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]() {
188 if (TRACE_msg_vm_is_enabled()) {
189 container_t container = PJ_container_get(vm->name().c_str());
190 PJ_container_remove_from_parent(container);
191 PJ_container_free(container);
195 /** @brief Start a vm (i.e., boot the guest operating system)
198 * If the VM cannot be started, an exception is generated.
200 void MSG_vm_start(msg_vm_t vm)
202 simcall_vm_start(vm);
204 if (TRACE_msg_vm_is_enabled()) {
205 container_t vm_container = PJ_container_get(vm->name().c_str());
206 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
207 val_t value = PJ_value_get_or_new("start", "0 0 1", type); // start is blue
208 new_pajePushState(MSG_get_clock(), vm_container, type, value);
212 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
215 * FIXME: No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep() or something.
216 * I'm not quite sure.
218 void MSG_vm_shutdown(msg_vm_t vm)
220 simcall_vm_shutdown(vm);
221 MSG_process_sleep(0.); // Make sure that the processes in the VM are killed in this scheduling round before processing
222 // (eg with the VM destroy)
225 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
226 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
227 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
229 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
232 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
234 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
237 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
239 return bprintf("__pr_mig_tx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
242 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
244 return bprintf("__pr_mig_rx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
247 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
249 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm->cname(), src_pm->cname(), dst_pm->cname());
252 struct migration_session {
257 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
258 * know the completion of the migration. */
260 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
264 static int migration_rx_fun(int argc, char *argv[])
266 XBT_DEBUG("mig: rx_start");
268 // The structure has been created in the do_migration function and should only be freed in the same place ;)
269 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
271 bool received_finalize = false;
273 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
274 while (!received_finalize) {
275 msg_task_t task = nullptr;
276 int ret = MSG_task_recv(&task, ms->mbox);
279 // An error occurred, clean the code and return
280 // The owner did not change, hence the task should be only destroyed on the other side
281 xbt_free(finalize_task_name);
285 if (strcmp(task->name, finalize_task_name) == 0)
286 received_finalize = 1;
288 MSG_task_destroy(task);
290 xbt_free(finalize_task_name);
292 // Here Stage 1, 2 and 3 have been performed.
293 // Hence complete the migration
295 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
296 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
297 // structure at the beginning and free it at the end of each function)
298 simgrid::s4u::VirtualMachine* vm = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm);
299 msg_host_t src_pm = ms->src_pm;
300 msg_host_t dst_pm = ms->dst_pm;
302 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
303 simgrid::simix::kernelImmediate([vm, src_pm, dst_pm]() {
304 /* Update the vm location */
305 /* precopy migration makes the VM temporally paused */
306 xbt_assert(static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
308 /* jump to vm_ws_xigrate(). this will update the vm location. */
309 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->migrate(dst_pm);
316 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
317 vm->pimpl_vm_->isMigrating = false;
318 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
319 sg_host_get_name(ms->dst_pm));
321 if (TRACE_msg_vm_is_enabled()) {
322 static long long int counter = 0;
323 char key[INSTR_DEFAULT_STR_SIZE];
324 snprintf(key, INSTR_DEFAULT_STR_SIZE, "%lld", counter++);
327 container_t msg = PJ_container_get(vm->name().c_str());
328 type_t type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
329 new_pajeStartLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
331 // destroy existing container of this vm
332 container_t existing_container = PJ_container_get(vm->name().c_str());
333 PJ_container_remove_from_parent(existing_container);
334 PJ_container_free(existing_container);
336 // create new container on the new_host location
337 PJ_container_new(vm->name().c_str(), INSTR_MSG_VM, PJ_container_get(sg_host_get_name(ms->dst_pm)));
340 msg = PJ_container_get(vm->name().c_str());
341 type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
342 new_pajeEndLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
345 // Inform the SRC that the migration has been correctly performed
347 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
348 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
349 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
350 // xbt_assert(ret == MSG_OK);
351 if(ret == MSG_HOST_FAILURE){
352 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
353 // has been correctly migrated on the DST node
354 // TODO What does it mean ? What should we do ?
355 MSG_task_destroy(task);
356 } else if(ret == MSG_TRANSFER_FAILURE){
357 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
358 MSG_task_destroy(task);
363 XBT_DEBUG("mig: rx_done");
367 static void start_dirty_page_tracking(msg_vm_t vm)
369 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
371 pimpl->dp_enabled = 1;
376 xbt_dict_cursor_t cursor = nullptr;
377 dirty_page_t dp = nullptr;
378 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
379 double remaining = MSG_task_get_flops_amount(dp->task);
380 dp->prev_clock = MSG_get_clock();
381 dp->prev_remaining = remaining;
383 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
387 static void stop_dirty_page_tracking(msg_vm_t vm)
389 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
392 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
394 double computed = dp->prev_remaining - remaining;
395 double duration = clock - dp->prev_clock;
397 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
398 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
403 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
405 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
409 xbt_dict_cursor_t cursor = nullptr;
410 dirty_page_t dp = nullptr;
411 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
412 double remaining = MSG_task_get_flops_amount(dp->task);
414 double clock = MSG_get_clock();
416 // total += calc_updated_pages(key, vm, dp, remaining, clock);
417 total += get_computed(key, vm, dp, remaining, clock);
419 dp->prev_remaining = remaining;
420 dp->prev_clock = clock;
423 total += pimpl->dp_updated_by_deleted_tasks;
425 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
426 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
428 pimpl->dp_updated_by_deleted_tasks = 0;
433 // TODO Is this code redundant with the information provided by
434 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
435 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
436 void MSG_host_add_task(msg_host_t host, msg_task_t task)
438 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
441 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
443 double remaining = MSG_task_get_flops_amount(task);
444 char *key = bprintf("%s-%p", task->name, task);
446 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
448 if (pimpl->dp_enabled) {
449 dp->prev_clock = MSG_get_clock();
450 dp->prev_remaining = remaining;
453 pimpl->dp_objs = xbt_dict_new();
454 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
455 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
456 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, pimpl->dp_enabled);
461 void MSG_host_del_task(msg_host_t host, msg_task_t task)
463 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
466 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
468 char *key = bprintf("%s-%p", task->name, task);
469 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
470 xbt_assert(dp->task == task);
472 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
473 * the information for the lookup_() function that will called soon. */
474 if (pimpl->dp_enabled) {
475 double remaining = MSG_task_get_flops_amount(task);
476 double clock = MSG_get_clock();
477 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
478 double updated = get_computed(key, host, dp, remaining, clock);
480 pimpl->dp_updated_by_deleted_tasks += updated;
483 xbt_dict_remove(pimpl->dp_objs, key);
486 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
490 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,
491 int stage, int stage2_round, double mig_speed, double timeout)
494 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
495 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
499 double clock_sta = MSG_get_clock();
503 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
505 ret = MSG_task_send(task, mbox);
511 } else if (ret == MSG_TIMEOUT) {
512 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
513 sent = size - remaining;
514 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
517 /* FIXME: why try-and-catch is used here? */
518 if(ret == MSG_HOST_FAILURE){
519 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
520 MSG_task_destroy(task);
521 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
522 }else if(ret == MSG_TRANSFER_FAILURE){
523 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
524 MSG_task_destroy(task);
525 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
528 double clock_end = MSG_get_clock();
529 double duration = clock_end - clock_sta;
530 double actual_speed = size / duration;
533 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
534 actual_speed, mig_speed);
536 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
542 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
544 double updated_size = computed * dp_rate;
545 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
546 if (updated_size > dp_cap) {
547 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
549 updated_size = dp_cap;
552 return (sg_size_t) updated_size;
555 static double send_stage1(struct migration_session* ms, sg_size_t ramsize, double mig_speed, double dp_rate,
558 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
559 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
560 sg_size_t remaining = ramsize;
561 double computed_total = 0;
563 while (remaining > 0) {
564 sg_size_t datasize = chunksize;
565 if (remaining < chunksize)
566 datasize = remaining;
568 remaining -= datasize;
569 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
570 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
571 computed_total += computed;
574 return computed_total;
577 static double get_threshold_value(double bandwidth, double max_downtime)
579 return max_downtime * bandwidth;
582 static int migration_tx_fun(int argc, char *argv[])
584 XBT_DEBUG("mig: tx_start");
586 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
587 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
589 s_vm_params_t params;
590 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
591 const sg_size_t ramsize = params.ramsize;
592 const sg_size_t devsize = params.devsize;
593 const int skip_stage1 = params.skip_stage1;
594 int skip_stage2 = params.skip_stage2;
595 const double dp_rate = params.dp_rate;
596 const double dp_cap = params.dp_cap;
597 const double mig_speed = params.mig_speed;
598 double max_downtime = params.max_downtime;
600 /* hard code it temporally. Fix Me */
601 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
602 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
604 double remaining_size = (double) (ramsize + devsize);
605 double threshold = 0.0;
607 /* check parameters */
609 XBT_WARN("migrate a VM, but ramsize is zero");
611 if (max_downtime <= 0) {
612 XBT_WARN("use the default max_downtime value 30ms");
616 /* Stage1: send all memory pages to the destination. */
617 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
618 start_dirty_page_tracking(ms->vm);
620 double computed_during_stage1 = 0;
622 double clock_prev_send = MSG_get_clock();
625 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
626 * transfer exceeds the timeout value. */
627 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
628 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
629 remaining_size -= sent;
630 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
632 if (sent < ramsize) {
633 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
635 } else if (sent > ramsize)
640 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
641 // Stop the dirty page tracking an return (there is no memory space to release)
642 stop_dirty_page_tracking(ms->vm);
646 double clock_post_send = MSG_get_clock();
647 mig_timeout -= (clock_post_send - clock_prev_send);
648 if (mig_timeout < 0) {
649 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
650 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
654 /* estimate bandwidth */
655 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
656 threshold = get_threshold_value(bandwidth, max_downtime);
657 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
661 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
664 int stage2_round = 0;
667 sg_size_t updated_size = 0;
668 if (stage2_round == 0) {
669 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
670 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
672 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
673 updated_size = get_updated_size(computed, dp_rate, dp_cap);
676 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
677 updated_size, computed_during_stage1, dp_rate, dp_cap);
679 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
680 remaining_size += updated_size;
681 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
682 (remaining_size < threshold) ? "<" : ">", threshold);
683 if (remaining_size < threshold)
687 double clock_prev_send = MSG_get_clock();
689 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
690 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
692 } catch (xbt_ex& e) {
693 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
695 // Stop the dirty page tracking an return (there is no memory space to release)
696 stop_dirty_page_tracking(ms->vm);
699 double clock_post_send = MSG_get_clock();
701 if (sent == updated_size) {
702 /* timeout did not happen */
703 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
704 threshold = get_threshold_value(bandwidth, max_downtime);
705 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
706 remaining_size -= sent;
708 mig_timeout -= (clock_post_send - clock_prev_send);
709 xbt_assert(mig_timeout > 0);
711 } else if (sent < updated_size) {
712 /* When timeout happens, we move to stage 3. The size of memory pages
713 * updated before timeout must be added to the remaining size. */
714 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
715 updated_size, (clock_post_send - clock_prev_send));
716 remaining_size -= sent;
718 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
719 updated_size = get_updated_size(computed, dp_rate, dp_cap);
720 remaining_size += updated_size;
727 /* Stage3: stop the VM and copy the rest of states. */
728 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
729 simcall_vm_suspend(ms->vm);
730 stop_dirty_page_tracking(ms->vm);
733 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
734 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
737 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
738 // Stop the dirty page tracking an return (there is no memory space to release)
739 simcall_vm_resume(ms->vm);
743 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
744 // effectively the VM on the DST node.
745 XBT_DEBUG("mig: tx_done");
750 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
752 struct migration_session *ms = xbt_new(struct migration_session, 1);
756 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
757 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
759 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
760 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
762 char** argv = xbt_new(char*, 2);
763 argv[0] = pr_rx_name;
765 MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
767 argv = xbt_new(char*, 2);
768 argv[0] = pr_tx_name;
770 MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
772 /* wait until the migration have finished or on error has occurred */
773 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
774 msg_task_t task = nullptr;
775 msg_error_t ret = MSG_TIMEOUT;
776 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) // Wait while you receive the message o
777 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
779 xbt_free(ms->mbox_ctl);
783 // xbt_assert(ret == MSG_OK);
784 if (ret == MSG_HOST_FAILURE) {
785 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
786 // other side. Hence, just throw the execption
787 XBT_ERROR("SRC crashes, throw an exception (m-control)");
788 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
791 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
792 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
793 XBT_ERROR("DST crashes, throw an exception (m-control)");
797 char* expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
798 xbt_assert(strcmp(task->name, expected_task_name) == 0);
799 xbt_free(expected_task_name);
800 MSG_task_destroy(task);
804 /** @brief Migrate the VM to the given host.
807 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
808 * depending on whether you want to do cold or hot migration.
810 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
813 * - One approach is ...
814 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
815 * receive the state of the source
816 * VM over network. We will finally destroy the source VM.
817 * - This behavior is similar to the way of migration in the real world.
818 * Even before a migration is completed, we will see a destination VM, consuming resources.
819 * - We have to relocate all processes. The existing process migration code will work for this?
820 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
822 * - Another one is ...
823 * We update the information of the given VM to place it to the destination physical host.
825 * The second one would be easier.
828 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
829 msg_host_t old_pm = pimpl->getPm();
832 THROWF(vm_error, 0, "Cannot start a migration from host '%s', which is offline.", sg_host_get_name(old_pm));
835 THROWF(vm_error, 0, "Cannot start a migration to host '%s', which is offline.", sg_host_get_name(new_pm));
837 if (!MSG_vm_is_running(vm))
838 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
840 if (MSG_vm_is_migrating(vm))
841 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
843 pimpl->isMigrating = 1;
845 int ret = do_migration(vm, old_pm, new_pm);
847 pimpl->isMigrating = 0;
848 THROWF(host_error, 0, "SRC host failed during migration");
849 } else if (ret == -2) {
850 pimpl->isMigrating = 0;
851 THROWF(host_error, 0, "DST host failed during migration");
855 /** @brief Immediately suspend the execution of all processes within the given VM.
858 * This function stops the execution of the VM. All the processes on this VM
859 * will pause. The state of the VM is preserved. We can later resume it again.
861 * No suspension cost occurs.
863 void MSG_vm_suspend(msg_vm_t vm)
865 if (MSG_vm_is_migrating(vm))
866 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
868 simcall_vm_suspend(vm);
870 XBT_DEBUG("vm_suspend done");
872 if (TRACE_msg_vm_is_enabled()) {
873 container_t vm_container = PJ_container_get(vm->name().c_str());
874 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
875 val_t value = PJ_value_get_or_new("suspend", "1 0 0", type); // suspend is red
876 new_pajePushState(MSG_get_clock(), vm_container, type, value);
880 /** @brief Resume the execution of the VM. All processes on the VM run again.
883 * No resume cost occurs.
885 void MSG_vm_resume(msg_vm_t vm)
887 simcall_vm_resume(vm);
889 if (TRACE_msg_vm_is_enabled()) {
890 container_t vm_container = PJ_container_get(vm->name().c_str());
891 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
892 new_pajePopState(MSG_get_clock(), vm_container, type);
897 /** @brief Immediately save the execution of all processes within the given VM.
900 * This function stops the execution of the VM. All the processes on this VM
901 * will pause. The state of the VM is preserved. We can later resume it again.
903 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
904 * or after, depending on the exact semantic of VM save to you.
906 void MSG_vm_save(msg_vm_t vm)
908 if (MSG_vm_is_migrating(vm))
909 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
913 if (TRACE_msg_vm_is_enabled()) {
914 container_t vm_container = PJ_container_get(vm->name().c_str());
915 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
916 val_t value = PJ_value_get_or_new("save", "0 1 0", type); // save is green
917 new_pajePushState(MSG_get_clock(), vm_container, type, value);
921 /** @brief Restore the execution of the VM. All processes on the VM run again.
924 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
925 * after, depending on the exact semantic of VM restore to you.
927 void MSG_vm_restore(msg_vm_t vm)
929 simcall_vm_restore(vm);
931 if (TRACE_msg_vm_is_enabled()) {
932 container_t vm_container = PJ_container_get(vm->name().c_str());
933 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
934 new_pajePopState(MSG_get_clock(), vm_container, type);
938 /** @brief Get the physical host of a given VM.
941 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
943 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getPm();
946 /** @brief Set a CPU bound for a given VM.
949 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
952 * On PM0, there are Task1 and VM0.
953 * On VM0, there is Task2.
954 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
956 * Task1\@PM0 gets 50%.
957 * Task2\@VM0 gets 25%.
958 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
960 * 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
961 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
963 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
964 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
965 * on the VM in share_resource() or we have to trap both the start and end of task execution.
967 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
969 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
970 * computation task (or a VM) never exceeds the capacity of a CPU core.
972 void MSG_vm_set_bound(msg_vm_t vm, double bound)
974 simcall_vm_set_bound(vm, bound);