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 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
26 /* **** ******** GENERAL ********* **** */
28 /** \ingroup m_vm_management
29 * \brief Set the parameters of a given host
32 * \param params a parameter object
34 void MSG_vm_set_params(msg_vm_t vm, vm_params_t params)
36 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(params);
39 /** \ingroup m_vm_management
40 * \brief Get the parameters of a given host
43 * \param params a prameter object
45 void MSG_vm_get_params(msg_vm_t vm, vm_params_t params)
47 static_cast<simgrid::s4u::VirtualMachine*>(vm)->parameters(params);
50 /* **** Check state of a VM **** */
51 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
53 return simcall_vm_get_state(vm) == state;
56 /** @brief Returns whether the given VM has just created, not running.
59 int MSG_vm_is_created(msg_vm_t vm)
61 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
64 /** @brief Returns whether the given VM is currently running
67 int MSG_vm_is_running(msg_vm_t vm)
69 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
72 /** @brief Returns whether the given VM is currently migrating
75 int MSG_vm_is_migrating(msg_vm_t vm)
77 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->isMigrating();
80 /** @brief Returns whether the given VM is currently suspended, not running.
83 int MSG_vm_is_suspended(msg_vm_t vm)
85 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
88 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
91 int MSG_vm_is_saving(msg_vm_t vm)
93 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
96 /** @brief Returns whether the given VM has been saved, not running.
99 int MSG_vm_is_saved(msg_vm_t vm)
101 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
104 /** @brief Returns whether the given VM is being restored, not running.
107 int MSG_vm_is_restoring(msg_vm_t vm)
109 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
112 /* **** ******** MSG vm actions ********* **** */
113 /** @brief Create a new VM with specified parameters.
115 * @param pm Physical machine that will host the VM
117 * @param ncpus [TODO]
118 * @param ramsize [TODO]
119 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
120 * @param disk_path (unused) Path to the image that boots
121 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
122 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
123 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
125 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name, int ncpus, int ramsize, int net_cap, char *disk_path,
126 int disksize, int mig_netspeed, int dp_intensity)
128 /* For the moment, intensity_rate is the percentage against the migration
130 double host_speed = MSG_host_get_speed(pm);
131 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
133 msg_vm_t vm = MSG_vm_create_core(pm, name);
134 s_vm_params_t params;
135 memset(¶ms, 0, sizeof(params));
136 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
137 //params.overcommit = 0;
139 params.skip_stage2 = 0;
140 params.max_downtime = 0.03;
141 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
142 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
143 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
145 //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate,
146 // params.mig_speed, dp_intensity, update_speed, host_speed);
147 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(¶ms);
152 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
155 * A VM is treated as a host. The name of the VM must be unique among all hosts.
157 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
159 /* make sure the VM of the same name does not exit */
161 simgrid::s4u::Host* ind_host_tmp = sg_host_by_name(name);
162 if (ind_host_tmp != nullptr && sg_host_simix(ind_host_tmp) != nullptr) {
163 XBT_ERROR("host %s already exits", name);
168 /* Note: ind_vm and vm_workstation point to the same elm object. */
169 /* Ask the SIMIX layer to create the surf vm resource */
170 sg_host_t ind_vm_workstation = simcall_vm_create(name, ind_pm);
172 msg_vm_t ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
174 XBT_DEBUG("A new VM (%s) has been created", name);
176 TRACE_msg_vm_create(name, ind_pm);
181 /** @brief Destroy a VM. Destroy the VM object from the simulation.
184 void MSG_vm_destroy(msg_vm_t vm)
186 if (MSG_vm_is_migrating(vm))
187 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
189 /* First, terminate all processes on the VM if necessary */
190 if (MSG_vm_is_running(vm))
191 simcall_vm_shutdown(vm);
193 if (!MSG_vm_is_created(vm)) {
194 XBT_CRITICAL("shutdown the given VM before destroying it");
198 /* Then, destroy the VM object */
199 simcall_vm_destroy(vm);
201 TRACE_msg_vm_end(vm);
204 /** @brief Start a vm (i.e., boot the guest operating system)
207 * If the VM cannot be started, an exception is generated.
209 void MSG_vm_start(msg_vm_t vm)
211 simcall_vm_start(vm);
213 TRACE_msg_vm_start(vm);
216 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
219 * FIXME: No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep() or something.
220 * I'm not quite sure.
222 void MSG_vm_shutdown(msg_vm_t vm)
224 /* msg_vm_t equals to msg_host_t */
225 simcall_vm_shutdown(vm);
227 // TRACE_msg_vm_(vm);
230 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
231 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
232 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
234 const char *vm_name = sg_host_get_name(vm);
235 const char *src_pm_name = sg_host_get_name(src_pm);
236 const char *dst_pm_name = sg_host_get_name(dst_pm);
238 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
241 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
243 const char *vm_name = sg_host_get_name(vm);
244 const char *src_pm_name = sg_host_get_name(src_pm);
245 const char *dst_pm_name = sg_host_get_name(dst_pm);
247 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
250 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
252 const char *vm_name = sg_host_get_name(vm);
253 const char *src_pm_name = sg_host_get_name(src_pm);
254 const char *dst_pm_name = sg_host_get_name(dst_pm);
256 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
259 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
261 const char *vm_name = sg_host_get_name(vm);
262 const char *src_pm_name = sg_host_get_name(src_pm);
263 const char *dst_pm_name = sg_host_get_name(dst_pm);
265 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
268 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
270 const char *vm_name = sg_host_get_name(vm);
271 const char *src_pm_name = sg_host_get_name(src_pm);
272 const char *dst_pm_name = sg_host_get_name(dst_pm);
274 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
277 struct migration_session {
282 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
283 * know the completion of the migration. */
285 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
289 static int migration_rx_fun(int argc, char *argv[])
291 XBT_DEBUG("mig: rx_start");
293 // The structure has been created in the do_migration function and should only be freed in the same place ;)
294 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
296 s_vm_params_t params;
297 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
301 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
305 msg_task_t task = nullptr;
306 ret = MSG_task_recv(&task, ms->mbox);
309 // An error occurred, clean the code and return
310 // The owner did not change, hence the task should be only destroyed on the other side
311 xbt_free(finalize_task_name);
316 if (strcmp(task->name, finalize_task_name) == 0)
319 MSG_task_destroy(task);
325 // Here Stage 1, 2 and 3 have been performed.
326 // Hence complete the migration
328 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
329 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
330 // structure at the beginning and free it at the end of each function)
331 simgrid::s4u::VirtualMachine* vm = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm);
332 msg_host_t src_pm = ms->src_pm;
333 msg_host_t dst_pm = ms->dst_pm;
335 // TODO: we have an issue, if the DST node is turning off during the three next calls, then the VM is in an
337 // state. I should check with Takahiro in order to make this portion of code atomic
339 // /* Update the vm location */
340 // simcall_vm_migrate(vm, dst_pm);
342 // /* Resume the VM */
343 // simcall_vm_resume(vm);
345 simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
348 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
349 vm->pimpl_vm_->isMigrating = false;
350 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
351 sg_host_get_name(ms->dst_pm));
352 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
354 // Inform the SRC that the migration has been correctly performed
356 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
357 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
358 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
359 // xbt_assert(ret == MSG_OK);
360 if(ret == MSG_HOST_FAILURE){
361 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
362 // has been correctly migrated on the DST node
363 // TODO What does it mean ? What should we do ?
364 MSG_task_destroy(task);
365 } else if(ret == MSG_TRANSFER_FAILURE){
366 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
367 MSG_task_destroy(task);
372 xbt_free(finalize_task_name);
374 XBT_DEBUG("mig: rx_done");
378 static void reset_dirty_pages(msg_vm_t vm)
380 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
383 xbt_dict_cursor_t cursor = nullptr;
384 dirty_page_t dp = nullptr;
387 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
388 double remaining = MSG_task_get_flops_amount(dp->task);
389 dp->prev_clock = MSG_get_clock();
390 dp->prev_remaining = remaining;
392 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
396 static void start_dirty_page_tracking(msg_vm_t vm)
398 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 1;
400 reset_dirty_pages(vm);
403 static void stop_dirty_page_tracking(msg_vm_t vm)
405 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
408 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
410 double computed = dp->prev_remaining - remaining;
411 double duration = clock - dp->prev_clock;
413 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
414 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
419 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
421 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
425 xbt_dict_cursor_t cursor = nullptr;
426 dirty_page_t dp = nullptr;
427 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
428 double remaining = MSG_task_get_flops_amount(dp->task);
430 double clock = MSG_get_clock();
432 // total += calc_updated_pages(key, vm, dp, remaining, clock);
433 total += get_computed(key, vm, dp, remaining, clock);
435 dp->prev_remaining = remaining;
436 dp->prev_clock = clock;
439 total += pimpl->dp_updated_by_deleted_tasks;
441 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
442 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
444 pimpl->dp_updated_by_deleted_tasks = 0;
449 // TODO Is this code redundant with the information provided by
450 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
451 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
452 void MSG_host_add_task(msg_host_t host, msg_task_t task)
454 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
457 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
459 double remaining = MSG_task_get_flops_amount(task);
460 char *key = bprintf("%s-%p", task->name, task);
462 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
464 if (pimpl->dp_enabled) {
465 dp->prev_clock = MSG_get_clock();
466 dp->prev_remaining = remaining;
469 pimpl->dp_objs = xbt_dict_new();
470 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
471 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
472 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, pimpl->dp_enabled);
477 void MSG_host_del_task(msg_host_t host, msg_task_t task)
479 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
482 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
484 char *key = bprintf("%s-%p", task->name, task);
485 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
486 xbt_assert(dp->task == task);
488 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
489 * the information for the lookup_() function that will called soon. */
490 if (pimpl->dp_enabled) {
491 double remaining = MSG_task_get_flops_amount(task);
492 double clock = MSG_get_clock();
493 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
494 double updated = get_computed(key, host, dp, remaining, clock);
496 pimpl->dp_updated_by_deleted_tasks += updated;
499 xbt_dict_remove(pimpl->dp_objs, key);
502 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
506 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,
507 int stage, int stage2_round, double mig_speed, double timeout)
510 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
511 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
515 double clock_sta = MSG_get_clock();
519 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
521 ret = MSG_task_send(task, mbox);
527 } else if (ret == MSG_TIMEOUT) {
528 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
529 sent = size - remaining;
530 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
533 /* FIXME: why try-and-catch is used here? */
534 if(ret == MSG_HOST_FAILURE){
535 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
536 MSG_task_destroy(task);
537 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
538 }else if(ret == MSG_TRANSFER_FAILURE){
539 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
540 MSG_task_destroy(task);
541 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
544 double clock_end = MSG_get_clock();
545 double duration = clock_end - clock_sta;
546 double actual_speed = size / duration;
549 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)",
550 stage, stage2_round, size, duration, actual_speed, mig_speed);
552 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)",
553 stage, size, duration, actual_speed, mig_speed);
558 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
560 double updated_size = computed * dp_rate;
561 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
562 if (updated_size > dp_cap) {
563 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
565 updated_size = dp_cap;
568 return (sg_size_t) updated_size;
571 static double send_stage1(struct migration_session *ms, sg_size_t ramsize, double mig_speed, double dp_rate,
574 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
575 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
576 sg_size_t remaining = ramsize;
577 double computed_total = 0;
579 while (remaining > 0) {
580 sg_size_t datasize = chunksize;
581 if (remaining < chunksize)
582 datasize = remaining;
584 remaining -= datasize;
585 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
586 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
587 computed_total += computed;
590 return computed_total;
593 static double get_threshold_value(double bandwidth, double max_downtime)
595 return max_downtime * bandwidth;
598 static int migration_tx_fun(int argc, char *argv[])
600 XBT_DEBUG("mig: tx_start");
602 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
603 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
605 s_vm_params_t params;
606 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
607 const sg_size_t ramsize = params.ramsize;
608 const sg_size_t devsize = params.devsize;
609 const int skip_stage1 = params.skip_stage1;
610 int skip_stage2 = params.skip_stage2;
611 const double dp_rate = params.dp_rate;
612 const double dp_cap = params.dp_cap;
613 const double mig_speed = params.mig_speed;
614 double max_downtime = params.max_downtime;
616 /* hard code it temporally. Fix Me */
617 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
618 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
620 double remaining_size = (double) (ramsize + devsize);
621 double threshold = 0.0;
623 /* check parameters */
625 XBT_WARN("migrate a VM, but ramsize is zero");
627 if (max_downtime <= 0) {
628 XBT_WARN("use the default max_downtime value 30ms");
632 /* Stage1: send all memory pages to the destination. */
633 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
634 start_dirty_page_tracking(ms->vm);
636 double computed_during_stage1 = 0;
638 double clock_prev_send = MSG_get_clock();
641 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
642 * transfer exceeds the timeout value. */
643 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
644 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
645 remaining_size -= sent;
646 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
648 if (sent < ramsize) {
649 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
651 } else if (sent > ramsize)
656 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
657 // Stop the dirty page tracking an return (there is no memory space to release)
658 stop_dirty_page_tracking(ms->vm);
662 double clock_post_send = MSG_get_clock();
663 mig_timeout -= (clock_post_send - clock_prev_send);
664 if (mig_timeout < 0) {
665 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
666 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
670 /* estimate bandwidth */
671 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
672 threshold = get_threshold_value(bandwidth, max_downtime);
673 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
677 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
680 int stage2_round = 0;
683 sg_size_t updated_size = 0;
684 if (stage2_round == 0) {
685 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
686 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
688 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
689 updated_size = get_updated_size(computed, dp_rate, dp_cap);
692 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
693 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
695 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
696 remaining_size += updated_size;
697 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
698 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
699 if (remaining_size < threshold)
703 double clock_prev_send = MSG_get_clock();
705 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
706 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
710 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
711 // Stop the dirty page tracking an return (there is no memory space to release)
712 stop_dirty_page_tracking(ms->vm);
715 double clock_post_send = MSG_get_clock();
717 if (sent == updated_size) {
718 /* timeout did not happen */
719 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
720 threshold = get_threshold_value(bandwidth, max_downtime);
721 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
722 remaining_size -= sent;
724 mig_timeout -= (clock_post_send - clock_prev_send);
725 xbt_assert(mig_timeout > 0);
727 } else if (sent < updated_size) {
728 /* When timeout happens, we move to stage 3. The size of memory pages
729 * updated before timeout must be added to the remaining size. */
730 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
731 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
732 remaining_size -= sent;
734 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
735 updated_size = get_updated_size(computed, dp_rate, dp_cap);
736 remaining_size += updated_size;
743 /* Stage3: stop the VM and copy the rest of states. */
744 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
745 simcall_vm_suspend(ms->vm);
746 stop_dirty_page_tracking(ms->vm);
749 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
750 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
753 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
754 // Stop the dirty page tracking an return (there is no memory space to release)
755 simcall_vm_resume(ms->vm);
759 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
760 // effectively the VM on the DST node.
761 XBT_DEBUG("mig: tx_done");
766 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
768 struct migration_session *ms = xbt_new(struct migration_session, 1);
772 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
773 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
775 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
776 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
778 // msg_process_t tx_process, rx_process;
779 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
780 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
783 char **argv = xbt_new(char *, 2);
784 argv[0] = pr_rx_name;
786 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
789 char **argv = xbt_new(char *, 2);
790 argv[0] = pr_tx_name;
792 /* tx_process = */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 */
798 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
799 msg_task_t task = nullptr;
800 msg_error_t ret = MSG_TIMEOUT;
801 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
802 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
804 xbt_free(ms->mbox_ctl);
808 //xbt_assert(ret == MSG_OK);
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
817 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){
818 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
819 XBT_ERROR("DST crashes, throw an exception (m-control)");
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);
831 /** @brief Migrate the VM to the given host.
834 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
835 * depending on whether you want to do cold or hot migration.
837 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
840 * - One approach is ...
841 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
842 * receive the state of the source
843 * VM over network. We will finally destroy the source VM.
844 * - This behavior is similar to the way of migration in the real world.
845 * Even before a migration is completed, we will see a destination VM, consuming resources.
846 * - We have to relocate all processes. The existing process migration code will work for this?
847 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
849 * - Another one is ...
850 * We update the information of the given VM to place it to the destination physical host.
852 * The second one would be easier.
855 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
857 if(MSG_host_is_off(old_pm))
858 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
860 if(MSG_host_is_off(new_pm))
861 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
863 if (!MSG_vm_is_running(vm))
864 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
866 if (MSG_vm_is_migrating(vm))
867 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
869 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
870 pimpl->isMigrating = 1;
873 int ret = do_migration(vm, old_pm, new_pm);
875 pimpl->isMigrating = 0;
876 THROWF(host_error, 0, "SRC host failed during migration");
879 pimpl->isMigrating = 0;
880 THROWF(host_error, 0, "DST host failed during migration");
884 // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration
885 // process. In that case, the VM has been already assigned to the DST node.
886 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
887 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
890 /** @brief Immediately suspend the execution of all processes within the given VM.
893 * This function stops the execution of the VM. All the processes on this VM
894 * will pause. The state of the VM is preserved. We can later resume it again.
896 * No suspension cost occurs.
898 void MSG_vm_suspend(msg_vm_t vm)
900 if (MSG_vm_is_migrating(vm))
901 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
903 simcall_vm_suspend(vm);
905 XBT_DEBUG("vm_suspend done");
907 TRACE_msg_vm_suspend(vm);
910 /** @brief Resume the execution of the VM. All processes on the VM run again.
913 * No resume cost occurs.
915 void MSG_vm_resume(msg_vm_t vm)
917 simcall_vm_resume(vm);
919 TRACE_msg_vm_resume(vm);
923 /** @brief Immediately save the execution of all processes within the given VM.
926 * This function stops the execution of the VM. All the processes on this VM
927 * will pause. The state of the VM is preserved. We can later resume it again.
929 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
930 * or after, depending on the exact semantic of VM save to you.
932 void MSG_vm_save(msg_vm_t vm)
934 if (MSG_vm_is_migrating(vm))
935 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
938 TRACE_msg_vm_save(vm);
941 /** @brief Restore the execution of the VM. All processes on the VM run again.
944 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
945 * after, depending on the exact semantic of VM restore to you.
947 void MSG_vm_restore(msg_vm_t vm)
949 simcall_vm_restore(vm);
951 TRACE_msg_vm_restore(vm);
954 /** @brief Get the physical host of a given VM.
957 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
959 return (msg_host_t) simcall_vm_get_pm(vm);
962 /** @brief Set a CPU bound for a given VM.
965 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
968 * On PM0, there are Task1 and VM0.
969 * On VM0, there is Task2.
970 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
972 * Task1\@PM0 gets 50%.
973 * Task2\@VM0 gets 25%.
974 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
976 * 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
977 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
979 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
980 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
981 * on the VM in share_resource() or we have to trap both the start and end of task execution.
983 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
985 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
986 * computation task (or a VM) never exceeds the capacity of a CPU core.
988 void MSG_vm_set_bound(msg_vm_t vm, double bound)
990 simcall_vm_set_bound(vm, bound);