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 /* **** ******** 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 void MSG_vm_set_ramsize(msg_vm_t vm, size_t size)
66 size_t MSG_vm_get_ramsize(msg_vm_t vm)
68 return vm->getRamsize();
71 /* **** Check state of a VM **** */
72 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
74 return vm->pimpl_vm_ != nullptr && vm->pimpl_vm_->getState() == state;
77 /** @brief Returns whether the given VM has just created, not running.
80 int MSG_vm_is_created(msg_vm_t vm)
82 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
85 /** @brief Returns whether the given VM is currently running
88 int MSG_vm_is_running(msg_vm_t vm)
90 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
93 /** @brief Returns whether the given VM is currently migrating
96 int MSG_vm_is_migrating(msg_vm_t vm)
98 return vm->isMigrating();
101 /** @brief Returns whether the given VM is currently suspended, not running.
104 int MSG_vm_is_suspended(msg_vm_t vm)
106 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
109 /* **** ******** MSG vm actions ********* **** */
110 /** @brief Create a new VM with specified parameters.
112 * @param pm Physical machine that will host the VM
113 * @param name Must be unique
114 * @param coreAmount Must be >= 1
115 * @param ramsize [TODO]
116 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
117 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
119 msg_vm_t MSG_vm_create(msg_host_t pm, const char* name, int coreAmount, int ramsize, int mig_netspeed, int dp_intensity)
121 simgrid::vm::VmHostExt::ensureVmExtInstalled();
123 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
125 msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, coreAmount, static_cast<sg_size_t>(ramsize) * 1024 * 1024);
126 s_vm_params_t params;
128 params.skip_stage1 = 0;
129 params.skip_stage2 = 0;
130 params.max_downtime = 0.03;
131 params.mig_speed = static_cast<double>(mig_netspeed) * 1024 * 1024; // mig_speed
132 params.dp_intensity = static_cast<double>(dp_intensity) / 100;
133 params.dp_cap = vm->getRamsize() * 0.9; // assume working set memory is 90% of ramsize
135 XBT_DEBUG("migspeed : %f intensity mem : %d", params.mig_speed, dp_intensity);
136 vm->setParameters(¶ms);
141 /** @brief Create a new VM object with the default parameters
144 * A VM is treated as a host. The name of the VM must be unique among all hosts.
146 msg_vm_t MSG_vm_create_core(msg_host_t pm, const char* name)
148 xbt_assert(sg_host_by_name(name) == nullptr,
149 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
151 msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, 1);
152 s_vm_params_t params;
153 memset(¶ms, 0, sizeof(params));
154 vm->setParameters(¶ms);
157 /** @brief Create a new VM object with the default parameters, but with a specified amount of cores
160 * A VM is treated as a host. The name of the VM must be unique among all hosts.
162 msg_vm_t MSG_vm_create_multicore(msg_host_t pm, const char* name, int coreAmount)
164 xbt_assert(sg_host_by_name(name) == nullptr,
165 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
167 msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, coreAmount);
168 s_vm_params_t params;
169 memset(¶ms, 0, sizeof(params));
170 vm->setParameters(¶ms);
174 /** @brief Destroy a VM. Destroy the VM object from the simulation.
177 void MSG_vm_destroy(msg_vm_t vm)
179 if (vm->isMigrating())
180 THROWF(vm_error, 0, "Cannot destroy VM '%s', which is migrating.", vm->getCname());
182 /* First, terminate all processes on the VM if necessary */
183 if (MSG_vm_is_running(vm))
186 /* Then, destroy the VM object */
187 simgrid::simix::kernelImmediate([vm]() { vm->destroy(); });
189 if (TRACE_msg_vm_is_enabled()) {
190 container_t container = simgrid::instr::Container::byName(vm->getName());
191 container->removeFromParent();
196 /** @brief Start a vm (i.e., boot the guest operating system)
199 * If the VM cannot be started (because of memory over-provisioning), an exception is generated.
201 void MSG_vm_start(msg_vm_t vm)
204 if (TRACE_msg_vm_is_enabled()) {
205 simgrid::instr::StateType* state = simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE");
206 state->addEntityValue("start", "0 0 1"); // start is blue
207 state->pushEvent("start");
211 /** @brief Immediately kills all processes within the given VM.
214 * Any memory that they allocated will be leaked, unless you used #MSG_process_on_exit().
216 * No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep().
218 void MSG_vm_shutdown(msg_vm_t vm)
220 smx_actor_t issuer = SIMIX_process_self();
221 simgrid::simix::kernelImmediate([vm, issuer]() { vm->pimpl_vm_->shutdown(issuer); });
223 // Make sure that processes in the VM are killed in this scheduling round before processing (eg with the VM destroy)
224 MSG_process_sleep(0.);
227 static std::string get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
229 return std::string("__pr_mig_tx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
232 static std::string get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
234 return std::string("__pr_mig_rx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
237 static std::string get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
239 return std::string("__task_mig_stage") + std::to_string(stage) + ":" + vm->getCname() + "(" + src_pm->getCname() +
240 "-" + dst_pm->getCname() + ")";
243 struct migration_session {
248 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
249 * know the completion of the migration. */
250 std::string mbox_ctl;
251 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
255 static int migration_rx_fun(int argc, char *argv[])
257 XBT_DEBUG("mig: rx_start");
259 // The structure has been created in the do_migration function and should only be freed in the same place ;)
260 migration_session* ms = static_cast<migration_session*>(MSG_process_get_data(MSG_process_self()));
262 bool received_finalize = false;
264 std::string finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
265 while (not received_finalize) {
266 msg_task_t task = nullptr;
267 int ret = MSG_task_recv(&task, ms->mbox.c_str());
270 // An error occurred, clean the code and return
271 // The owner did not change, hence the task should be only destroyed on the other side
275 if (finalize_task_name == task->name)
276 received_finalize = 1;
278 MSG_task_destroy(task);
281 // Here Stage 1, 2 and 3 have been performed.
282 // Hence complete the migration
284 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
285 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
286 // structure at the beginning and free it at the end of each function)
287 simgrid::s4u::VirtualMachine* vm = ms->vm;
288 msg_host_t dst_pm = ms->dst_pm;
290 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
291 simgrid::simix::kernelImmediate([vm, dst_pm]() {
292 /* Update the vm location */
293 /* precopy migration makes the VM temporally paused */
294 xbt_assert(vm->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
296 /* Update the vm location and resume it */
297 vm->pimpl_vm_->setPm(dst_pm);
298 vm->pimpl_vm_->resume();
302 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
303 vm->pimpl_vm_->isMigrating = false;
304 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->getCname(), ms->src_pm->getCname(), ms->dst_pm->getCname());
306 if (TRACE_msg_vm_is_enabled()) {
307 static long long int counter = 0;
308 std::string key = std::to_string(counter);
312 container_t msg = simgrid::instr::Container::byName(vm->getName());
313 simgrid::instr::Container::getRoot()->getLink("MSG_VM_LINK")->startEvent(msg, "M", key);
315 // destroy existing container of this vm
316 container_t existing_container = simgrid::instr::Container::byName(vm->getName());
317 existing_container->removeFromParent();
318 delete existing_container;
320 // create new container on the new_host location
321 new simgrid::instr::Container(vm->getCname(), "MSG_VM", simgrid::instr::Container::byName(ms->dst_pm->getName()));
324 msg = simgrid::instr::Container::byName(vm->getName());
325 simgrid::instr::Container::getRoot()->getLink("MSG_VM_LINK")->endEvent(msg, "M", key);
328 // Inform the SRC that the migration has been correctly performed
329 std::string task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
330 msg_task_t task = MSG_task_create(task_name.c_str(), 0, 0, nullptr);
331 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl.c_str());
332 if(ret == MSG_HOST_FAILURE){
333 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
334 // has been correctly migrated on the DST node
335 // TODO What does it mean ? What should we do ?
336 MSG_task_destroy(task);
337 } else if(ret == MSG_TRANSFER_FAILURE){
338 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
339 MSG_task_destroy(task);
342 XBT_DEBUG("mig: rx_done");
346 static void start_dirty_page_tracking(msg_vm_t vm)
348 vm->pimpl_vm_->dp_enabled = 1;
349 if (vm->pimpl_vm_->dp_objs.empty())
352 for (auto const& elm : vm->pimpl_vm_->dp_objs) {
353 dirty_page_t dp = elm.second;
354 double remaining = MSG_task_get_remaining_work_ratio(dp->task);
355 dp->prev_clock = MSG_get_clock();
356 dp->prev_remaining = remaining;
357 XBT_DEBUG("%s@%s remaining %f", elm.first.c_str(), vm->getCname(), remaining);
361 static void stop_dirty_page_tracking(msg_vm_t vm)
363 vm->pimpl_vm_->dp_enabled = 0;
366 static double get_computed(const std::string& key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
368 double computed = dp->prev_remaining - remaining;
369 double duration = clock - dp->prev_clock;
371 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", key.c_str(), vm->getCname(), computed,
372 dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
377 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
381 for (auto const& elm : vm->pimpl_vm_->dp_objs) {
382 const std::string& key = elm.first;
383 dirty_page_t dp = elm.second;
384 double remaining = MSG_task_get_remaining_work_ratio(dp->task);
386 double clock = MSG_get_clock();
388 total += get_computed(key, vm, dp, remaining, clock);
390 dp->prev_remaining = remaining;
391 dp->prev_clock = clock;
394 total += vm->pimpl_vm_->dp_updated_by_deleted_tasks;
396 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
397 stage2_round_for_fancy_debug, total, vm->pimpl_vm_->dp_updated_by_deleted_tasks);
399 vm->pimpl_vm_->dp_updated_by_deleted_tasks = 0;
404 // TODO Is this code redundant with the information provided by
405 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
406 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
407 void MSG_host_add_task(msg_host_t host, msg_task_t task)
409 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
413 double remaining = MSG_task_get_initial_flops_amount(task);
414 std::string key = simgrid::xbt::string_printf("%s-%p", task->name, task);
416 dirty_page_t dp = new s_dirty_page;
418 if (vm->pimpl_vm_->dp_enabled) {
419 dp->prev_clock = MSG_get_clock();
420 dp->prev_remaining = remaining;
422 vm->pimpl_vm_->dp_objs.insert({key, dp});
423 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key.c_str(), host->getCname(), remaining,
424 vm->pimpl_vm_->dp_enabled);
427 void MSG_host_del_task(msg_host_t host, msg_task_t task)
429 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
433 std::string key = simgrid::xbt::string_printf("%s-%p", task->name, task);
434 dirty_page_t dp = nullptr;
435 auto dp_obj = vm->pimpl_vm_->dp_objs.find(key);
436 if (dp_obj != vm->pimpl_vm_->dp_objs.end())
438 xbt_assert(dp && dp->task == task);
440 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
441 * the information for the lookup_() function that will called soon. */
442 if (vm->pimpl_vm_->dp_enabled) {
443 double remaining = MSG_task_get_remaining_work_ratio(task);
444 double clock = MSG_get_clock();
445 double updated = get_computed(key, vm, dp, remaining, clock); // was host instead of vm
447 vm->pimpl_vm_->dp_updated_by_deleted_tasks += updated;
450 vm->pimpl_vm_->dp_objs.erase(key);
453 XBT_DEBUG("del %s on %s", key.c_str(), host->getCname());
456 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,
457 const std::string& mbox, int stage, int stage2_round, double mig_speed,
461 std::string task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
462 msg_task_t task = MSG_task_create(task_name.c_str(), 0, static_cast<double>(size), nullptr);
464 double clock_sta = MSG_get_clock();
468 ret = MSG_task_send_with_timeout_bounded(task, mbox.c_str(), timeout, mig_speed);
470 ret = MSG_task_send(task, mbox.c_str());
474 } else if (ret == MSG_TIMEOUT) {
475 sg_size_t remaining = static_cast<sg_size_t>(MSG_task_get_remaining_communication(task));
476 sent = size - remaining;
477 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
480 /* FIXME: why try-and-catch is used here? */
481 if(ret == MSG_HOST_FAILURE){
482 XBT_DEBUG("SRC host failed during migration of %s (stage %d)", vm->getCname(), stage);
483 MSG_task_destroy(task);
484 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", vm->getCname(), stage);
485 }else if(ret == MSG_TRANSFER_FAILURE){
486 XBT_DEBUG("DST host failed during migration of %s (stage %d)", vm->getCname(), stage);
487 MSG_task_destroy(task);
488 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", vm->getCname(), stage);
491 double clock_end = MSG_get_clock();
492 double duration = clock_end - clock_sta;
493 double actual_speed = size / duration;
496 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
497 actual_speed, mig_speed);
499 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
505 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
507 double updated_size = computed * dp_rate;
508 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
509 if (updated_size > dp_cap) {
510 updated_size = dp_cap;
513 return static_cast<sg_size_t>(updated_size);
516 static int migration_tx_fun(int argc, char *argv[])
518 XBT_DEBUG("mig: tx_start");
520 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
521 migration_session* ms = static_cast<migration_session*>(MSG_process_get_data(MSG_process_self()));
523 double host_speed = ms->vm->pimpl_vm_->getPm()->getSpeed();
524 s_vm_params_t params;
525 ms->vm->getParameters(¶ms);
526 const sg_size_t ramsize = ms->vm->getRamsize();
527 const sg_size_t devsize = params.devsize;
528 const int skip_stage1 = params.skip_stage1;
529 int skip_stage2 = params.skip_stage2;
530 const double dp_rate = host_speed ? (params.mig_speed * params.dp_intensity) / host_speed : 1;
531 const double dp_cap = params.dp_cap;
532 const double mig_speed = params.mig_speed;
533 double max_downtime = params.max_downtime;
535 double mig_timeout = 10000000.0;
537 double remaining_size = static_cast<double>(ramsize + devsize);
538 double threshold = 0.0;
540 /* check parameters */
542 XBT_WARN("migrate a VM, but ramsize is zero");
544 if (max_downtime <= 0) {
545 XBT_WARN("use the default max_downtime value 30ms");
549 /* Stage1: send all memory pages to the destination. */
550 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
551 start_dirty_page_tracking(ms->vm);
553 double computed_during_stage1 = 0;
554 if (not skip_stage1) {
555 double clock_prev_send = MSG_get_clock();
558 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
559 * transfer exceeds the timeout value. */
560 XBT_VERB("Stage 1: Gonna send %llu bytes", ramsize);
561 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
562 remaining_size -= sent;
563 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
565 if (sent < ramsize) {
566 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
568 } else if (sent > ramsize)
573 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
574 // Stop the dirty page tracking an return (there is no memory space to release)
575 stop_dirty_page_tracking(ms->vm);
579 double clock_post_send = MSG_get_clock();
580 mig_timeout -= (clock_post_send - clock_prev_send);
581 if (mig_timeout < 0) {
582 XBT_VERB("The duration of stage 1 exceeds the timeout value, skip stage 2");
586 /* estimate bandwidth */
587 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
588 threshold = bandwidth * max_downtime;
589 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
593 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
594 if (not skip_stage2) {
596 int stage2_round = 0;
599 sg_size_t updated_size = 0;
600 if (stage2_round == 0) {
601 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
602 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
604 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
605 updated_size = get_updated_size(computed, dp_rate, dp_cap);
608 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
609 updated_size, computed_during_stage1, dp_rate, dp_cap);
611 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
612 remaining_size += updated_size;
613 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
614 (remaining_size < threshold) ? "<" : ">", threshold);
615 if (remaining_size < threshold)
619 double clock_prev_send = MSG_get_clock();
621 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
622 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
624 } catch (xbt_ex& e) {
625 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
627 // Stop the dirty page tracking an return (there is no memory space to release)
628 stop_dirty_page_tracking(ms->vm);
631 double clock_post_send = MSG_get_clock();
633 if (sent == updated_size) {
634 /* timeout did not happen */
635 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
636 threshold = bandwidth * max_downtime;
637 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
638 remaining_size -= sent;
640 mig_timeout -= (clock_post_send - clock_prev_send);
641 xbt_assert(mig_timeout > 0);
643 } else if (sent < updated_size) {
644 /* When timeout happens, we move to stage 3. The size of memory pages
645 * updated before timeout must be added to the remaining size. */
646 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
647 updated_size, (clock_post_send - clock_prev_send));
648 remaining_size -= sent;
650 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
651 updated_size = get_updated_size(computed, dp_rate, dp_cap);
652 remaining_size += updated_size;
659 /* Stage3: stop the VM and copy the rest of states. */
660 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
661 simgrid::vm::VirtualMachineImpl* pimpl = ms->vm->pimpl_vm_;
662 pimpl->setState(SURF_VM_STATE_RUNNING); // FIXME: this bypass of the checks in suspend() is not nice
663 pimpl->isMigrating = false; // FIXME: this bypass of the checks in suspend() is not nice
664 pimpl->suspend(SIMIX_process_self());
665 stop_dirty_page_tracking(ms->vm);
668 XBT_DEBUG("Stage 3: Gonna send %f bytes", remaining_size);
669 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, static_cast<sg_size_t>(remaining_size), ms->mbox, 3, 0,
673 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
674 // Stop the dirty page tracking an return (there is no memory space to release)
675 ms->vm->pimpl_vm_->resume();
679 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
680 // effectively the VM on the DST node.
681 XBT_DEBUG("mig: tx_done");
686 /** @brief Migrate the VM to the given host.
689 void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm)
692 * - One approach is ...
693 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
694 * receive the state of the source
695 * VM over network. We will finally destroy the source VM.
696 * - This behavior is similar to the way of migration in the real world.
697 * Even before a migration is completed, we will see a destination VM, consuming resources.
698 * - We have to relocate all processes. The existing process migration code will work for this?
699 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
701 * - Another one is ...
702 * We update the information of the given VM to place it to the destination physical host.
704 * The second one would be easier.
707 msg_host_t src_pm = vm->pimpl_vm_->getPm();
710 THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->getCname(), src_pm->getCname());
712 THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->getCname(), dst_pm->getCname());
713 if (not MSG_vm_is_running(vm))
714 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->getCname());
715 if (vm->isMigrating())
716 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->getCname());
718 vm->pimpl_vm_->isMigrating = true;
720 migration_session ms;
725 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
726 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
728 simgrid::xbt::string_printf("__mbox_mig_ctl:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
729 ms.mbox = simgrid::xbt::string_printf("__mbox_mig_src_dst:%s(%s-%s)", vm->getCname(), src_pm->getCname(),
732 std::string pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
733 std::string pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
735 MSG_process_create(pr_rx_name.c_str(), migration_rx_fun, &ms, dst_pm);
737 MSG_process_create(pr_tx_name.c_str(), migration_tx_fun, &ms, src_pm);
739 /* wait until the migration have finished or on error has occurred */
740 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
741 msg_task_t task = nullptr;
742 msg_error_t ret = MSG_task_receive(&task, ms.mbox_ctl.c_str());
744 vm->pimpl_vm_->isMigrating = false;
746 if (ret == MSG_HOST_FAILURE) {
747 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
748 // other side. Hence, just throw the execption
749 XBT_ERROR("SRC crashes, throw an exception (m-control)");
750 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
752 THROWF(host_error, 0, "Source host '%s' failed during the migration of VM '%s'.", src_pm->getCname(),
754 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
755 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
756 XBT_ERROR("DST crashes, throw an exception (m-control)");
757 THROWF(host_error, 0, "Destination host '%s' failed during the migration of VM '%s'.", dst_pm->getCname(),
761 xbt_assert(get_mig_task_name(vm, src_pm, dst_pm, 4) == task->name);
762 MSG_task_destroy(task);
765 /** @brief Immediately suspend the execution of all processes within the given VM.
768 * This function stops the execution of the VM. All the processes on this VM
769 * will pause. The state of the VM is preserved. We can later resume it again.
771 * No suspension cost occurs.
773 void MSG_vm_suspend(msg_vm_t vm)
775 smx_actor_t issuer = SIMIX_process_self();
776 simgrid::simix::kernelImmediate([vm, issuer]() { vm->pimpl_vm_->suspend(issuer); });
778 XBT_DEBUG("vm_suspend done");
780 if (TRACE_msg_vm_is_enabled()) {
781 simgrid::instr::StateType* state = simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE");
782 state->addEntityValue("suspend", "1 0 0"); // suspend is red
783 state->pushEvent("suspend");
787 /** @brief Resume the execution of the VM. All processes on the VM run again.
790 * No resume cost occurs.
792 void MSG_vm_resume(msg_vm_t vm)
794 vm->pimpl_vm_->resume();
796 if (TRACE_msg_vm_is_enabled())
797 simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE")->popEvent();
800 /** @brief Get the physical host of a given VM.
803 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
808 /** @brief Set a CPU bound for a given VM.
811 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
814 * On PM0, there are Task1 and VM0.
815 * On VM0, there is Task2.
816 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
818 * Task1\@PM0 gets 50%.
819 * Task2\@VM0 gets 25%.
820 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
822 * 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
823 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
825 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
826 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
827 * on the VM in share_resource() or we have to trap both the start and end of task execution.
829 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
831 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
832 * computation task (or a VM) never exceeds the capacity of a CPU core.
834 void MSG_vm_set_bound(msg_vm_t vm, double bound)
836 simgrid::simix::kernelImmediate([vm, bound]() { vm->pimpl_vm_->setBound(bound); });