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 pm, const char* name)
163 xbt_assert(sg_host_by_name(name) == nullptr,
164 "Cannot create a VM named %s: this name is already used by an host or a VM", name);
166 return new simgrid::s4u::VirtualMachine(name, pm);
169 /** @brief Destroy a VM. Destroy the VM object from the simulation.
172 void MSG_vm_destroy(msg_vm_t vm)
174 if (MSG_vm_is_migrating(vm))
175 THROWF(vm_error, 0, "VM(%s) is migrating", vm->name().c_str());
177 /* First, terminate all processes on the VM if necessary */
178 if (MSG_vm_is_running(vm))
181 xbt_assert(MSG_vm_is_created(vm), "shutdown the given VM before destroying it");
183 /* Then, destroy the VM object */
184 simgrid::simix::kernelImmediate([vm]() {
185 /* this code basically performs a similar thing like SIMIX_host_destroy() */
186 XBT_DEBUG("destroy %s", vm->name().c_str());
188 /* FIXME: this is really strange that everything fails if the next line is removed.
189 * This is as if we shared these data with the PM, which definitely should not be the case...
191 * We need to test that suspending a VM does not suspends the processes running on its PM, for example.
192 * Or we need to simplify this code enough to make it actually readable (but this sounds harder than testing)
194 vm->extension_set<simgrid::simix::Host>(nullptr);
196 /* Don't free these things twice: they are the ones of my physical host */
197 vm->pimpl_cpu = nullptr;
198 vm->pimpl_netcard = nullptr;
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);
228 MSG_process_sleep(0.); // Make sure that the processes in the VM are killed in this scheduling round before processing
229 // (eg with the VM destroy)
230 // TRACE_msg_vm_(vm);
233 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
234 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
235 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
237 const char *vm_name = sg_host_get_name(vm);
238 const char *src_pm_name = sg_host_get_name(src_pm);
239 const char *dst_pm_name = sg_host_get_name(dst_pm);
241 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
244 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
246 const char *vm_name = sg_host_get_name(vm);
247 const char *src_pm_name = sg_host_get_name(src_pm);
248 const char *dst_pm_name = sg_host_get_name(dst_pm);
250 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
253 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
255 const char *vm_name = sg_host_get_name(vm);
256 const char *src_pm_name = sg_host_get_name(src_pm);
257 const char *dst_pm_name = sg_host_get_name(dst_pm);
259 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
262 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
264 const char *vm_name = sg_host_get_name(vm);
265 const char *src_pm_name = sg_host_get_name(src_pm);
266 const char *dst_pm_name = sg_host_get_name(dst_pm);
268 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
271 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
273 const char *vm_name = sg_host_get_name(vm);
274 const char *src_pm_name = sg_host_get_name(src_pm);
275 const char *dst_pm_name = sg_host_get_name(dst_pm);
277 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
280 struct migration_session {
285 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
286 * know the completion of the migration. */
288 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
292 static int migration_rx_fun(int argc, char *argv[])
294 XBT_DEBUG("mig: rx_start");
296 // The structure has been created in the do_migration function and should only be freed in the same place ;)
297 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
299 s_vm_params_t params;
300 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
304 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
308 msg_task_t task = nullptr;
309 ret = MSG_task_recv(&task, ms->mbox);
312 // An error occurred, clean the code and return
313 // The owner did not change, hence the task should be only destroyed on the other side
314 xbt_free(finalize_task_name);
319 if (strcmp(task->name, finalize_task_name) == 0)
322 MSG_task_destroy(task);
328 // Here Stage 1, 2 and 3 have been performed.
329 // Hence complete the migration
331 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
332 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
333 // structure at the beginning and free it at the end of each function)
334 simgrid::s4u::VirtualMachine* vm = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm);
335 msg_host_t src_pm = ms->src_pm;
336 msg_host_t dst_pm = ms->dst_pm;
338 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
339 simgrid::simix::kernelImmediate([vm, src_pm, dst_pm]() {
340 /* Update the vm location */
341 /* precopy migration makes the VM temporally paused */
342 xbt_assert(static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
344 /* jump to vm_ws_xigrate(). this will update the vm location. */
345 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->migrate(dst_pm);
352 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
353 vm->pimpl_vm_->isMigrating = false;
354 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
355 sg_host_get_name(ms->dst_pm));
356 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
358 // Inform the SRC that the migration has been correctly performed
360 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
361 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
362 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
363 // xbt_assert(ret == MSG_OK);
364 if(ret == MSG_HOST_FAILURE){
365 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
366 // has been correctly migrated on the DST node
367 // TODO What does it mean ? What should we do ?
368 MSG_task_destroy(task);
369 } else if(ret == MSG_TRANSFER_FAILURE){
370 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
371 MSG_task_destroy(task);
376 xbt_free(finalize_task_name);
378 XBT_DEBUG("mig: rx_done");
382 static void reset_dirty_pages(msg_vm_t vm)
384 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
387 xbt_dict_cursor_t cursor = nullptr;
388 dirty_page_t dp = nullptr;
391 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
392 double remaining = MSG_task_get_flops_amount(dp->task);
393 dp->prev_clock = MSG_get_clock();
394 dp->prev_remaining = remaining;
396 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
400 static void start_dirty_page_tracking(msg_vm_t vm)
402 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 1;
404 reset_dirty_pages(vm);
407 static void stop_dirty_page_tracking(msg_vm_t vm)
409 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
412 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
414 double computed = dp->prev_remaining - remaining;
415 double duration = clock - dp->prev_clock;
417 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
418 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
423 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
425 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
429 xbt_dict_cursor_t cursor = nullptr;
430 dirty_page_t dp = nullptr;
431 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
432 double remaining = MSG_task_get_flops_amount(dp->task);
434 double clock = MSG_get_clock();
436 // total += calc_updated_pages(key, vm, dp, remaining, clock);
437 total += get_computed(key, vm, dp, remaining, clock);
439 dp->prev_remaining = remaining;
440 dp->prev_clock = clock;
443 total += pimpl->dp_updated_by_deleted_tasks;
445 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
446 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
448 pimpl->dp_updated_by_deleted_tasks = 0;
453 // TODO Is this code redundant with the information provided by
454 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
455 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
456 void MSG_host_add_task(msg_host_t host, msg_task_t task)
458 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
461 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
463 double remaining = MSG_task_get_flops_amount(task);
464 char *key = bprintf("%s-%p", task->name, task);
466 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
468 if (pimpl->dp_enabled) {
469 dp->prev_clock = MSG_get_clock();
470 dp->prev_remaining = remaining;
473 pimpl->dp_objs = xbt_dict_new();
474 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
475 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
476 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, pimpl->dp_enabled);
481 void MSG_host_del_task(msg_host_t host, msg_task_t task)
483 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
486 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
488 char *key = bprintf("%s-%p", task->name, task);
489 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
490 xbt_assert(dp->task == task);
492 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
493 * the information for the lookup_() function that will called soon. */
494 if (pimpl->dp_enabled) {
495 double remaining = MSG_task_get_flops_amount(task);
496 double clock = MSG_get_clock();
497 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
498 double updated = get_computed(key, host, dp, remaining, clock);
500 pimpl->dp_updated_by_deleted_tasks += updated;
503 xbt_dict_remove(pimpl->dp_objs, key);
506 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
510 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,
511 int stage, int stage2_round, double mig_speed, double timeout)
514 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
515 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
519 double clock_sta = MSG_get_clock();
523 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
525 ret = MSG_task_send(task, mbox);
531 } else if (ret == MSG_TIMEOUT) {
532 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
533 sent = size - remaining;
534 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
537 /* FIXME: why try-and-catch is used here? */
538 if(ret == MSG_HOST_FAILURE){
539 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
540 MSG_task_destroy(task);
541 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
542 }else if(ret == MSG_TRANSFER_FAILURE){
543 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
544 MSG_task_destroy(task);
545 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
548 double clock_end = MSG_get_clock();
549 double duration = clock_end - clock_sta;
550 double actual_speed = size / duration;
553 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)",
554 stage, stage2_round, size, duration, actual_speed, mig_speed);
556 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)",
557 stage, size, duration, actual_speed, mig_speed);
562 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
564 double updated_size = computed * dp_rate;
565 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
566 if (updated_size > dp_cap) {
567 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
569 updated_size = dp_cap;
572 return (sg_size_t) updated_size;
575 static double send_stage1(struct migration_session *ms, sg_size_t ramsize, double mig_speed, double dp_rate,
578 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
579 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
580 sg_size_t remaining = ramsize;
581 double computed_total = 0;
583 while (remaining > 0) {
584 sg_size_t datasize = chunksize;
585 if (remaining < chunksize)
586 datasize = remaining;
588 remaining -= datasize;
589 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
590 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
591 computed_total += computed;
594 return computed_total;
597 static double get_threshold_value(double bandwidth, double max_downtime)
599 return max_downtime * bandwidth;
602 static int migration_tx_fun(int argc, char *argv[])
604 XBT_DEBUG("mig: tx_start");
606 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
607 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
609 s_vm_params_t params;
610 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
611 const sg_size_t ramsize = params.ramsize;
612 const sg_size_t devsize = params.devsize;
613 const int skip_stage1 = params.skip_stage1;
614 int skip_stage2 = params.skip_stage2;
615 const double dp_rate = params.dp_rate;
616 const double dp_cap = params.dp_cap;
617 const double mig_speed = params.mig_speed;
618 double max_downtime = params.max_downtime;
620 /* hard code it temporally. Fix Me */
621 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
622 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
624 double remaining_size = (double) (ramsize + devsize);
625 double threshold = 0.0;
627 /* check parameters */
629 XBT_WARN("migrate a VM, but ramsize is zero");
631 if (max_downtime <= 0) {
632 XBT_WARN("use the default max_downtime value 30ms");
636 /* Stage1: send all memory pages to the destination. */
637 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
638 start_dirty_page_tracking(ms->vm);
640 double computed_during_stage1 = 0;
642 double clock_prev_send = MSG_get_clock();
645 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
646 * transfer exceeds the timeout value. */
647 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
648 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
649 remaining_size -= sent;
650 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
652 if (sent < ramsize) {
653 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
655 } else if (sent > ramsize)
660 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
661 // Stop the dirty page tracking an return (there is no memory space to release)
662 stop_dirty_page_tracking(ms->vm);
666 double clock_post_send = MSG_get_clock();
667 mig_timeout -= (clock_post_send - clock_prev_send);
668 if (mig_timeout < 0) {
669 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
670 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
674 /* estimate bandwidth */
675 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
676 threshold = get_threshold_value(bandwidth, max_downtime);
677 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
681 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
684 int stage2_round = 0;
687 sg_size_t updated_size = 0;
688 if (stage2_round == 0) {
689 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
690 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
692 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
693 updated_size = get_updated_size(computed, dp_rate, dp_cap);
696 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
697 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
699 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
700 remaining_size += updated_size;
701 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
702 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
703 if (remaining_size < threshold)
707 double clock_prev_send = MSG_get_clock();
709 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
710 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
714 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
715 // Stop the dirty page tracking an return (there is no memory space to release)
716 stop_dirty_page_tracking(ms->vm);
719 double clock_post_send = MSG_get_clock();
721 if (sent == updated_size) {
722 /* timeout did not happen */
723 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
724 threshold = get_threshold_value(bandwidth, max_downtime);
725 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
726 remaining_size -= sent;
728 mig_timeout -= (clock_post_send - clock_prev_send);
729 xbt_assert(mig_timeout > 0);
731 } else if (sent < updated_size) {
732 /* When timeout happens, we move to stage 3. The size of memory pages
733 * updated before timeout must be added to the remaining size. */
734 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
735 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
736 remaining_size -= sent;
738 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
739 updated_size = get_updated_size(computed, dp_rate, dp_cap);
740 remaining_size += updated_size;
747 /* Stage3: stop the VM and copy the rest of states. */
748 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
749 simcall_vm_suspend(ms->vm);
750 stop_dirty_page_tracking(ms->vm);
753 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
754 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
757 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
758 // Stop the dirty page tracking an return (there is no memory space to release)
759 simcall_vm_resume(ms->vm);
763 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
764 // effectively the VM on the DST node.
765 XBT_DEBUG("mig: tx_done");
770 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
772 struct migration_session *ms = xbt_new(struct migration_session, 1);
776 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
777 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
779 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
780 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
782 // msg_process_t tx_process, rx_process;
783 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
784 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
787 char **argv = xbt_new(char *, 2);
788 argv[0] = pr_rx_name;
790 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
793 char **argv = xbt_new(char *, 2);
794 argv[0] = pr_tx_name;
796 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
800 /* wait until the migration have finished or on error has occurred */
802 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
803 msg_task_t task = nullptr;
804 msg_error_t ret = MSG_TIMEOUT;
805 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
806 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
808 xbt_free(ms->mbox_ctl);
812 //xbt_assert(ret == MSG_OK);
813 if(ret == MSG_HOST_FAILURE){
814 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
815 // other side. Hence, just throw the execption
816 XBT_ERROR("SRC crashes, throw an exception (m-control)");
817 //MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
821 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){
822 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
823 XBT_ERROR("DST crashes, throw an exception (m-control)");
827 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
828 xbt_assert(strcmp(task->name, expected_task_name) == 0);
829 xbt_free(expected_task_name);
830 MSG_task_destroy(task);
835 /** @brief Migrate the VM to the given host.
838 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
839 * depending on whether you want to do cold or hot migration.
841 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
844 * - One approach is ...
845 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
846 * receive the state of the source
847 * VM over network. We will finally destroy the source VM.
848 * - This behavior is similar to the way of migration in the real world.
849 * Even before a migration is completed, we will see a destination VM, consuming resources.
850 * - We have to relocate all processes. The existing process migration code will work for this?
851 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
853 * - Another one is ...
854 * We update the information of the given VM to place it to the destination physical host.
856 * The second one would be easier.
859 msg_host_t old_pm = (msg_host_t) simcall_vm_get_pm(vm);
861 if(MSG_host_is_off(old_pm))
862 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_get_name(old_pm));
864 if(MSG_host_is_off(new_pm))
865 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_get_name(new_pm));
867 if (!MSG_vm_is_running(vm))
868 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
870 if (MSG_vm_is_migrating(vm))
871 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
873 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::surf::VirtualMachineImpl*>(vm->pimpl_);
874 pimpl->isMigrating = 1;
877 int ret = do_migration(vm, old_pm, new_pm);
879 pimpl->isMigrating = 0;
880 THROWF(host_error, 0, "SRC host failed during migration");
883 pimpl->isMigrating = 0;
884 THROWF(host_error, 0, "DST host failed during migration");
888 // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration
889 // process. In that case, the VM has been already assigned to the DST node.
890 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
891 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
894 /** @brief Immediately suspend the execution of all processes within the given VM.
897 * This function stops the execution of the VM. All the processes on this VM
898 * will pause. The state of the VM is preserved. We can later resume it again.
900 * No suspension cost occurs.
902 void MSG_vm_suspend(msg_vm_t vm)
904 if (MSG_vm_is_migrating(vm))
905 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
907 simcall_vm_suspend(vm);
909 XBT_DEBUG("vm_suspend done");
911 TRACE_msg_vm_suspend(vm);
914 /** @brief Resume the execution of the VM. All processes on the VM run again.
917 * No resume cost occurs.
919 void MSG_vm_resume(msg_vm_t vm)
921 simcall_vm_resume(vm);
923 TRACE_msg_vm_resume(vm);
927 /** @brief Immediately save the execution of all processes within the given VM.
930 * This function stops the execution of the VM. All the processes on this VM
931 * will pause. The state of the VM is preserved. We can later resume it again.
933 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
934 * or after, depending on the exact semantic of VM save to you.
936 void MSG_vm_save(msg_vm_t vm)
938 if (MSG_vm_is_migrating(vm))
939 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
942 TRACE_msg_vm_save(vm);
945 /** @brief Restore the execution of the VM. All processes on the VM run again.
948 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
949 * after, depending on the exact semantic of VM restore to you.
951 void MSG_vm_restore(msg_vm_t vm)
953 simcall_vm_restore(vm);
955 TRACE_msg_vm_restore(vm);
958 /** @brief Get the physical host of a given VM.
961 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
963 return (msg_host_t) simcall_vm_get_pm(vm);
966 /** @brief Set a CPU bound for a given VM.
969 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
972 * On PM0, there are Task1 and VM0.
973 * On VM0, there is Task2.
974 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
976 * Task1\@PM0 gets 50%.
977 * Task2\@VM0 gets 25%.
978 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
980 * 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
981 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
983 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
984 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
985 * on the VM in share_resource() or we have to trap both the start and end of task execution.
987 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
989 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
990 * computation task (or a VM) never exceeds the capacity of a CPU core.
992 void MSG_vm_set_bound(msg_vm_t vm, double bound)
994 simcall_vm_set_bound(vm, bound);