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 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getState() == state;
64 /** @brief Returns whether the given VM has just created, not running.
67 int MSG_vm_is_created(msg_vm_t vm)
69 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
72 /** @brief Returns whether the given VM is currently running
75 int MSG_vm_is_running(msg_vm_t vm)
77 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
80 /** @brief Returns whether the given VM is currently migrating
83 int MSG_vm_is_migrating(msg_vm_t vm)
85 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->isMigrating();
88 /** @brief Returns whether the given VM is currently suspended, not running.
91 int MSG_vm_is_suspended(msg_vm_t vm)
93 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
96 /* **** ******** MSG vm actions ********* **** */
97 /** @brief Create a new VM with specified parameters.
99 * @param pm Physical machine that will host the VM
100 * @param name Must be unique
101 * @param ramsize [TODO]
102 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
103 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
105 msg_vm_t MSG_vm_create(msg_host_t pm, const char* name, int ramsize, int mig_netspeed, int dp_intensity)
107 simgrid::vm::VmHostExt::ensureVmExtInstalled();
109 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
110 double host_speed = MSG_host_get_speed(pm);
111 double update_speed = (static_cast<double>(dp_intensity)/100) * mig_netspeed;
113 msg_vm_t vm = MSG_vm_create_core(pm, name);
114 s_vm_params_t params;
115 memset(¶ms, 0, sizeof(params));
116 params.ramsize = static_cast<sg_size_t>(ramsize) * 1024 * 1024;
118 params.skip_stage2 = 0;
119 params.max_downtime = 0.03;
120 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
121 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
122 params.mig_speed = static_cast<double>(mig_netspeed) * 1024 * 1024; // mig_speed
124 XBT_DEBUG("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f", params.dp_rate,
125 params.mig_speed, dp_intensity, update_speed, host_speed);
126 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(¶ms);
131 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
134 * A VM is treated as a host. The name of the VM must be unique among all hosts.
136 msg_vm_t MSG_vm_create_core(msg_host_t pm, const char* name)
138 xbt_assert(sg_host_by_name(name) == nullptr,
139 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
141 return new simgrid::s4u::VirtualMachine(name, pm);
144 /** @brief Destroy a VM. Destroy the VM object from the simulation.
147 void MSG_vm_destroy(msg_vm_t vm)
149 if (MSG_vm_is_migrating(vm))
150 THROWF(vm_error, 0, "Cannot destroy VM '%s', which is migrating.", vm->cname());
152 /* First, terminate all processes on the VM if necessary */
153 if (MSG_vm_is_running(vm))
156 xbt_assert(MSG_vm_is_created(vm) || __MSG_vm_is_state(vm, SURF_VM_STATE_DESTROYED),
157 "shutdown the given VM before destroying it");
159 /* Then, destroy the VM object */
160 simgrid::simix::kernelImmediate([vm]() {
164 if (TRACE_msg_vm_is_enabled()) {
165 container_t container = PJ_container_get(vm->cname());
166 PJ_container_remove_from_parent(container);
167 PJ_container_free(container);
171 /** @brief Start a vm (i.e., boot the guest operating system)
174 * If the VM cannot be started (because of memory overprovisionning), an exception is generated.
176 void MSG_vm_start(msg_vm_t vm)
178 simgrid::simix::kernelImmediate([vm]() {
179 simgrid::vm::VmHostExt::ensureVmExtInstalled();
181 simgrid::s4u::VirtualMachine* typedVM = static_cast<simgrid::s4u::VirtualMachine*>(vm);
182 simgrid::s4u::Host* pm = typedVM->pimpl_vm_->getPm();
183 if (pm->extension<simgrid::vm::VmHostExt>() == nullptr)
184 pm->extension_set(new simgrid::vm::VmHostExt());
186 long pm_ramsize = pm->extension<simgrid::vm::VmHostExt>()->ramsize;
187 int pm_overcommit = pm->extension<simgrid::vm::VmHostExt>()->overcommit;
188 long vm_ramsize = typedVM->getRamsize();
190 if (pm_ramsize && !pm_overcommit) { /* Only verify that we don't overcommit on need */
191 /* Retrieve the memory occupied by the VMs on that host. Yep, we have to traverse all VMs of all hosts for that */
192 long total_ramsize_of_vms = 0;
193 for (simgrid::s4u::VirtualMachine* ws_vm : simgrid::vm::VirtualMachineImpl::allVms_)
194 if (pm == ws_vm->pimpl_vm_->getPm())
195 total_ramsize_of_vms += ws_vm->pimpl_vm_->getRamsize();
197 if (vm_ramsize > pm_ramsize - total_ramsize_of_vms) {
198 XBT_WARN("cannnot start %s@%s due to memory shortage: vm_ramsize %ld, free %ld, pm_ramsize %ld (bytes).",
199 vm->cname(), pm->cname(), vm_ramsize, pm_ramsize - total_ramsize_of_vms, pm_ramsize);
200 THROWF(vm_error, 0, "Memory shortage on host '%s', VM '%s' cannot be started", pm->cname(), vm->cname());
204 typedVM->pimpl_vm_->setState(SURF_VM_STATE_RUNNING);
207 if (TRACE_msg_vm_is_enabled()) {
208 container_t vm_container = PJ_container_get(vm->cname());
209 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
210 val_t value = PJ_value_get_or_new("start", "0 0 1", type); // start is blue
211 new_pajePushState(MSG_get_clock(), vm_container, type, value);
215 /** @brief Immediately kills all processes within the given VM.
218 * Any memory that they allocated will be leaked, unless you used #MSG_process_on_exit().
220 * No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep().
222 void MSG_vm_shutdown(msg_vm_t vm)
224 smx_actor_t issuer=SIMIX_process_self();
225 simgrid::simix::kernelImmediate([vm,issuer]() {
226 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->shutdown(issuer);
229 // Make sure that the processes in the VM are killed in this scheduling round before processing
230 // (eg with the VM destroy)
231 MSG_process_sleep(0.);
234 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
236 return bprintf("__pr_mig_tx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
239 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
241 return bprintf("__pr_mig_rx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
244 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
246 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm->cname(), src_pm->cname(), dst_pm->cname());
249 struct migration_session {
254 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
255 * know the completion of the migration. */
257 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
261 static int migration_rx_fun(int argc, char *argv[])
263 XBT_DEBUG("mig: rx_start");
265 // The structure has been created in the do_migration function and should only be freed in the same place ;)
266 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
268 bool received_finalize = false;
270 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
271 while (!received_finalize) {
272 msg_task_t task = nullptr;
273 int ret = MSG_task_recv(&task, ms->mbox);
276 // An error occurred, clean the code and return
277 // The owner did not change, hence the task should be only destroyed on the other side
278 xbt_free(finalize_task_name);
282 if (strcmp(task->name, finalize_task_name) == 0)
283 received_finalize = 1;
285 MSG_task_destroy(task);
287 xbt_free(finalize_task_name);
289 // Here Stage 1, 2 and 3 have been performed.
290 // Hence complete the migration
292 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
293 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
294 // structure at the beginning and free it at the end of each function)
295 simgrid::s4u::VirtualMachine* vm = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm);
296 msg_host_t dst_pm = ms->dst_pm;
298 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
299 simgrid::simix::kernelImmediate([vm, dst_pm]() {
300 /* Update the vm location */
301 /* precopy migration makes the VM temporally paused */
302 xbt_assert(vm->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
304 /* Update the vm location and resume it */
305 vm->pimpl_vm_->setPm(dst_pm);
306 vm->pimpl_vm_->resume();
310 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
311 vm->pimpl_vm_->isMigrating = false;
312 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->cname(), ms->src_pm->cname(), ms->dst_pm->cname());
314 if (TRACE_msg_vm_is_enabled()) {
315 static long long int counter = 0;
316 char key[INSTR_DEFAULT_STR_SIZE];
317 snprintf(key, INSTR_DEFAULT_STR_SIZE, "%lld", counter);
321 container_t msg = PJ_container_get(vm->cname());
322 type_t type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
323 new_pajeStartLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
325 // destroy existing container of this vm
326 container_t existing_container = PJ_container_get(vm->cname());
327 PJ_container_remove_from_parent(existing_container);
328 PJ_container_free(existing_container);
330 // create new container on the new_host location
331 PJ_container_new(vm->cname(), INSTR_MSG_VM, PJ_container_get(ms->dst_pm->cname()));
334 msg = PJ_container_get(vm->cname());
335 type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
336 new_pajeEndLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
339 // Inform the SRC that the migration has been correctly performed
340 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
341 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
342 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
343 // xbt_assert(ret == MSG_OK);
344 if(ret == MSG_HOST_FAILURE){
345 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
346 // has been correctly migrated on the DST node
347 // TODO What does it mean ? What should we do ?
348 MSG_task_destroy(task);
349 } else if(ret == MSG_TRANSFER_FAILURE){
350 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
351 MSG_task_destroy(task);
355 XBT_DEBUG("mig: rx_done");
359 static void start_dirty_page_tracking(msg_vm_t vm)
361 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
363 pimpl->dp_enabled = 1;
368 xbt_dict_cursor_t cursor = nullptr;
369 dirty_page_t dp = nullptr;
370 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
371 double remaining = MSG_task_get_flops_amount(dp->task);
372 dp->prev_clock = MSG_get_clock();
373 dp->prev_remaining = remaining;
375 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
379 static void stop_dirty_page_tracking(msg_vm_t vm)
381 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
384 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
386 double computed = dp->prev_remaining - remaining;
387 double duration = clock - dp->prev_clock;
389 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", key, vm->cname(), computed,
390 dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
395 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
397 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
401 xbt_dict_cursor_t cursor = nullptr;
402 dirty_page_t dp = nullptr;
403 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
404 double remaining = MSG_task_get_flops_amount(dp->task);
406 double clock = MSG_get_clock();
408 // total += calc_updated_pages(key, vm, dp, remaining, clock);
409 total += get_computed(key, vm, dp, remaining, clock);
411 dp->prev_remaining = remaining;
412 dp->prev_clock = clock;
415 total += pimpl->dp_updated_by_deleted_tasks;
417 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
418 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
420 pimpl->dp_updated_by_deleted_tasks = 0;
425 // TODO Is this code redundant with the information provided by
426 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
427 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
428 void MSG_host_add_task(msg_host_t host, msg_task_t task)
430 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
433 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
435 double remaining = MSG_task_get_flops_amount(task);
436 char *key = bprintf("%s-%p", task->name, task);
438 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
440 if (pimpl->dp_enabled) {
441 dp->prev_clock = MSG_get_clock();
442 dp->prev_remaining = remaining;
445 pimpl->dp_objs = xbt_dict_new_homogeneous(nullptr);
446 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
447 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
448 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, host->cname(), remaining, pimpl->dp_enabled);
453 void MSG_host_del_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 char *key = bprintf("%s-%p", task->name, task);
461 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
462 xbt_assert(dp->task == task);
464 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
465 * the information for the lookup_() function that will called soon. */
466 if (pimpl->dp_enabled) {
467 double remaining = MSG_task_get_flops_amount(task);
468 double clock = MSG_get_clock();
469 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
470 double updated = get_computed(key, host, dp, remaining, clock);
472 pimpl->dp_updated_by_deleted_tasks += updated;
475 xbt_dict_remove(pimpl->dp_objs, key);
478 XBT_DEBUG("del %s on %s", key, host->cname());
482 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,
483 int stage, int stage2_round, double mig_speed, double timeout)
486 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
487 msg_task_t task = MSG_task_create(task_name, 0, static_cast<double>(size), nullptr);
491 double clock_sta = MSG_get_clock();
495 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
497 ret = MSG_task_send(task, mbox);
503 } else if (ret == MSG_TIMEOUT) {
504 sg_size_t remaining = static_cast<sg_size_t>(MSG_task_get_remaining_communication(task));
505 sent = size - remaining;
506 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
509 /* FIXME: why try-and-catch is used here? */
510 if(ret == MSG_HOST_FAILURE){
511 XBT_DEBUG("SRC host failed during migration of %s (stage %d)", vm->cname(), stage);
512 MSG_task_destroy(task);
513 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", vm->cname(), stage);
514 }else if(ret == MSG_TRANSFER_FAILURE){
515 XBT_DEBUG("DST host failed during migration of %s (stage %d)", vm->cname(), stage);
516 MSG_task_destroy(task);
517 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", vm->cname(), stage);
520 double clock_end = MSG_get_clock();
521 double duration = clock_end - clock_sta;
522 double actual_speed = size / duration;
525 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
526 actual_speed, mig_speed);
528 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
534 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
536 double updated_size = computed * dp_rate;
537 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
538 if (updated_size > dp_cap) {
539 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
541 updated_size = dp_cap;
544 return static_cast<sg_size_t>(updated_size);
547 static int migration_tx_fun(int argc, char *argv[])
549 XBT_DEBUG("mig: tx_start");
551 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
552 migration_session *ms = static_cast<migration_session *>(MSG_process_get_data(MSG_process_self()));
554 s_vm_params_t params;
555 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
556 const sg_size_t ramsize = params.ramsize;
557 const sg_size_t devsize = params.devsize;
558 const int skip_stage1 = params.skip_stage1;
559 int skip_stage2 = params.skip_stage2;
560 const double dp_rate = params.dp_rate;
561 const double dp_cap = params.dp_cap;
562 const double mig_speed = params.mig_speed;
563 double max_downtime = params.max_downtime;
565 double mig_timeout = 10000000.0;
567 double remaining_size = static_cast<double>(ramsize + devsize);
568 double threshold = 0.0;
570 /* check parameters */
572 XBT_WARN("migrate a VM, but ramsize is zero");
574 if (max_downtime <= 0) {
575 XBT_WARN("use the default max_downtime value 30ms");
579 /* Stage1: send all memory pages to the destination. */
580 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
581 start_dirty_page_tracking(ms->vm);
583 double computed_during_stage1 = 0;
585 double clock_prev_send = MSG_get_clock();
588 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
589 * transfer exceeds the timeout value. */
590 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
591 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
592 remaining_size -= sent;
593 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
595 if (sent < ramsize) {
596 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
598 } else if (sent > ramsize)
603 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
604 // Stop the dirty page tracking an return (there is no memory space to release)
605 stop_dirty_page_tracking(ms->vm);
609 double clock_post_send = MSG_get_clock();
610 mig_timeout -= (clock_post_send - clock_prev_send);
611 if (mig_timeout < 0) {
612 XBT_VERB("The duration of stage 1 exceeds the timeout value, skip stage 2");
616 /* estimate bandwidth */
617 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
618 threshold = bandwidth * max_downtime;
619 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
623 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
626 int stage2_round = 0;
629 sg_size_t updated_size = 0;
630 if (stage2_round == 0) {
631 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
632 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
634 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
635 updated_size = get_updated_size(computed, dp_rate, dp_cap);
638 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
639 updated_size, computed_during_stage1, dp_rate, dp_cap);
641 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
642 remaining_size += updated_size;
643 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
644 (remaining_size < threshold) ? "<" : ">", threshold);
645 if (remaining_size < threshold)
649 double clock_prev_send = MSG_get_clock();
651 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
652 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
654 } catch (xbt_ex& e) {
655 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
657 // Stop the dirty page tracking an return (there is no memory space to release)
658 stop_dirty_page_tracking(ms->vm);
661 double clock_post_send = MSG_get_clock();
663 if (sent == updated_size) {
664 /* timeout did not happen */
665 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
666 threshold = bandwidth * max_downtime;
667 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
668 remaining_size -= sent;
670 mig_timeout -= (clock_post_send - clock_prev_send);
671 xbt_assert(mig_timeout > 0);
673 } else if (sent < updated_size) {
674 /* When timeout happens, we move to stage 3. The size of memory pages
675 * updated before timeout must be added to the remaining size. */
676 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
677 updated_size, (clock_post_send - clock_prev_send));
678 remaining_size -= sent;
680 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
681 updated_size = get_updated_size(computed, dp_rate, dp_cap);
682 remaining_size += updated_size;
689 /* Stage3: stop the VM and copy the rest of states. */
690 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
691 simgrid::vm::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->pimpl_vm_;
692 pimpl->setState(SURF_VM_STATE_RUNNING); // FIXME: this bypass of the checks in suspend() is not nice
693 pimpl->isMigrating = false; // FIXME: this bypass of the checks in suspend() is not nice
694 pimpl->suspend(SIMIX_process_self());
695 stop_dirty_page_tracking(ms->vm);
698 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
699 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, static_cast<sg_size_t>(remaining_size), ms->mbox, 3, 0,
703 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
704 // Stop the dirty page tracking an return (there is no memory space to release)
705 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->pimpl_vm_->resume();
709 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
710 // effectively the VM on the DST node.
711 XBT_DEBUG("mig: tx_done");
716 /** @brief Migrate the VM to the given host.
719 void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm)
722 * - One approach is ...
723 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
724 * receive the state of the source
725 * VM over network. We will finally destroy the source VM.
726 * - This behavior is similar to the way of migration in the real world.
727 * Even before a migration is completed, we will see a destination VM, consuming resources.
728 * - We have to relocate all processes. The existing process migration code will work for this?
729 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
731 * - Another one is ...
732 * We update the information of the given VM to place it to the destination physical host.
734 * The second one would be easier.
737 simgrid::s4u::VirtualMachine* typedVm = static_cast<simgrid::s4u::VirtualMachine*>(vm);
738 simgrid::vm::VirtualMachineImpl* pimpl = typedVm->pimpl_vm_;
739 msg_host_t src_pm = pimpl->getPm();
742 THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->cname(), src_pm->cname());
744 THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->cname(), dst_pm->cname());
745 if (!MSG_vm_is_running(vm))
746 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->cname());
747 if (typedVm->isMigrating())
748 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->cname());
750 pimpl->isMigrating = true;
752 struct migration_session *ms = xbt_new(struct migration_session, 1);
757 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
758 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
759 ms->mbox_ctl = bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
760 ms->mbox = bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
762 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
763 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
765 char** argv = xbt_new(char*, 2);
766 argv[0] = pr_rx_name;
768 MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
770 argv = xbt_new(char*, 2);
771 argv[0] = pr_tx_name;
773 MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
775 /* wait until the migration have finished or on error has occurred */
776 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
777 msg_task_t task = nullptr;
778 msg_error_t ret = MSG_task_receive(&task, ms->mbox_ctl);
780 pimpl->isMigrating = false;
782 xbt_free(ms->mbox_ctl);
786 if (ret == MSG_HOST_FAILURE) {
787 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
788 // other side. Hence, just throw the execption
789 XBT_ERROR("SRC crashes, throw an exception (m-control)");
790 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
792 THROWF(host_error, 0, "Source host '%s' failed during the migration of VM '%s'.", src_pm->cname(), vm->cname());
793 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
794 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
795 XBT_ERROR("DST crashes, throw an exception (m-control)");
796 THROWF(host_error, 0, "Destination host '%s' failed during the migration of VM '%s'.", dst_pm->cname(),
800 char* expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
801 xbt_assert(strcmp(task->name, expected_task_name) == 0);
802 xbt_free(expected_task_name);
803 MSG_task_destroy(task);
806 /** @brief Immediately suspend the execution of all processes within the given VM.
809 * This function stops the execution of the VM. All the processes on this VM
810 * will pause. The state of the VM is preserved. We can later resume it again.
812 * No suspension cost occurs.
814 void MSG_vm_suspend(msg_vm_t vm)
816 smx_actor_t issuer = SIMIX_process_self();
817 simgrid::simix::kernelImmediate([vm,issuer]() {
818 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->suspend(issuer);
821 XBT_DEBUG("vm_suspend done");
823 if (TRACE_msg_vm_is_enabled()) {
824 container_t vm_container = PJ_container_get(vm->cname());
825 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
826 val_t value = PJ_value_get_or_new("suspend", "1 0 0", type); // suspend is red
827 new_pajePushState(MSG_get_clock(), vm_container, type, value);
831 /** @brief Resume the execution of the VM. All processes on the VM run again.
834 * No resume cost occurs.
836 void MSG_vm_resume(msg_vm_t vm)
838 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->resume();
840 if (TRACE_msg_vm_is_enabled()) {
841 container_t vm_container = PJ_container_get(vm->cname());
842 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
843 new_pajePopState(MSG_get_clock(), vm_container, type);
847 /** @brief Get the physical host of a given VM.
850 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
852 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->pm();
855 /** @brief Set a CPU bound for a given VM.
858 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
861 * On PM0, there are Task1 and VM0.
862 * On VM0, there is Task2.
863 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
865 * Task1\@PM0 gets 50%.
866 * Task2\@VM0 gets 25%.
867 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
869 * 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
870 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
872 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
873 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
874 * on the VM in share_resource() or we have to trap both the start and end of task execution.
876 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
878 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
879 * computation task (or a VM) never exceeds the capacity of a CPU core.
881 void MSG_vm_set_bound(msg_vm_t vm, double bound)
883 simgrid::simix::kernelImmediate(
884 [vm, bound]() { static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->setBound(bound); });