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 <simgrid/s4u/host.hpp>
16 #include "msg_private.h"
17 #include "xbt/sysdep.h"
19 #include "simgrid/host.h"
21 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
24 /* **** ******** GENERAL ********* **** */
25 /** \ingroup m_vm_management
26 * \brief Returns the value of a given vm property
29 * \param name a property name
30 * \return value of a property (or nullptr if property not set)
32 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
34 return MSG_host_get_property_value(vm, name);
37 /** \ingroup m_vm_management
38 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
41 * \return a dict containing the properties
43 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
45 xbt_assert((vm != nullptr), "Invalid parameters (vm is nullptr)");
46 return vm->properties();
49 /** \ingroup m_host_management
50 * \brief Change the value of a given host property
53 * \param name a property name
54 * \param value what to change the property to
55 * \param free_ctn the freeing function to use to kill the value on need
57 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
59 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
62 /** \ingroup msg_vm_management
63 * \brief Finds a msg_vm_t using its name.
65 * This is a name directory service
66 * \param name the name of a vm.
67 * \return the corresponding vm
69 * Please note that a VM is a specific host. Hence, you should give a different name for each VM/PM.
71 msg_vm_t MSG_vm_get_by_name(const char *name)
73 return MSG_host_by_name(name);
76 /** \ingroup m_vm_management
78 * \brief Return the name of the #msg_host_t.
80 * This functions checks whether \a host is a valid pointer or not and return its name.
82 const char *MSG_vm_get_name(msg_vm_t vm)
84 return MSG_host_get_name(vm);
87 /* **** Check state of a VM **** */
88 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
90 return simcall_vm_get_state(vm) == state;
93 /** @brief Returns whether the given VM has just created, not running.
96 int MSG_vm_is_created(msg_vm_t vm)
98 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
101 /** @brief Returns whether the given VM is currently running
104 int MSG_vm_is_running(msg_vm_t vm)
106 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
109 /** @brief Returns whether the given VM is currently migrating
112 int MSG_vm_is_migrating(msg_vm_t vm)
114 msg_host_priv_t priv = sg_host_msg(vm);
115 return priv->is_migrating;
118 /** @brief Returns whether the given VM is currently suspended, not running.
121 int MSG_vm_is_suspended(msg_vm_t vm)
123 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
126 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
129 int MSG_vm_is_saving(msg_vm_t vm)
131 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
134 /** @brief Returns whether the given VM has been saved, not running.
137 int MSG_vm_is_saved(msg_vm_t vm)
139 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
142 /** @brief Returns whether the given VM is being restored, not running.
145 int MSG_vm_is_restoring(msg_vm_t vm)
147 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
150 /* **** ******** MSG vm actions ********* **** */
151 /** @brief Create a new VM with specified parameters.
153 * @param pm Physical machine that will host the VM
155 * @param ncpus [TODO]
156 * @param ramsize [TODO]
157 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
158 * @param disk_path (unused) Path to the image that boots
159 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
160 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
161 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
163 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name, int ncpus, int ramsize, int net_cap, char *disk_path,
164 int disksize, int mig_netspeed, int dp_intensity)
166 /* For the moment, intensity_rate is the percentage against the migration
168 double host_speed = MSG_host_get_speed(pm);
169 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
171 msg_vm_t vm = MSG_vm_create_core(pm, name);
172 s_vm_params_t params;
173 memset(¶ms, 0, sizeof(params));
174 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
175 //params.overcommit = 0;
177 params.skip_stage2 = 0;
178 params.max_downtime = 0.03;
179 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
180 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
181 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
183 //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate,
184 // params.mig_speed, dp_intensity, update_speed, host_speed);
185 vm->setParameters(¶ms);
190 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
193 * A VM is treated as a host. The name of the VM must be unique among all hosts.
195 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
197 /* make sure the VM of the same name does not exit */
199 simgrid::s4u::Host* ind_host_tmp = (simgrid::s4u::Host*) xbt_dict_get_or_null(host_list, name);
200 if (ind_host_tmp != nullptr && sg_host_simix(ind_host_tmp) != nullptr) {
201 XBT_ERROR("host %s already exits", name);
206 /* Note: ind_vm and vm_workstation point to the same elm object. */
207 /* Ask the SIMIX layer to create the surf vm resource */
208 sg_host_t ind_vm_workstation = (sg_host_t) simcall_vm_create(name, ind_pm);
210 msg_vm_t ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
212 XBT_DEBUG("A new VM (%s) has been created", name);
214 TRACE_msg_vm_create(name, ind_pm);
219 /** @brief Destroy a VM. Destroy the VM object from the simulation.
222 void MSG_vm_destroy(msg_vm_t vm)
224 if (MSG_vm_is_migrating(vm))
225 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
227 /* First, terminate all processes on the VM if necessary */
228 if (MSG_vm_is_running(vm))
229 simcall_vm_shutdown(vm);
231 if (!MSG_vm_is_created(vm)) {
232 XBT_CRITICAL("shutdown the given VM before destroying it");
236 /* Then, destroy the VM object */
237 simcall_vm_destroy(vm);
239 TRACE_msg_vm_end(vm);
242 /** @brief Start a vm (i.e., boot the guest operating system)
245 * If the VM cannot be started, an exception is generated.
247 void MSG_vm_start(msg_vm_t vm)
249 simcall_vm_start(vm);
251 TRACE_msg_vm_start(vm);
254 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
257 * FIXME: No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep() or something.
258 * I'm not quite sure.
260 void MSG_vm_shutdown(msg_vm_t vm)
262 /* msg_vm_t equals to msg_host_t */
263 simcall_vm_shutdown(vm);
265 // TRACE_msg_vm_(vm);
268 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
269 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
270 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
272 const char *vm_name = sg_host_get_name(vm);
273 const char *src_pm_name = sg_host_get_name(src_pm);
274 const char *dst_pm_name = sg_host_get_name(dst_pm);
276 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
279 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
281 const char *vm_name = sg_host_get_name(vm);
282 const char *src_pm_name = sg_host_get_name(src_pm);
283 const char *dst_pm_name = sg_host_get_name(dst_pm);
285 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
288 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
290 const char *vm_name = sg_host_get_name(vm);
291 const char *src_pm_name = sg_host_get_name(src_pm);
292 const char *dst_pm_name = sg_host_get_name(dst_pm);
294 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
297 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
299 const char *vm_name = sg_host_get_name(vm);
300 const char *src_pm_name = sg_host_get_name(src_pm);
301 const char *dst_pm_name = sg_host_get_name(dst_pm);
303 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
306 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
308 const char *vm_name = sg_host_get_name(vm);
309 const char *src_pm_name = sg_host_get_name(src_pm);
310 const char *dst_pm_name = sg_host_get_name(dst_pm);
312 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
315 struct migration_session {
320 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
321 * know the completion of the migration. */
323 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
327 static int migration_rx_fun(int argc, char *argv[])
329 XBT_DEBUG("mig: rx_start");
331 // The structure has been created in the do_migration function and should only be freed in the same place ;)
332 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
334 s_vm_params_t params;
335 ms->vm->parameters(¶ms);
339 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
343 msg_task_t task = nullptr;
344 ret = MSG_task_recv(&task, ms->mbox);
347 // An error occurred, clean the code and return
348 // The owner did not change, hence the task should be only destroyed on the other side
349 xbt_free(finalize_task_name);
354 if (strcmp(task->name, finalize_task_name) == 0)
357 MSG_task_destroy(task);
363 // Here Stage 1, 2 and 3 have been performed.
364 // Hence complete the migration
366 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
367 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
368 // structure at the beginning and free it at the end of each function)
369 msg_vm_t vm = ms->vm;
370 msg_host_t src_pm = ms->src_pm;
371 msg_host_t dst_pm = ms-> dst_pm;
372 msg_host_priv_t priv = sg_host_msg(vm);
374 // TODO: we have an issue, if the DST node is turning off during the three next calls, then the VM is in an inconsistent
375 // state. I should check with Takahiro in order to make this portion of code atomic
376 // /* deinstall the current affinity setting for the CPU */
377 // simcall_vm_set_affinity(vm, src_pm, 0);
379 // /* Update the vm location */
380 // simcall_vm_migrate(vm, dst_pm);
382 // /* Resume the VM */
383 // simcall_vm_resume(vm);
385 simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
387 /* install the affinity setting of the VM on the destination pm */
390 unsigned long affinity_mask =
391 (unsigned long)(uintptr_t) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *)dst_pm, sizeof(msg_host_t));
392 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
393 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
397 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
398 msg_host_priv_t priv = sg_host_msg(vm);
399 priv->is_migrating = 0;
400 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
401 sg_host_get_name(ms->dst_pm));
402 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
404 // Inform the SRC that the migration has been correctly performed
406 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
407 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
408 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
409 // xbt_assert(ret == MSG_OK);
410 if(ret == MSG_HOST_FAILURE){
411 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
412 // has been correctly migrated on the DST node
413 // TODO What does it mean ? What should we do ?
414 MSG_task_destroy(task);
415 } else if(ret == MSG_TRANSFER_FAILURE){
416 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
417 MSG_task_destroy(task);
422 xbt_free(finalize_task_name);
424 XBT_DEBUG("mig: rx_done");
428 static void reset_dirty_pages(msg_vm_t vm)
430 msg_host_priv_t priv = sg_host_msg(vm);
433 xbt_dict_cursor_t cursor = nullptr;
434 dirty_page_t dp = nullptr;
435 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
436 double remaining = MSG_task_get_flops_amount(dp->task);
437 dp->prev_clock = MSG_get_clock();
438 dp->prev_remaining = remaining;
440 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
444 static void start_dirty_page_tracking(msg_vm_t vm)
446 msg_host_priv_t priv = sg_host_msg(vm);
447 priv->dp_enabled = 1;
449 reset_dirty_pages(vm);
452 static void stop_dirty_page_tracking(msg_vm_t vm)
454 msg_host_priv_t priv = sg_host_msg(vm);
455 priv->dp_enabled = 0;
458 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
460 double computed = dp->prev_remaining - remaining;
461 double duration = clock - dp->prev_clock;
463 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
464 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
469 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
471 msg_host_priv_t priv = sg_host_msg(vm);
475 xbt_dict_cursor_t cursor = nullptr;
476 dirty_page_t dp = nullptr;
477 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
478 double remaining = MSG_task_get_flops_amount(dp->task);
480 double clock = MSG_get_clock();
482 // total += calc_updated_pages(key, vm, dp, remaining, clock);
483 total += get_computed(key, vm, dp, remaining, clock);
485 dp->prev_remaining = remaining;
486 dp->prev_clock = clock;
489 total += priv->dp_updated_by_deleted_tasks;
491 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
492 stage_for_fancy_debug, stage2_round_for_fancy_debug, total, priv->dp_updated_by_deleted_tasks);
494 priv->dp_updated_by_deleted_tasks = 0;
499 // TODO Is this code redundant with the information provided by
500 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
501 void MSG_host_add_task(msg_host_t host, msg_task_t task)
503 msg_host_priv_t priv = sg_host_msg(host);
504 double remaining = MSG_task_get_flops_amount(task);
505 char *key = bprintf("%s-%p", task->name, task);
507 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
510 /* It should be okay that we add a task onto a migrating VM. */
511 if (priv->dp_enabled) {
512 dp->prev_clock = MSG_get_clock();
513 dp->prev_remaining = remaining;
516 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == nullptr);
517 xbt_dict_set(priv->dp_objs, key, dp, nullptr);
518 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, priv->dp_enabled);
523 void MSG_host_del_task(msg_host_t host, msg_task_t task)
525 msg_host_priv_t priv = sg_host_msg(host);
527 char *key = bprintf("%s-%p", task->name, task);
529 dirty_page_t dp = (dirty_page_t) xbt_dict_get_or_null(priv->dp_objs, key);
530 xbt_assert(dp->task == task);
532 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
533 * the information for the lookup_() function that will called soon. */
534 if (priv->dp_enabled) {
535 double remaining = MSG_task_get_flops_amount(task);
536 double clock = MSG_get_clock();
537 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
538 double updated = get_computed(key, host, dp, remaining, clock);
540 priv->dp_updated_by_deleted_tasks += updated;
543 xbt_dict_remove(priv->dp_objs, key);
546 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
551 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,
552 int stage, int stage2_round, double mig_speed, double timeout)
555 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
556 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
560 double clock_sta = MSG_get_clock();
564 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
566 ret = MSG_task_send(task, mbox);
572 } else if (ret == MSG_TIMEOUT) {
573 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
574 sent = size - remaining;
575 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
578 /* FIXME: why try-and-catch is used here? */
579 if(ret == MSG_HOST_FAILURE){
580 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
581 MSG_task_destroy(task);
582 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
583 }else if(ret == MSG_TRANSFER_FAILURE){
584 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
585 MSG_task_destroy(task);
586 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
589 double clock_end = MSG_get_clock();
590 double duration = clock_end - clock_sta;
591 double actual_speed = size / duration;
594 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)",
595 stage, stage2_round, size, duration, actual_speed, mig_speed);
597 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)",
598 stage, size, duration, actual_speed, mig_speed);
603 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
605 double updated_size = computed * dp_rate;
606 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
607 if (updated_size > dp_cap) {
608 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
610 updated_size = dp_cap;
613 return (sg_size_t) updated_size;
616 static double send_stage1(struct migration_session *ms, sg_size_t ramsize, double mig_speed, double dp_rate,
619 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
620 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
621 sg_size_t remaining = ramsize;
622 double computed_total = 0;
624 while (remaining > 0) {
625 sg_size_t datasize = chunksize;
626 if (remaining < chunksize)
627 datasize = remaining;
629 remaining -= datasize;
630 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
631 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
632 computed_total += computed;
635 return computed_total;
638 static double get_threshold_value(double bandwidth, double max_downtime)
640 return max_downtime * bandwidth;
643 static int migration_tx_fun(int argc, char *argv[])
645 XBT_DEBUG("mig: tx_start");
647 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
648 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
650 s_vm_params_t params;
651 ms->vm->parameters(¶ms);
652 const sg_size_t ramsize = params.ramsize;
653 const sg_size_t devsize = params.devsize;
654 const int skip_stage1 = params.skip_stage1;
655 int skip_stage2 = params.skip_stage2;
656 const double dp_rate = params.dp_rate;
657 const double dp_cap = params.dp_cap;
658 const double mig_speed = params.mig_speed;
659 double max_downtime = params.max_downtime;
661 /* hard code it temporally. Fix Me */
662 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
663 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
665 double remaining_size = (double) (ramsize + devsize);
666 double threshold = 0.0;
668 /* check parameters */
670 XBT_WARN("migrate a VM, but ramsize is zero");
672 if (max_downtime == 0) {
673 XBT_WARN("use the default max_downtime value 30ms");
677 /* Stage1: send all memory pages to the destination. */
678 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
679 start_dirty_page_tracking(ms->vm);
681 double computed_during_stage1 = 0;
683 double clock_prev_send = MSG_get_clock();
686 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
687 * transfer exceeds the timeout value. */
688 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
689 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
690 remaining_size -= sent;
691 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
693 if (sent < ramsize) {
694 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
696 } else if (sent > ramsize)
701 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
702 // Stop the dirty page tracking an return (there is no memory space to release)
703 stop_dirty_page_tracking(ms->vm);
707 double clock_post_send = MSG_get_clock();
708 mig_timeout -= (clock_post_send - clock_prev_send);
709 if (mig_timeout < 0) {
710 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
711 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
715 /* estimate bandwidth */
716 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
717 threshold = get_threshold_value(bandwidth, max_downtime);
718 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
722 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
725 int stage2_round = 0;
728 sg_size_t updated_size = 0;
729 if (stage2_round == 0) {
730 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
731 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
733 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
734 updated_size = get_updated_size(computed, dp_rate, dp_cap);
737 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
738 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
740 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
741 remaining_size += updated_size;
742 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
743 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
744 if (remaining_size < threshold)
748 double clock_prev_send = MSG_get_clock();
750 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
751 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
755 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
756 // Stop the dirty page tracking an return (there is no memory space to release)
757 stop_dirty_page_tracking(ms->vm);
760 double clock_post_send = MSG_get_clock();
762 if (sent == updated_size) {
763 /* timeout did not happen */
764 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
765 threshold = get_threshold_value(bandwidth, max_downtime);
766 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
767 remaining_size -= sent;
769 mig_timeout -= (clock_post_send - clock_prev_send);
770 xbt_assert(mig_timeout > 0);
772 } else if (sent < updated_size) {
773 /* When timeout happens, we move to stage 3. The size of memory pages
774 * updated before timeout must be added to the remaining size. */
775 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
776 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
777 remaining_size -= sent;
779 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
780 updated_size = get_updated_size(computed, dp_rate, dp_cap);
781 remaining_size += updated_size;
788 /* Stage3: stop the VM and copy the rest of states. */
789 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
790 simcall_vm_suspend(ms->vm);
791 stop_dirty_page_tracking(ms->vm);
794 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
795 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
798 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
799 // Stop the dirty page tracking an return (there is no memory space to release)
800 simcall_vm_resume(ms->vm);
804 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
805 // effectively the VM on the DST node.
806 XBT_DEBUG("mig: tx_done");
811 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
813 struct migration_session *ms = xbt_new(struct migration_session, 1);
817 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
818 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
820 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
821 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
823 // msg_process_t tx_process, rx_process;
824 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
825 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
828 char **argv = xbt_new(char *, 2);
829 argv[0] = pr_rx_name;
831 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
834 char **argv = xbt_new(char *, 2);
835 argv[0] = pr_tx_name;
837 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
841 /* wait until the migration have finished or on error has occurred */
843 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
844 msg_task_t task = nullptr;
845 msg_error_t ret = MSG_TIMEOUT;
846 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
847 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
849 xbt_free(ms->mbox_ctl);
853 //xbt_assert(ret == MSG_OK);
854 if(ret == MSG_HOST_FAILURE){
855 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
856 // other side. Hence, just throw the execption
857 XBT_ERROR("SRC crashes, throw an exception (m-control)");
858 //MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
862 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){
863 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
864 XBT_ERROR("DST crashes, throw an exception (m-control)");
868 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
869 xbt_assert(strcmp(task->name, expected_task_name) == 0);
870 xbt_free(expected_task_name);
871 MSG_task_destroy(task);
876 /** @brief Migrate the VM to the given host.
879 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
880 * depending on whether you want to do cold or hot migration.
882 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
885 * - One approach is ...
886 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
887 * receive the state of the source
888 * VM over network. We will finally destroy the source VM.
889 * - This behavior is similar to the way of migration in the real world.
890 * Even before a migration is completed, we will see a destination VM, consuming resources.
891 * - We have to relocate all processes. The existing process migration code will work for this?
892 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
894 * - Another one is ...
895 * We update the information of the given VM to place it to the destination physical host.
897 * The second one would be easier.
900 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
902 if(MSG_host_is_off(old_pm))
903 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
905 if(MSG_host_is_off(new_pm))
906 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
908 if (!MSG_vm_is_running(vm))
909 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
911 if (MSG_vm_is_migrating(vm))
912 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
914 msg_host_priv_t priv = sg_host_msg(vm);
915 priv->is_migrating = 1;
918 int ret = do_migration(vm, old_pm, new_pm);
920 priv->is_migrating = 0;
921 THROWF(host_error, 0, "SRC host failed during migration");
924 priv->is_migrating = 0;
925 THROWF(host_error, 0, "DST host failed during migration");
929 // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration
930 // process. In that case, the VM has been already assigned to the DST node.
931 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
932 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
935 /** @brief Immediately suspend the execution of all processes within the given VM.
938 * This function stops the execution of the VM. All the processes on this VM
939 * will pause. The state of the VM is preserved. We can later resume it again.
941 * No suspension cost occurs.
943 void MSG_vm_suspend(msg_vm_t vm)
945 if (MSG_vm_is_migrating(vm))
946 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
948 simcall_vm_suspend(vm);
950 XBT_DEBUG("vm_suspend done");
952 TRACE_msg_vm_suspend(vm);
955 /** @brief Resume the execution of the VM. All processes on the VM run again.
958 * No resume cost occurs.
960 void MSG_vm_resume(msg_vm_t vm)
962 simcall_vm_resume(vm);
964 TRACE_msg_vm_resume(vm);
968 /** @brief Immediately save the execution of all processes within the given VM.
971 * This function stops the execution of the VM. All the processes on this VM
972 * will pause. The state of the VM is preserved. We can later resume it again.
974 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
975 * or after, depending on the exact semantic of VM save to you.
977 void MSG_vm_save(msg_vm_t vm)
979 if (MSG_vm_is_migrating(vm))
980 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
983 TRACE_msg_vm_save(vm);
986 /** @brief Restore the execution of the VM. All processes on the VM run again.
989 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
990 * after, depending on the exact semantic of VM restore to you.
992 void MSG_vm_restore(msg_vm_t vm)
994 simcall_vm_restore(vm);
996 TRACE_msg_vm_restore(vm);
999 /** @brief Get the physical host of a given VM.
1002 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1004 return (msg_host_t) simcall_vm_get_pm(vm);
1007 /** @brief Set a CPU bound for a given VM.
1010 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1013 * On PM0, there are Task1 and VM0.
1014 * On VM0, there is Task2.
1015 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1017 * Task1\@PM0 gets 50%.
1018 * Task2\@VM0 gets 25%.
1019 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1021 * 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
1022 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
1024 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
1025 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
1026 * on the VM in share_resource() or we have to trap both the start and end of task execution.
1028 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
1030 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
1031 * computation task (or a VM) never exceeds the capacity of a CPU core.
1033 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1035 simcall_vm_set_bound(vm, bound);
1038 /** @brief Set the CPU affinity of a given VM.
1041 * This function changes the CPU affinity of a given VM. Usage is the same as
1042 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1044 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1046 msg_host_priv_t priv = sg_host_msg(vm);
1049 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1051 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *)(uintptr_t) mask, nullptr);
1053 msg_host_t pm_now = MSG_vm_get_pm(vm);
1055 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1056 simcall_vm_set_affinity(vm, pm, mask);
1058 XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));