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"
21 #include "xbt/string.hpp"
26 double prev_clock = 0.0;
27 double prev_remaining = 0.0;
28 msg_task_t task = nullptr;
30 typedef s_dirty_page* dirty_page_t;
32 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
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 void MSG_vm_set_ramsize(msg_vm_t vm, size_t size)
65 size_t MSG_vm_get_ramsize(msg_vm_t vm)
67 return vm->getRamsize();
70 /* **** Check state of a VM **** */
71 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
73 return vm->pimpl_vm_ != nullptr && vm->getState() == state;
76 /** @brief Returns whether the given VM has just created, not running.
79 int MSG_vm_is_created(msg_vm_t vm)
81 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
84 /** @brief Returns whether the given VM is currently running
87 int MSG_vm_is_running(msg_vm_t vm)
89 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
92 /** @brief Returns whether the given VM is currently migrating
95 int MSG_vm_is_migrating(msg_vm_t vm)
97 return vm->isMigrating();
100 /** @brief Returns whether the given VM is currently suspended, not running.
103 int MSG_vm_is_suspended(msg_vm_t vm)
105 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
108 /* **** ******** MSG vm actions ********* **** */
109 /** @brief Create a new VM with specified parameters.
111 * @param pm Physical machine that will host the VM
112 * @param name Must be unique
113 * @param coreAmount Must be >= 1
114 * @param ramsize [TODO]
115 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
116 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
118 msg_vm_t MSG_vm_create(msg_host_t pm, const char* name, int coreAmount, int ramsize, int mig_netspeed, int dp_intensity)
120 simgrid::vm::VmHostExt::ensureVmExtInstalled();
122 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
124 msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, coreAmount, static_cast<sg_size_t>(ramsize) * 1024 * 1024);
125 s_vm_params_t params;
126 params.max_downtime = 0.03;
127 params.mig_speed = static_cast<double>(mig_netspeed) * 1024 * 1024; // mig_speed
128 params.dp_intensity = static_cast<double>(dp_intensity) / 100;
129 params.dp_cap = vm->getRamsize() * 0.9; // assume working set memory is 90% of ramsize
131 XBT_DEBUG("migspeed : %f intensity mem : %d", params.mig_speed, dp_intensity);
132 vm->setParameters(¶ms);
137 /** @brief Create a new VM object with the default parameters
140 * A VM is treated as a host. The name of the VM must be unique among all hosts.
142 msg_vm_t MSG_vm_create_core(msg_host_t pm, const char* name)
144 xbt_assert(sg_host_by_name(name) == nullptr,
145 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
147 msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, 1);
148 s_vm_params_t params;
149 memset(¶ms, 0, sizeof(params));
150 vm->setParameters(¶ms);
153 /** @brief Create a new VM object with the default parameters, but with a specified amount of cores
156 * A VM is treated as a host. The name of the VM must be unique among all hosts.
158 msg_vm_t MSG_vm_create_multicore(msg_host_t pm, const char* name, int coreAmount)
160 xbt_assert(sg_host_by_name(name) == nullptr,
161 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
163 msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, coreAmount);
164 s_vm_params_t params;
165 memset(¶ms, 0, sizeof(params));
166 vm->setParameters(¶ms);
170 /** @brief Destroy a VM. Destroy the VM object from the simulation.
173 void MSG_vm_destroy(msg_vm_t vm)
175 if (vm->isMigrating())
176 THROWF(vm_error, 0, "Cannot destroy VM '%s', which is migrating.", vm->getCname());
178 /* First, terminate all processes on the VM if necessary */
181 /* Then, destroy the VM object */
184 if (TRACE_msg_vm_is_enabled()) {
185 container_t container = simgrid::instr::Container::byName(vm->getName());
186 container->removeFromParent();
191 /** @brief Start a vm (i.e., boot the guest operating system)
194 * If the VM cannot be started (because of memory over-provisioning), an exception is generated.
196 void MSG_vm_start(msg_vm_t vm)
199 if (TRACE_msg_vm_is_enabled()) {
200 simgrid::instr::StateType* state = simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE");
201 state->addEntityValue("start", "0 0 1"); // start is blue
202 state->pushEvent("start");
206 /** @brief Immediately kills all processes within the given VM.
209 * Any memory that they allocated will be leaked, unless you used #MSG_process_on_exit().
211 * No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep().
213 void MSG_vm_shutdown(msg_vm_t vm)
218 static std::string get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
220 return std::string("__pr_mig_tx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
223 static std::string get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
225 return std::string("__pr_mig_rx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
228 static std::string get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
230 return std::string("__task_mig_stage") + std::to_string(stage) + ":" + vm->getCname() + "(" + src_pm->getCname() +
231 "-" + dst_pm->getCname() + ")";
234 struct migration_session {
239 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
240 * know the completion of the migration. */
241 std::string mbox_ctl;
242 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
246 static int migration_rx_fun(int argc, char *argv[])
248 XBT_DEBUG("mig: rx_start");
250 // The structure has been created in the do_migration function and should only be freed in the same place ;)
251 migration_session* ms = static_cast<migration_session*>(MSG_process_get_data(MSG_process_self()));
253 bool received_finalize = false;
255 std::string finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
256 while (not received_finalize) {
257 msg_task_t task = nullptr;
258 int ret = MSG_task_recv(&task, ms->mbox.c_str());
261 // An error occurred, clean the code and return
262 // The owner did not change, hence the task should be only destroyed on the other side
266 if (finalize_task_name == task->name)
267 received_finalize = 1;
269 MSG_task_destroy(task);
272 // Here Stage 1, 2 and 3 have been performed.
273 // Hence complete the migration
275 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
276 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
277 // structure at the beginning and free it at the end of each function)
278 simgrid::s4u::VirtualMachine* vm = ms->vm;
279 msg_host_t dst_pm = ms->dst_pm;
281 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
282 simgrid::simix::kernelImmediate([vm, dst_pm]() {
283 /* Update the vm location */
284 /* precopy migration makes the VM temporally paused */
285 xbt_assert(vm->getState() == SURF_VM_STATE_SUSPENDED);
287 /* Update the vm location and resume it */
288 vm->pimpl_vm_->setPm(dst_pm);
293 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
294 vm->pimpl_vm_->isMigrating = false;
295 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->getCname(), ms->src_pm->getCname(), ms->dst_pm->getCname());
297 if (TRACE_msg_vm_is_enabled()) {
298 static long long int counter = 0;
299 std::string key = std::to_string(counter);
303 container_t msg = simgrid::instr::Container::byName(vm->getName());
304 simgrid::instr::Container::getRoot()->getLink("MSG_VM_LINK")->startEvent(msg, "M", key);
306 // destroy existing container of this vm
307 container_t existing_container = simgrid::instr::Container::byName(vm->getName());
308 existing_container->removeFromParent();
309 delete existing_container;
311 // create new container on the new_host location
312 new simgrid::instr::Container(vm->getCname(), "MSG_VM", simgrid::instr::Container::byName(ms->dst_pm->getName()));
315 msg = simgrid::instr::Container::byName(vm->getName());
316 simgrid::instr::Container::getRoot()->getLink("MSG_VM_LINK")->endEvent(msg, "M", key);
319 // Inform the SRC that the migration has been correctly performed
320 std::string task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
321 msg_task_t task = MSG_task_create(task_name.c_str(), 0, 0, nullptr);
322 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl.c_str());
323 if(ret == MSG_HOST_FAILURE){
324 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
325 // has been correctly migrated on the DST node
326 // TODO What does it mean ? What should we do ?
327 MSG_task_destroy(task);
328 } else if(ret == MSG_TRANSFER_FAILURE){
329 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
330 MSG_task_destroy(task);
333 XBT_DEBUG("mig: rx_done");
337 static void start_dirty_page_tracking(msg_vm_t vm)
339 vm->pimpl_vm_->dp_enabled = true;
340 if (vm->pimpl_vm_->dp_objs.empty())
343 for (auto const& elm : vm->pimpl_vm_->dp_objs) {
344 dirty_page_t dp = elm.second;
345 double remaining = MSG_task_get_flops_amount(dp->task);
346 dp->prev_clock = MSG_get_clock();
347 dp->prev_remaining = remaining;
348 XBT_DEBUG("%s@%s remaining %f", elm.first.c_str(), vm->getCname(), remaining);
352 static void stop_dirty_page_tracking(msg_vm_t vm)
354 vm->pimpl_vm_->dp_enabled = false;
357 static double get_computed(const std::string& key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
359 double computed = dp->prev_remaining - remaining;
360 double duration = clock - dp->prev_clock;
362 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", key.c_str(), vm->getCname(), computed,
363 dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
368 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
372 for (auto const& elm : vm->pimpl_vm_->dp_objs) {
373 const std::string& key = elm.first;
374 dirty_page_t dp = elm.second;
375 double remaining = MSG_task_get_flops_amount(dp->task);
377 double clock = MSG_get_clock();
379 total += get_computed(key, vm, dp, remaining, clock);
381 dp->prev_remaining = remaining;
382 dp->prev_clock = clock;
385 total += vm->pimpl_vm_->dp_updated_by_deleted_tasks;
387 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
388 stage2_round_for_fancy_debug, total, vm->pimpl_vm_->dp_updated_by_deleted_tasks);
390 vm->pimpl_vm_->dp_updated_by_deleted_tasks = 0;
395 // TODO Is this code redundant with the information provided by
396 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
397 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
398 void MSG_host_add_task(msg_host_t host, msg_task_t task)
400 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
404 double remaining = MSG_task_get_flops_amount(task);
405 std::string key = simgrid::xbt::string_printf("%s-%p", task->name, task);
407 dirty_page_t dp = new s_dirty_page;
409 if (vm->pimpl_vm_->dp_enabled) {
410 dp->prev_clock = MSG_get_clock();
411 dp->prev_remaining = remaining;
413 vm->pimpl_vm_->dp_objs.insert({key, dp});
414 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key.c_str(), host->getCname(), remaining,
415 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 std::string key = simgrid::xbt::string_printf("%s-%p", task->name, task);
425 dirty_page_t dp = nullptr;
426 auto dp_obj = vm->pimpl_vm_->dp_objs.find(key);
427 if (dp_obj != vm->pimpl_vm_->dp_objs.end())
429 xbt_assert(dp && dp->task == task);
431 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
432 * the information for the lookup_() function that will called soon. */
433 if (vm->pimpl_vm_->dp_enabled) {
434 double remaining = MSG_task_get_flops_amount(task);
435 double clock = MSG_get_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.c_str(), host->getCname());
447 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,
448 const std::string& mbox, int stage, int stage2_round, double mig_speed,
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.c_str(), timeout, mig_speed);
461 ret = MSG_task_send(task, mbox.c_str());
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 updated_size = dp_cap;
504 return static_cast<sg_size_t>(updated_size);
507 static int migration_tx_fun(int argc, char *argv[])
509 XBT_DEBUG("mig: tx_start");
511 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
512 migration_session* ms = static_cast<migration_session*>(MSG_process_get_data(MSG_process_self()));
514 double host_speed = ms->vm->getPm()->getSpeed();
515 s_vm_params_t params;
516 ms->vm->getParameters(¶ms);
517 const sg_size_t ramsize = ms->vm->getRamsize();
518 const double dp_rate = host_speed ? (params.mig_speed * params.dp_intensity) / host_speed : 1;
519 const double dp_cap = params.dp_cap;
520 const double mig_speed = params.mig_speed;
521 double max_downtime = params.max_downtime;
523 double mig_timeout = 10000000.0;
524 bool skip_stage2 = false;
526 size_t remaining_size = ramsize;
527 size_t threshold = 0.0;
529 /* check parameters */
531 XBT_WARN("migrate a VM, but ramsize is zero");
533 if (max_downtime <= 0) {
534 XBT_WARN("use the default max_downtime value 30ms");
538 /* Stage1: send all memory pages to the destination. */
539 XBT_DEBUG("mig-stage1: remaining_size %zu", remaining_size);
540 start_dirty_page_tracking(ms->vm);
542 double computed_during_stage1 = 0;
543 double clock_prev_send = MSG_get_clock();
546 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
547 * transfer exceeds the timeout value. */
548 XBT_VERB("Stage 1: Gonna send %llu bytes", ramsize);
549 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
550 remaining_size -= sent;
551 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
553 if (sent < ramsize) {
554 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
556 } else if (sent > ramsize)
559 } catch (xbt_ex& e) {
560 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
561 // Stop the dirty page tracking an return (there is no memory space to release)
562 stop_dirty_page_tracking(ms->vm);
566 double clock_post_send = MSG_get_clock();
567 mig_timeout -= (clock_post_send - clock_prev_send);
568 if (mig_timeout < 0) {
569 XBT_VERB("The duration of stage 1 exceeds the timeout value, skip stage 2");
573 /* estimate bandwidth */
574 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
575 threshold = bandwidth * max_downtime;
576 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %zu", bandwidth / 1024 / 1024, threshold);
578 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
579 if (not skip_stage2) {
581 int stage2_round = 0;
584 sg_size_t updated_size = 0;
585 if (stage2_round == 0) {
586 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
587 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
589 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
590 updated_size = get_updated_size(computed, dp_rate, dp_cap);
593 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
594 updated_size, computed_during_stage1, dp_rate, dp_cap);
596 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
597 remaining_size += updated_size;
598 XBT_DEBUG("mig-stage2.%d: remaining_size %zu (%s threshold %zu)", stage2_round, remaining_size,
599 (remaining_size < threshold) ? "<" : ">", threshold);
600 if (remaining_size < threshold)
604 double clock_prev_send = MSG_get_clock();
606 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
607 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
609 } catch (xbt_ex& e) {
610 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
612 // Stop the dirty page tracking an return (there is no memory space to release)
613 stop_dirty_page_tracking(ms->vm);
616 double clock_post_send = MSG_get_clock();
618 if (sent == updated_size) {
619 /* timeout did not happen */
620 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
621 threshold = bandwidth * max_downtime;
622 XBT_DEBUG("actual bandwidth %f, threshold %zu", bandwidth / 1024 / 1024, threshold);
623 remaining_size -= sent;
625 mig_timeout -= (clock_post_send - clock_prev_send);
626 xbt_assert(mig_timeout > 0);
628 } else if (sent < updated_size) {
629 /* When timeout happens, we move to stage 3. The size of memory pages
630 * updated before timeout must be added to the remaining size. */
631 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
632 updated_size, (clock_post_send - clock_prev_send));
633 remaining_size -= sent;
635 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
636 updated_size = get_updated_size(computed, dp_rate, dp_cap);
637 remaining_size += updated_size;
644 /* Stage3: stop the VM and copy the rest of states. */
645 XBT_DEBUG("mig-stage3: remaining_size %zu", remaining_size);
646 simgrid::vm::VirtualMachineImpl* pimpl = ms->vm->pimpl_vm_;
647 pimpl->setState(SURF_VM_STATE_RUNNING); // FIXME: this bypass of the checks in suspend() is not nice
648 pimpl->isMigrating = false; // FIXME: this bypass of the checks in suspend() is not nice
649 pimpl->suspend(SIMIX_process_self());
650 stop_dirty_page_tracking(ms->vm);
653 XBT_DEBUG("Stage 3: Gonna send %zu bytes", remaining_size);
654 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, remaining_size, ms->mbox, 3, 0, mig_speed, -1);
657 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
658 // Stop the dirty page tracking an return (there is no memory space to release)
663 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
664 // effectively the VM on the DST node.
665 XBT_DEBUG("mig: tx_done");
670 /** @brief Migrate the VM to the given host.
673 void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm)
676 * - One approach is ...
677 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
678 * receive the state of the source
679 * VM over network. We will finally destroy the source VM.
680 * - This behavior is similar to the way of migration in the real world.
681 * Even before a migration is completed, we will see a destination VM, consuming resources.
682 * - We have to relocate all processes. The existing process migration code will work for this?
683 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
685 * - Another one is ...
686 * We update the information of the given VM to place it to the destination physical host.
688 * The second one would be easier.
691 msg_host_t src_pm = vm->getPm();
694 THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->getCname(), src_pm->getCname());
696 THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->getCname(), dst_pm->getCname());
697 if (not MSG_vm_is_running(vm))
698 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->getCname());
699 if (vm->isMigrating())
700 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->getCname());
702 vm->pimpl_vm_->isMigrating = true;
704 migration_session ms;
709 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
710 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
712 simgrid::xbt::string_printf("__mbox_mig_ctl:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
713 ms.mbox = simgrid::xbt::string_printf("__mbox_mig_src_dst:%s(%s-%s)", vm->getCname(), src_pm->getCname(),
716 std::string pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
717 std::string pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
719 MSG_process_create(pr_rx_name.c_str(), migration_rx_fun, &ms, dst_pm);
721 MSG_process_create(pr_tx_name.c_str(), migration_tx_fun, &ms, src_pm);
723 /* wait until the migration have finished or on error has occurred */
724 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
725 msg_task_t task = nullptr;
726 msg_error_t ret = MSG_task_receive(&task, ms.mbox_ctl.c_str());
728 vm->pimpl_vm_->isMigrating = false;
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)
760 if (TRACE_msg_vm_is_enabled()) {
761 simgrid::instr::StateType* state = simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE");
762 state->addEntityValue("suspend", "1 0 0"); // suspend is red
763 state->pushEvent("suspend");
767 /** @brief Resume the execution of the VM. All processes on the VM run again.
770 * No resume cost occurs.
772 void MSG_vm_resume(msg_vm_t vm)
775 if (TRACE_msg_vm_is_enabled())
776 simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE")->popEvent();
779 /** @brief Get the physical host of a given VM.
782 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
787 void MSG_vm_set_bound(msg_vm_t vm, double bound)