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 ********* **** */
36 /** \ingroup m_vm_management
37 * \brief Set the parameters of a given host
40 * \param params a parameter object
42 void MSG_vm_set_params(msg_vm_t vm, vm_params_t params)
44 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(params);
47 /** \ingroup m_vm_management
48 * \brief Get the parameters of a given host
50 * \param vm the vm you are interested into
51 * \param params a prameter object
53 void MSG_vm_get_params(msg_vm_t vm, vm_params_t params)
55 static_cast<simgrid::s4u::VirtualMachine*>(vm)->parameters(params);
58 /* **** Check state of a VM **** */
59 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
61 simgrid::s4u::VirtualMachine* castedVm = static_cast<simgrid::s4u::VirtualMachine*>(vm);
62 return castedVm->pimpl_vm_ != nullptr && castedVm->pimpl_vm_->getState() == state;
65 /** @brief Returns whether the given VM has just created, not running.
68 int MSG_vm_is_created(msg_vm_t vm)
70 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
73 /** @brief Returns whether the given VM is currently running
76 int MSG_vm_is_running(msg_vm_t vm)
78 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
81 /** @brief Returns whether the given VM is currently migrating
84 int MSG_vm_is_migrating(msg_vm_t vm)
86 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->isMigrating();
89 /** @brief Returns whether the given VM is currently suspended, not running.
92 int MSG_vm_is_suspended(msg_vm_t vm)
94 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
97 /* **** ******** MSG vm actions ********* **** */
98 /** @brief Create a new VM with specified parameters.
100 * @param pm Physical machine that will host the VM
101 * @param name Must be unique
102 * @param ramsize [TODO]
103 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
104 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
106 msg_vm_t MSG_vm_create(msg_host_t pm, const char* name, int ramsize, int mig_netspeed, int dp_intensity)
108 simgrid::vm::VmHostExt::ensureVmExtInstalled();
110 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
111 double host_speed = MSG_host_get_speed(pm);
112 double update_speed = (static_cast<double>(dp_intensity)/100) * mig_netspeed;
114 msg_vm_t vm = MSG_vm_create_core(pm, name);
115 s_vm_params_t params;
116 memset(¶ms, 0, sizeof(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.dp_rate = (update_speed * 1024 * 1024) / host_speed;
122 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
123 params.mig_speed = static_cast<double>(mig_netspeed) * 1024 * 1024; // mig_speed
125 XBT_DEBUG("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f", params.dp_rate,
126 params.mig_speed, dp_intensity, update_speed, host_speed);
127 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(¶ms);
132 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
135 * A VM is treated as a host. The name of the VM must be unique among all hosts.
137 msg_vm_t MSG_vm_create_core(msg_host_t pm, const char* name)
139 xbt_assert(sg_host_by_name(name) == nullptr,
140 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
142 return new simgrid::s4u::VirtualMachine(name, pm);
145 /** @brief Destroy a VM. Destroy the VM object from the simulation.
148 void MSG_vm_destroy(msg_vm_t vm)
150 if (MSG_vm_is_migrating(vm))
151 THROWF(vm_error, 0, "Cannot destroy VM '%s', which is migrating.", vm->cname());
153 /* First, terminate all processes on the VM if necessary */
154 if (MSG_vm_is_running(vm))
157 /* Then, destroy the VM object */
158 simgrid::simix::kernelImmediate([vm]() {
162 if (TRACE_msg_vm_is_enabled()) {
163 container_t container = PJ_container_get(vm->cname());
164 PJ_container_remove_from_parent(container);
165 PJ_container_free(container);
169 /** @brief Start a vm (i.e., boot the guest operating system)
172 * If the VM cannot be started (because of memory over-provisioning), an exception is generated.
174 void MSG_vm_start(msg_vm_t vm)
176 simgrid::simix::kernelImmediate([vm]() {
177 simgrid::vm::VmHostExt::ensureVmExtInstalled();
179 simgrid::s4u::VirtualMachine* typedVM = static_cast<simgrid::s4u::VirtualMachine*>(vm);
180 simgrid::s4u::Host* pm = typedVM->pimpl_vm_->getPm();
181 if (pm->extension<simgrid::vm::VmHostExt>() == nullptr)
182 pm->extension_set(new simgrid::vm::VmHostExt());
184 long pm_ramsize = pm->extension<simgrid::vm::VmHostExt>()->ramsize;
185 int pm_overcommit = pm->extension<simgrid::vm::VmHostExt>()->overcommit;
186 long vm_ramsize = typedVM->getRamsize();
188 if (pm_ramsize && not pm_overcommit) { /* Only verify that we don't overcommit on need */
189 /* Retrieve the memory occupied by the VMs on that host. Yep, we have to traverse all VMs of all hosts for that */
190 long total_ramsize_of_vms = 0;
191 for (simgrid::s4u::VirtualMachine* ws_vm : simgrid::vm::VirtualMachineImpl::allVms_)
192 if (pm == ws_vm->pimpl_vm_->getPm())
193 total_ramsize_of_vms += ws_vm->pimpl_vm_->getRamsize();
195 if (vm_ramsize > pm_ramsize - total_ramsize_of_vms) {
196 XBT_WARN("cannnot start %s@%s due to memory shortage: vm_ramsize %ld, free %ld, pm_ramsize %ld (bytes).",
197 vm->cname(), pm->cname(), vm_ramsize, pm_ramsize - total_ramsize_of_vms, pm_ramsize);
198 THROWF(vm_error, 0, "Memory shortage on host '%s', VM '%s' cannot be started", pm->cname(), vm->cname());
202 typedVM->pimpl_vm_->setState(SURF_VM_STATE_RUNNING);
205 if (TRACE_msg_vm_is_enabled()) {
206 container_t vm_container = PJ_container_get(vm->cname());
207 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
208 val_t value = PJ_value_get_or_new("start", "0 0 1", type); // start is blue
209 new PushStateEvent(MSG_get_clock(), vm_container, type, value);
213 /** @brief Immediately kills all processes within the given VM.
216 * Any memory that they allocated will be leaked, unless you used #MSG_process_on_exit().
218 * No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep().
220 void MSG_vm_shutdown(msg_vm_t vm)
222 smx_actor_t issuer=SIMIX_process_self();
223 simgrid::simix::kernelImmediate([vm,issuer]() {
224 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->shutdown(issuer);
227 // Make sure that the processes in the VM are killed in this scheduling round before processing
228 // (eg with the VM destroy)
229 MSG_process_sleep(0.);
232 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
234 return bprintf("__pr_mig_tx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
237 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
239 return bprintf("__pr_mig_rx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
242 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
244 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm->cname(), src_pm->cname(), dst_pm->cname());
247 struct migration_session {
252 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
253 * know the completion of the migration. */
255 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
259 static int migration_rx_fun(int argc, char *argv[])
261 XBT_DEBUG("mig: rx_start");
263 // The structure has been created in the do_migration function and should only be freed in the same place ;)
264 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
266 bool received_finalize = false;
268 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
269 while (not received_finalize) {
270 msg_task_t task = nullptr;
271 int ret = MSG_task_recv(&task, ms->mbox);
274 // An error occurred, clean the code and return
275 // The owner did not change, hence the task should be only destroyed on the other side
276 xbt_free(finalize_task_name);
280 if (strcmp(task->name, finalize_task_name) == 0)
281 received_finalize = 1;
283 MSG_task_destroy(task);
285 xbt_free(finalize_task_name);
287 // Here Stage 1, 2 and 3 have been performed.
288 // Hence complete the migration
290 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
291 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
292 // structure at the beginning and free it at the end of each function)
293 simgrid::s4u::VirtualMachine* vm = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm);
294 msg_host_t dst_pm = ms->dst_pm;
296 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
297 simgrid::simix::kernelImmediate([vm, dst_pm]() {
298 /* Update the vm location */
299 /* precopy migration makes the VM temporally paused */
300 xbt_assert(vm->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
302 /* Update the vm location and resume it */
303 vm->pimpl_vm_->setPm(dst_pm);
304 vm->pimpl_vm_->resume();
308 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
309 vm->pimpl_vm_->isMigrating = false;
310 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->cname(), ms->src_pm->cname(), ms->dst_pm->cname());
312 if (TRACE_msg_vm_is_enabled()) {
313 static long long int counter = 0;
314 char key[INSTR_DEFAULT_STR_SIZE];
315 snprintf(key, INSTR_DEFAULT_STR_SIZE, "%lld", counter);
319 container_t msg = PJ_container_get(vm->cname());
320 type_t type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
321 new StartLinkEvent(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
323 // destroy existing container of this vm
324 container_t existing_container = PJ_container_get(vm->cname());
325 PJ_container_remove_from_parent(existing_container);
326 PJ_container_free(existing_container);
328 // create new container on the new_host location
329 PJ_container_new(vm->cname(), INSTR_MSG_VM, PJ_container_get(ms->dst_pm->cname()));
332 msg = PJ_container_get(vm->cname());
333 type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
334 new EndLinkEvent(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
337 // Inform the SRC that the migration has been correctly performed
338 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
339 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
340 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
341 // xbt_assert(ret == MSG_OK);
342 if(ret == MSG_HOST_FAILURE){
343 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
344 // has been correctly migrated on the DST node
345 // TODO What does it mean ? What should we do ?
346 MSG_task_destroy(task);
347 } else if(ret == MSG_TRANSFER_FAILURE){
348 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
349 MSG_task_destroy(task);
353 XBT_DEBUG("mig: rx_done");
357 static void start_dirty_page_tracking(msg_vm_t vm)
359 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
361 pimpl->dp_enabled = 1;
362 if (not pimpl->dp_objs)
366 xbt_dict_cursor_t cursor = nullptr;
367 dirty_page_t dp = nullptr;
368 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
369 double remaining = MSG_task_get_flops_amount(dp->task);
370 dp->prev_clock = MSG_get_clock();
371 dp->prev_remaining = remaining;
373 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
377 static void stop_dirty_page_tracking(msg_vm_t vm)
379 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
382 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
384 double computed = dp->prev_remaining - remaining;
385 double duration = clock - dp->prev_clock;
387 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", key, vm->cname(), computed,
388 dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
393 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
395 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
399 xbt_dict_cursor_t cursor = nullptr;
400 dirty_page_t dp = nullptr;
401 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
402 double remaining = MSG_task_get_flops_amount(dp->task);
404 double clock = MSG_get_clock();
406 // total += calc_updated_pages(key, vm, dp, remaining, clock);
407 total += get_computed(key, vm, dp, remaining, clock);
409 dp->prev_remaining = remaining;
410 dp->prev_clock = clock;
413 total += pimpl->dp_updated_by_deleted_tasks;
415 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
416 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
418 pimpl->dp_updated_by_deleted_tasks = 0;
423 // TODO Is this code redundant with the information provided by
424 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
425 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
426 void MSG_host_add_task(msg_host_t host, msg_task_t task)
428 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
431 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
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 (pimpl->dp_enabled) {
439 dp->prev_clock = MSG_get_clock();
440 dp->prev_remaining = remaining;
442 if (not pimpl->dp_objs)
443 pimpl->dp_objs = xbt_dict_new_homogeneous(nullptr);
444 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
445 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
446 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, host->cname(), remaining, pimpl->dp_enabled);
451 void MSG_host_del_task(msg_host_t host, msg_task_t task)
453 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
456 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
458 char *key = bprintf("%s-%p", task->name, task);
459 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
460 xbt_assert(dp->task == task);
462 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
463 * the information for the lookup_() function that will called soon. */
464 if (pimpl->dp_enabled) {
465 double remaining = MSG_task_get_flops_amount(task);
466 double clock = MSG_get_clock();
467 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
468 double updated = get_computed(key, host, dp, remaining, clock);
470 pimpl->dp_updated_by_deleted_tasks += updated;
473 xbt_dict_remove(pimpl->dp_objs, key);
476 XBT_DEBUG("del %s on %s", key, host->cname());
480 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,
481 int stage, int stage2_round, double mig_speed, double timeout)
484 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
485 msg_task_t task = MSG_task_create(task_name, 0, static_cast<double>(size), nullptr);
489 double clock_sta = MSG_get_clock();
493 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
495 ret = MSG_task_send(task, mbox);
501 } else if (ret == MSG_TIMEOUT) {
502 sg_size_t remaining = static_cast<sg_size_t>(MSG_task_get_remaining_communication(task));
503 sent = size - remaining;
504 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
507 /* FIXME: why try-and-catch is used here? */
508 if(ret == MSG_HOST_FAILURE){
509 XBT_DEBUG("SRC host failed during migration of %s (stage %d)", vm->cname(), stage);
510 MSG_task_destroy(task);
511 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", vm->cname(), stage);
512 }else if(ret == MSG_TRANSFER_FAILURE){
513 XBT_DEBUG("DST host failed during migration of %s (stage %d)", vm->cname(), stage);
514 MSG_task_destroy(task);
515 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", vm->cname(), stage);
518 double clock_end = MSG_get_clock();
519 double duration = clock_end - clock_sta;
520 double actual_speed = size / duration;
523 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
524 actual_speed, mig_speed);
526 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
532 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
534 double updated_size = computed * dp_rate;
535 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
536 if (updated_size > dp_cap) {
537 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
539 updated_size = dp_cap;
542 return static_cast<sg_size_t>(updated_size);
545 static int migration_tx_fun(int argc, char *argv[])
547 XBT_DEBUG("mig: tx_start");
549 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
550 migration_session *ms = static_cast<migration_session *>(MSG_process_get_data(MSG_process_self()));
552 s_vm_params_t params;
553 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
554 const sg_size_t ramsize = params.ramsize;
555 const sg_size_t devsize = params.devsize;
556 const int skip_stage1 = params.skip_stage1;
557 int skip_stage2 = params.skip_stage2;
558 const double dp_rate = params.dp_rate;
559 const double dp_cap = params.dp_cap;
560 const double mig_speed = params.mig_speed;
561 double max_downtime = params.max_downtime;
563 double mig_timeout = 10000000.0;
565 double remaining_size = static_cast<double>(ramsize + devsize);
566 double threshold = 0.0;
568 /* check parameters */
570 XBT_WARN("migrate a VM, but ramsize is zero");
572 if (max_downtime <= 0) {
573 XBT_WARN("use the default max_downtime value 30ms");
577 /* Stage1: send all memory pages to the destination. */
578 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
579 start_dirty_page_tracking(ms->vm);
581 double computed_during_stage1 = 0;
582 if (not skip_stage1) {
583 double clock_prev_send = MSG_get_clock();
586 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
587 * transfer exceeds the timeout value. */
588 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
589 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
590 remaining_size -= sent;
591 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
593 if (sent < ramsize) {
594 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
596 } else if (sent > ramsize)
601 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
602 // Stop the dirty page tracking an return (there is no memory space to release)
603 stop_dirty_page_tracking(ms->vm);
607 double clock_post_send = MSG_get_clock();
608 mig_timeout -= (clock_post_send - clock_prev_send);
609 if (mig_timeout < 0) {
610 XBT_VERB("The duration of stage 1 exceeds the timeout value, skip stage 2");
614 /* estimate bandwidth */
615 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
616 threshold = bandwidth * max_downtime;
617 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
621 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
622 if (not skip_stage2) {
624 int stage2_round = 0;
627 sg_size_t updated_size = 0;
628 if (stage2_round == 0) {
629 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
630 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
632 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
633 updated_size = get_updated_size(computed, dp_rate, dp_cap);
636 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
637 updated_size, computed_during_stage1, dp_rate, dp_cap);
639 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
640 remaining_size += updated_size;
641 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
642 (remaining_size < threshold) ? "<" : ">", threshold);
643 if (remaining_size < threshold)
647 double clock_prev_send = MSG_get_clock();
649 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
650 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
652 } catch (xbt_ex& e) {
653 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
655 // Stop the dirty page tracking an return (there is no memory space to release)
656 stop_dirty_page_tracking(ms->vm);
659 double clock_post_send = MSG_get_clock();
661 if (sent == updated_size) {
662 /* timeout did not happen */
663 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
664 threshold = bandwidth * max_downtime;
665 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
666 remaining_size -= sent;
668 mig_timeout -= (clock_post_send - clock_prev_send);
669 xbt_assert(mig_timeout > 0);
671 } else if (sent < updated_size) {
672 /* When timeout happens, we move to stage 3. The size of memory pages
673 * updated before timeout must be added to the remaining size. */
674 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
675 updated_size, (clock_post_send - clock_prev_send));
676 remaining_size -= sent;
678 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
679 updated_size = get_updated_size(computed, dp_rate, dp_cap);
680 remaining_size += updated_size;
687 /* Stage3: stop the VM and copy the rest of states. */
688 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
689 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->pimpl_vm_;
690 pimpl->setState(SURF_VM_STATE_RUNNING); // FIXME: this bypass of the checks in suspend() is not nice
691 pimpl->isMigrating = false; // FIXME: this bypass of the checks in suspend() is not nice
692 pimpl->suspend(SIMIX_process_self());
693 stop_dirty_page_tracking(ms->vm);
696 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
697 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, static_cast<sg_size_t>(remaining_size), ms->mbox, 3, 0,
701 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
702 // Stop the dirty page tracking an return (there is no memory space to release)
703 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->pimpl_vm_->resume();
707 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
708 // effectively the VM on the DST node.
709 XBT_DEBUG("mig: tx_done");
714 /** @brief Migrate the VM to the given host.
717 void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm)
720 * - One approach is ...
721 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
722 * receive the state of the source
723 * VM over network. We will finally destroy the source VM.
724 * - This behavior is similar to the way of migration in the real world.
725 * Even before a migration is completed, we will see a destination VM, consuming resources.
726 * - We have to relocate all processes. The existing process migration code will work for this?
727 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
729 * - Another one is ...
730 * We update the information of the given VM to place it to the destination physical host.
732 * The second one would be easier.
735 simgrid::s4u::VirtualMachine* typedVm = static_cast<simgrid::s4u::VirtualMachine*>(vm);
736 simgrid::vm::VirtualMachineImpl* pimpl = typedVm->pimpl_vm_;
737 msg_host_t src_pm = pimpl->getPm();
740 THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->cname(), src_pm->cname());
742 THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->cname(), dst_pm->cname());
743 if (not MSG_vm_is_running(vm))
744 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->cname());
745 if (typedVm->isMigrating())
746 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->cname());
748 pimpl->isMigrating = true;
750 struct migration_session *ms = xbt_new(struct migration_session, 1);
755 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
756 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
757 ms->mbox_ctl = bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
758 ms->mbox = bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
760 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
761 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
763 char** argv = xbt_new(char*, 2);
764 argv[0] = pr_rx_name;
766 MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
768 argv = xbt_new(char*, 2);
769 argv[0] = pr_tx_name;
771 MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
773 /* wait until the migration have finished or on error has occurred */
774 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
775 msg_task_t task = nullptr;
776 msg_error_t ret = MSG_task_receive(&task, ms->mbox_ctl);
778 pimpl->isMigrating = false;
780 xbt_free(ms->mbox_ctl);
784 if (ret == MSG_HOST_FAILURE) {
785 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
786 // other side. Hence, just throw the execption
787 XBT_ERROR("SRC crashes, throw an exception (m-control)");
788 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
790 THROWF(host_error, 0, "Source host '%s' failed during the migration of VM '%s'.", src_pm->cname(), vm->cname());
791 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
792 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
793 XBT_ERROR("DST crashes, throw an exception (m-control)");
794 THROWF(host_error, 0, "Destination host '%s' failed during the migration of VM '%s'.", dst_pm->cname(),
798 char* expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
799 xbt_assert(strcmp(task->name, expected_task_name) == 0);
800 xbt_free(expected_task_name);
801 MSG_task_destroy(task);
804 /** @brief Immediately suspend the execution of all processes within the given VM.
807 * This function stops the execution of the VM. All the processes on this VM
808 * will pause. The state of the VM is preserved. We can later resume it again.
810 * No suspension cost occurs.
812 void MSG_vm_suspend(msg_vm_t vm)
814 smx_actor_t issuer = SIMIX_process_self();
815 simgrid::simix::kernelImmediate([vm,issuer]() {
816 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->suspend(issuer);
819 XBT_DEBUG("vm_suspend done");
821 if (TRACE_msg_vm_is_enabled()) {
822 container_t vm_container = PJ_container_get(vm->cname());
823 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
824 val_t value = PJ_value_get_or_new("suspend", "1 0 0", type); // suspend is red
825 new PushStateEvent(MSG_get_clock(), vm_container, type, value);
829 /** @brief Resume the execution of the VM. All processes on the VM run again.
832 * No resume cost occurs.
834 void MSG_vm_resume(msg_vm_t vm)
836 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->resume();
838 if (TRACE_msg_vm_is_enabled()) {
839 container_t vm_container = PJ_container_get(vm->cname());
840 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
841 new PopStateEvent(MSG_get_clock(), vm_container, type);
845 /** @brief Get the physical host of a given VM.
848 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
850 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->pm();
853 /** @brief Set a CPU bound for a given VM.
856 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
859 * On PM0, there are Task1 and VM0.
860 * On VM0, there is Task2.
861 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
863 * Task1\@PM0 gets 50%.
864 * Task2\@VM0 gets 25%.
865 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
867 * 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
868 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
870 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
871 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
872 * on the VM in share_resource() or we have to trap both the start and end of task execution.
874 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
876 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
877 * computation task (or a VM) never exceeds the capacity of a CPU core.
879 void MSG_vm_set_bound(msg_vm_t vm, double bound)
881 simgrid::simix::kernelImmediate(
882 [vm, bound]() { static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->setBound(bound); });