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)
187 simgrid::simix::kernelImmediate([vm]() {
188 simgrid::vm::VmHostExt::ensureVmExtInstalled();
190 simgrid::s4u::Host* pm = vm->pimpl_vm_->getPm();
191 if (pm->extension<simgrid::vm::VmHostExt>() == nullptr)
192 pm->extension_set(new simgrid::vm::VmHostExt());
194 long pm_ramsize = pm->extension<simgrid::vm::VmHostExt>()->ramsize;
195 int pm_overcommit = pm->extension<simgrid::vm::VmHostExt>()->overcommit;
196 long vm_ramsize = vm->getRamsize();
198 if (pm_ramsize && not pm_overcommit) { /* Only verify that we don't overcommit on need */
199 /* Retrieve the memory occupied by the VMs on that host. Yep, we have to traverse all VMs of all hosts for that */
200 long total_ramsize_of_vms = 0;
201 for (simgrid::s4u::VirtualMachine* ws_vm : simgrid::vm::VirtualMachineImpl::allVms_)
202 if (pm == ws_vm->pimpl_vm_->getPm())
203 total_ramsize_of_vms += ws_vm->pimpl_vm_->getRamsize();
205 if (vm_ramsize > pm_ramsize - total_ramsize_of_vms) {
206 XBT_WARN("cannnot start %s@%s due to memory shortage: vm_ramsize %ld, free %ld, pm_ramsize %ld (bytes).",
207 vm->getCname(), pm->getCname(), vm_ramsize, pm_ramsize - total_ramsize_of_vms, pm_ramsize);
208 THROWF(vm_error, 0, "Memory shortage on host '%s', VM '%s' cannot be started", pm->getCname(), vm->getCname());
212 vm->pimpl_vm_->setState(SURF_VM_STATE_RUNNING);
215 if (TRACE_msg_vm_is_enabled()) {
216 container_t vm_container = PJ_container_get(vm->getCname());
217 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
218 val_t value = PJ_value_get_or_new("start", "0 0 1", type); // start is blue
219 new PushStateEvent(MSG_get_clock(), vm_container, type, value);
223 /** @brief Immediately kills all processes within the given VM.
226 * Any memory that they allocated will be leaked, unless you used #MSG_process_on_exit().
228 * No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep().
230 void MSG_vm_shutdown(msg_vm_t vm)
232 smx_actor_t issuer=SIMIX_process_self();
233 simgrid::simix::kernelImmediate([vm, issuer]() { vm->pimpl_vm_->shutdown(issuer); });
235 // Make sure that the processes in the VM are killed in this scheduling round before processing
236 // (eg with the VM destroy)
237 MSG_process_sleep(0.);
240 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
242 return bprintf("__pr_mig_tx:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
245 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
247 return bprintf("__pr_mig_rx:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
250 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
252 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm->getCname(), src_pm->getCname(), dst_pm->getCname());
255 struct migration_session {
260 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
261 * know the completion of the migration. */
263 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
267 static int migration_rx_fun(int argc, char *argv[])
269 XBT_DEBUG("mig: rx_start");
271 // The structure has been created in the do_migration function and should only be freed in the same place ;)
272 struct migration_session* ms = static_cast<migration_session*>(MSG_process_get_data(MSG_process_self()));
274 bool received_finalize = false;
276 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
277 while (not received_finalize) {
278 msg_task_t task = nullptr;
279 int ret = MSG_task_recv(&task, ms->mbox);
282 // An error occurred, clean the code and return
283 // The owner did not change, hence the task should be only destroyed on the other side
284 xbt_free(finalize_task_name);
288 if (strcmp(task->name, finalize_task_name) == 0)
289 received_finalize = 1;
291 MSG_task_destroy(task);
293 xbt_free(finalize_task_name);
295 // Here Stage 1, 2 and 3 have been performed.
296 // Hence complete the migration
298 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
299 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
300 // structure at the beginning and free it at the end of each function)
301 simgrid::s4u::VirtualMachine* vm = ms->vm;
302 msg_host_t dst_pm = ms->dst_pm;
304 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
305 simgrid::simix::kernelImmediate([vm, dst_pm]() {
306 /* Update the vm location */
307 /* precopy migration makes the VM temporally paused */
308 xbt_assert(vm->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
310 /* Update the vm location and resume it */
311 vm->pimpl_vm_->setPm(dst_pm);
312 vm->pimpl_vm_->resume();
316 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
317 vm->pimpl_vm_->isMigrating = false;
318 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->getCname(), ms->src_pm->getCname(), ms->dst_pm->getCname());
320 if (TRACE_msg_vm_is_enabled()) {
321 static long long int counter = 0;
322 char key[INSTR_DEFAULT_STR_SIZE];
323 snprintf(key, INSTR_DEFAULT_STR_SIZE, "%lld", counter);
327 container_t msg = PJ_container_get(vm->getCname());
328 type_t type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
329 new StartLinkEvent(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
331 // destroy existing container of this vm
332 container_t existing_container = PJ_container_get(vm->getCname());
333 PJ_container_remove_from_parent(existing_container);
334 PJ_container_free(existing_container);
336 // create new container on the new_host location
337 PJ_container_new(vm->getCname(), INSTR_MSG_VM, PJ_container_get(ms->dst_pm->getCname()));
340 msg = PJ_container_get(vm->getCname());
341 type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
342 new EndLinkEvent(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
345 // Inform the SRC that the migration has been correctly performed
346 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
347 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
348 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
349 // xbt_assert(ret == MSG_OK);
350 if(ret == MSG_HOST_FAILURE){
351 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
352 // has been correctly migrated on the DST node
353 // TODO What does it mean ? What should we do ?
354 MSG_task_destroy(task);
355 } else if(ret == MSG_TRANSFER_FAILURE){
356 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
357 MSG_task_destroy(task);
361 XBT_DEBUG("mig: rx_done");
365 static void start_dirty_page_tracking(msg_vm_t vm)
367 vm->pimpl_vm_->dp_enabled = 1;
368 if (vm->pimpl_vm_->dp_objs.empty())
371 for (auto elm : vm->pimpl_vm_->dp_objs) {
372 dirty_page_t dp = elm.second;
373 double remaining = MSG_task_get_flops_amount(dp->task);
374 dp->prev_clock = MSG_get_clock();
375 dp->prev_remaining = remaining;
376 XBT_DEBUG("%s@%s remaining %f", elm.first.c_str(), vm->getCname(), remaining);
380 static void stop_dirty_page_tracking(msg_vm_t vm)
382 vm->pimpl_vm_->dp_enabled = 0;
385 static double get_computed(const char* key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
387 double computed = dp->prev_remaining - remaining;
388 double duration = clock - dp->prev_clock;
390 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", key, vm->getCname(), computed,
391 dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
396 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
400 for (auto elm : vm->pimpl_vm_->dp_objs) {
401 const char* key = elm.first.c_str();
402 dirty_page_t dp = elm.second;
403 double remaining = MSG_task_get_flops_amount(dp->task);
405 double clock = MSG_get_clock();
407 // total += calc_updated_pages(key, vm, dp, remaining, clock);
408 total += get_computed(key, vm, dp, remaining, clock);
410 dp->prev_remaining = remaining;
411 dp->prev_clock = clock;
414 total += vm->pimpl_vm_->dp_updated_by_deleted_tasks;
416 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
417 stage2_round_for_fancy_debug, total, vm->pimpl_vm_->dp_updated_by_deleted_tasks);
419 vm->pimpl_vm_->dp_updated_by_deleted_tasks = 0;
424 // TODO Is this code redundant with the information provided by
425 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
426 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
427 void MSG_host_add_task(msg_host_t host, msg_task_t task)
429 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
433 double remaining = MSG_task_get_flops_amount(task);
434 char *key = bprintf("%s-%p", task->name, task);
436 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
438 if (vm->pimpl_vm_->dp_enabled) {
439 dp->prev_clock = MSG_get_clock();
440 dp->prev_remaining = remaining;
442 vm->pimpl_vm_->dp_objs.insert({key, dp});
443 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, host->getCname(), remaining, vm->pimpl_vm_->dp_enabled);
448 void MSG_host_del_task(msg_host_t host, msg_task_t task)
450 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
454 char *key = bprintf("%s-%p", task->name, task);
455 dirty_page_t dp = nullptr;
456 if (vm->pimpl_vm_->dp_objs.find(key) != vm->pimpl_vm_->dp_objs.end())
457 dp = vm->pimpl_vm_->dp_objs.at(key);
458 xbt_assert(dp && dp->task == task);
460 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
461 * the information for the lookup_() function that will called soon. */
462 if (vm->pimpl_vm_->dp_enabled) {
463 double remaining = MSG_task_get_flops_amount(task);
464 double clock = MSG_get_clock();
465 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
466 double updated = get_computed(key, vm, dp, remaining, clock); // was host instead of vm
468 vm->pimpl_vm_->dp_updated_by_deleted_tasks += updated;
471 vm->pimpl_vm_->dp_objs.erase(key);
474 XBT_DEBUG("del %s on %s", key, host->getCname());
478 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,
479 int stage, int stage2_round, double mig_speed, double timeout)
482 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
483 msg_task_t task = MSG_task_create(task_name, 0, static_cast<double>(size), nullptr);
485 double clock_sta = MSG_get_clock();
489 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
491 ret = MSG_task_send(task, mbox);
497 } else if (ret == MSG_TIMEOUT) {
498 sg_size_t remaining = static_cast<sg_size_t>(MSG_task_get_remaining_communication(task));
499 sent = size - remaining;
500 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
503 /* FIXME: why try-and-catch is used here? */
504 if(ret == MSG_HOST_FAILURE){
505 XBT_DEBUG("SRC host failed during migration of %s (stage %d)", vm->getCname(), stage);
506 MSG_task_destroy(task);
507 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", vm->getCname(), stage);
508 }else if(ret == MSG_TRANSFER_FAILURE){
509 XBT_DEBUG("DST host failed during migration of %s (stage %d)", vm->getCname(), stage);
510 MSG_task_destroy(task);
511 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", vm->getCname(), stage);
514 double clock_end = MSG_get_clock();
515 double duration = clock_end - clock_sta;
516 double actual_speed = size / duration;
519 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
520 actual_speed, mig_speed);
522 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
528 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
530 double updated_size = computed * dp_rate;
531 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
532 if (updated_size > dp_cap) {
533 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
535 updated_size = dp_cap;
538 return static_cast<sg_size_t>(updated_size);
541 static int migration_tx_fun(int argc, char *argv[])
543 XBT_DEBUG("mig: tx_start");
545 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
546 migration_session *ms = static_cast<migration_session *>(MSG_process_get_data(MSG_process_self()));
548 double host_speed = ms->vm->pimpl_vm_->getPm()->getSpeed();
549 s_vm_params_t params;
550 ms->vm->getParameters(¶ms);
551 const sg_size_t ramsize = params.ramsize;
552 const sg_size_t devsize = params.devsize;
553 const int skip_stage1 = params.skip_stage1;
554 int skip_stage2 = params.skip_stage2;
555 const double dp_rate = host_speed ? (params.mig_speed * params.dp_intensity) / host_speed : 1;
556 const double dp_cap = params.dp_cap;
557 const double mig_speed = params.mig_speed;
558 double max_downtime = params.max_downtime;
560 double mig_timeout = 10000000.0;
562 double remaining_size = static_cast<double>(ramsize + devsize);
563 double threshold = 0.0;
565 /* check parameters */
567 XBT_WARN("migrate a VM, but ramsize is zero");
569 if (max_downtime <= 0) {
570 XBT_WARN("use the default max_downtime value 30ms");
574 /* Stage1: send all memory pages to the destination. */
575 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
576 start_dirty_page_tracking(ms->vm);
578 double computed_during_stage1 = 0;
579 if (not skip_stage1) {
580 double clock_prev_send = MSG_get_clock();
583 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
584 * transfer exceeds the timeout value. */
585 XBT_VERB("Stage 1: Gonna send %llu bytes", ramsize);
586 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
587 remaining_size -= sent;
588 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
590 if (sent < ramsize) {
591 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
593 } else if (sent > ramsize)
598 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
599 // Stop the dirty page tracking an return (there is no memory space to release)
600 stop_dirty_page_tracking(ms->vm);
604 double clock_post_send = MSG_get_clock();
605 mig_timeout -= (clock_post_send - clock_prev_send);
606 if (mig_timeout < 0) {
607 XBT_VERB("The duration of stage 1 exceeds the timeout value, skip stage 2");
611 /* estimate bandwidth */
612 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
613 threshold = bandwidth * max_downtime;
614 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
618 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
619 if (not skip_stage2) {
621 int stage2_round = 0;
624 sg_size_t updated_size = 0;
625 if (stage2_round == 0) {
626 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
627 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
629 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
630 updated_size = get_updated_size(computed, dp_rate, dp_cap);
633 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
634 updated_size, computed_during_stage1, dp_rate, dp_cap);
636 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
637 remaining_size += updated_size;
638 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
639 (remaining_size < threshold) ? "<" : ">", threshold);
640 if (remaining_size < threshold)
644 double clock_prev_send = MSG_get_clock();
646 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
647 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
649 } catch (xbt_ex& e) {
650 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
652 // Stop the dirty page tracking an return (there is no memory space to release)
653 stop_dirty_page_tracking(ms->vm);
656 double clock_post_send = MSG_get_clock();
658 if (sent == updated_size) {
659 /* timeout did not happen */
660 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
661 threshold = bandwidth * max_downtime;
662 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
663 remaining_size -= sent;
665 mig_timeout -= (clock_post_send - clock_prev_send);
666 xbt_assert(mig_timeout > 0);
668 } else if (sent < updated_size) {
669 /* When timeout happens, we move to stage 3. The size of memory pages
670 * updated before timeout must be added to the remaining size. */
671 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
672 updated_size, (clock_post_send - clock_prev_send));
673 remaining_size -= sent;
675 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
676 updated_size = get_updated_size(computed, dp_rate, dp_cap);
677 remaining_size += updated_size;
684 /* Stage3: stop the VM and copy the rest of states. */
685 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
686 simgrid::vm::VirtualMachineImpl* pimpl = ms->vm->pimpl_vm_;
687 pimpl->setState(SURF_VM_STATE_RUNNING); // FIXME: this bypass of the checks in suspend() is not nice
688 pimpl->isMigrating = false; // FIXME: this bypass of the checks in suspend() is not nice
689 pimpl->suspend(SIMIX_process_self());
690 stop_dirty_page_tracking(ms->vm);
693 XBT_DEBUG("Stage 3: Gonna send %f bytes", remaining_size);
694 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, static_cast<sg_size_t>(remaining_size), ms->mbox, 3, 0,
698 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
699 // Stop the dirty page tracking an return (there is no memory space to release)
700 ms->vm->pimpl_vm_->resume();
704 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
705 // effectively the VM on the DST node.
706 XBT_DEBUG("mig: tx_done");
711 /** @brief Migrate the VM to the given host.
714 void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm)
717 * - One approach is ...
718 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
719 * receive the state of the source
720 * VM over network. We will finally destroy the source VM.
721 * - This behavior is similar to the way of migration in the real world.
722 * Even before a migration is completed, we will see a destination VM, consuming resources.
723 * - We have to relocate all processes. The existing process migration code will work for this?
724 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
726 * - Another one is ...
727 * We update the information of the given VM to place it to the destination physical host.
729 * The second one would be easier.
732 msg_host_t src_pm = vm->pimpl_vm_->getPm();
735 THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->getCname(), src_pm->getCname());
737 THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->getCname(), dst_pm->getCname());
738 if (not MSG_vm_is_running(vm))
739 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->getCname());
740 if (vm->isMigrating())
741 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->getCname());
743 vm->pimpl_vm_->isMigrating = true;
745 struct migration_session *ms = xbt_new(struct migration_session, 1);
750 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
751 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
752 ms->mbox_ctl = bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
753 ms->mbox = bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
755 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
756 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
758 char** argv = xbt_new(char*, 2);
759 argv[0] = pr_rx_name;
761 MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
763 argv = xbt_new(char*, 2);
764 argv[0] = pr_tx_name;
766 MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
768 /* wait until the migration have finished or on error has occurred */
769 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
770 msg_task_t task = nullptr;
771 msg_error_t ret = MSG_task_receive(&task, ms->mbox_ctl);
773 vm->pimpl_vm_->isMigrating = false;
775 xbt_free(ms->mbox_ctl);
779 if (ret == MSG_HOST_FAILURE) {
780 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
781 // other side. Hence, just throw the execption
782 XBT_ERROR("SRC crashes, throw an exception (m-control)");
783 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
785 THROWF(host_error, 0, "Source host '%s' failed during the migration of VM '%s'.", src_pm->getCname(),
787 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
788 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
789 XBT_ERROR("DST crashes, throw an exception (m-control)");
790 THROWF(host_error, 0, "Destination host '%s' failed during the migration of VM '%s'.", dst_pm->getCname(),
794 char* expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
795 xbt_assert(strcmp(task->name, expected_task_name) == 0);
796 xbt_free(expected_task_name);
797 MSG_task_destroy(task);
800 /** @brief Immediately suspend the execution of all processes within the given VM.
803 * This function stops the execution of the VM. All the processes on this VM
804 * will pause. The state of the VM is preserved. We can later resume it again.
806 * No suspension cost occurs.
808 void MSG_vm_suspend(msg_vm_t vm)
810 smx_actor_t issuer = SIMIX_process_self();
811 simgrid::simix::kernelImmediate([vm, issuer]() { vm->pimpl_vm_->suspend(issuer); });
813 XBT_DEBUG("vm_suspend done");
815 if (TRACE_msg_vm_is_enabled()) {
816 container_t vm_container = PJ_container_get(vm->getCname());
817 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
818 val_t value = PJ_value_get_or_new("suspend", "1 0 0", type); // suspend is red
819 new PushStateEvent(MSG_get_clock(), vm_container, type, value);
823 /** @brief Resume the execution of the VM. All processes on the VM run again.
826 * No resume cost occurs.
828 void MSG_vm_resume(msg_vm_t vm)
830 vm->pimpl_vm_->resume();
832 if (TRACE_msg_vm_is_enabled()) {
833 container_t vm_container = PJ_container_get(vm->getCname());
834 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
835 new PopStateEvent(MSG_get_clock(), vm_container, type);
839 /** @brief Get the physical host of a given VM.
842 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
847 /** @brief Set a CPU bound for a given VM.
850 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
853 * On PM0, there are Task1 and VM0.
854 * On VM0, there is Task2.
855 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
857 * Task1\@PM0 gets 50%.
858 * Task2\@VM0 gets 25%.
859 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
861 * 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
862 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
864 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
865 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
866 * on the VM in share_resource() or we have to trap both the start and end of task execution.
868 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
870 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
871 * computation task (or a VM) never exceeds the capacity of a CPU core.
873 void MSG_vm_set_bound(msg_vm_t vm, double bound)
875 simgrid::simix::kernelImmediate([vm, bound]() { vm->pimpl_vm_->setBound(bound); });