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/plugins/vm/VirtualMachineImpl.hpp"
15 #include "src/plugins/vm/VmHostExt.hpp"
16 #include "src/simix/ActorImpl.hpp"
17 #include <simgrid/s4u/VirtualMachine.hpp>
18 #include <simgrid/s4u/host.hpp>
20 #include "msg_private.h"
21 #include "xbt/sysdep.h"
23 #include "simgrid/host.h"
25 #include "src/simix/smx_host_private.h" /* don't ask me why the VM functions are in there (FIXME:KILLME) */
27 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
30 /* **** ******** GENERAL ********* **** */
32 /** \ingroup m_vm_management
33 * \brief Set the parameters of a given host
36 * \param params a parameter object
38 void MSG_vm_set_params(msg_vm_t vm, vm_params_t params)
40 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(params);
43 /** \ingroup m_vm_management
44 * \brief Get the parameters of a given host
47 * \param params a prameter object
49 void MSG_vm_get_params(msg_vm_t vm, vm_params_t params)
51 static_cast<simgrid::s4u::VirtualMachine*>(vm)->parameters(params);
54 /* **** Check state of a VM **** */
55 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
57 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getState() == state;
60 /** @brief Returns whether the given VM has just created, not running.
63 int MSG_vm_is_created(msg_vm_t vm)
65 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
68 /** @brief Returns whether the given VM is currently running
71 int MSG_vm_is_running(msg_vm_t vm)
73 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
76 /** @brief Returns whether the given VM is currently migrating
79 int MSG_vm_is_migrating(msg_vm_t vm)
81 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->isMigrating();
84 /** @brief Returns whether the given VM is currently suspended, not running.
87 int MSG_vm_is_suspended(msg_vm_t vm)
89 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
92 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
95 int MSG_vm_is_saving(msg_vm_t vm)
97 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
100 /** @brief Returns whether the given VM has been saved, not running.
103 int MSG_vm_is_saved(msg_vm_t vm)
105 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
108 /** @brief Returns whether the given VM is being restored, not running.
111 int MSG_vm_is_restoring(msg_vm_t vm)
113 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
116 /* **** ******** MSG vm actions ********* **** */
117 /** @brief Create a new VM with specified parameters.
119 * @param pm Physical machine that will host the VM
121 * @param ncpus [TODO]
122 * @param ramsize [TODO]
123 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
124 * @param disk_path (unused) Path to the image that boots
125 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
126 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
127 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
129 msg_vm_t MSG_vm_create(msg_host_t pm, const char* name, int ncpus, int ramsize, int net_cap, char* disk_path,
130 int disksize, int mig_netspeed, int dp_intensity)
132 simgrid::vm::VmHostExt::ensureVmExtInstalled();
134 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
135 double host_speed = MSG_host_get_speed(pm);
136 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
138 msg_vm_t vm = MSG_vm_create_core(pm, name);
139 s_vm_params_t params;
140 memset(¶ms, 0, sizeof(params));
141 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
142 //params.overcommit = 0;
144 params.skip_stage2 = 0;
145 params.max_downtime = 0.03;
146 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
147 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
148 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
150 //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate,
151 // params.mig_speed, dp_intensity, update_speed, host_speed);
152 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(¶ms);
157 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
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_core(msg_host_t pm, const char* name)
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 return new simgrid::s4u::VirtualMachine(name, pm);
170 /** @brief Destroy a VM. Destroy the VM object from the simulation.
173 void MSG_vm_destroy(msg_vm_t vm)
175 if (MSG_vm_is_migrating(vm))
176 THROWF(vm_error, 0, "Cannot destroy VM '%s', which is migrating.", vm->cname());
178 /* First, terminate all processes on the VM if necessary */
179 if (MSG_vm_is_running(vm))
182 xbt_assert(MSG_vm_is_created(vm), "shutdown the given VM before destroying it");
184 /* Then, destroy the VM object */
185 simgrid::simix::kernelImmediate([vm]() {
189 if (TRACE_msg_vm_is_enabled()) {
190 container_t container = PJ_container_get(vm->name().c_str());
191 PJ_container_remove_from_parent(container);
192 PJ_container_free(container);
196 /** @brief Start a vm (i.e., boot the guest operating system)
199 * If the VM cannot be started (because of memory overprovisionning), an exception is generated.
201 void MSG_vm_start(msg_vm_t vm)
203 simgrid::simix::kernelImmediate([vm]() {
204 simgrid::vm::VmHostExt::ensureVmExtInstalled();
206 simgrid::s4u::VirtualMachine* typedVM = static_cast<simgrid::s4u::VirtualMachine*>(vm);
207 simgrid::s4u::Host* pm = typedVM->pimpl_vm_->getPm();
208 if (pm->extension<simgrid::vm::VmHostExt>() == nullptr)
209 pm->extension_set(new simgrid::vm::VmHostExt());
211 long pm_ramsize = pm->extension<simgrid::vm::VmHostExt>()->ramsize;
212 int pm_overcommit = pm->extension<simgrid::vm::VmHostExt>()->overcommit;
213 long vm_ramsize = typedVM->getRamsize();
215 if (pm_ramsize && !pm_overcommit) { /* Only verify that we don't overcommit on need */
216 /* Retrieve the memory occupied by the VMs on that host. Yep, we have to traverse all VMs of all hosts for that */
217 long total_ramsize_of_vms = 0;
218 for (simgrid::s4u::VirtualMachine* ws_vm : simgrid::vm::VirtualMachineImpl::allVms_)
219 if (pm == ws_vm->pimpl_vm_->getPm())
220 total_ramsize_of_vms += ws_vm->pimpl_vm_->getRamsize();
222 if (vm_ramsize > pm_ramsize - total_ramsize_of_vms) {
223 XBT_WARN("cannnot start %s@%s due to memory shortage: vm_ramsize %ld, free %ld, pm_ramsize %ld (bytes).",
224 vm->cname(), pm->cname(), vm_ramsize, pm_ramsize - total_ramsize_of_vms, pm_ramsize);
225 THROWF(vm_error, 0, "Memory shortage on host '%s', VM '%s' cannot be started", pm->cname(), vm->cname());
229 typedVM->pimpl_vm_->setState(SURF_VM_STATE_RUNNING);
232 if (TRACE_msg_vm_is_enabled()) {
233 container_t vm_container = PJ_container_get(vm->name().c_str());
234 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
235 val_t value = PJ_value_get_or_new("start", "0 0 1", type); // start is blue
236 new_pajePushState(MSG_get_clock(), vm_container, type, value);
240 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
243 * FIXME: No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep() or something.
244 * I'm not quite sure.
246 void MSG_vm_shutdown(msg_vm_t vm)
248 simcall_vm_shutdown(vm);
250 // Make sure that the processes in the VM are killed in this scheduling round before processing
251 // (eg with the VM destroy)
252 MSG_process_sleep(0.);
255 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
257 return bprintf("__pr_mig_tx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
260 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
262 return bprintf("__pr_mig_rx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
265 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
267 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm->cname(), src_pm->cname(), dst_pm->cname());
270 struct migration_session {
275 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
276 * know the completion of the migration. */
278 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
282 static int migration_rx_fun(int argc, char *argv[])
284 XBT_DEBUG("mig: rx_start");
286 // The structure has been created in the do_migration function and should only be freed in the same place ;)
287 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
289 bool received_finalize = false;
291 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
292 while (!received_finalize) {
293 msg_task_t task = nullptr;
294 int ret = MSG_task_recv(&task, ms->mbox);
297 // An error occurred, clean the code and return
298 // The owner did not change, hence the task should be only destroyed on the other side
299 xbt_free(finalize_task_name);
303 if (strcmp(task->name, finalize_task_name) == 0)
304 received_finalize = 1;
306 MSG_task_destroy(task);
308 xbt_free(finalize_task_name);
310 // Here Stage 1, 2 and 3 have been performed.
311 // Hence complete the migration
313 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
314 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
315 // structure at the beginning and free it at the end of each function)
316 simgrid::s4u::VirtualMachine* vm = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm);
317 msg_host_t src_pm = ms->src_pm;
318 msg_host_t dst_pm = ms->dst_pm;
320 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
321 simgrid::simix::kernelImmediate([vm, src_pm, dst_pm]() {
322 /* Update the vm location */
323 /* precopy migration makes the VM temporally paused */
324 xbt_assert(static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
326 /* jump to vm_ws_xigrate(). this will update the vm location. */
327 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->migrate(dst_pm);
334 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
335 vm->pimpl_vm_->isMigrating = false;
336 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->cname(), ms->src_pm->cname(), ms->dst_pm->cname());
338 if (TRACE_msg_vm_is_enabled()) {
339 static long long int counter = 0;
340 char key[INSTR_DEFAULT_STR_SIZE];
341 snprintf(key, INSTR_DEFAULT_STR_SIZE, "%lld", counter++);
344 container_t msg = PJ_container_get(vm->cname());
345 type_t type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
346 new_pajeStartLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
348 // destroy existing container of this vm
349 container_t existing_container = PJ_container_get(vm->cname());
350 PJ_container_remove_from_parent(existing_container);
351 PJ_container_free(existing_container);
353 // create new container on the new_host location
354 PJ_container_new(vm->cname(), INSTR_MSG_VM, PJ_container_get(ms->dst_pm->cname()));
357 msg = PJ_container_get(vm->name().c_str());
358 type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
359 new_pajeEndLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
362 // Inform the SRC that the migration has been correctly performed
364 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
365 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
366 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
367 // xbt_assert(ret == MSG_OK);
368 if(ret == MSG_HOST_FAILURE){
369 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
370 // has been correctly migrated on the DST node
371 // TODO What does it mean ? What should we do ?
372 MSG_task_destroy(task);
373 } else if(ret == MSG_TRANSFER_FAILURE){
374 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
375 MSG_task_destroy(task);
380 XBT_DEBUG("mig: rx_done");
384 static void start_dirty_page_tracking(msg_vm_t vm)
386 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
388 pimpl->dp_enabled = 1;
393 xbt_dict_cursor_t cursor = nullptr;
394 dirty_page_t dp = nullptr;
395 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
396 double remaining = MSG_task_get_flops_amount(dp->task);
397 dp->prev_clock = MSG_get_clock();
398 dp->prev_remaining = remaining;
400 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
404 static void stop_dirty_page_tracking(msg_vm_t vm)
406 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
409 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
411 double computed = dp->prev_remaining - remaining;
412 double duration = clock - dp->prev_clock;
414 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", key, vm->cname(), computed,
415 dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
420 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
422 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
426 xbt_dict_cursor_t cursor = nullptr;
427 dirty_page_t dp = nullptr;
428 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
429 double remaining = MSG_task_get_flops_amount(dp->task);
431 double clock = MSG_get_clock();
433 // total += calc_updated_pages(key, vm, dp, remaining, clock);
434 total += get_computed(key, vm, dp, remaining, clock);
436 dp->prev_remaining = remaining;
437 dp->prev_clock = clock;
440 total += pimpl->dp_updated_by_deleted_tasks;
442 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
443 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
445 pimpl->dp_updated_by_deleted_tasks = 0;
450 // TODO Is this code redundant with the information provided by
451 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
452 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
453 void MSG_host_add_task(msg_host_t host, msg_task_t task)
455 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
458 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
460 double remaining = MSG_task_get_flops_amount(task);
461 char *key = bprintf("%s-%p", task->name, task);
463 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
465 if (pimpl->dp_enabled) {
466 dp->prev_clock = MSG_get_clock();
467 dp->prev_remaining = remaining;
470 pimpl->dp_objs = xbt_dict_new();
471 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
472 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
473 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, host->cname(), remaining, pimpl->dp_enabled);
478 void MSG_host_del_task(msg_host_t host, msg_task_t task)
480 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
483 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
485 char *key = bprintf("%s-%p", task->name, task);
486 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
487 xbt_assert(dp->task == task);
489 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
490 * the information for the lookup_() function that will called soon. */
491 if (pimpl->dp_enabled) {
492 double remaining = MSG_task_get_flops_amount(task);
493 double clock = MSG_get_clock();
494 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
495 double updated = get_computed(key, host, dp, remaining, clock);
497 pimpl->dp_updated_by_deleted_tasks += updated;
500 xbt_dict_remove(pimpl->dp_objs, key);
503 XBT_DEBUG("del %s on %s", key, host->cname());
507 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,
508 int stage, int stage2_round, double mig_speed, double timeout)
511 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
512 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
516 double clock_sta = MSG_get_clock();
520 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
522 ret = MSG_task_send(task, mbox);
528 } else if (ret == MSG_TIMEOUT) {
529 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
530 sent = size - remaining;
531 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
534 /* FIXME: why try-and-catch is used here? */
535 if(ret == MSG_HOST_FAILURE){
536 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
537 MSG_task_destroy(task);
538 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", vm->cname(), stage);
539 }else if(ret == MSG_TRANSFER_FAILURE){
540 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
541 MSG_task_destroy(task);
542 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", vm->cname(), stage);
545 double clock_end = MSG_get_clock();
546 double duration = clock_end - clock_sta;
547 double actual_speed = size / duration;
550 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
551 actual_speed, mig_speed);
553 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
559 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
561 double updated_size = computed * dp_rate;
562 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
563 if (updated_size > dp_cap) {
564 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
566 updated_size = dp_cap;
569 return (sg_size_t) updated_size;
572 static int migration_tx_fun(int argc, char *argv[])
574 XBT_DEBUG("mig: tx_start");
576 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
577 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
579 s_vm_params_t params;
580 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
581 const sg_size_t ramsize = params.ramsize;
582 const sg_size_t devsize = params.devsize;
583 const int skip_stage1 = params.skip_stage1;
584 int skip_stage2 = params.skip_stage2;
585 const double dp_rate = params.dp_rate;
586 const double dp_cap = params.dp_cap;
587 const double mig_speed = params.mig_speed;
588 double max_downtime = params.max_downtime;
590 double mig_timeout = 10000000.0;
592 double remaining_size = (double) (ramsize + devsize);
593 double threshold = 0.0;
595 /* check parameters */
597 XBT_WARN("migrate a VM, but ramsize is zero");
599 if (max_downtime <= 0) {
600 XBT_WARN("use the default max_downtime value 30ms");
604 /* Stage1: send all memory pages to the destination. */
605 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
606 start_dirty_page_tracking(ms->vm);
608 double computed_during_stage1 = 0;
610 double clock_prev_send = MSG_get_clock();
613 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
614 * transfer exceeds the timeout value. */
615 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
616 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
617 remaining_size -= sent;
618 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
620 if (sent < ramsize) {
621 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
623 } else if (sent > ramsize)
628 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
629 // Stop the dirty page tracking an return (there is no memory space to release)
630 stop_dirty_page_tracking(ms->vm);
634 double clock_post_send = MSG_get_clock();
635 mig_timeout -= (clock_post_send - clock_prev_send);
636 if (mig_timeout < 0) {
637 XBT_VERB("The duration of stage 1 exceeds the timeout value, skip stage 2");
641 /* estimate bandwidth */
642 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
643 threshold = bandwidth * max_downtime;
644 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
648 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
651 int stage2_round = 0;
654 sg_size_t updated_size = 0;
655 if (stage2_round == 0) {
656 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
657 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
659 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
660 updated_size = get_updated_size(computed, dp_rate, dp_cap);
663 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
664 updated_size, computed_during_stage1, dp_rate, dp_cap);
666 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
667 remaining_size += updated_size;
668 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
669 (remaining_size < threshold) ? "<" : ">", threshold);
670 if (remaining_size < threshold)
674 double clock_prev_send = MSG_get_clock();
676 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
677 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
679 } catch (xbt_ex& e) {
680 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
682 // Stop the dirty page tracking an return (there is no memory space to release)
683 stop_dirty_page_tracking(ms->vm);
686 double clock_post_send = MSG_get_clock();
688 if (sent == updated_size) {
689 /* timeout did not happen */
690 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
691 threshold = bandwidth * max_downtime;
692 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
693 remaining_size -= sent;
695 mig_timeout -= (clock_post_send - clock_prev_send);
696 xbt_assert(mig_timeout > 0);
698 } else if (sent < updated_size) {
699 /* When timeout happens, we move to stage 3. The size of memory pages
700 * updated before timeout must be added to the remaining size. */
701 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
702 updated_size, (clock_post_send - clock_prev_send));
703 remaining_size -= sent;
705 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
706 updated_size = get_updated_size(computed, dp_rate, dp_cap);
707 remaining_size += updated_size;
714 /* Stage3: stop the VM and copy the rest of states. */
715 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
716 simcall_vm_suspend(ms->vm);
717 stop_dirty_page_tracking(ms->vm);
720 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
721 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
724 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
725 // Stop the dirty page tracking an return (there is no memory space to release)
726 simcall_vm_resume(ms->vm);
730 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
731 // effectively the VM on the DST node.
732 XBT_DEBUG("mig: tx_done");
737 /** @brief Migrate the VM to the given host.
740 void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm)
743 * - One approach is ...
744 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
745 * receive the state of the source
746 * VM over network. We will finally destroy the source VM.
747 * - This behavior is similar to the way of migration in the real world.
748 * Even before a migration is completed, we will see a destination VM, consuming resources.
749 * - We have to relocate all processes. The existing process migration code will work for this?
750 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
752 * - Another one is ...
753 * We update the information of the given VM to place it to the destination physical host.
755 * The second one would be easier.
758 simgrid::s4u::VirtualMachine* typedVm = static_cast<simgrid::s4u::VirtualMachine*>(vm);
759 simgrid::vm::VirtualMachineImpl* pimpl = typedVm->pimpl_vm_;
760 msg_host_t src_pm = pimpl->getPm();
763 THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->cname(), src_pm->cname());
765 THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->cname(), dst_pm->cname());
766 if (!MSG_vm_is_running(vm))
767 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->cname());
768 if (typedVm->isMigrating())
769 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->cname());
771 pimpl->isMigrating = true;
773 struct migration_session *ms = xbt_new(struct migration_session, 1);
778 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
779 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
780 ms->mbox_ctl = bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
781 ms->mbox = bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
783 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
784 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
786 char** argv = xbt_new(char*, 2);
787 argv[0] = pr_rx_name;
789 MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
791 argv = xbt_new(char*, 2);
792 argv[0] = pr_tx_name;
794 MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
796 /* wait until the migration have finished or on error has occurred */
797 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
798 msg_task_t task = nullptr;
799 msg_error_t ret = MSG_TIMEOUT;
800 while (ret == MSG_TIMEOUT && dst_pm->isOn()) // The rx will tell me when he gots the VM
801 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
803 pimpl->isMigrating = false;
805 xbt_free(ms->mbox_ctl);
809 if (ret == MSG_HOST_FAILURE) {
810 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
811 // other side. Hence, just throw the execption
812 XBT_ERROR("SRC crashes, throw an exception (m-control)");
813 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
815 THROWF(host_error, 0, "Source host '%s' failed during the migration of VM '%s'.", src_pm->cname(), vm->cname());
816 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
817 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
818 XBT_ERROR("DST crashes, throw an exception (m-control)");
819 THROWF(host_error, 0, "Destination host '%s' failed during the migration of VM '%s'.", dst_pm->cname(),
823 char* expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
824 xbt_assert(strcmp(task->name, expected_task_name) == 0);
825 xbt_free(expected_task_name);
826 MSG_task_destroy(task);
829 /** @brief Immediately suspend the execution of all processes within the given VM.
832 * This function stops the execution of the VM. All the processes on this VM
833 * will pause. The state of the VM is preserved. We can later resume it again.
835 * No suspension cost occurs.
837 void MSG_vm_suspend(msg_vm_t vm)
839 if (MSG_vm_is_migrating(vm))
840 THROWF(vm_error, 0, "Cannot suspend VM '%s', which is migrating", vm->cname());
842 simcall_vm_suspend(vm);
844 XBT_DEBUG("vm_suspend done");
846 if (TRACE_msg_vm_is_enabled()) {
847 container_t vm_container = PJ_container_get(vm->cname());
848 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
849 val_t value = PJ_value_get_or_new("suspend", "1 0 0", type); // suspend is red
850 new_pajePushState(MSG_get_clock(), vm_container, type, value);
854 /** @brief Resume the execution of the VM. All processes on the VM run again.
857 * No resume cost occurs.
859 void MSG_vm_resume(msg_vm_t vm)
861 simcall_vm_resume(vm);
863 if (TRACE_msg_vm_is_enabled()) {
864 container_t vm_container = PJ_container_get(vm->cname());
865 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
866 new_pajePopState(MSG_get_clock(), vm_container, type);
871 /** @brief Immediately save the execution of all processes within the given VM.
874 * This function stops the execution of the VM. All the processes on this VM
875 * will pause. The state of the VM is preserved. We can later resume it again.
877 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
878 * or after, depending on the exact semantic of VM save to you.
880 void MSG_vm_save(msg_vm_t vm)
882 if (MSG_vm_is_migrating(vm))
883 THROWF(vm_error, 0, "Cannot save VM '%s', which is migrating.", vm->cname());
887 if (TRACE_msg_vm_is_enabled()) {
888 container_t vm_container = PJ_container_get(vm->cname());
889 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
890 val_t value = PJ_value_get_or_new("save", "0 1 0", type); // save is green
891 new_pajePushState(MSG_get_clock(), vm_container, type, value);
895 /** @brief Restore the execution of the VM. All processes on the VM run again.
898 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
899 * after, depending on the exact semantic of VM restore to you.
901 void MSG_vm_restore(msg_vm_t vm)
903 simgrid::simix::kernelImmediate([vm]() {
904 if (static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getState() != SURF_VM_STATE_SAVED)
905 THROWF(vm_error, 0, "VM(%s) was not saved", vm->name().c_str());
907 XBT_DEBUG("restore VM(%s), where %d processes exist", vm->name().c_str(),
908 xbt_swag_size(sg_host_simix(vm)->process_list));
910 /* jump to vm_ws_restore() */
911 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->restore();
913 smx_actor_t smx_process, smx_process_safe;
914 xbt_swag_foreach_safe(smx_process, smx_process_safe, sg_host_simix(vm)->process_list)
916 XBT_DEBUG("resume %s", smx_process->name.c_str());
917 SIMIX_process_resume(smx_process);
921 if (TRACE_msg_vm_is_enabled()) {
922 container_t vm_container = PJ_container_get(vm->cname());
923 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
924 new_pajePopState(MSG_get_clock(), vm_container, type);
928 /** @brief Get the physical host of a given VM.
931 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
933 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getPm();
936 /** @brief Set a CPU bound for a given VM.
939 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
942 * On PM0, there are Task1 and VM0.
943 * On VM0, there is Task2.
944 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
946 * Task1\@PM0 gets 50%.
947 * Task2\@VM0 gets 25%.
948 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
950 * 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
951 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
953 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
954 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
955 * on the VM in share_resource() or we have to trap both the start and end of task execution.
957 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
959 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
960 * computation task (or a VM) never exceeds the capacity of a CPU core.
962 void MSG_vm_set_bound(msg_vm_t vm, double bound)
964 simgrid::simix::kernelImmediate(
965 [vm, bound]() { static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->setBound(bound); });