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 <simgrid/s4u/VirtualMachine.hpp>
16 #include <simgrid/s4u/host.hpp>
18 #include "msg_private.h"
19 #include "xbt/sysdep.h"
21 #include "simgrid/host.h"
23 #include "src/simix/smx_host_private.h" /* don't ask me why the VM functions are in there (FIXME:KILLME) */
25 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
28 /* **** ******** GENERAL ********* **** */
30 /** \ingroup m_vm_management
31 * \brief Set the parameters of a given host
34 * \param params a parameter object
36 void MSG_vm_set_params(msg_vm_t vm, vm_params_t params)
38 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(params);
41 /** \ingroup m_vm_management
42 * \brief Get the parameters of a given host
45 * \param params a prameter object
47 void MSG_vm_get_params(msg_vm_t vm, vm_params_t params)
49 static_cast<simgrid::s4u::VirtualMachine*>(vm)->parameters(params);
52 /* **** Check state of a VM **** */
53 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
55 return simcall_vm_get_state(vm) == state;
58 /** @brief Returns whether the given VM has just created, not running.
61 int MSG_vm_is_created(msg_vm_t vm)
63 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
66 /** @brief Returns whether the given VM is currently running
69 int MSG_vm_is_running(msg_vm_t vm)
71 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
74 /** @brief Returns whether the given VM is currently migrating
77 int MSG_vm_is_migrating(msg_vm_t vm)
79 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->isMigrating();
82 /** @brief Returns whether the given VM is currently suspended, not running.
85 int MSG_vm_is_suspended(msg_vm_t vm)
87 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
90 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
93 int MSG_vm_is_saving(msg_vm_t vm)
95 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
98 /** @brief Returns whether the given VM has been saved, not running.
101 int MSG_vm_is_saved(msg_vm_t vm)
103 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
106 /** @brief Returns whether the given VM is being restored, not running.
109 int MSG_vm_is_restoring(msg_vm_t vm)
111 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
114 /* **** ******** MSG vm actions ********* **** */
115 /** @brief Create a new VM with specified parameters.
117 * @param pm Physical machine that will host the VM
119 * @param ncpus [TODO]
120 * @param ramsize [TODO]
121 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
122 * @param disk_path (unused) Path to the image that boots
123 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
124 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
125 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
127 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name, int ncpus, int ramsize, int net_cap, char *disk_path,
128 int disksize, int mig_netspeed, int dp_intensity)
130 /* For the moment, intensity_rate is the percentage against the migration
132 double host_speed = MSG_host_get_speed(pm);
133 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
135 msg_vm_t vm = MSG_vm_create_core(pm, name);
136 s_vm_params_t params;
137 memset(¶ms, 0, sizeof(params));
138 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
139 //params.overcommit = 0;
141 params.skip_stage2 = 0;
142 params.max_downtime = 0.03;
143 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
144 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
145 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
147 //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate,
148 // params.mig_speed, dp_intensity, update_speed, host_speed);
149 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(¶ms);
154 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
157 * A VM is treated as a host. The name of the VM must be unique among all hosts.
159 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
161 /* make sure the VM of the same name does not exit */
163 simgrid::s4u::Host* ind_host_tmp = sg_host_by_name(name);
164 if (ind_host_tmp != nullptr && sg_host_simix(ind_host_tmp) != nullptr) {
165 XBT_ERROR("host %s already exits", name);
170 /* Note: ind_vm and vm_workstation point to the same elm object. */
171 /* Ask the SIMIX layer to create the surf vm resource */
172 sg_host_t ind_vm_workstation = simcall_vm_create(name, ind_pm);
174 msg_vm_t ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
176 XBT_DEBUG("A new VM (%s) has been created", name);
178 TRACE_msg_vm_create(name, ind_pm);
183 /** @brief Destroy a VM. Destroy the VM object from the simulation.
186 void MSG_vm_destroy(msg_vm_t vm)
188 if (MSG_vm_is_migrating(vm))
189 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
191 /* First, terminate all processes on the VM if necessary */
192 if (MSG_vm_is_running(vm))
193 simcall_vm_shutdown(vm);
195 if (!MSG_vm_is_created(vm)) {
196 XBT_CRITICAL("shutdown the given VM before destroying it");
200 /* Then, destroy the VM object */
201 simcall_vm_destroy(vm);
203 TRACE_msg_vm_end(vm);
206 /** @brief Start a vm (i.e., boot the guest operating system)
209 * If the VM cannot be started, an exception is generated.
211 void MSG_vm_start(msg_vm_t vm)
213 simcall_vm_start(vm);
215 TRACE_msg_vm_start(vm);
218 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
221 * FIXME: No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep() or something.
222 * I'm not quite sure.
224 void MSG_vm_shutdown(msg_vm_t vm)
226 /* msg_vm_t equals to msg_host_t */
227 simcall_vm_shutdown(vm);
229 // TRACE_msg_vm_(vm);
232 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
233 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
234 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
236 const char *vm_name = sg_host_get_name(vm);
237 const char *src_pm_name = sg_host_get_name(src_pm);
238 const char *dst_pm_name = sg_host_get_name(dst_pm);
240 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
243 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
245 const char *vm_name = sg_host_get_name(vm);
246 const char *src_pm_name = sg_host_get_name(src_pm);
247 const char *dst_pm_name = sg_host_get_name(dst_pm);
249 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
252 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
254 const char *vm_name = sg_host_get_name(vm);
255 const char *src_pm_name = sg_host_get_name(src_pm);
256 const char *dst_pm_name = sg_host_get_name(dst_pm);
258 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
261 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
263 const char *vm_name = sg_host_get_name(vm);
264 const char *src_pm_name = sg_host_get_name(src_pm);
265 const char *dst_pm_name = sg_host_get_name(dst_pm);
267 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
270 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
272 const char *vm_name = sg_host_get_name(vm);
273 const char *src_pm_name = sg_host_get_name(src_pm);
274 const char *dst_pm_name = sg_host_get_name(dst_pm);
276 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
279 struct migration_session {
284 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
285 * know the completion of the migration. */
287 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
291 static int migration_rx_fun(int argc, char *argv[])
293 XBT_DEBUG("mig: rx_start");
295 // The structure has been created in the do_migration function and should only be freed in the same place ;)
296 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
298 s_vm_params_t params;
299 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
303 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
307 msg_task_t task = nullptr;
308 ret = MSG_task_recv(&task, ms->mbox);
311 // An error occurred, clean the code and return
312 // The owner did not change, hence the task should be only destroyed on the other side
313 xbt_free(finalize_task_name);
318 if (strcmp(task->name, finalize_task_name) == 0)
321 MSG_task_destroy(task);
327 // Here Stage 1, 2 and 3 have been performed.
328 // Hence complete the migration
330 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
331 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
332 // structure at the beginning and free it at the end of each function)
333 simgrid::s4u::VirtualMachine* vm = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm);
334 msg_host_t src_pm = ms->src_pm;
335 msg_host_t dst_pm = ms->dst_pm;
337 // TODO: we have an issue, if the DST node is turning off during the three next calls, then the VM is in an
339 // state. I should check with Takahiro in order to make this portion of code atomic
341 // /* Update the vm location */
342 // simcall_vm_migrate(vm, dst_pm);
344 // /* Resume the VM */
345 // simcall_vm_resume(vm);
347 simgrid::simix::kernelImmediate([vm, src_pm, dst_pm]() {
348 /* Update the vm location */
349 SIMIX_vm_migrate(vm, dst_pm);
356 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
357 vm->pimpl_vm_->isMigrating = false;
358 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
359 sg_host_get_name(ms->dst_pm));
360 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
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_free(finalize_task_name);
382 XBT_DEBUG("mig: rx_done");
386 static void reset_dirty_pages(msg_vm_t vm)
388 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
391 xbt_dict_cursor_t cursor = nullptr;
392 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 start_dirty_page_tracking(msg_vm_t vm)
406 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 1;
408 reset_dirty_pages(vm);
411 static void stop_dirty_page_tracking(msg_vm_t vm)
413 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
416 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
418 double computed = dp->prev_remaining - remaining;
419 double duration = clock - dp->prev_clock;
421 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
422 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
427 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
429 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
433 xbt_dict_cursor_t cursor = nullptr;
434 dirty_page_t dp = nullptr;
435 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
436 double remaining = MSG_task_get_flops_amount(dp->task);
438 double clock = MSG_get_clock();
440 // total += calc_updated_pages(key, vm, dp, remaining, clock);
441 total += get_computed(key, vm, dp, remaining, clock);
443 dp->prev_remaining = remaining;
444 dp->prev_clock = clock;
447 total += pimpl->dp_updated_by_deleted_tasks;
449 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
450 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
452 pimpl->dp_updated_by_deleted_tasks = 0;
457 // TODO Is this code redundant with the information provided by
458 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
459 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
460 void MSG_host_add_task(msg_host_t host, msg_task_t task)
462 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
465 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
467 double remaining = MSG_task_get_flops_amount(task);
468 char *key = bprintf("%s-%p", task->name, task);
470 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
472 if (pimpl->dp_enabled) {
473 dp->prev_clock = MSG_get_clock();
474 dp->prev_remaining = remaining;
477 pimpl->dp_objs = xbt_dict_new();
478 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
479 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
480 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, pimpl->dp_enabled);
485 void MSG_host_del_task(msg_host_t host, msg_task_t task)
487 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
490 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
492 char *key = bprintf("%s-%p", task->name, task);
493 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
494 xbt_assert(dp->task == task);
496 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
497 * the information for the lookup_() function that will called soon. */
498 if (pimpl->dp_enabled) {
499 double remaining = MSG_task_get_flops_amount(task);
500 double clock = MSG_get_clock();
501 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
502 double updated = get_computed(key, host, dp, remaining, clock);
504 pimpl->dp_updated_by_deleted_tasks += updated;
507 xbt_dict_remove(pimpl->dp_objs, key);
510 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
514 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,
515 int stage, int stage2_round, double mig_speed, double timeout)
518 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
519 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
523 double clock_sta = MSG_get_clock();
527 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
529 ret = MSG_task_send(task, mbox);
535 } else if (ret == MSG_TIMEOUT) {
536 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
537 sent = size - remaining;
538 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
541 /* FIXME: why try-and-catch is used here? */
542 if(ret == MSG_HOST_FAILURE){
543 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
544 MSG_task_destroy(task);
545 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
546 }else if(ret == MSG_TRANSFER_FAILURE){
547 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
548 MSG_task_destroy(task);
549 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
552 double clock_end = MSG_get_clock();
553 double duration = clock_end - clock_sta;
554 double actual_speed = size / duration;
557 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)",
558 stage, stage2_round, size, duration, actual_speed, mig_speed);
560 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)",
561 stage, size, duration, actual_speed, mig_speed);
566 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
568 double updated_size = computed * dp_rate;
569 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
570 if (updated_size > dp_cap) {
571 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
573 updated_size = dp_cap;
576 return (sg_size_t) updated_size;
579 static double send_stage1(struct migration_session *ms, sg_size_t ramsize, double mig_speed, double dp_rate,
582 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
583 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
584 sg_size_t remaining = ramsize;
585 double computed_total = 0;
587 while (remaining > 0) {
588 sg_size_t datasize = chunksize;
589 if (remaining < chunksize)
590 datasize = remaining;
592 remaining -= datasize;
593 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
594 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
595 computed_total += computed;
598 return computed_total;
601 static double get_threshold_value(double bandwidth, double max_downtime)
603 return max_downtime * bandwidth;
606 static int migration_tx_fun(int argc, char *argv[])
608 XBT_DEBUG("mig: tx_start");
610 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
611 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
613 s_vm_params_t params;
614 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
615 const sg_size_t ramsize = params.ramsize;
616 const sg_size_t devsize = params.devsize;
617 const int skip_stage1 = params.skip_stage1;
618 int skip_stage2 = params.skip_stage2;
619 const double dp_rate = params.dp_rate;
620 const double dp_cap = params.dp_cap;
621 const double mig_speed = params.mig_speed;
622 double max_downtime = params.max_downtime;
624 /* hard code it temporally. Fix Me */
625 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
626 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
628 double remaining_size = (double) (ramsize + devsize);
629 double threshold = 0.0;
631 /* check parameters */
633 XBT_WARN("migrate a VM, but ramsize is zero");
635 if (max_downtime <= 0) {
636 XBT_WARN("use the default max_downtime value 30ms");
640 /* Stage1: send all memory pages to the destination. */
641 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
642 start_dirty_page_tracking(ms->vm);
644 double computed_during_stage1 = 0;
646 double clock_prev_send = MSG_get_clock();
649 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
650 * transfer exceeds the timeout value. */
651 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
652 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
653 remaining_size -= sent;
654 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
656 if (sent < ramsize) {
657 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
659 } else if (sent > ramsize)
664 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
665 // Stop the dirty page tracking an return (there is no memory space to release)
666 stop_dirty_page_tracking(ms->vm);
670 double clock_post_send = MSG_get_clock();
671 mig_timeout -= (clock_post_send - clock_prev_send);
672 if (mig_timeout < 0) {
673 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
674 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
678 /* estimate bandwidth */
679 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
680 threshold = get_threshold_value(bandwidth, max_downtime);
681 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
685 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
688 int stage2_round = 0;
691 sg_size_t updated_size = 0;
692 if (stage2_round == 0) {
693 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
694 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
696 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
697 updated_size = get_updated_size(computed, dp_rate, dp_cap);
700 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
701 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
703 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
704 remaining_size += updated_size;
705 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
706 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
707 if (remaining_size < threshold)
711 double clock_prev_send = MSG_get_clock();
713 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
714 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
718 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
719 // Stop the dirty page tracking an return (there is no memory space to release)
720 stop_dirty_page_tracking(ms->vm);
723 double clock_post_send = MSG_get_clock();
725 if (sent == updated_size) {
726 /* timeout did not happen */
727 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
728 threshold = get_threshold_value(bandwidth, max_downtime);
729 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
730 remaining_size -= sent;
732 mig_timeout -= (clock_post_send - clock_prev_send);
733 xbt_assert(mig_timeout > 0);
735 } else if (sent < updated_size) {
736 /* When timeout happens, we move to stage 3. The size of memory pages
737 * updated before timeout must be added to the remaining size. */
738 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
739 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
740 remaining_size -= sent;
742 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
743 updated_size = get_updated_size(computed, dp_rate, dp_cap);
744 remaining_size += updated_size;
751 /* Stage3: stop the VM and copy the rest of states. */
752 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
753 simcall_vm_suspend(ms->vm);
754 stop_dirty_page_tracking(ms->vm);
757 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
758 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
761 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
762 // Stop the dirty page tracking an return (there is no memory space to release)
763 simcall_vm_resume(ms->vm);
767 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
768 // effectively the VM on the DST node.
769 XBT_DEBUG("mig: tx_done");
774 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
776 struct migration_session *ms = xbt_new(struct migration_session, 1);
780 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
781 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
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 // msg_process_t tx_process, rx_process;
787 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
788 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
791 char **argv = xbt_new(char *, 2);
792 argv[0] = pr_rx_name;
794 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
797 char **argv = xbt_new(char *, 2);
798 argv[0] = pr_tx_name;
800 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
804 /* wait until the migration have finished or on error has occurred */
806 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
807 msg_task_t task = nullptr;
808 msg_error_t ret = MSG_TIMEOUT;
809 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
810 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
812 xbt_free(ms->mbox_ctl);
816 //xbt_assert(ret == MSG_OK);
817 if(ret == MSG_HOST_FAILURE){
818 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
819 // other side. Hence, just throw the execption
820 XBT_ERROR("SRC crashes, throw an exception (m-control)");
821 //MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
825 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){
826 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
827 XBT_ERROR("DST crashes, throw an exception (m-control)");
831 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
832 xbt_assert(strcmp(task->name, expected_task_name) == 0);
833 xbt_free(expected_task_name);
834 MSG_task_destroy(task);
839 /** @brief Migrate the VM to the given host.
842 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
843 * depending on whether you want to do cold or hot migration.
845 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
848 * - One approach is ...
849 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
850 * receive the state of the source
851 * VM over network. We will finally destroy the source VM.
852 * - This behavior is similar to the way of migration in the real world.
853 * Even before a migration is completed, we will see a destination VM, consuming resources.
854 * - We have to relocate all processes. The existing process migration code will work for this?
855 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
857 * - Another one is ...
858 * We update the information of the given VM to place it to the destination physical host.
860 * The second one would be easier.
863 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
865 if(MSG_host_is_off(old_pm))
866 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
868 if(MSG_host_is_off(new_pm))
869 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
871 if (!MSG_vm_is_running(vm))
872 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
874 if (MSG_vm_is_migrating(vm))
875 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
877 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
878 pimpl->isMigrating = 1;
881 int ret = do_migration(vm, old_pm, new_pm);
883 pimpl->isMigrating = 0;
884 THROWF(host_error, 0, "SRC host failed during migration");
887 pimpl->isMigrating = 0;
888 THROWF(host_error, 0, "DST host failed during migration");
892 // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration
893 // process. In that case, the VM has been already assigned to the DST node.
894 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
895 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
898 /** @brief Immediately suspend the execution of all processes within the given VM.
901 * This function stops the execution of the VM. All the processes on this VM
902 * will pause. The state of the VM is preserved. We can later resume it again.
904 * No suspension cost occurs.
906 void MSG_vm_suspend(msg_vm_t vm)
908 if (MSG_vm_is_migrating(vm))
909 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
911 simcall_vm_suspend(vm);
913 XBT_DEBUG("vm_suspend done");
915 TRACE_msg_vm_suspend(vm);
918 /** @brief Resume the execution of the VM. All processes on the VM run again.
921 * No resume cost occurs.
923 void MSG_vm_resume(msg_vm_t vm)
925 simcall_vm_resume(vm);
927 TRACE_msg_vm_resume(vm);
931 /** @brief Immediately save the execution of all processes within the given VM.
934 * This function stops the execution of the VM. All the processes on this VM
935 * will pause. The state of the VM is preserved. We can later resume it again.
937 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
938 * or after, depending on the exact semantic of VM save to you.
940 void MSG_vm_save(msg_vm_t vm)
942 if (MSG_vm_is_migrating(vm))
943 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
946 TRACE_msg_vm_save(vm);
949 /** @brief Restore the execution of the VM. All processes on the VM run again.
952 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
953 * after, depending on the exact semantic of VM restore to you.
955 void MSG_vm_restore(msg_vm_t vm)
957 simcall_vm_restore(vm);
959 TRACE_msg_vm_restore(vm);
962 /** @brief Get the physical host of a given VM.
965 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
967 return (msg_host_t) simcall_vm_get_pm(vm);
970 /** @brief Set a CPU bound for a given VM.
973 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
976 * On PM0, there are Task1 and VM0.
977 * On VM0, there is Task2.
978 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
980 * Task1\@PM0 gets 50%.
981 * Task2\@VM0 gets 25%.
982 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
984 * 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
985 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
987 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
988 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
989 * on the VM in share_resource() or we have to trap both the start and end of task execution.
991 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
993 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
994 * computation task (or a VM) never exceeds the capacity of a CPU core.
996 void MSG_vm_set_bound(msg_vm_t vm, double bound)
998 simcall_vm_set_bound(vm, bound);