1 /* Copyright (c) 2012-2017. 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.hpp"
15 #include "src/msg/msg_private.hpp"
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 s_dirty_page* dirty_page_t;
31 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
33 /* **** ******** GENERAL ********* **** */
34 const char* MSG_vm_get_name(msg_vm_t vm)
36 return vm->getCname();
39 /** \ingroup m_vm_management
40 * \brief Set the parameters of a given host
43 * \param params a parameter object
45 void MSG_vm_set_params(msg_vm_t vm, vm_params_t params)
47 vm->setParameters(params);
50 /** \ingroup m_vm_management
51 * \brief Get the parameters of a given host
53 * \param vm the vm you are interested into
54 * \param params a prameter object
56 void MSG_vm_get_params(msg_vm_t vm, vm_params_t params)
58 vm->getParameters(params);
61 /* **** Check state of a VM **** */
62 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
64 return vm->pimpl_vm_ != nullptr && vm->pimpl_vm_->getState() == state;
67 /** @brief Returns whether the given VM has just created, not running.
70 int MSG_vm_is_created(msg_vm_t vm)
72 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
75 /** @brief Returns whether the given VM is currently running
78 int MSG_vm_is_running(msg_vm_t vm)
80 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
83 /** @brief Returns whether the given VM is currently migrating
86 int MSG_vm_is_migrating(msg_vm_t vm)
88 return vm->isMigrating();
91 /** @brief Returns whether the given VM is currently suspended, not running.
94 int MSG_vm_is_suspended(msg_vm_t vm)
96 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
99 /* **** ******** MSG vm actions ********* **** */
100 /** @brief Create a new VM with specified parameters.
102 * @param pm Physical machine that will host the VM
103 * @param name Must be unique
104 * @param coreAmount Must be >= 1
105 * @param ramsize [TODO]
106 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
107 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
109 msg_vm_t MSG_vm_create(msg_host_t pm, const char* name, int coreAmount, int ramsize, int mig_netspeed, int dp_intensity)
111 simgrid::vm::VmHostExt::ensureVmExtInstalled();
113 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
115 msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, coreAmount);
116 s_vm_params_t params{};
117 params.ramsize = static_cast<sg_size_t>(ramsize) * 1024 * 1024;
119 params.skip_stage2 = 0;
120 params.max_downtime = 0.03;
121 params.mig_speed = static_cast<double>(mig_netspeed) * 1024 * 1024; // mig_speed
122 params.dp_intensity = static_cast<double>(dp_intensity) / 100;
123 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
125 XBT_DEBUG("migspeed : %f intensity mem : %d", params.mig_speed, dp_intensity);
126 vm->setParameters(¶ms);
131 /** @brief Create a new VM object with the default parameters
134 * A VM is treated as a host. The name of the VM must be unique among all hosts.
136 msg_vm_t MSG_vm_create_core(msg_host_t pm, const char* name)
138 xbt_assert(sg_host_by_name(name) == nullptr,
139 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
141 return new simgrid::s4u::VirtualMachine(name, pm, 1);
143 /** @brief Create a new VM object with the default parameters, but with a specified amount of cores
146 * A VM is treated as a host. The name of the VM must be unique among all hosts.
148 msg_vm_t MSG_vm_create_multicore(msg_host_t pm, const char* name, int coreAmount)
150 xbt_assert(sg_host_by_name(name) == nullptr,
151 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
153 return new simgrid::s4u::VirtualMachine(name, pm, coreAmount);
156 /** @brief Destroy a VM. Destroy the VM object from the simulation.
159 void MSG_vm_destroy(msg_vm_t vm)
161 if (vm->isMigrating())
162 THROWF(vm_error, 0, "Cannot destroy VM '%s', which is migrating.", vm->getCname());
164 /* First, terminate all processes on the VM if necessary */
165 if (MSG_vm_is_running(vm))
168 /* Then, destroy the VM object */
169 simgrid::simix::kernelImmediate([vm]() { vm->destroy(); });
171 if (TRACE_msg_vm_is_enabled()) {
172 container_t container = simgrid::instr::Container::byName(vm->getName());
173 container->removeFromParent();
178 /** @brief Start a vm (i.e., boot the guest operating system)
181 * If the VM cannot be started (because of memory over-provisioning), an exception is generated.
183 void MSG_vm_start(msg_vm_t vm)
186 if (TRACE_msg_vm_is_enabled()) {
187 simgrid::instr::StateType* state = simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE");
188 state->addEntityValue("start", "0 0 1"); // start is blue
189 state->pushEvent("start");
193 /** @brief Immediately kills all processes within the given VM.
196 * Any memory that they allocated will be leaked, unless you used #MSG_process_on_exit().
198 * No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep().
200 void MSG_vm_shutdown(msg_vm_t vm)
202 smx_actor_t issuer = SIMIX_process_self();
203 simgrid::simix::kernelImmediate([vm, issuer]() { vm->pimpl_vm_->shutdown(issuer); });
205 // Make sure that processes in the VM are killed in this scheduling round before processing (eg with the VM destroy)
206 MSG_process_sleep(0.);
209 static std::string get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
211 return std::string("__pr_mig_tx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
214 static std::string get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
216 return std::string("__pr_mig_rx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
219 static std::string get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
221 return std::string("__task_mig_stage") + std::to_string(stage) + ":" + vm->getCname() + "(" + src_pm->getCname() +
222 "-" + dst_pm->getCname() + ")";
225 struct migration_session {
230 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
231 * know the completion of the migration. */
233 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
237 static int migration_rx_fun(int argc, char *argv[])
239 XBT_DEBUG("mig: rx_start");
241 // The structure has been created in the do_migration function and should only be freed in the same place ;)
242 migration_session* ms = static_cast<migration_session*>(MSG_process_get_data(MSG_process_self()));
244 bool received_finalize = false;
246 std::string finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
247 while (not received_finalize) {
248 msg_task_t task = nullptr;
249 int ret = MSG_task_recv(&task, ms->mbox);
252 // An error occurred, clean the code and return
253 // The owner did not change, hence the task should be only destroyed on the other side
257 if (finalize_task_name == task->name)
258 received_finalize = 1;
260 MSG_task_destroy(task);
263 // Here Stage 1, 2 and 3 have been performed.
264 // Hence complete the migration
266 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
267 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
268 // structure at the beginning and free it at the end of each function)
269 simgrid::s4u::VirtualMachine* vm = ms->vm;
270 msg_host_t dst_pm = ms->dst_pm;
272 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
273 simgrid::simix::kernelImmediate([vm, dst_pm]() {
274 /* Update the vm location */
275 /* precopy migration makes the VM temporally paused */
276 xbt_assert(vm->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
278 /* Update the vm location and resume it */
279 vm->pimpl_vm_->setPm(dst_pm);
280 vm->pimpl_vm_->resume();
284 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
285 vm->pimpl_vm_->isMigrating = false;
286 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->getCname(), ms->src_pm->getCname(), ms->dst_pm->getCname());
288 if (TRACE_msg_vm_is_enabled()) {
289 static long long int counter = 0;
290 std::string key = std::to_string(counter);
294 container_t msg = simgrid::instr::Container::byName(vm->getName());
295 simgrid::instr::Container::getRoot()->getLink("MSG_VM_LINK")->startEvent(msg, "M", key);
297 // destroy existing container of this vm
298 container_t existing_container = simgrid::instr::Container::byName(vm->getName());
299 existing_container->removeFromParent();
300 delete existing_container;
302 // create new container on the new_host location
303 new simgrid::instr::Container(vm->getCname(), "MSG_VM", simgrid::instr::Container::byName(ms->dst_pm->getName()));
306 msg = simgrid::instr::Container::byName(vm->getName());
307 simgrid::instr::Container::getRoot()->getLink("MSG_VM_LINK")->endEvent(msg, "M", key);
310 // Inform the SRC that the migration has been correctly performed
311 std::string task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
312 msg_task_t task = MSG_task_create(task_name.c_str(), 0, 0, nullptr);
313 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
314 if(ret == MSG_HOST_FAILURE){
315 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
316 // has been correctly migrated on the DST node
317 // TODO What does it mean ? What should we do ?
318 MSG_task_destroy(task);
319 } else if(ret == MSG_TRANSFER_FAILURE){
320 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
321 MSG_task_destroy(task);
324 XBT_DEBUG("mig: rx_done");
328 static void start_dirty_page_tracking(msg_vm_t vm)
330 vm->pimpl_vm_->dp_enabled = 1;
331 if (vm->pimpl_vm_->dp_objs.empty())
334 for (auto const& elm : vm->pimpl_vm_->dp_objs) {
335 dirty_page_t dp = elm.second;
336 double remaining = MSG_task_get_flops_amount(dp->task);
337 dp->prev_clock = MSG_get_clock();
338 dp->prev_remaining = remaining;
339 XBT_DEBUG("%s@%s remaining %f", elm.first.c_str(), vm->getCname(), remaining);
343 static void stop_dirty_page_tracking(msg_vm_t vm)
345 vm->pimpl_vm_->dp_enabled = 0;
348 static double get_computed(const char* key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
350 double computed = dp->prev_remaining - remaining;
351 double duration = clock - dp->prev_clock;
353 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", key, vm->getCname(), computed,
354 dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
359 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
363 for (auto const& elm : vm->pimpl_vm_->dp_objs) {
364 const char* key = elm.first.c_str();
365 dirty_page_t dp = elm.second;
366 double remaining = MSG_task_get_flops_amount(dp->task);
368 double clock = MSG_get_clock();
370 total += get_computed(key, vm, dp, remaining, clock);
372 dp->prev_remaining = remaining;
373 dp->prev_clock = clock;
376 total += vm->pimpl_vm_->dp_updated_by_deleted_tasks;
378 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
379 stage2_round_for_fancy_debug, total, vm->pimpl_vm_->dp_updated_by_deleted_tasks);
381 vm->pimpl_vm_->dp_updated_by_deleted_tasks = 0;
386 // TODO Is this code redundant with the information provided by
387 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
388 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
389 void MSG_host_add_task(msg_host_t host, msg_task_t task)
391 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
395 double remaining = MSG_task_get_flops_amount(task);
396 char *key = bprintf("%s-%p", task->name, task);
398 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
400 if (vm->pimpl_vm_->dp_enabled) {
401 dp->prev_clock = MSG_get_clock();
402 dp->prev_remaining = remaining;
404 vm->pimpl_vm_->dp_objs.insert({key, dp});
405 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, host->getCname(), remaining, vm->pimpl_vm_->dp_enabled);
410 void MSG_host_del_task(msg_host_t host, msg_task_t task)
412 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
416 char *key = bprintf("%s-%p", task->name, task);
417 dirty_page_t dp = nullptr;
418 if (vm->pimpl_vm_->dp_objs.find(key) != vm->pimpl_vm_->dp_objs.end())
419 dp = vm->pimpl_vm_->dp_objs.at(key);
420 xbt_assert(dp && dp->task == task);
422 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
423 * the information for the lookup_() function that will called soon. */
424 if (vm->pimpl_vm_->dp_enabled) {
425 double remaining = MSG_task_get_flops_amount(task);
426 double clock = MSG_get_clock();
427 double updated = get_computed(key, vm, dp, remaining, clock); // was host instead of vm
429 vm->pimpl_vm_->dp_updated_by_deleted_tasks += updated;
432 vm->pimpl_vm_->dp_objs.erase(key);
435 XBT_DEBUG("del %s on %s", key, host->getCname());
439 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,
440 int stage, int stage2_round, double mig_speed, double timeout)
443 std::string task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
444 msg_task_t task = MSG_task_create(task_name.c_str(), 0, static_cast<double>(size), nullptr);
446 double clock_sta = MSG_get_clock();
450 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
452 ret = MSG_task_send(task, mbox);
456 } else if (ret == MSG_TIMEOUT) {
457 sg_size_t remaining = static_cast<sg_size_t>(MSG_task_get_remaining_communication(task));
458 sent = size - remaining;
459 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
462 /* FIXME: why try-and-catch is used here? */
463 if(ret == MSG_HOST_FAILURE){
464 XBT_DEBUG("SRC host failed during migration of %s (stage %d)", vm->getCname(), stage);
465 MSG_task_destroy(task);
466 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", vm->getCname(), stage);
467 }else if(ret == MSG_TRANSFER_FAILURE){
468 XBT_DEBUG("DST host failed during migration of %s (stage %d)", vm->getCname(), stage);
469 MSG_task_destroy(task);
470 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", vm->getCname(), stage);
473 double clock_end = MSG_get_clock();
474 double duration = clock_end - clock_sta;
475 double actual_speed = size / duration;
478 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
479 actual_speed, mig_speed);
481 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
487 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
489 double updated_size = computed * dp_rate;
490 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
491 if (updated_size > dp_cap) {
492 updated_size = dp_cap;
495 return static_cast<sg_size_t>(updated_size);
498 static int migration_tx_fun(int argc, char *argv[])
500 XBT_DEBUG("mig: tx_start");
502 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
503 migration_session* ms = static_cast<migration_session*>(MSG_process_get_data(MSG_process_self()));
505 double host_speed = ms->vm->pimpl_vm_->getPm()->getSpeed();
506 s_vm_params_t params;
507 ms->vm->getParameters(¶ms);
508 const sg_size_t ramsize = params.ramsize;
509 const sg_size_t devsize = params.devsize;
510 const int skip_stage1 = params.skip_stage1;
511 int skip_stage2 = params.skip_stage2;
512 const double dp_rate = host_speed ? (params.mig_speed * params.dp_intensity) / host_speed : 1;
513 const double dp_cap = params.dp_cap;
514 const double mig_speed = params.mig_speed;
515 double max_downtime = params.max_downtime;
517 double mig_timeout = 10000000.0;
519 double remaining_size = static_cast<double>(ramsize + devsize);
520 double threshold = 0.0;
522 /* check parameters */
524 XBT_WARN("migrate a VM, but ramsize is zero");
526 if (max_downtime <= 0) {
527 XBT_WARN("use the default max_downtime value 30ms");
531 /* Stage1: send all memory pages to the destination. */
532 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
533 start_dirty_page_tracking(ms->vm);
535 double computed_during_stage1 = 0;
536 if (not skip_stage1) {
537 double clock_prev_send = MSG_get_clock();
540 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
541 * transfer exceeds the timeout value. */
542 XBT_VERB("Stage 1: Gonna send %llu bytes", ramsize);
543 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
544 remaining_size -= sent;
545 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
547 if (sent < ramsize) {
548 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
550 } else if (sent > ramsize)
555 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
556 // Stop the dirty page tracking an return (there is no memory space to release)
557 stop_dirty_page_tracking(ms->vm);
561 double clock_post_send = MSG_get_clock();
562 mig_timeout -= (clock_post_send - clock_prev_send);
563 if (mig_timeout < 0) {
564 XBT_VERB("The duration of stage 1 exceeds the timeout value, skip stage 2");
568 /* estimate bandwidth */
569 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
570 threshold = bandwidth * max_downtime;
571 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
575 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
576 if (not skip_stage2) {
578 int stage2_round = 0;
581 sg_size_t updated_size = 0;
582 if (stage2_round == 0) {
583 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
584 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
586 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
587 updated_size = get_updated_size(computed, dp_rate, dp_cap);
590 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
591 updated_size, computed_during_stage1, dp_rate, dp_cap);
593 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
594 remaining_size += updated_size;
595 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
596 (remaining_size < threshold) ? "<" : ">", threshold);
597 if (remaining_size < threshold)
601 double clock_prev_send = MSG_get_clock();
603 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
604 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
606 } catch (xbt_ex& e) {
607 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
609 // Stop the dirty page tracking an return (there is no memory space to release)
610 stop_dirty_page_tracking(ms->vm);
613 double clock_post_send = MSG_get_clock();
615 if (sent == updated_size) {
616 /* timeout did not happen */
617 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
618 threshold = bandwidth * max_downtime;
619 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
620 remaining_size -= sent;
622 mig_timeout -= (clock_post_send - clock_prev_send);
623 xbt_assert(mig_timeout > 0);
625 } else if (sent < updated_size) {
626 /* When timeout happens, we move to stage 3. The size of memory pages
627 * updated before timeout must be added to the remaining size. */
628 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
629 updated_size, (clock_post_send - clock_prev_send));
630 remaining_size -= sent;
632 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
633 updated_size = get_updated_size(computed, dp_rate, dp_cap);
634 remaining_size += updated_size;
641 /* Stage3: stop the VM and copy the rest of states. */
642 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
643 simgrid::vm::VirtualMachineImpl* pimpl = ms->vm->pimpl_vm_;
644 pimpl->setState(SURF_VM_STATE_RUNNING); // FIXME: this bypass of the checks in suspend() is not nice
645 pimpl->isMigrating = false; // FIXME: this bypass of the checks in suspend() is not nice
646 pimpl->suspend(SIMIX_process_self());
647 stop_dirty_page_tracking(ms->vm);
650 XBT_DEBUG("Stage 3: Gonna send %f bytes", remaining_size);
651 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, static_cast<sg_size_t>(remaining_size), ms->mbox, 3, 0,
655 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
656 // Stop the dirty page tracking an return (there is no memory space to release)
657 ms->vm->pimpl_vm_->resume();
661 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
662 // effectively the VM on the DST node.
663 XBT_DEBUG("mig: tx_done");
668 /** @brief Migrate the VM to the given host.
671 void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm)
674 * - One approach is ...
675 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
676 * receive the state of the source
677 * VM over network. We will finally destroy the source VM.
678 * - This behavior is similar to the way of migration in the real world.
679 * Even before a migration is completed, we will see a destination VM, consuming resources.
680 * - We have to relocate all processes. The existing process migration code will work for this?
681 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
683 * - Another one is ...
684 * We update the information of the given VM to place it to the destination physical host.
686 * The second one would be easier.
689 msg_host_t src_pm = vm->pimpl_vm_->getPm();
692 THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->getCname(), src_pm->getCname());
694 THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->getCname(), dst_pm->getCname());
695 if (not MSG_vm_is_running(vm))
696 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->getCname());
697 if (vm->isMigrating())
698 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->getCname());
700 vm->pimpl_vm_->isMigrating = true;
702 migration_session* ms = xbt_new(migration_session, 1);
707 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
708 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
709 ms->mbox_ctl = bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
710 ms->mbox = bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
712 std::string pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
713 std::string pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
715 MSG_process_create(pr_rx_name.c_str(), migration_rx_fun, ms, dst_pm);
717 MSG_process_create(pr_tx_name.c_str(), migration_tx_fun, ms, src_pm);
719 /* wait until the migration have finished or on error has occurred */
720 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
721 msg_task_t task = nullptr;
722 msg_error_t ret = MSG_task_receive(&task, ms->mbox_ctl);
724 vm->pimpl_vm_->isMigrating = false;
726 xbt_free(ms->mbox_ctl);
730 if (ret == MSG_HOST_FAILURE) {
731 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
732 // other side. Hence, just throw the execption
733 XBT_ERROR("SRC crashes, throw an exception (m-control)");
734 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
736 THROWF(host_error, 0, "Source host '%s' failed during the migration of VM '%s'.", src_pm->getCname(),
738 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
739 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
740 XBT_ERROR("DST crashes, throw an exception (m-control)");
741 THROWF(host_error, 0, "Destination host '%s' failed during the migration of VM '%s'.", dst_pm->getCname(),
745 xbt_assert(get_mig_task_name(vm, src_pm, dst_pm, 4) == task->name);
746 MSG_task_destroy(task);
749 /** @brief Immediately suspend the execution of all processes within the given VM.
752 * This function stops the execution of the VM. All the processes on this VM
753 * will pause. The state of the VM is preserved. We can later resume it again.
755 * No suspension cost occurs.
757 void MSG_vm_suspend(msg_vm_t vm)
759 smx_actor_t issuer = SIMIX_process_self();
760 simgrid::simix::kernelImmediate([vm, issuer]() { vm->pimpl_vm_->suspend(issuer); });
762 XBT_DEBUG("vm_suspend done");
764 if (TRACE_msg_vm_is_enabled()) {
765 simgrid::instr::StateType* state = simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE");
766 state->addEntityValue("suspend", "1 0 0"); // suspend is red
767 state->pushEvent("suspend");
771 /** @brief Resume the execution of the VM. All processes on the VM run again.
774 * No resume cost occurs.
776 void MSG_vm_resume(msg_vm_t vm)
778 vm->pimpl_vm_->resume();
780 if (TRACE_msg_vm_is_enabled())
781 simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE")->popEvent();
784 /** @brief Get the physical host of a given VM.
787 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
792 /** @brief Set a CPU bound for a given VM.
795 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
798 * On PM0, there are Task1 and VM0.
799 * On VM0, there is Task2.
800 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
802 * Task1\@PM0 gets 50%.
803 * Task2\@VM0 gets 25%.
804 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
806 * 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
807 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
809 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
810 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
811 * on the VM in share_resource() or we have to trap both the start and end of task execution.
813 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
815 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
816 * computation task (or a VM) never exceeds the capacity of a CPU core.
818 void MSG_vm_set_bound(msg_vm_t vm, double bound)
820 simgrid::simix::kernelImmediate([vm, bound]() { vm->pimpl_vm_->setBound(bound); });