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 "src/instr/instr_private.h"
15 #include "src/msg/msg_private.h"
16 #include "src/plugins/vm/VirtualMachineImpl.hpp"
17 #include "src/plugins/vm/VmHostExt.hpp"
19 #include "simgrid/host.h"
20 #include "simgrid/simix.hpp"
26 double prev_remaining;
29 typedef struct dirty_page s_dirty_page;
30 typedef struct dirty_page* dirty_page_t;
32 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
34 /* **** ******** GENERAL ********* **** */
35 const char* MSG_vm_get_name(msg_vm_t vm)
37 return vm->getCname();
40 /** \ingroup m_vm_management
41 * \brief Set the parameters of a given host
44 * \param params a parameter object
46 void MSG_vm_set_params(msg_vm_t vm, vm_params_t params)
48 vm->setParameters(params);
51 /** \ingroup m_vm_management
52 * \brief Get the parameters of a given host
54 * \param vm the vm you are interested into
55 * \param params a prameter object
57 void MSG_vm_get_params(msg_vm_t vm, vm_params_t params)
59 vm->getParameters(params);
62 /* **** Check state of a VM **** */
63 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
65 return vm->pimpl_vm_ != nullptr && vm->pimpl_vm_->getState() == state;
68 /** @brief Returns whether the given VM has just created, not running.
71 int MSG_vm_is_created(msg_vm_t vm)
73 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
76 /** @brief Returns whether the given VM is currently running
79 int MSG_vm_is_running(msg_vm_t vm)
81 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
84 /** @brief Returns whether the given VM is currently migrating
87 int MSG_vm_is_migrating(msg_vm_t vm)
89 return vm->isMigrating();
92 /** @brief Returns whether the given VM is currently suspended, not running.
95 int MSG_vm_is_suspended(msg_vm_t vm)
97 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
100 /* **** ******** MSG vm actions ********* **** */
101 /** @brief Create a new VM with specified parameters.
103 * @param pm Physical machine that will host the VM
104 * @param name Must be unique
105 * @param coreAmount Must be >= 1
106 * @param ramsize [TODO]
107 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
108 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
110 msg_vm_t MSG_vm_create(msg_host_t pm, const char* name, int coreAmount, int ramsize, int mig_netspeed, int dp_intensity)
112 simgrid::vm::VmHostExt::ensureVmExtInstalled();
114 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
116 msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, coreAmount);
117 s_vm_params_t params;
118 memset(¶ms, 0, sizeof(params));
119 params.ramsize = static_cast<sg_size_t>(ramsize) * 1024 * 1024;
121 params.skip_stage2 = 0;
122 params.max_downtime = 0.03;
123 params.mig_speed = static_cast<double>(mig_netspeed) * 1024 * 1024; // mig_speed
124 params.dp_intensity = static_cast<double>(dp_intensity) / 100;
125 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
127 XBT_DEBUG("migspeed : %f intensity mem : %d", params.mig_speed, dp_intensity);
128 vm->setParameters(¶ms);
133 /** @brief Create a new VM object with the default parameters
136 * A VM is treated as a host. The name of the VM must be unique among all hosts.
138 msg_vm_t MSG_vm_create_core(msg_host_t pm, const char* name)
140 xbt_assert(sg_host_by_name(name) == nullptr,
141 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
143 return new simgrid::s4u::VirtualMachine(name, pm, 1);
145 /** @brief Create a new VM object with the default parameters, but with a specified amount of cores
148 * A VM is treated as a host. The name of the VM must be unique among all hosts.
150 msg_vm_t MSG_vm_create_multicore(msg_host_t pm, const char* name, int coreAmount)
152 xbt_assert(sg_host_by_name(name) == nullptr,
153 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
155 return new simgrid::s4u::VirtualMachine(name, pm, coreAmount);
158 /** @brief Destroy a VM. Destroy the VM object from the simulation.
161 void MSG_vm_destroy(msg_vm_t vm)
163 if (vm->isMigrating())
164 THROWF(vm_error, 0, "Cannot destroy VM '%s', which is migrating.", vm->getCname());
166 /* First, terminate all processes on the VM if necessary */
167 if (MSG_vm_is_running(vm))
170 /* Then, destroy the VM object */
171 simgrid::simix::kernelImmediate([vm]() { vm->destroy(); });
173 if (TRACE_msg_vm_is_enabled()) {
174 container_t container = PJ_container_get(vm->getCname());
175 PJ_container_remove_from_parent(container);
176 PJ_container_free(container);
180 /** @brief Start a vm (i.e., boot the guest operating system)
183 * If the VM cannot be started (because of memory over-provisioning), an exception is generated.
185 void MSG_vm_start(msg_vm_t vm)
188 if (TRACE_msg_vm_is_enabled()) {
189 container_t vm_container = PJ_container_get(vm->getCname());
190 simgrid::instr::Type* type = PJ_type_get("MSG_VM_STATE", vm_container->type_);
191 simgrid::instr::Value* val = simgrid::instr::Value::get_or_new("start", "0 0 1", type); // start is blue
192 new simgrid::instr::PushStateEvent(MSG_get_clock(), vm_container, type, val);
196 /** @brief Immediately kills all processes within the given VM.
199 * Any memory that they allocated will be leaked, unless you used #MSG_process_on_exit().
201 * No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep().
203 void MSG_vm_shutdown(msg_vm_t vm)
205 smx_actor_t issuer = SIMIX_process_self();
206 simgrid::simix::kernelImmediate([vm, issuer]() { vm->pimpl_vm_->shutdown(issuer); });
208 // Make sure that processes in the VM are killed in this scheduling round before processing (eg with the VM destroy)
209 MSG_process_sleep(0.);
212 static std::string get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
214 return std::string("__pr_mig_tx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
217 static std::string get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
219 return std::string("__pr_mig_rx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
222 static std::string get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
224 return std::string("__task_mig_stage") + std::to_string(stage) + ":" + vm->getCname() + "(" + src_pm->getCname() +
225 "-" + dst_pm->getCname() + ")";
228 struct migration_session {
233 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
234 * know the completion of the migration. */
236 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
240 static int migration_rx_fun(int argc, char *argv[])
242 XBT_DEBUG("mig: rx_start");
244 // The structure has been created in the do_migration function and should only be freed in the same place ;)
245 struct migration_session* ms = static_cast<migration_session*>(MSG_process_get_data(MSG_process_self()));
247 bool received_finalize = false;
249 std::string finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
250 while (not received_finalize) {
251 msg_task_t task = nullptr;
252 int ret = MSG_task_recv(&task, ms->mbox);
255 // An error occurred, clean the code and return
256 // The owner did not change, hence the task should be only destroyed on the other side
260 if (finalize_task_name == task->name)
261 received_finalize = 1;
263 MSG_task_destroy(task);
266 // Here Stage 1, 2 and 3 have been performed.
267 // Hence complete the migration
269 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
270 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
271 // structure at the beginning and free it at the end of each function)
272 simgrid::s4u::VirtualMachine* vm = ms->vm;
273 msg_host_t dst_pm = ms->dst_pm;
275 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
276 simgrid::simix::kernelImmediate([vm, dst_pm]() {
277 /* Update the vm location */
278 /* precopy migration makes the VM temporally paused */
279 xbt_assert(vm->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
281 /* Update the vm location and resume it */
282 vm->pimpl_vm_->setPm(dst_pm);
283 vm->pimpl_vm_->resume();
287 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
288 vm->pimpl_vm_->isMigrating = false;
289 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->getCname(), ms->src_pm->getCname(), ms->dst_pm->getCname());
291 if (TRACE_msg_vm_is_enabled()) {
292 static long long int counter = 0;
293 char key[INSTR_DEFAULT_STR_SIZE];
294 snprintf(key, INSTR_DEFAULT_STR_SIZE, "%lld", counter);
298 container_t msg = PJ_container_get(vm->getCname());
299 simgrid::instr::Type* type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
300 new simgrid::instr::StartLinkEvent(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
302 // destroy existing container of this vm
303 container_t existing_container = PJ_container_get(vm->getCname());
304 PJ_container_remove_from_parent(existing_container);
305 PJ_container_free(existing_container);
307 // create new container on the new_host location
308 PJ_container_new(vm->getCname(), simgrid::instr::INSTR_MSG_VM, PJ_container_get(ms->dst_pm->getCname()));
311 msg = PJ_container_get(vm->getCname());
312 type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
313 new simgrid::instr::EndLinkEvent(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
316 // Inform the SRC that the migration has been correctly performed
317 std::string task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
318 msg_task_t task = MSG_task_create(task_name.c_str(), 0, 0, nullptr);
319 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
320 // xbt_assert(ret == MSG_OK);
321 if(ret == MSG_HOST_FAILURE){
322 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
323 // has been correctly migrated on the DST node
324 // TODO What does it mean ? What should we do ?
325 MSG_task_destroy(task);
326 } else if(ret == MSG_TRANSFER_FAILURE){
327 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
328 MSG_task_destroy(task);
331 XBT_DEBUG("mig: rx_done");
335 static void start_dirty_page_tracking(msg_vm_t vm)
337 vm->pimpl_vm_->dp_enabled = 1;
338 if (vm->pimpl_vm_->dp_objs.empty())
341 for (auto const& elm : vm->pimpl_vm_->dp_objs) {
342 dirty_page_t dp = elm.second;
343 double remaining = MSG_task_get_flops_amount(dp->task);
344 dp->prev_clock = MSG_get_clock();
345 dp->prev_remaining = remaining;
346 XBT_DEBUG("%s@%s remaining %f", elm.first.c_str(), vm->getCname(), remaining);
350 static void stop_dirty_page_tracking(msg_vm_t vm)
352 vm->pimpl_vm_->dp_enabled = 0;
355 static double get_computed(const char* key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
357 double computed = dp->prev_remaining - remaining;
358 double duration = clock - dp->prev_clock;
360 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", key, vm->getCname(), computed,
361 dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
366 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
370 for (auto const& elm : vm->pimpl_vm_->dp_objs) {
371 const char* key = elm.first.c_str();
372 dirty_page_t dp = elm.second;
373 double remaining = MSG_task_get_flops_amount(dp->task);
375 double clock = MSG_get_clock();
377 // total += calc_updated_pages(key, vm, dp, remaining, clock);
378 total += get_computed(key, vm, dp, remaining, clock);
380 dp->prev_remaining = remaining;
381 dp->prev_clock = clock;
384 total += vm->pimpl_vm_->dp_updated_by_deleted_tasks;
386 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
387 stage2_round_for_fancy_debug, total, vm->pimpl_vm_->dp_updated_by_deleted_tasks);
389 vm->pimpl_vm_->dp_updated_by_deleted_tasks = 0;
394 // TODO Is this code redundant with the information provided by
395 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
396 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
397 void MSG_host_add_task(msg_host_t host, msg_task_t task)
399 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
403 double remaining = MSG_task_get_flops_amount(task);
404 char *key = bprintf("%s-%p", task->name, task);
406 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
408 if (vm->pimpl_vm_->dp_enabled) {
409 dp->prev_clock = MSG_get_clock();
410 dp->prev_remaining = remaining;
412 vm->pimpl_vm_->dp_objs.insert({key, dp});
413 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, host->getCname(), remaining, vm->pimpl_vm_->dp_enabled);
418 void MSG_host_del_task(msg_host_t host, msg_task_t task)
420 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
424 char *key = bprintf("%s-%p", task->name, task);
425 dirty_page_t dp = nullptr;
426 if (vm->pimpl_vm_->dp_objs.find(key) != vm->pimpl_vm_->dp_objs.end())
427 dp = vm->pimpl_vm_->dp_objs.at(key);
428 xbt_assert(dp && dp->task == task);
430 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
431 * the information for the lookup_() function that will called soon. */
432 if (vm->pimpl_vm_->dp_enabled) {
433 double remaining = MSG_task_get_flops_amount(task);
434 double clock = MSG_get_clock();
435 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
436 double updated = get_computed(key, vm, dp, remaining, clock); // was host instead of vm
438 vm->pimpl_vm_->dp_updated_by_deleted_tasks += updated;
441 vm->pimpl_vm_->dp_objs.erase(key);
444 XBT_DEBUG("del %s on %s", key, host->getCname());
448 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,
449 int stage, int stage2_round, double mig_speed, double timeout)
452 std::string task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
453 msg_task_t task = MSG_task_create(task_name.c_str(), 0, static_cast<double>(size), nullptr);
455 double clock_sta = MSG_get_clock();
459 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
461 ret = MSG_task_send(task, mbox);
465 } else if (ret == MSG_TIMEOUT) {
466 sg_size_t remaining = static_cast<sg_size_t>(MSG_task_get_remaining_communication(task));
467 sent = size - remaining;
468 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
471 /* FIXME: why try-and-catch is used here? */
472 if(ret == MSG_HOST_FAILURE){
473 XBT_DEBUG("SRC host failed during migration of %s (stage %d)", vm->getCname(), stage);
474 MSG_task_destroy(task);
475 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", vm->getCname(), stage);
476 }else if(ret == MSG_TRANSFER_FAILURE){
477 XBT_DEBUG("DST host failed during migration of %s (stage %d)", vm->getCname(), stage);
478 MSG_task_destroy(task);
479 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", vm->getCname(), stage);
482 double clock_end = MSG_get_clock();
483 double duration = clock_end - clock_sta;
484 double actual_speed = size / duration;
487 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
488 actual_speed, mig_speed);
490 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
496 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
498 double updated_size = computed * dp_rate;
499 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
500 if (updated_size > dp_cap) {
501 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
503 updated_size = dp_cap;
506 return static_cast<sg_size_t>(updated_size);
509 static int migration_tx_fun(int argc, char *argv[])
511 XBT_DEBUG("mig: tx_start");
513 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
514 migration_session *ms = static_cast<migration_session *>(MSG_process_get_data(MSG_process_self()));
516 double host_speed = ms->vm->pimpl_vm_->getPm()->getSpeed();
517 s_vm_params_t params;
518 ms->vm->getParameters(¶ms);
519 const sg_size_t ramsize = params.ramsize;
520 const sg_size_t devsize = params.devsize;
521 const int skip_stage1 = params.skip_stage1;
522 int skip_stage2 = params.skip_stage2;
523 const double dp_rate = host_speed ? (params.mig_speed * params.dp_intensity) / host_speed : 1;
524 const double dp_cap = params.dp_cap;
525 const double mig_speed = params.mig_speed;
526 double max_downtime = params.max_downtime;
528 double mig_timeout = 10000000.0;
530 double remaining_size = static_cast<double>(ramsize + devsize);
531 double threshold = 0.0;
533 /* check parameters */
535 XBT_WARN("migrate a VM, but ramsize is zero");
537 if (max_downtime <= 0) {
538 XBT_WARN("use the default max_downtime value 30ms");
542 /* Stage1: send all memory pages to the destination. */
543 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
544 start_dirty_page_tracking(ms->vm);
546 double computed_during_stage1 = 0;
547 if (not skip_stage1) {
548 double clock_prev_send = MSG_get_clock();
551 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
552 * transfer exceeds the timeout value. */
553 XBT_VERB("Stage 1: Gonna send %llu bytes", ramsize);
554 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
555 remaining_size -= sent;
556 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
558 if (sent < ramsize) {
559 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
561 } else if (sent > ramsize)
566 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
567 // Stop the dirty page tracking an return (there is no memory space to release)
568 stop_dirty_page_tracking(ms->vm);
572 double clock_post_send = MSG_get_clock();
573 mig_timeout -= (clock_post_send - clock_prev_send);
574 if (mig_timeout < 0) {
575 XBT_VERB("The duration of stage 1 exceeds the timeout value, skip stage 2");
579 /* estimate bandwidth */
580 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
581 threshold = bandwidth * max_downtime;
582 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
586 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
587 if (not skip_stage2) {
589 int stage2_round = 0;
592 sg_size_t updated_size = 0;
593 if (stage2_round == 0) {
594 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
595 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
597 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
598 updated_size = get_updated_size(computed, dp_rate, dp_cap);
601 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
602 updated_size, computed_during_stage1, dp_rate, dp_cap);
604 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
605 remaining_size += updated_size;
606 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
607 (remaining_size < threshold) ? "<" : ">", threshold);
608 if (remaining_size < threshold)
612 double clock_prev_send = MSG_get_clock();
614 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
615 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
617 } catch (xbt_ex& e) {
618 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
620 // Stop the dirty page tracking an return (there is no memory space to release)
621 stop_dirty_page_tracking(ms->vm);
624 double clock_post_send = MSG_get_clock();
626 if (sent == updated_size) {
627 /* timeout did not happen */
628 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
629 threshold = bandwidth * max_downtime;
630 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
631 remaining_size -= sent;
633 mig_timeout -= (clock_post_send - clock_prev_send);
634 xbt_assert(mig_timeout > 0);
636 } else if (sent < updated_size) {
637 /* When timeout happens, we move to stage 3. The size of memory pages
638 * updated before timeout must be added to the remaining size. */
639 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
640 updated_size, (clock_post_send - clock_prev_send));
641 remaining_size -= sent;
643 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
644 updated_size = get_updated_size(computed, dp_rate, dp_cap);
645 remaining_size += updated_size;
652 /* Stage3: stop the VM and copy the rest of states. */
653 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
654 simgrid::vm::VirtualMachineImpl* pimpl = ms->vm->pimpl_vm_;
655 pimpl->setState(SURF_VM_STATE_RUNNING); // FIXME: this bypass of the checks in suspend() is not nice
656 pimpl->isMigrating = false; // FIXME: this bypass of the checks in suspend() is not nice
657 pimpl->suspend(SIMIX_process_self());
658 stop_dirty_page_tracking(ms->vm);
661 XBT_DEBUG("Stage 3: Gonna send %f bytes", remaining_size);
662 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, static_cast<sg_size_t>(remaining_size), ms->mbox, 3, 0,
666 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
667 // Stop the dirty page tracking an return (there is no memory space to release)
668 ms->vm->pimpl_vm_->resume();
672 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
673 // effectively the VM on the DST node.
674 XBT_DEBUG("mig: tx_done");
679 /** @brief Migrate the VM to the given host.
682 void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm)
685 * - One approach is ...
686 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
687 * receive the state of the source
688 * VM over network. We will finally destroy the source VM.
689 * - This behavior is similar to the way of migration in the real world.
690 * Even before a migration is completed, we will see a destination VM, consuming resources.
691 * - We have to relocate all processes. The existing process migration code will work for this?
692 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
694 * - Another one is ...
695 * We update the information of the given VM to place it to the destination physical host.
697 * The second one would be easier.
700 msg_host_t src_pm = vm->pimpl_vm_->getPm();
703 THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->getCname(), src_pm->getCname());
705 THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->getCname(), dst_pm->getCname());
706 if (not MSG_vm_is_running(vm))
707 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->getCname());
708 if (vm->isMigrating())
709 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->getCname());
711 vm->pimpl_vm_->isMigrating = true;
713 struct migration_session *ms = xbt_new(struct migration_session, 1);
718 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
719 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
720 ms->mbox_ctl = bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
721 ms->mbox = bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
723 std::string pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
724 std::string pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
726 MSG_process_create(pr_rx_name.c_str(), migration_rx_fun, ms, dst_pm);
728 MSG_process_create(pr_tx_name.c_str(), migration_tx_fun, ms, src_pm);
730 /* wait until the migration have finished or on error has occurred */
731 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
732 msg_task_t task = nullptr;
733 msg_error_t ret = MSG_task_receive(&task, ms->mbox_ctl);
735 vm->pimpl_vm_->isMigrating = false;
737 xbt_free(ms->mbox_ctl);
741 if (ret == MSG_HOST_FAILURE) {
742 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
743 // other side. Hence, just throw the execption
744 XBT_ERROR("SRC crashes, throw an exception (m-control)");
745 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
747 THROWF(host_error, 0, "Source host '%s' failed during the migration of VM '%s'.", src_pm->getCname(),
749 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
750 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
751 XBT_ERROR("DST crashes, throw an exception (m-control)");
752 THROWF(host_error, 0, "Destination host '%s' failed during the migration of VM '%s'.", dst_pm->getCname(),
756 xbt_assert(get_mig_task_name(vm, src_pm, dst_pm, 4) == task->name);
757 MSG_task_destroy(task);
760 /** @brief Immediately suspend the execution of all processes within the given VM.
763 * This function stops the execution of the VM. All the processes on this VM
764 * will pause. The state of the VM is preserved. We can later resume it again.
766 * No suspension cost occurs.
768 void MSG_vm_suspend(msg_vm_t vm)
770 smx_actor_t issuer = SIMIX_process_self();
771 simgrid::simix::kernelImmediate([vm, issuer]() { vm->pimpl_vm_->suspend(issuer); });
773 XBT_DEBUG("vm_suspend done");
775 if (TRACE_msg_vm_is_enabled()) {
776 container_t vm_container = PJ_container_get(vm->getCname());
777 simgrid::instr::Type* type = PJ_type_get("MSG_VM_STATE", vm_container->type_);
778 simgrid::instr::Value* val = simgrid::instr::Value::get_or_new("suspend", "1 0 0", type); // suspend is red
779 new simgrid::instr::PushStateEvent(MSG_get_clock(), vm_container, type, val);
783 /** @brief Resume the execution of the VM. All processes on the VM run again.
786 * No resume cost occurs.
788 void MSG_vm_resume(msg_vm_t vm)
790 vm->pimpl_vm_->resume();
792 if (TRACE_msg_vm_is_enabled()) {
793 container_t vm_container = PJ_container_get(vm->getCname());
794 simgrid::instr::Type* type = PJ_type_get("MSG_VM_STATE", vm_container->type_);
795 new simgrid::instr::PopStateEvent(MSG_get_clock(), vm_container, type);
799 /** @brief Get the physical host of a given VM.
802 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
807 /** @brief Set a CPU bound for a given VM.
810 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
813 * On PM0, there are Task1 and VM0.
814 * On VM0, there is Task2.
815 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
817 * Task1\@PM0 gets 50%.
818 * Task2\@VM0 gets 25%.
819 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
821 * 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
822 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
824 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
825 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
826 * on the VM in share_resource() or we have to trap both the start and end of task execution.
828 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
830 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
831 * computation task (or a VM) never exceeds the capacity of a CPU core.
833 void MSG_vm_set_bound(msg_vm_t vm, double bound)
835 simgrid::simix::kernelImmediate([vm, bound]() { vm->pimpl_vm_->setBound(bound); });