm_management
67 * \brief Finds a msg_vm_t using its name.
69 * This is a name directory service
70 * \param name the name of a vm.
71 * \return the corresponding vm
73 * Please note that a VM is a specific host. Hence, you should give a different name
77 msg_vm_t MSG_vm_get_by_name(const char *name)
79 return MSG_host_by_name(name);
82 /** \ingroup m_vm_management
84 * \brief Return the name of the #msg_host_t.
86 * This functions checks whether \a host is a valid pointer or not and return
89 const char *MSG_vm_get_name(msg_vm_t vm)
91 return MSG_host_get_name(vm);
95 /* **** Check state of a VM **** */
96 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
98 return simcall_vm_get_state(vm) == state;
101 /** @brief Returns whether the given VM has just created, not running.
104 int MSG_vm_is_created(msg_vm_t vm)
106 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
109 /** @brief Returns whether the given VM is currently running
112 int MSG_vm_is_running(msg_vm_t vm)
114 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
117 /** @brief Returns whether the given VM is currently migrating
120 int MSG_vm_is_migrating(msg_vm_t vm)
122 msg_host_priv_t priv = sg_host_msg(vm);
123 return priv->is_migrating;
126 /** @brief Returns whether the given VM is currently suspended, not running.
129 int MSG_vm_is_suspended(msg_vm_t vm)
131 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
134 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
137 int MSG_vm_is_saving(msg_vm_t vm)
139 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
142 /** @brief Returns whether the given VM has been saved, not running.
145 int MSG_vm_is_saved(msg_vm_t vm)
147 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
150 /** @brief Returns whether the given VM is being restored, not running.
153 int MSG_vm_is_restoring(msg_vm_t vm)
155 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
160 /* ------------------------------------------------------------------------- */
161 /* ------------------------------------------------------------------------- */
163 /* **** ******** MSG vm actions ********* **** */
165 /** @brief Create a new VM with specified parameters.
167 * @param pm Physical machine that will host the VM
168 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
169 * @param disk_path (unused) Path to the image that boots
170 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
171 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
172 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
175 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name,
176 int ncpus, int ramsize,
177 int net_cap, char *disk_path, int disksize,
178 int mig_netspeed, int dp_intensity)
180 /* For the moment, intensity_rate is the percentage against the migration
182 double host_speed = MSG_get_host_speed(pm);
183 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
185 msg_vm_t vm = MSG_vm_create_core(pm, name);
186 s_vm_params_t params;
187 memset(¶ms, 0, sizeof(params));
188 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
189 //params.overcommit = 0;
191 params.skip_stage2 = 0;
192 params.max_downtime = 0.03;
193 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
194 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
195 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
197 //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);
198 simcall_host_set_params(vm, ¶ms);
204 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
207 * A VM is treated as a host. The name of the VM must be unique among all hosts.
209 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
211 /* make sure the VM of the same name does not exit */
213 xbt_dictelm_t ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
214 if (ind_host_tmp && sg_host_simix(ind_host_tmp) != NULL) {
215 XBT_ERROR("host %s already exits", name);
220 /* Note: ind_vm and vm_workstation point to the same elm object. */
221 msg_vm_t ind_vm = NULL;
222 void *ind_vm_workstation = NULL;
224 /* Ask the SIMIX layer to create the surf vm resource */
225 ind_vm_workstation = simcall_vm_create(name, ind_pm);
226 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
228 XBT_DEBUG("A new VM (%s) has been created", name);
230 TRACE_msg_vm_create(name, ind_pm);
235 /** @brief Destroy a VM. Destroy the VM object from the simulation.
238 void MSG_vm_destroy(msg_vm_t vm)
240 if (MSG_vm_is_migrating(vm))
241 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
243 /* First, terminate all processes on the VM if necessary */
244 if (MSG_vm_is_running(vm))
245 simcall_vm_shutdown(vm);
247 if (!MSG_vm_is_created(vm)) {
248 XBT_CRITICAL("shutdown the given VM before destroying it");
252 /* Then, destroy the VM object */
253 simcall_vm_destroy(vm);
255 __MSG_host_destroy(vm);
257 TRACE_msg_vm_end(vm);
261 /** @brief Start a vm (i.e., boot the guest operating system)
264 * If the VM cannot be started, an exception is generated.
267 void MSG_vm_start(msg_vm_t vm)
269 simcall_vm_start(vm);
271 TRACE_msg_vm_start(vm);
276 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
279 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
280 * use a #MSG_process_sleep() or something. I'm not quite sure.
282 void MSG_vm_shutdown(msg_vm_t vm)
284 /* msg_vm_t equals to msg_host_t */
285 simcall_vm_shutdown(vm);
287 // TRACE_msg_vm_(vm);
292 /* We have two mailboxes. mbox is used to transfer migration data between
293 * source and destination PMs. mbox_ctl is used to detect the completion of a
294 * migration. The names of these mailboxes must not conflict with others. */
295 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
297 char *vm_name = sg_host_get_name(vm);
298 char *src_pm_name = sg_host_get_name(src_pm);
299 char *dst_pm_name = sg_host_get_name(dst_pm);
301 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
304 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
306 char *vm_name = sg_host_get_name(vm);
307 char *src_pm_name = sg_host_get_name(src_pm);
308 char *dst_pm_name = sg_host_get_name(dst_pm);
310 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
313 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
315 char *vm_name = sg_host_get_name(vm);
316 char *src_pm_name = sg_host_get_name(src_pm);
317 char *dst_pm_name = sg_host_get_name(dst_pm);
319 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
322 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
324 char *vm_name = sg_host_get_name(vm);
325 char *src_pm_name = sg_host_get_name(src_pm);
326 char *dst_pm_name = sg_host_get_name(dst_pm);
328 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
331 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
333 char *vm_name = sg_host_get_name(vm);
334 char *src_pm_name = sg_host_get_name(src_pm);
335 char *dst_pm_name = sg_host_get_name(dst_pm);
337 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
341 struct migration_session {
346 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
347 * know the completion of the migration. */
349 /* The migration_rx and migration_tx processes use mbox to transfer migration
355 static int migration_rx_fun(int argc, char *argv[])
357 XBT_DEBUG("mig: rx_start");
359 // The structure has been created in the do_migration function and should only be freed in the same place ;)
360 struct migration_session *ms = MSG_process_get_data(MSG_process_self());
362 s_vm_params_t params;
363 simcall_host_get_params(ms->vm, ¶ms);
367 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
371 msg_task_t task = NULL;
372 ret = MSG_task_recv(&task, ms->mbox);
375 // An error occured, clean the code and return
376 // The owner did not change, hence the task should be only destroyed on the other side
377 xbt_free(finalize_task_name);
382 if (strcmp(task->name, finalize_task_name) == 0)
385 MSG_task_destroy(task);
391 // Here Stage 1, 2 and 3 have been performed.
392 // Hence complete the migration
394 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
395 // 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)
396 msg_vm_t vm = ms->vm;
397 msg_host_t src_pm = ms->src_pm;
398 msg_host_t dst_pm = ms-> dst_pm;
399 msg_host_priv_t priv = sg_host_msg(vm);
401 // // 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
402 // // I should check with Takahiro in order to make this portion of code atomic
403 // /* deinstall the current affinity setting for the CPU */
404 // simcall_vm_set_affinity(vm, src_pm, 0);
406 // /* Update the vm location */
407 // simcall_vm_migrate(vm, dst_pm);
409 // /* Resume the VM */
410 // simcall_vm_resume(vm);
412 simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
414 /* install the affinity setting of the VM on the destination pm */
417 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));
418 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
419 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
424 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
425 msg_host_priv_t priv = sg_host_msg(vm);
426 priv->is_migrating = 0;
427 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->key, ms->src_pm->key, ms->dst_pm->key);
428 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
430 // Inform the SRC that the migration has been correctly performed
432 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
433 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
434 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
435 // xbt_assert(ret == MSG_OK);
436 if(ret == MSG_HOST_FAILURE){
437 // 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
438 // TODO What does it mean ? What should we do ?
439 MSG_task_destroy(task);
440 } else if(ret == MSG_TRANSFER_FAILURE){
441 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
442 MSG_task_destroy(task);
449 xbt_free(finalize_task_name);
451 XBT_DEBUG("mig: rx_done");
456 static void reset_dirty_pages(msg_vm_t vm)
458 msg_host_priv_t priv = sg_host_msg(vm);
461 xbt_dict_cursor_t cursor = NULL;
462 dirty_page_t dp = NULL;
463 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
464 double remaining = MSG_task_get_flops_amount(dp->task);
465 dp->prev_clock = MSG_get_clock();
466 dp->prev_remaining = remaining;
468 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
472 static void start_dirty_page_tracking(msg_vm_t vm)
474 msg_host_priv_t priv = sg_host_msg(vm);
475 priv->dp_enabled = 1;
477 reset_dirty_pages(vm);
480 static void stop_dirty_page_tracking(msg_vm_t vm)
482 msg_host_priv_t priv = sg_host_msg(vm);
483 priv->dp_enabled = 0;
487 /* It might be natural that we define dp_rate for each task. But, we will also
488 * have to care about how each task behavior affects the memory update behavior
489 * at the operating system level. It may not be easy to model it with a simple algorithm. */
490 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
492 double computed = dp->prev_remaining - remaining;
493 double duration = clock - dp->prev_clock;
494 double updated = dp->task->dp_rate * computed;
496 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
497 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
498 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
499 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
505 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
507 double computed = dp->prev_remaining - remaining;
508 double duration = clock - dp->prev_clock;
510 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
511 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
516 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
518 msg_host_priv_t priv = sg_host_msg(vm);
522 xbt_dict_cursor_t cursor = NULL;
523 dirty_page_t dp = NULL;
524 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
525 double remaining = MSG_task_get_flops_amount(dp->task);
527 double clock = MSG_get_clock();
529 // total += calc_updated_pages(key, vm, dp, remaining, clock);
530 total += get_computed(key, vm, dp, remaining, clock);
532 dp->prev_remaining = remaining;
533 dp->prev_clock = clock;
536 total += priv->dp_updated_by_deleted_tasks;
538 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
539 stage_for_fancy_debug,
540 stage2_round_for_fancy_debug,
541 total, priv->dp_updated_by_deleted_tasks);
545 priv->dp_updated_by_deleted_tasks = 0;
551 // TODO Is this code redundant with the information provided by
552 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
553 void MSG_host_add_task(msg_host_t host, msg_task_t task)
555 msg_host_priv_t priv = sg_host_msg(host);
556 double remaining = MSG_task_get_flops_amount(task);
557 char *key = bprintf("%s-%p", task->name, task);
559 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
562 /* It should be okay that we add a task onto a migrating VM. */
563 if (priv->dp_enabled) {
564 dp->prev_clock = MSG_get_clock();
565 dp->prev_remaining = remaining;
568 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
569 xbt_dict_set(priv->dp_objs, key, dp, NULL);
570 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, priv->dp_enabled);
575 void MSG_host_del_task(msg_host_t host, msg_task_t task)
577 msg_host_priv_t priv = sg_host_msg(host);
579 char *key = bprintf("%s-%p", task->name, task);
581 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
582 xbt_assert(dp->task == task);
584 /* If we are in the middle of dirty page tracking, we record how much
585 * computation has been done until now, and keep the information for the
586 * lookup_() function that will called soon. */
587 if (priv->dp_enabled) {
588 double remaining = MSG_task_get_flops_amount(task);
589 double clock = MSG_get_clock();
590 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
591 double updated = get_computed(key, host, dp, remaining, clock);
593 priv->dp_updated_by_deleted_tasks += updated;
596 xbt_dict_remove(priv->dp_objs, key);
599 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
607 static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm,
608 sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double timeout)
611 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
612 msg_task_t task = MSG_task_create(task_name, 0, (double)size, NULL);
616 double clock_sta = MSG_get_clock();
620 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
622 ret = MSG_task_send(task, mbox);
628 } else if (ret == MSG_TIMEOUT) {
629 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
630 sent = size - remaining;
631 XBT_INFO("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu",
632 timeout, remaining, size);
635 /* FIXME: why try-and-catch is used here? */
636 if(ret == MSG_HOST_FAILURE){
637 //XBT_INFO("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
638 MSG_task_destroy(task);
639 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
640 }else if(ret == MSG_TRANSFER_FAILURE){
641 //XBT_INFO("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
642 MSG_task_destroy(task);
643 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
646 double clock_end = MSG_get_clock();
647 double duration = clock_end - clock_sta;
648 double actual_speed = size / duration;
651 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration, actual_speed, mig_speed);
653 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed, mig_speed);
658 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
660 double updated_size = computed * dp_rate;
661 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
662 if (updated_size > dp_cap) {
663 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
664 updated_size = dp_cap;
667 return (sg_size_t) updated_size;
670 static double send_stage1(struct migration_session *ms,
671 sg_size_t ramsize, double mig_speed, double dp_rate, double dp_cap)
674 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
675 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
676 sg_size_t remaining = ramsize;
677 double computed_total = 0;
679 while (remaining > 0) {
680 sg_size_t datasize = chunksize;
681 if (remaining < chunksize)
682 datasize = remaining;
684 remaining -= datasize;
685 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
686 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
687 computed_total += computed;
690 return computed_total;
695 static double get_threshold_value(double bandwidth, double max_downtime)
697 return max_downtime * bandwidth;
700 static int migration_tx_fun(int argc, char *argv[])
702 XBT_DEBUG("mig: tx_start");
704 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
705 struct migration_session *ms = MSG_process_get_data(MSG_process_self());
707 s_vm_params_t params;
708 simcall_host_get_params(ms->vm, ¶ms);
709 const sg_size_t ramsize = params.ramsize;
710 const sg_size_t devsize = params.devsize;
711 const int skip_stage1 = params.skip_stage1;
712 int skip_stage2 = params.skip_stage2;
713 const double dp_rate = params.dp_rate;
714 const double dp_cap = params.dp_cap;
715 const double mig_speed = params.mig_speed;
716 double max_downtime = params.max_downtime;
718 /* hard code it temporally. Fix Me */
719 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
720 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
722 double remaining_size = (double) (ramsize + devsize);
723 double threshold = 0.0;
725 /* check parameters */
727 XBT_WARN("migrate a VM, but ramsize is zero");
729 if (max_downtime == 0) {
730 XBT_WARN("use the default max_downtime value 30ms");
734 /* Stage1: send all memory pages to the destination. */
735 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
736 start_dirty_page_tracking(ms->vm);
738 double computed_during_stage1 = 0;
740 double clock_prev_send = MSG_get_clock();
743 /* At stage 1, we do not need timeout. We have to send all the memory
744 * pages even though the duration of this tranfer exceeds the timeout
746 XBT_INFO("Stage 1: Gonna send %llu", ramsize);
747 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
748 remaining_size -= sent;
749 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
751 if (sent < ramsize) {
752 XBT_INFO("mig-stage1: timeout, force moving to stage 3");
754 } else if (sent > ramsize)
758 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
759 // Stop the dirty page tracking an return (there is no memory space to release)
760 stop_dirty_page_tracking(ms->vm);
764 double clock_post_send = MSG_get_clock();
765 mig_timeout -= (clock_post_send - clock_prev_send);
766 if (mig_timeout < 0) {
767 XBT_INFO("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
768 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
772 /* estimate bandwidth */
773 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
774 threshold = get_threshold_value(bandwidth, max_downtime);
775 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
779 /* Stage2: send update pages iteratively until the size of remaining states
780 * becomes smaller than the threshold value. */
785 int stage2_round = 0;
788 sg_size_t updated_size = 0;
789 if (stage2_round == 0) {
790 /* just after stage1, nothing has been updated. But, we have to send the
791 * data updated during stage1 */
792 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
794 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
795 updated_size = get_updated_size(computed, dp_rate, dp_cap);
798 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
799 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
802 /* Check whether the remaining size is below the threshold value. If so,
803 * move to stage 3. */
804 remaining_size += updated_size;
805 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
806 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
807 if (remaining_size < threshold)
812 double clock_prev_send = MSG_get_clock();
814 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
815 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed, mig_timeout);
817 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
818 // Stop the dirty page tracking an return (there is no memory space to release)
819 stop_dirty_page_tracking(ms->vm);
822 double clock_post_send = MSG_get_clock();
824 if (sent == updated_size) {
825 /* timeout did not happen */
826 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
827 threshold = get_threshold_value(bandwidth, max_downtime);
828 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
829 remaining_size -= sent;
831 mig_timeout -= (clock_post_send - clock_prev_send);
832 xbt_assert(mig_timeout > 0);
834 } else if (sent < updated_size) {
835 /* When timeout happens, we move to stage 3. The size of memory pages
836 * updated before timeout must be added to the remaining size. */
837 XBT_INFO("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
838 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
839 remaining_size -= sent;
841 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
842 updated_size = get_updated_size(computed, dp_rate, dp_cap);
843 remaining_size += updated_size;
852 /* Stage3: stop the VM and copy the rest of states. */
853 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
854 simcall_vm_suspend(ms->vm);
855 stop_dirty_page_tracking(ms->vm);
858 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
859 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
861 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
862 // Stop the dirty page tracking an return (there is no memory space to release)
863 simcall_vm_resume(ms->vm);
867 // 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.
868 XBT_DEBUG("mig: tx_done");
875 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
877 struct migration_session *ms = xbt_new(struct migration_session, 1);
881 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
882 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
885 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
886 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
888 // msg_process_t tx_process, rx_process;
889 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
890 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
893 char **argv = xbt_new(char *, 2);
894 argv[0] = pr_rx_name;
896 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
899 char **argv = xbt_new(char *, 2);
900 argv[0] = pr_tx_name;
902 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
906 /* wait until the migration have finished or on error has occured */
908 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
909 msg_task_t task = NULL;
910 msg_error_t ret = MSG_TIMEOUT;
911 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
912 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
914 xbt_free(ms->mbox_ctl);
918 //xbt_assert(ret == MSG_OK);
919 if(ret == MSG_HOST_FAILURE){
920 // Note that since the communication failed, the owner did not change and the task should be destroyed on the other side.
921 // Hence, just throw the execption
922 XBT_INFO("SRC crashes, throw an exception (m-control)");
923 //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
926 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){ // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
927 XBT_INFO("DST crashes, throw an exception (m-control)");
932 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
933 xbt_assert(strcmp(task->name, expected_task_name) == 0);
934 xbt_free(expected_task_name);
935 MSG_task_destroy(task);
943 /** @brief Migrate the VM to the given host.
946 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
947 * MSG_task_send() before or after, depending on whether you want to do cold or hot
950 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
953 * - One approach is ...
954 * We first create a new VM (i.e., destination VM) on the destination
955 * physical host. The destination VM will receive the state of the source
956 * VM over network. We will finally destroy the source VM.
957 * - This behavior is similar to the way of migration in the real world.
958 * Even before a migration is completed, we will see a destination VM,
959 * consuming resources.
960 * - We have to relocate all processes. The existing process migraion code
961 * will work for this?
962 * - The name of the VM is a somewhat unique ID in the code. It is tricky
963 * for the destination VM?
965 * - Another one is ...
966 * We update the information of the given VM to place it to the destination
969 * The second one would be easier.
973 msg_host_t old_pm = simcall_vm_get_pm(vm);
975 if(MSG_host_is_off(old_pm))
976 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
978 if(MSG_host_is_off(new_pm))
979 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
981 if (!MSG_vm_is_running(vm))
982 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
984 if (MSG_vm_is_migrating(vm))
985 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
987 msg_host_priv_t priv = sg_host_msg(vm);
988 priv->is_migrating = 1;
992 int ret = do_migration(vm, old_pm, new_pm);
994 priv->is_migrating = 0;
995 THROWF(host_error, 0, "SRC host failed during migration");
998 priv->is_migrating = 0;
999 THROWF(host_error, 0, "DST host failed during migration");
1003 // 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
1004 // In that case, the VM has been already assigned to the DST node.
1005 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1006 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1010 /** @brief Immediately suspend the execution of all processes within the given VM.
1013 * This function stops the execution of the VM. All the processes on this VM
1014 * will pause. The state of the VM is preserved. We can later resume it again.
1016 * No suspension cost occurs.
1018 void MSG_vm_suspend(msg_vm_t vm)
1020 if (MSG_vm_is_migrating(vm))
1021 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
1023 simcall_vm_suspend(vm);
1025 XBT_DEBUG("vm_suspend done");
1027 TRACE_msg_vm_suspend(vm);
1031 /** @brief Resume the execution of the VM. All processes on the VM run again.
1034 * No resume cost occurs.
1036 void MSG_vm_resume(msg_vm_t vm)
1038 simcall_vm_resume(vm);
1040 TRACE_msg_vm_resume(vm);
1044 /** @brief Immediately save the execution of all processes within the given VM.
1047 * This function stops the execution of the VM. All the processes on this VM
1048 * will pause. The state of the VM is preserved. We can later resume it again.
1050 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1051 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1052 * of VM save to you.
1054 void MSG_vm_save(msg_vm_t vm)
1056 if (MSG_vm_is_migrating(vm))
1057 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
1059 simcall_vm_save(vm);
1060 TRACE_msg_vm_save(vm);
1063 /** @brief Restore the execution of the VM. All processes on the VM run again.
1066 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1067 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1068 * of VM restore to you.
1070 void MSG_vm_restore(msg_vm_t vm)
1072 simcall_vm_restore(vm);
1074 TRACE_msg_vm_restore(vm);
1078 /** @brief Get the physical host of a given VM.
1081 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1083 return simcall_vm_get_pm(vm);
1087 /** @brief Set a CPU bound for a given VM.
1091 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1094 * On PM0, there are Task1 and VM0.
1095 * On VM0, there is Task2.
1096 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1098 * Task1\@PM0 gets 50%.
1099 * Task2\@VM0 gets 25%.
1100 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1102 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1103 * task on the VM does not affect the bound of the dummy CPU action. The bound
1104 * of the dummy CPU action is unlimited.
1106 * There are some solutions for this problem. One option is to update the bound
1107 * of the dummy CPU action automatically. It should be the sum of all tasks on
1108 * the VM. But, this solution might be costly, because we have to scan all tasks
1109 * on the VM in share_resource() or we have to trap both the start and end of
1112 * The current solution is to use MSG_vm_set_bound(), which allows us to
1113 * directly set the bound of the dummy CPU action.
1117 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1118 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1119 * exceeds the capacity of a CPU core.
1121 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1123 simcall_vm_set_bound(vm, bound);
1127 /** @brief Set the CPU affinity of a given VM.
1130 * This function changes the CPU affinity of a given VM. Usage is the same as
1131 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1133 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1135 msg_host_priv_t priv = sg_host_msg(vm);
1138 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1140 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *)(uintptr_t) mask, NULL);
1142 msg_host_t pm_now = MSG_vm_get_pm(vm);
1144 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1145 simcall_vm_set_affinity(vm, pm, mask);
1147 XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));