1 /* Copyright (c) 2012-2015. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
8 * 1. add the support of trace
9 * 2. use parallel tasks to simulate CPU overhead and remove the experimental code generating micro computation tasks
14 #include "src/plugins/vm/VirtualMachineImpl.hpp"
15 #include "src/plugins/vm/VmHostExt.hpp"
16 #include <simgrid/s4u/VirtualMachine.hpp>
17 #include <simgrid/s4u/host.hpp>
19 #include "msg_private.h"
20 #include "xbt/sysdep.h"
22 #include "simgrid/host.h"
24 #include "src/simix/smx_host_private.h" /* don't ask me why the VM functions are in there (FIXME:KILLME) */
26 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
29 /* **** ******** GENERAL ********* **** */
31 /** \ingroup m_vm_management
32 * \brief Set the parameters of a given host
35 * \param params a parameter object
37 void MSG_vm_set_params(msg_vm_t vm, vm_params_t params)
39 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(params);
42 /** \ingroup m_vm_management
43 * \brief Get the parameters of a given host
46 * \param params a prameter object
48 void MSG_vm_get_params(msg_vm_t vm, vm_params_t params)
50 static_cast<simgrid::s4u::VirtualMachine*>(vm)->parameters(params);
53 /* **** Check state of a VM **** */
54 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
56 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getState() == state;
59 /** @brief Returns whether the given VM has just created, not running.
62 int MSG_vm_is_created(msg_vm_t vm)
64 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
67 /** @brief Returns whether the given VM is currently running
70 int MSG_vm_is_running(msg_vm_t vm)
72 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
75 /** @brief Returns whether the given VM is currently migrating
78 int MSG_vm_is_migrating(msg_vm_t vm)
80 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->isMigrating();
83 /** @brief Returns whether the given VM is currently suspended, not running.
86 int MSG_vm_is_suspended(msg_vm_t vm)
88 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
91 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
94 int MSG_vm_is_saving(msg_vm_t vm)
96 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
99 /** @brief Returns whether the given VM has been saved, not running.
102 int MSG_vm_is_saved(msg_vm_t vm)
104 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
107 /** @brief Returns whether the given VM is being restored, not running.
110 int MSG_vm_is_restoring(msg_vm_t vm)
112 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
115 /* **** ******** MSG vm actions ********* **** */
116 /** @brief Create a new VM with specified parameters.
118 * @param pm Physical machine that will host the VM
120 * @param ncpus [TODO]
121 * @param ramsize [TODO]
122 * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s)
123 * @param disk_path (unused) Path to the image that boots
124 * @param disksize (unused) will represent the size of the VM (will be used during migrations)
125 * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure.
126 * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure.
128 msg_vm_t MSG_vm_create(msg_host_t pm, const char *name, int ncpus, int ramsize, int net_cap, char *disk_path,
129 int disksize, int mig_netspeed, int dp_intensity)
131 simgrid::vm::VmHostExt::ensureVmExtInstalled();
133 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
134 double host_speed = MSG_host_get_speed(pm);
135 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
137 msg_vm_t vm = MSG_vm_create_core(pm, name);
138 s_vm_params_t params;
139 memset(¶ms, 0, sizeof(params));
140 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
141 //params.overcommit = 0;
143 params.skip_stage2 = 0;
144 params.max_downtime = 0.03;
145 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
146 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
147 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
149 //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate,
150 // params.mig_speed, dp_intensity, update_speed, host_speed);
151 static_cast<simgrid::s4u::VirtualMachine*>(vm)->setParameters(¶ms);
156 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
159 * A VM is treated as a host. The name of the VM must be unique among all hosts.
161 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
163 /* make sure the VM of the same name does not exit */
165 simgrid::s4u::Host* ind_host_tmp = sg_host_by_name(name);
166 if (ind_host_tmp != nullptr && sg_host_simix(ind_host_tmp) != nullptr) {
167 XBT_ERROR("host %s already exits", name);
172 /* Note: ind_vm and vm_workstation point to the same elm object. */
173 /* Ask the SIMIX layer to create the surf vm resource */
174 sg_host_t ind_vm_workstation = simcall_vm_create(name, ind_pm);
176 msg_vm_t ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
178 XBT_DEBUG("A new VM (%s) has been created", name);
180 TRACE_msg_vm_create(name, ind_pm);
185 /** @brief Destroy a VM. Destroy the VM object from the simulation.
188 void MSG_vm_destroy(msg_vm_t vm)
190 if (MSG_vm_is_migrating(vm))
191 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
193 /* First, terminate all processes on the VM if necessary */
194 if (MSG_vm_is_running(vm))
195 simcall_vm_shutdown(vm);
197 if (!MSG_vm_is_created(vm)) {
198 XBT_CRITICAL("shutdown the given VM before destroying it");
202 /* Then, destroy the VM object */
203 simcall_vm_destroy(vm);
205 TRACE_msg_vm_end(vm);
208 /** @brief Start a vm (i.e., boot the guest operating system)
211 * If the VM cannot be started, an exception is generated.
213 void MSG_vm_start(msg_vm_t vm)
215 simcall_vm_start(vm);
217 TRACE_msg_vm_start(vm);
220 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
223 * FIXME: No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep() or something.
224 * I'm not quite sure.
226 void MSG_vm_shutdown(msg_vm_t vm)
228 /* msg_vm_t equals to msg_host_t */
229 simcall_vm_shutdown(vm);
230 MSG_process_sleep(0.); // Make sure that the processes in the VM are killed in this scheduling round before processing
231 // (eg with the VM destroy)
232 // TRACE_msg_vm_(vm);
235 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
236 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
237 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
239 const char *vm_name = sg_host_get_name(vm);
240 const char *src_pm_name = sg_host_get_name(src_pm);
241 const char *dst_pm_name = sg_host_get_name(dst_pm);
243 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
246 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
248 const char *vm_name = sg_host_get_name(vm);
249 const char *src_pm_name = sg_host_get_name(src_pm);
250 const char *dst_pm_name = sg_host_get_name(dst_pm);
252 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
255 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
257 const char *vm_name = sg_host_get_name(vm);
258 const char *src_pm_name = sg_host_get_name(src_pm);
259 const char *dst_pm_name = sg_host_get_name(dst_pm);
261 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
264 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
266 const char *vm_name = sg_host_get_name(vm);
267 const char *src_pm_name = sg_host_get_name(src_pm);
268 const char *dst_pm_name = sg_host_get_name(dst_pm);
270 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
273 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
275 const char *vm_name = sg_host_get_name(vm);
276 const char *src_pm_name = sg_host_get_name(src_pm);
277 const char *dst_pm_name = sg_host_get_name(dst_pm);
279 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
282 struct migration_session {
287 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
288 * know the completion of the migration. */
290 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
294 static int migration_rx_fun(int argc, char *argv[])
296 XBT_DEBUG("mig: rx_start");
298 // The structure has been created in the do_migration function and should only be freed in the same place ;)
299 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
301 s_vm_params_t params;
302 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
306 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
310 msg_task_t task = nullptr;
311 ret = MSG_task_recv(&task, ms->mbox);
314 // An error occurred, clean the code and return
315 // The owner did not change, hence the task should be only destroyed on the other side
316 xbt_free(finalize_task_name);
321 if (strcmp(task->name, finalize_task_name) == 0)
324 MSG_task_destroy(task);
330 // Here Stage 1, 2 and 3 have been performed.
331 // Hence complete the migration
333 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
334 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
335 // structure at the beginning and free it at the end of each function)
336 simgrid::s4u::VirtualMachine* vm = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm);
337 msg_host_t src_pm = ms->src_pm;
338 msg_host_t dst_pm = ms->dst_pm;
340 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
341 simgrid::simix::kernelImmediate([vm, src_pm, dst_pm]() {
342 /* Update the vm location */
343 /* precopy migration makes the VM temporally paused */
344 xbt_assert(static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
346 /* jump to vm_ws_xigrate(). this will update the vm location. */
347 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->migrate(dst_pm);
354 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
355 vm->pimpl_vm_->isMigrating = false;
356 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
357 sg_host_get_name(ms->dst_pm));
358 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
360 // Inform the SRC that the migration has been correctly performed
362 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
363 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
364 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
365 // xbt_assert(ret == MSG_OK);
366 if(ret == MSG_HOST_FAILURE){
367 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
368 // has been correctly migrated on the DST node
369 // TODO What does it mean ? What should we do ?
370 MSG_task_destroy(task);
371 } else if(ret == MSG_TRANSFER_FAILURE){
372 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
373 MSG_task_destroy(task);
378 xbt_free(finalize_task_name);
380 XBT_DEBUG("mig: rx_done");
384 static void reset_dirty_pages(msg_vm_t vm)
386 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
389 xbt_dict_cursor_t cursor = nullptr;
390 dirty_page_t dp = nullptr;
393 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
394 double remaining = MSG_task_get_flops_amount(dp->task);
395 dp->prev_clock = MSG_get_clock();
396 dp->prev_remaining = remaining;
398 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
402 static void start_dirty_page_tracking(msg_vm_t vm)
404 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 1;
406 reset_dirty_pages(vm);
409 static void stop_dirty_page_tracking(msg_vm_t vm)
411 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
414 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
416 double computed = dp->prev_remaining - remaining;
417 double duration = clock - dp->prev_clock;
419 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
420 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
425 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
427 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
431 xbt_dict_cursor_t cursor = nullptr;
432 dirty_page_t dp = nullptr;
433 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
434 double remaining = MSG_task_get_flops_amount(dp->task);
436 double clock = MSG_get_clock();
438 // total += calc_updated_pages(key, vm, dp, remaining, clock);
439 total += get_computed(key, vm, dp, remaining, clock);
441 dp->prev_remaining = remaining;
442 dp->prev_clock = clock;
445 total += pimpl->dp_updated_by_deleted_tasks;
447 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
448 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
450 pimpl->dp_updated_by_deleted_tasks = 0;
455 // TODO Is this code redundant with the information provided by
456 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
457 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
458 void MSG_host_add_task(msg_host_t host, msg_task_t task)
460 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
463 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
465 double remaining = MSG_task_get_flops_amount(task);
466 char *key = bprintf("%s-%p", task->name, task);
468 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
470 if (pimpl->dp_enabled) {
471 dp->prev_clock = MSG_get_clock();
472 dp->prev_remaining = remaining;
475 pimpl->dp_objs = xbt_dict_new();
476 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
477 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
478 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, pimpl->dp_enabled);
483 void MSG_host_del_task(msg_host_t host, msg_task_t task)
485 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
488 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
490 char *key = bprintf("%s-%p", task->name, task);
491 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
492 xbt_assert(dp->task == task);
494 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
495 * the information for the lookup_() function that will called soon. */
496 if (pimpl->dp_enabled) {
497 double remaining = MSG_task_get_flops_amount(task);
498 double clock = MSG_get_clock();
499 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
500 double updated = get_computed(key, host, dp, remaining, clock);
502 pimpl->dp_updated_by_deleted_tasks += updated;
505 xbt_dict_remove(pimpl->dp_objs, key);
508 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
512 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,
513 int stage, int stage2_round, double mig_speed, double timeout)
516 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
517 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
521 double clock_sta = MSG_get_clock();
525 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
527 ret = MSG_task_send(task, mbox);
533 } else if (ret == MSG_TIMEOUT) {
534 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
535 sent = size - remaining;
536 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
539 /* FIXME: why try-and-catch is used here? */
540 if(ret == MSG_HOST_FAILURE){
541 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
542 MSG_task_destroy(task);
543 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
544 }else if(ret == MSG_TRANSFER_FAILURE){
545 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
546 MSG_task_destroy(task);
547 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
550 double clock_end = MSG_get_clock();
551 double duration = clock_end - clock_sta;
552 double actual_speed = size / duration;
555 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)",
556 stage, stage2_round, size, duration, actual_speed, mig_speed);
558 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)",
559 stage, size, duration, actual_speed, mig_speed);
564 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
566 double updated_size = computed * dp_rate;
567 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
568 if (updated_size > dp_cap) {
569 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
571 updated_size = dp_cap;
574 return (sg_size_t) updated_size;
577 static double send_stage1(struct migration_session *ms, sg_size_t ramsize, double mig_speed, double dp_rate,
580 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
581 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
582 sg_size_t remaining = ramsize;
583 double computed_total = 0;
585 while (remaining > 0) {
586 sg_size_t datasize = chunksize;
587 if (remaining < chunksize)
588 datasize = remaining;
590 remaining -= datasize;
591 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
592 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
593 computed_total += computed;
596 return computed_total;
599 static double get_threshold_value(double bandwidth, double max_downtime)
601 return max_downtime * bandwidth;
604 static int migration_tx_fun(int argc, char *argv[])
606 XBT_DEBUG("mig: tx_start");
608 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
609 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
611 s_vm_params_t params;
612 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
613 const sg_size_t ramsize = params.ramsize;
614 const sg_size_t devsize = params.devsize;
615 const int skip_stage1 = params.skip_stage1;
616 int skip_stage2 = params.skip_stage2;
617 const double dp_rate = params.dp_rate;
618 const double dp_cap = params.dp_cap;
619 const double mig_speed = params.mig_speed;
620 double max_downtime = params.max_downtime;
622 /* hard code it temporally. Fix Me */
623 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
624 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
626 double remaining_size = (double) (ramsize + devsize);
627 double threshold = 0.0;
629 /* check parameters */
631 XBT_WARN("migrate a VM, but ramsize is zero");
633 if (max_downtime <= 0) {
634 XBT_WARN("use the default max_downtime value 30ms");
638 /* Stage1: send all memory pages to the destination. */
639 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
640 start_dirty_page_tracking(ms->vm);
642 double computed_during_stage1 = 0;
644 double clock_prev_send = MSG_get_clock();
647 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
648 * transfer exceeds the timeout value. */
649 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
650 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
651 remaining_size -= sent;
652 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
654 if (sent < ramsize) {
655 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
657 } else if (sent > ramsize)
662 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
663 // Stop the dirty page tracking an return (there is no memory space to release)
664 stop_dirty_page_tracking(ms->vm);
668 double clock_post_send = MSG_get_clock();
669 mig_timeout -= (clock_post_send - clock_prev_send);
670 if (mig_timeout < 0) {
671 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
672 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
676 /* estimate bandwidth */
677 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
678 threshold = get_threshold_value(bandwidth, max_downtime);
679 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
683 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
686 int stage2_round = 0;
689 sg_size_t updated_size = 0;
690 if (stage2_round == 0) {
691 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
692 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
694 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
695 updated_size = get_updated_size(computed, dp_rate, dp_cap);
698 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
699 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
701 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
702 remaining_size += updated_size;
703 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
704 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
705 if (remaining_size < threshold)
709 double clock_prev_send = MSG_get_clock();
711 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
712 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
716 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
717 // Stop the dirty page tracking an return (there is no memory space to release)
718 stop_dirty_page_tracking(ms->vm);
721 double clock_post_send = MSG_get_clock();
723 if (sent == updated_size) {
724 /* timeout did not happen */
725 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
726 threshold = get_threshold_value(bandwidth, max_downtime);
727 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
728 remaining_size -= sent;
730 mig_timeout -= (clock_post_send - clock_prev_send);
731 xbt_assert(mig_timeout > 0);
733 } else if (sent < updated_size) {
734 /* When timeout happens, we move to stage 3. The size of memory pages
735 * updated before timeout must be added to the remaining size. */
736 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
737 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
738 remaining_size -= sent;
740 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
741 updated_size = get_updated_size(computed, dp_rate, dp_cap);
742 remaining_size += updated_size;
749 /* Stage3: stop the VM and copy the rest of states. */
750 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
751 simcall_vm_suspend(ms->vm);
752 stop_dirty_page_tracking(ms->vm);
755 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
756 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
759 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
760 // Stop the dirty page tracking an return (there is no memory space to release)
761 simcall_vm_resume(ms->vm);
765 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
766 // effectively the VM on the DST node.
767 XBT_DEBUG("mig: tx_done");
772 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
774 struct migration_session *ms = xbt_new(struct migration_session, 1);
778 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
779 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
781 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
782 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
784 // msg_process_t tx_process, rx_process;
785 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
786 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
789 char **argv = xbt_new(char *, 2);
790 argv[0] = pr_rx_name;
792 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
795 char **argv = xbt_new(char *, 2);
796 argv[0] = pr_tx_name;
798 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
802 /* wait until the migration have finished or on error has occurred */
804 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
805 msg_task_t task = nullptr;
806 msg_error_t ret = MSG_TIMEOUT;
807 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
808 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
810 xbt_free(ms->mbox_ctl);
814 //xbt_assert(ret == MSG_OK);
815 if(ret == MSG_HOST_FAILURE){
816 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
817 // other side. Hence, just throw the execption
818 XBT_ERROR("SRC crashes, throw an exception (m-control)");
819 //MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
823 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){
824 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
825 XBT_ERROR("DST crashes, throw an exception (m-control)");
829 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
830 xbt_assert(strcmp(task->name, expected_task_name) == 0);
831 xbt_free(expected_task_name);
832 MSG_task_destroy(task);
837 /** @brief Migrate the VM to the given host.
840 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
841 * depending on whether you want to do cold or hot migration.
843 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
846 * - One approach is ...
847 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
848 * receive the state of the source
849 * VM over network. We will finally destroy the source VM.
850 * - This behavior is similar to the way of migration in the real world.
851 * Even before a migration is completed, we will see a destination VM, consuming resources.
852 * - We have to relocate all processes. The existing process migration code will work for this?
853 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
855 * - Another one is ...
856 * We update the information of the given VM to place it to the destination physical host.
858 * The second one would be easier.
861 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
863 if(MSG_host_is_off(old_pm))
864 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
866 if(MSG_host_is_off(new_pm))
867 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
869 if (!MSG_vm_is_running(vm))
870 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
872 if (MSG_vm_is_migrating(vm))
873 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
875 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
876 pimpl->isMigrating = 1;
879 int ret = do_migration(vm, old_pm, new_pm);
881 pimpl->isMigrating = 0;
882 THROWF(host_error, 0, "SRC host failed during migration");
885 pimpl->isMigrating = 0;
886 THROWF(host_error, 0, "DST host failed during migration");
890 // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration
891 // process. In that case, the VM has been already assigned to the DST node.
892 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
893 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
896 /** @brief Immediately suspend the execution of all processes within the given VM.
899 * This function stops the execution of the VM. All the processes on this VM
900 * will pause. The state of the VM is preserved. We can later resume it again.
902 * No suspension cost occurs.
904 void MSG_vm_suspend(msg_vm_t vm)
906 if (MSG_vm_is_migrating(vm))
907 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
909 simcall_vm_suspend(vm);
911 XBT_DEBUG("vm_suspend done");
913 TRACE_msg_vm_suspend(vm);
916 /** @brief Resume the execution of the VM. All processes on the VM run again.
919 * No resume cost occurs.
921 void MSG_vm_resume(msg_vm_t vm)
923 simcall_vm_resume(vm);
925 TRACE_msg_vm_resume(vm);
929 /** @brief Immediately save the execution of all processes within the given VM.
932 * This function stops the execution of the VM. All the processes on this VM
933 * will pause. The state of the VM is preserved. We can later resume it again.
935 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
936 * or after, depending on the exact semantic of VM save to you.
938 void MSG_vm_save(msg_vm_t vm)
940 if (MSG_vm_is_migrating(vm))
941 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
944 TRACE_msg_vm_save(vm);
947 /** @brief Restore the execution of the VM. All processes on the VM run again.
950 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
951 * after, depending on the exact semantic of VM restore to you.
953 void MSG_vm_restore(msg_vm_t vm)
955 simcall_vm_restore(vm);
957 TRACE_msg_vm_restore(vm);
960 /** @brief Get the physical host of a given VM.
963 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
965 return (msg_host_t) simcall_vm_get_pm(vm);
968 /** @brief Set a CPU bound for a given VM.
971 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
974 * On PM0, there are Task1 and VM0.
975 * On VM0, there is Task2.
976 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
978 * Task1\@PM0 gets 50%.
979 * Task2\@VM0 gets 25%.
980 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
982 * 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
983 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
985 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
986 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
987 * on the VM in share_resource() or we have to trap both the start and end of task execution.
989 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
991 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
992 * computation task (or a VM) never exceeds the capacity of a CPU core.
994 void MSG_vm_set_bound(msg_vm_t vm, double bound)
996 simcall_vm_set_bound(vm, bound);