1 /* Copyright (c) 2012-2015. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
8 * 1. add the support of trace
9 * 2. use parallel tasks to simulate CPU overhead and remove the experimental code generating micro computation tasks
14 #include "msg_private.h"
15 #include "xbt/sysdep.h"
17 #include "simgrid/host.h"
19 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
22 /* **** ******** GENERAL ********* **** */
23 /** \ingroup m_vm_management
24 * \brief Returns the value of a given vm property
27 * \param name a property name
28 * \return value of a property (or nullptr if property not set)
30 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
32 return MSG_host_get_property_value(vm, name);
35 /** \ingroup m_vm_management
36 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
39 * \return a dict containing the properties
41 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
43 xbt_assert((vm != nullptr), "Invalid parameters (vm is nullptr)");
44 return vm->properties();
47 /** \ingroup m_host_management
48 * \brief Change the value of a given host property
51 * \param name a property name
52 * \param value what to change the property to
53 * \param free_ctn the freeing function to use to kill the value on need
55 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
57 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
60 /** \ingroup msg_vm_management
61 * \brief Finds a msg_vm_t using its name.
63 * This is a name directory service
64 * \param name the name of a vm.
65 * \return the corresponding vm
67 * Please note that a VM is a specific host. Hence, you should give a different name for each VM/PM.
69 msg_vm_t MSG_vm_get_by_name(const char *name)
71 return MSG_host_by_name(name);
74 /** \ingroup m_vm_management
76 * \brief Return the name of the #msg_host_t.
78 * This functions checks whether \a host is a valid pointer or not and return its name.
80 const char *MSG_vm_get_name(msg_vm_t vm)
82 return MSG_host_get_name(vm);
85 /* **** Check state of a VM **** */
86 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
88 return simcall_vm_get_state(vm) == state;
91 /** @brief Returns whether the given VM has just created, not running.
94 int MSG_vm_is_created(msg_vm_t vm)
96 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
99 /** @brief Returns whether the given VM is currently running
102 int MSG_vm_is_running(msg_vm_t vm)
104 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
107 /** @brief Returns whether the given VM is currently migrating
110 int MSG_vm_is_migrating(msg_vm_t vm)
112 msg_host_priv_t priv = sg_host_msg(vm);
113 return priv->is_migrating;
116 /** @brief Returns whether the given VM is currently suspended, not running.
119 int MSG_vm_is_suspended(msg_vm_t vm)
121 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
124 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
127 int MSG_vm_is_saving(msg_vm_t vm)
129 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
132 /** @brief Returns whether the given VM has been saved, not running.
135 int MSG_vm_is_saved(msg_vm_t vm)
137 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
140 /** @brief Returns whether the given VM is being restored, not running.
143 int MSG_vm_is_restoring(msg_vm_t vm)
145 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
148 /* **** ******** MSG vm actions ********* **** */
149 /** @brief Create a new VM with specified parameters.
151 * @param pm Physical machine that will host the VM
153 * @param ncpus [TODO]
154 * @param ramsize [TODO]
155 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
156 * @param disk_path (unused) Path to the image that boots
157 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
158 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
159 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
161 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name, int ncpus, int ramsize, int net_cap, char *disk_path,
162 int disksize, int mig_netspeed, int dp_intensity)
164 /* For the moment, intensity_rate is the percentage against the migration
166 double host_speed = MSG_host_get_speed(pm);
167 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
169 msg_vm_t vm = MSG_vm_create_core(pm, name);
170 s_vm_params_t params;
171 memset(¶ms, 0, sizeof(params));
172 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
173 //params.overcommit = 0;
175 params.skip_stage2 = 0;
176 params.max_downtime = 0.03;
177 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
178 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
179 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
181 //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate,
182 // params.mig_speed, dp_intensity, update_speed, host_speed);
183 vm->setParameters(¶ms);
188 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
191 * A VM is treated as a host. The name of the VM must be unique among all hosts.
193 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
195 /* make sure the VM of the same name does not exit */
197 simgrid::s4u::Host* ind_host_tmp = (simgrid::s4u::Host*) xbt_dict_get_or_null(host_list, name);
198 if (ind_host_tmp != nullptr && sg_host_simix(ind_host_tmp) != nullptr) {
199 XBT_ERROR("host %s already exits", name);
204 /* Note: ind_vm and vm_workstation point to the same elm object. */
205 /* Ask the SIMIX layer to create the surf vm resource */
206 sg_host_t ind_vm_workstation = (sg_host_t) simcall_vm_create(name, ind_pm);
208 msg_vm_t ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
210 XBT_DEBUG("A new VM (%s) has been created", name);
212 TRACE_msg_vm_create(name, ind_pm);
217 /** @brief Destroy a VM. Destroy the VM object from the simulation.
220 void MSG_vm_destroy(msg_vm_t vm)
222 if (MSG_vm_is_migrating(vm))
223 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
225 /* First, terminate all processes on the VM if necessary */
226 if (MSG_vm_is_running(vm))
227 simcall_vm_shutdown(vm);
229 if (!MSG_vm_is_created(vm)) {
230 XBT_CRITICAL("shutdown the given VM before destroying it");
234 /* Then, destroy the VM object */
235 simcall_vm_destroy(vm);
237 TRACE_msg_vm_end(vm);
240 /** @brief Start a vm (i.e., boot the guest operating system)
243 * If the VM cannot be started, an exception is generated.
245 void MSG_vm_start(msg_vm_t vm)
247 simcall_vm_start(vm);
249 TRACE_msg_vm_start(vm);
252 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
255 * FIXME: No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep() or something.
256 * I'm not quite sure.
258 void MSG_vm_shutdown(msg_vm_t vm)
260 /* msg_vm_t equals to msg_host_t */
261 simcall_vm_shutdown(vm);
263 // TRACE_msg_vm_(vm);
266 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
267 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
268 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
270 const char *vm_name = sg_host_get_name(vm);
271 const char *src_pm_name = sg_host_get_name(src_pm);
272 const char *dst_pm_name = sg_host_get_name(dst_pm);
274 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
277 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
279 const char *vm_name = sg_host_get_name(vm);
280 const char *src_pm_name = sg_host_get_name(src_pm);
281 const char *dst_pm_name = sg_host_get_name(dst_pm);
283 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
286 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
288 const char *vm_name = sg_host_get_name(vm);
289 const char *src_pm_name = sg_host_get_name(src_pm);
290 const char *dst_pm_name = sg_host_get_name(dst_pm);
292 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
295 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
297 const char *vm_name = sg_host_get_name(vm);
298 const char *src_pm_name = sg_host_get_name(src_pm);
299 const char *dst_pm_name = sg_host_get_name(dst_pm);
301 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
304 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
306 const char *vm_name = sg_host_get_name(vm);
307 const char *src_pm_name = sg_host_get_name(src_pm);
308 const char *dst_pm_name = sg_host_get_name(dst_pm);
310 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
313 struct migration_session {
318 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
319 * know the completion of the migration. */
321 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
325 static int migration_rx_fun(int argc, char *argv[])
327 XBT_DEBUG("mig: rx_start");
329 // The structure has been created in the do_migration function and should only be freed in the same place ;)
330 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
332 s_vm_params_t params;
333 ms->vm->parameters(¶ms);
337 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
341 msg_task_t task = nullptr;
342 ret = MSG_task_recv(&task, ms->mbox);
345 // An error occurred, clean the code and return
346 // The owner did not change, hence the task should be only destroyed on the other side
347 xbt_free(finalize_task_name);
352 if (strcmp(task->name, finalize_task_name) == 0)
355 MSG_task_destroy(task);
361 // Here Stage 1, 2 and 3 have been performed.
362 // Hence complete the migration
364 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
365 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
366 // structure at the beginning and free it at the end of each function)
367 msg_vm_t vm = ms->vm;
368 msg_host_t src_pm = ms->src_pm;
369 msg_host_t dst_pm = ms-> dst_pm;
370 msg_host_priv_t priv = sg_host_msg(vm);
372 // TODO: we have an issue, if the DST node is turning off during the three next calls, then the VM is in an inconsistent
373 // state. I should check with Takahiro in order to make this portion of code atomic
374 // /* deinstall the current affinity setting for the CPU */
375 // simcall_vm_set_affinity(vm, src_pm, 0);
377 // /* Update the vm location */
378 // simcall_vm_migrate(vm, dst_pm);
380 // /* Resume the VM */
381 // simcall_vm_resume(vm);
383 simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
385 /* install the affinity setting of the VM on the destination pm */
388 unsigned long affinity_mask =
389 (unsigned long)(uintptr_t) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *)dst_pm, sizeof(msg_host_t));
390 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
391 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
395 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
396 msg_host_priv_t priv = sg_host_msg(vm);
397 priv->is_migrating = 0;
398 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
399 sg_host_get_name(ms->dst_pm));
400 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
402 // Inform the SRC that the migration has been correctly performed
404 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
405 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
406 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
407 // xbt_assert(ret == MSG_OK);
408 if(ret == MSG_HOST_FAILURE){
409 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
410 // has been correctly migrated on the DST node
411 // TODO What does it mean ? What should we do ?
412 MSG_task_destroy(task);
413 } else if(ret == MSG_TRANSFER_FAILURE){
414 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
415 MSG_task_destroy(task);
420 xbt_free(finalize_task_name);
422 XBT_DEBUG("mig: rx_done");
426 static void reset_dirty_pages(msg_vm_t vm)
428 msg_host_priv_t priv = sg_host_msg(vm);
431 xbt_dict_cursor_t cursor = nullptr;
432 dirty_page_t dp = nullptr;
433 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
434 double remaining = MSG_task_get_flops_amount(dp->task);
435 dp->prev_clock = MSG_get_clock();
436 dp->prev_remaining = remaining;
438 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
442 static void start_dirty_page_tracking(msg_vm_t vm)
444 msg_host_priv_t priv = sg_host_msg(vm);
445 priv->dp_enabled = 1;
447 reset_dirty_pages(vm);
450 static void stop_dirty_page_tracking(msg_vm_t vm)
452 msg_host_priv_t priv = sg_host_msg(vm);
453 priv->dp_enabled = 0;
456 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
458 double computed = dp->prev_remaining - remaining;
459 double duration = clock - dp->prev_clock;
461 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
462 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
467 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
469 msg_host_priv_t priv = sg_host_msg(vm);
473 xbt_dict_cursor_t cursor = nullptr;
474 dirty_page_t dp = nullptr;
475 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
476 double remaining = MSG_task_get_flops_amount(dp->task);
478 double clock = MSG_get_clock();
480 // total += calc_updated_pages(key, vm, dp, remaining, clock);
481 total += get_computed(key, vm, dp, remaining, clock);
483 dp->prev_remaining = remaining;
484 dp->prev_clock = clock;
487 total += priv->dp_updated_by_deleted_tasks;
489 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
490 stage_for_fancy_debug, stage2_round_for_fancy_debug, total, priv->dp_updated_by_deleted_tasks);
492 priv->dp_updated_by_deleted_tasks = 0;
497 // TODO Is this code redundant with the information provided by
498 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
499 void MSG_host_add_task(msg_host_t host, msg_task_t task)
501 msg_host_priv_t priv = sg_host_msg(host);
502 double remaining = MSG_task_get_flops_amount(task);
503 char *key = bprintf("%s-%p", task->name, task);
505 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
508 /* It should be okay that we add a task onto a migrating VM. */
509 if (priv->dp_enabled) {
510 dp->prev_clock = MSG_get_clock();
511 dp->prev_remaining = remaining;
514 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == nullptr);
515 xbt_dict_set(priv->dp_objs, key, dp, nullptr);
516 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, priv->dp_enabled);
521 void MSG_host_del_task(msg_host_t host, msg_task_t task)
523 msg_host_priv_t priv = sg_host_msg(host);
525 char *key = bprintf("%s-%p", task->name, task);
527 dirty_page_t dp = (dirty_page_t) xbt_dict_get_or_null(priv->dp_objs, key);
528 xbt_assert(dp->task == task);
530 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
531 * the information for the lookup_() function that will called soon. */
532 if (priv->dp_enabled) {
533 double remaining = MSG_task_get_flops_amount(task);
534 double clock = MSG_get_clock();
535 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
536 double updated = get_computed(key, host, dp, remaining, clock);
538 priv->dp_updated_by_deleted_tasks += updated;
541 xbt_dict_remove(priv->dp_objs, key);
544 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
549 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,
550 int stage, int stage2_round, double mig_speed, double timeout)
553 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
554 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
558 double clock_sta = MSG_get_clock();
562 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
564 ret = MSG_task_send(task, mbox);
570 } else if (ret == MSG_TIMEOUT) {
571 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
572 sent = size - remaining;
573 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
576 /* FIXME: why try-and-catch is used here? */
577 if(ret == MSG_HOST_FAILURE){
578 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
579 MSG_task_destroy(task);
580 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
581 }else if(ret == MSG_TRANSFER_FAILURE){
582 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
583 MSG_task_destroy(task);
584 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
587 double clock_end = MSG_get_clock();
588 double duration = clock_end - clock_sta;
589 double actual_speed = size / duration;
592 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)",
593 stage, stage2_round, size, duration, actual_speed, mig_speed);
595 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)",
596 stage, size, duration, actual_speed, mig_speed);
601 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
603 double updated_size = computed * dp_rate;
604 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
605 if (updated_size > dp_cap) {
606 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
608 updated_size = dp_cap;
611 return (sg_size_t) updated_size;
614 static double send_stage1(struct migration_session *ms, sg_size_t ramsize, double mig_speed, double dp_rate,
617 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
618 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
619 sg_size_t remaining = ramsize;
620 double computed_total = 0;
622 while (remaining > 0) {
623 sg_size_t datasize = chunksize;
624 if (remaining < chunksize)
625 datasize = remaining;
627 remaining -= datasize;
628 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
629 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
630 computed_total += computed;
633 return computed_total;
636 static double get_threshold_value(double bandwidth, double max_downtime)
638 return max_downtime * bandwidth;
641 static int migration_tx_fun(int argc, char *argv[])
643 XBT_DEBUG("mig: tx_start");
645 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
646 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
648 s_vm_params_t params;
649 ms->vm->parameters(¶ms);
650 const sg_size_t ramsize = params.ramsize;
651 const sg_size_t devsize = params.devsize;
652 const int skip_stage1 = params.skip_stage1;
653 int skip_stage2 = params.skip_stage2;
654 const double dp_rate = params.dp_rate;
655 const double dp_cap = params.dp_cap;
656 const double mig_speed = params.mig_speed;
657 double max_downtime = params.max_downtime;
659 /* hard code it temporally. Fix Me */
660 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
661 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
663 double remaining_size = (double) (ramsize + devsize);
664 double threshold = 0.0;
666 /* check parameters */
668 XBT_WARN("migrate a VM, but ramsize is zero");
670 if (max_downtime == 0) {
671 XBT_WARN("use the default max_downtime value 30ms");
675 /* Stage1: send all memory pages to the destination. */
676 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
677 start_dirty_page_tracking(ms->vm);
679 double computed_during_stage1 = 0;
681 double clock_prev_send = MSG_get_clock();
684 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
685 * transfer exceeds the timeout value. */
686 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
687 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
688 remaining_size -= sent;
689 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
691 if (sent < ramsize) {
692 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
694 } else if (sent > ramsize)
699 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
700 // Stop the dirty page tracking an return (there is no memory space to release)
701 stop_dirty_page_tracking(ms->vm);
705 double clock_post_send = MSG_get_clock();
706 mig_timeout -= (clock_post_send - clock_prev_send);
707 if (mig_timeout < 0) {
708 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
709 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
713 /* estimate bandwidth */
714 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
715 threshold = get_threshold_value(bandwidth, max_downtime);
716 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
720 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
723 int stage2_round = 0;
726 sg_size_t updated_size = 0;
727 if (stage2_round == 0) {
728 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
729 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
731 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
732 updated_size = get_updated_size(computed, dp_rate, dp_cap);
735 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
736 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
738 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
739 remaining_size += updated_size;
740 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
741 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
742 if (remaining_size < threshold)
746 double clock_prev_send = MSG_get_clock();
748 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
749 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
753 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
754 // Stop the dirty page tracking an return (there is no memory space to release)
755 stop_dirty_page_tracking(ms->vm);
758 double clock_post_send = MSG_get_clock();
760 if (sent == updated_size) {
761 /* timeout did not happen */
762 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
763 threshold = get_threshold_value(bandwidth, max_downtime);
764 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
765 remaining_size -= sent;
767 mig_timeout -= (clock_post_send - clock_prev_send);
768 xbt_assert(mig_timeout > 0);
770 } else if (sent < updated_size) {
771 /* When timeout happens, we move to stage 3. The size of memory pages
772 * updated before timeout must be added to the remaining size. */
773 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
774 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
775 remaining_size -= sent;
777 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
778 updated_size = get_updated_size(computed, dp_rate, dp_cap);
779 remaining_size += updated_size;
786 /* Stage3: stop the VM and copy the rest of states. */
787 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
788 simcall_vm_suspend(ms->vm);
789 stop_dirty_page_tracking(ms->vm);
792 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
793 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
796 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
797 // Stop the dirty page tracking an return (there is no memory space to release)
798 simcall_vm_resume(ms->vm);
802 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
803 // effectively the VM on the DST node.
804 XBT_DEBUG("mig: tx_done");
809 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
811 struct migration_session *ms = xbt_new(struct migration_session, 1);
815 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
816 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
818 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
819 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
821 // msg_process_t tx_process, rx_process;
822 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
823 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
826 char **argv = xbt_new(char *, 2);
827 argv[0] = pr_rx_name;
829 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
832 char **argv = xbt_new(char *, 2);
833 argv[0] = pr_tx_name;
835 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
839 /* wait until the migration have finished or on error has occurred */
841 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
842 msg_task_t task = nullptr;
843 msg_error_t ret = MSG_TIMEOUT;
844 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
845 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
847 xbt_free(ms->mbox_ctl);
851 //xbt_assert(ret == MSG_OK);
852 if(ret == MSG_HOST_FAILURE){
853 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
854 // other side. Hence, just throw the execption
855 XBT_ERROR("SRC crashes, throw an exception (m-control)");
856 //MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
860 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){
861 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
862 XBT_ERROR("DST crashes, throw an exception (m-control)");
866 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
867 xbt_assert(strcmp(task->name, expected_task_name) == 0);
868 xbt_free(expected_task_name);
869 MSG_task_destroy(task);
874 /** @brief Migrate the VM to the given host.
877 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
878 * depending on whether you want to do cold or hot migration.
880 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
883 * - One approach is ...
884 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
885 * receive the state of the source
886 * VM over network. We will finally destroy the source VM.
887 * - This behavior is similar to the way of migration in the real world.
888 * Even before a migration is completed, we will see a destination VM, consuming resources.
889 * - We have to relocate all processes. The existing process migration code will work for this?
890 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
892 * - Another one is ...
893 * We update the information of the given VM to place it to the destination physical host.
895 * The second one would be easier.
898 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
900 if(MSG_host_is_off(old_pm))
901 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
903 if(MSG_host_is_off(new_pm))
904 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
906 if (!MSG_vm_is_running(vm))
907 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
909 if (MSG_vm_is_migrating(vm))
910 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
912 msg_host_priv_t priv = sg_host_msg(vm);
913 priv->is_migrating = 1;
916 int ret = do_migration(vm, old_pm, new_pm);
918 priv->is_migrating = 0;
919 THROWF(host_error, 0, "SRC host failed during migration");
922 priv->is_migrating = 0;
923 THROWF(host_error, 0, "DST host failed during migration");
927 // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration
928 // process. In that case, the VM has been already assigned to the DST node.
929 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
930 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
933 /** @brief Immediately suspend the execution of all processes within the given VM.
936 * This function stops the execution of the VM. All the processes on this VM
937 * will pause. The state of the VM is preserved. We can later resume it again.
939 * No suspension cost occurs.
941 void MSG_vm_suspend(msg_vm_t vm)
943 if (MSG_vm_is_migrating(vm))
944 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
946 simcall_vm_suspend(vm);
948 XBT_DEBUG("vm_suspend done");
950 TRACE_msg_vm_suspend(vm);
953 /** @brief Resume the execution of the VM. All processes on the VM run again.
956 * No resume cost occurs.
958 void MSG_vm_resume(msg_vm_t vm)
960 simcall_vm_resume(vm);
962 TRACE_msg_vm_resume(vm);
966 /** @brief Immediately save the execution of all processes within the given VM.
969 * This function stops the execution of the VM. All the processes on this VM
970 * will pause. The state of the VM is preserved. We can later resume it again.
972 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
973 * or after, depending on the exact semantic of VM save to you.
975 void MSG_vm_save(msg_vm_t vm)
977 if (MSG_vm_is_migrating(vm))
978 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
981 TRACE_msg_vm_save(vm);
984 /** @brief Restore the execution of the VM. All processes on the VM run again.
987 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
988 * after, depending on the exact semantic of VM restore to you.
990 void MSG_vm_restore(msg_vm_t vm)
992 simcall_vm_restore(vm);
994 TRACE_msg_vm_restore(vm);
997 /** @brief Get the physical host of a given VM.
1000 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1002 return (msg_host_t) simcall_vm_get_pm(vm);
1005 /** @brief Set a CPU bound for a given VM.
1008 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1011 * On PM0, there are Task1 and VM0.
1012 * On VM0, there is Task2.
1013 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1015 * Task1\@PM0 gets 50%.
1016 * Task2\@VM0 gets 25%.
1017 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1019 * 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
1020 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
1022 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
1023 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
1024 * on the VM in share_resource() or we have to trap both the start and end of task execution.
1026 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
1028 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
1029 * computation task (or a VM) never exceeds the capacity of a CPU core.
1031 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1033 simcall_vm_set_bound(vm, bound);
1036 /** @brief Set the CPU affinity of a given VM.
1039 * This function changes the CPU affinity of a given VM. Usage is the same as
1040 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1042 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1044 msg_host_priv_t priv = sg_host_msg(vm);
1047 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1049 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *)(uintptr_t) mask, nullptr);
1051 msg_host_t pm_now = MSG_vm_get_pm(vm);
1053 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1054 simcall_vm_set_affinity(vm, pm, mask);
1056 XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));