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, "Cannot destroy VM '%s', which is migrating.", vm->cname());
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);
222 // Make sure that the processes in the VM are killed in this scheduling round before processing
223 // (eg with the VM destroy)
224 MSG_process_sleep(0.);
227 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
229 return bprintf("__pr_mig_tx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
232 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
234 return bprintf("__pr_mig_rx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
237 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
239 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm->cname(), src_pm->cname(), dst_pm->cname());
242 struct migration_session {
247 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
248 * know the completion of the migration. */
250 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
254 static int migration_rx_fun(int argc, char *argv[])
256 XBT_DEBUG("mig: rx_start");
258 // The structure has been created in the do_migration function and should only be freed in the same place ;)
259 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
261 bool received_finalize = false;
263 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
264 while (!received_finalize) {
265 msg_task_t task = nullptr;
266 int ret = MSG_task_recv(&task, ms->mbox);
269 // An error occurred, clean the code and return
270 // The owner did not change, hence the task should be only destroyed on the other side
271 xbt_free(finalize_task_name);
275 if (strcmp(task->name, finalize_task_name) == 0)
276 received_finalize = 1;
278 MSG_task_destroy(task);
280 xbt_free(finalize_task_name);
282 // Here Stage 1, 2 and 3 have been performed.
283 // Hence complete the migration
285 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
286 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
287 // structure at the beginning and free it at the end of each function)
288 simgrid::s4u::VirtualMachine* vm = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm);
289 msg_host_t src_pm = ms->src_pm;
290 msg_host_t dst_pm = ms->dst_pm;
292 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
293 simgrid::simix::kernelImmediate([vm, src_pm, dst_pm]() {
294 /* Update the vm location */
295 /* precopy migration makes the VM temporally paused */
296 xbt_assert(static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
298 /* jump to vm_ws_xigrate(). this will update the vm location. */
299 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->migrate(dst_pm);
306 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
307 vm->pimpl_vm_->isMigrating = false;
308 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->cname(), ms->src_pm->cname(), ms->dst_pm->cname());
310 if (TRACE_msg_vm_is_enabled()) {
311 static long long int counter = 0;
312 char key[INSTR_DEFAULT_STR_SIZE];
313 snprintf(key, INSTR_DEFAULT_STR_SIZE, "%lld", counter++);
316 container_t msg = PJ_container_get(vm->name().c_str());
317 type_t type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
318 new_pajeStartLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
320 // destroy existing container of this vm
321 container_t existing_container = PJ_container_get(vm->name().c_str());
322 PJ_container_remove_from_parent(existing_container);
323 PJ_container_free(existing_container);
325 // create new container on the new_host location
326 PJ_container_new(vm->cname(), INSTR_MSG_VM, PJ_container_get(sg_host_get_name(ms->dst_pm)));
329 msg = PJ_container_get(vm->name().c_str());
330 type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
331 new_pajeEndLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
334 // Inform the SRC that the migration has been correctly performed
336 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
337 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
338 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
339 // xbt_assert(ret == MSG_OK);
340 if(ret == MSG_HOST_FAILURE){
341 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
342 // has been correctly migrated on the DST node
343 // TODO What does it mean ? What should we do ?
344 MSG_task_destroy(task);
345 } else if(ret == MSG_TRANSFER_FAILURE){
346 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
347 MSG_task_destroy(task);
352 XBT_DEBUG("mig: rx_done");
356 static void start_dirty_page_tracking(msg_vm_t vm)
358 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
360 pimpl->dp_enabled = 1;
365 xbt_dict_cursor_t cursor = nullptr;
366 dirty_page_t dp = nullptr;
367 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
368 double remaining = MSG_task_get_flops_amount(dp->task);
369 dp->prev_clock = MSG_get_clock();
370 dp->prev_remaining = remaining;
372 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
376 static void stop_dirty_page_tracking(msg_vm_t vm)
378 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
381 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
383 double computed = dp->prev_remaining - remaining;
384 double duration = clock - dp->prev_clock;
386 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
387 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
392 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
394 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
398 xbt_dict_cursor_t cursor = nullptr;
399 dirty_page_t dp = nullptr;
400 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
401 double remaining = MSG_task_get_flops_amount(dp->task);
403 double clock = MSG_get_clock();
405 // total += calc_updated_pages(key, vm, dp, remaining, clock);
406 total += get_computed(key, vm, dp, remaining, clock);
408 dp->prev_remaining = remaining;
409 dp->prev_clock = clock;
412 total += pimpl->dp_updated_by_deleted_tasks;
414 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
415 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
417 pimpl->dp_updated_by_deleted_tasks = 0;
422 // TODO Is this code redundant with the information provided by
423 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
424 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
425 void MSG_host_add_task(msg_host_t host, msg_task_t task)
427 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
430 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
432 double remaining = MSG_task_get_flops_amount(task);
433 char *key = bprintf("%s-%p", task->name, task);
435 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
437 if (pimpl->dp_enabled) {
438 dp->prev_clock = MSG_get_clock();
439 dp->prev_remaining = remaining;
442 pimpl->dp_objs = xbt_dict_new();
443 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
444 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
445 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, pimpl->dp_enabled);
450 void MSG_host_del_task(msg_host_t host, msg_task_t task)
452 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
455 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
457 char *key = bprintf("%s-%p", task->name, task);
458 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
459 xbt_assert(dp->task == task);
461 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
462 * the information for the lookup_() function that will called soon. */
463 if (pimpl->dp_enabled) {
464 double remaining = MSG_task_get_flops_amount(task);
465 double clock = MSG_get_clock();
466 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
467 double updated = get_computed(key, host, dp, remaining, clock);
469 pimpl->dp_updated_by_deleted_tasks += updated;
472 xbt_dict_remove(pimpl->dp_objs, key);
475 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
479 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,
480 int stage, int stage2_round, double mig_speed, double timeout)
483 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
484 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
488 double clock_sta = MSG_get_clock();
492 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
494 ret = MSG_task_send(task, mbox);
500 } else if (ret == MSG_TIMEOUT) {
501 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
502 sent = size - remaining;
503 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
506 /* FIXME: why try-and-catch is used here? */
507 if(ret == MSG_HOST_FAILURE){
508 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
509 MSG_task_destroy(task);
510 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
511 }else if(ret == MSG_TRANSFER_FAILURE){
512 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
513 MSG_task_destroy(task);
514 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
517 double clock_end = MSG_get_clock();
518 double duration = clock_end - clock_sta;
519 double actual_speed = size / duration;
522 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
523 actual_speed, mig_speed);
525 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
531 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
533 double updated_size = computed * dp_rate;
534 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
535 if (updated_size > dp_cap) {
536 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
538 updated_size = dp_cap;
541 return (sg_size_t) updated_size;
544 static double send_stage1(struct migration_session* ms, sg_size_t ramsize, double mig_speed, double dp_rate,
547 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
548 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
549 sg_size_t remaining = ramsize;
550 double computed_total = 0;
552 while (remaining > 0) {
553 sg_size_t datasize = chunksize;
554 if (remaining < chunksize)
555 datasize = remaining;
557 remaining -= datasize;
558 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
559 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
560 computed_total += computed;
563 return computed_total;
566 static double get_threshold_value(double bandwidth, double max_downtime)
568 return max_downtime * bandwidth;
571 static int migration_tx_fun(int argc, char *argv[])
573 XBT_DEBUG("mig: tx_start");
575 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
576 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
578 s_vm_params_t params;
579 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
580 const sg_size_t ramsize = params.ramsize;
581 const sg_size_t devsize = params.devsize;
582 const int skip_stage1 = params.skip_stage1;
583 int skip_stage2 = params.skip_stage2;
584 const double dp_rate = params.dp_rate;
585 const double dp_cap = params.dp_cap;
586 const double mig_speed = params.mig_speed;
587 double max_downtime = params.max_downtime;
589 /* hard code it temporally. Fix Me */
590 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
591 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
593 double remaining_size = (double) (ramsize + devsize);
594 double threshold = 0.0;
596 /* check parameters */
598 XBT_WARN("migrate a VM, but ramsize is zero");
600 if (max_downtime <= 0) {
601 XBT_WARN("use the default max_downtime value 30ms");
605 /* Stage1: send all memory pages to the destination. */
606 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
607 start_dirty_page_tracking(ms->vm);
609 double computed_during_stage1 = 0;
611 double clock_prev_send = MSG_get_clock();
614 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
615 * transfer exceeds the timeout value. */
616 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
617 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
618 remaining_size -= sent;
619 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
621 if (sent < ramsize) {
622 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
624 } else if (sent > ramsize)
629 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
630 // Stop the dirty page tracking an return (there is no memory space to release)
631 stop_dirty_page_tracking(ms->vm);
635 double clock_post_send = MSG_get_clock();
636 mig_timeout -= (clock_post_send - clock_prev_send);
637 if (mig_timeout < 0) {
638 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
639 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
643 /* estimate bandwidth */
644 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
645 threshold = get_threshold_value(bandwidth, max_downtime);
646 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
650 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
653 int stage2_round = 0;
656 sg_size_t updated_size = 0;
657 if (stage2_round == 0) {
658 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
659 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
661 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
662 updated_size = get_updated_size(computed, dp_rate, dp_cap);
665 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
666 updated_size, computed_during_stage1, dp_rate, dp_cap);
668 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
669 remaining_size += updated_size;
670 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
671 (remaining_size < threshold) ? "<" : ">", threshold);
672 if (remaining_size < threshold)
676 double clock_prev_send = MSG_get_clock();
678 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
679 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
681 } catch (xbt_ex& e) {
682 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
684 // Stop the dirty page tracking an return (there is no memory space to release)
685 stop_dirty_page_tracking(ms->vm);
688 double clock_post_send = MSG_get_clock();
690 if (sent == updated_size) {
691 /* timeout did not happen */
692 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
693 threshold = get_threshold_value(bandwidth, max_downtime);
694 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
695 remaining_size -= sent;
697 mig_timeout -= (clock_post_send - clock_prev_send);
698 xbt_assert(mig_timeout > 0);
700 } else if (sent < updated_size) {
701 /* When timeout happens, we move to stage 3. The size of memory pages
702 * updated before timeout must be added to the remaining size. */
703 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
704 updated_size, (clock_post_send - clock_prev_send));
705 remaining_size -= sent;
707 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
708 updated_size = get_updated_size(computed, dp_rate, dp_cap);
709 remaining_size += updated_size;
716 /* Stage3: stop the VM and copy the rest of states. */
717 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
718 simcall_vm_suspend(ms->vm);
719 stop_dirty_page_tracking(ms->vm);
722 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
723 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
726 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
727 // Stop the dirty page tracking an return (there is no memory space to release)
728 simcall_vm_resume(ms->vm);
732 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
733 // effectively the VM on the DST node.
734 XBT_DEBUG("mig: tx_done");
739 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
741 struct migration_session *ms = xbt_new(struct migration_session, 1);
746 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
747 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
748 ms->mbox_ctl = bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
749 ms->mbox = bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
751 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
752 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
754 char** argv = xbt_new(char*, 2);
755 argv[0] = pr_rx_name;
757 MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
759 argv = xbt_new(char*, 2);
760 argv[0] = pr_tx_name;
762 MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
764 /* wait until the migration have finished or on error has occurred */
765 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
766 msg_task_t task = nullptr;
767 msg_error_t ret = MSG_TIMEOUT;
768 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) // Wait while you receive the message o
769 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
771 xbt_free(ms->mbox_ctl);
775 if (ret == MSG_HOST_FAILURE) {
776 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
777 // other side. Hence, just throw the execption
778 XBT_ERROR("SRC crashes, throw an exception (m-control)");
779 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
782 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
783 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
784 XBT_ERROR("DST crashes, throw an exception (m-control)");
788 char* expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
789 xbt_assert(strcmp(task->name, expected_task_name) == 0);
790 xbt_free(expected_task_name);
791 MSG_task_destroy(task);
795 /** @brief Migrate the VM to the given host.
798 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
799 * depending on whether you want to do cold or hot migration.
801 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
804 * - One approach is ...
805 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
806 * receive the state of the source
807 * VM over network. We will finally destroy the source VM.
808 * - This behavior is similar to the way of migration in the real world.
809 * Even before a migration is completed, we will see a destination VM, consuming resources.
810 * - We have to relocate all processes. The existing process migration code will work for this?
811 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
813 * - Another one is ...
814 * We update the information of the given VM to place it to the destination physical host.
816 * The second one would be easier.
819 simgrid::s4u::VirtualMachine* typedVm = static_cast<simgrid::s4u::VirtualMachine*>(vm);
820 simgrid::surf::VirtualMachineImpl* pimpl = typedVm->pimpl_vm_;
821 msg_host_t old_pm = pimpl->getPm();
824 THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->cname(), old_pm->cname());
827 THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->cname(), new_pm->cname());
829 if (!MSG_vm_is_running(vm))
830 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->cname());
832 if (typedVm->isMigrating())
833 THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->cname());
835 pimpl->isMigrating = true;
837 int ret = do_migration(vm, old_pm, new_pm);
839 pimpl->isMigrating = false;
840 THROWF(host_error, 0, "Source host '%s' failed during the migration of VM '%s'.", old_pm->cname(), vm->cname());
841 } else if (ret == -2) {
842 pimpl->isMigrating = false;
843 THROWF(host_error, 0, "Destination host '%s' failed during the migration of VM '%s'.", new_pm->cname(),
848 /** @brief Immediately suspend the execution of all processes within the given VM.
851 * This function stops the execution of the VM. All the processes on this VM
852 * will pause. The state of the VM is preserved. We can later resume it again.
854 * No suspension cost occurs.
856 void MSG_vm_suspend(msg_vm_t vm)
858 if (MSG_vm_is_migrating(vm))
859 THROWF(vm_error, 0, "Cannot suspend VM '%s', which is migrating", vm->cname());
861 simcall_vm_suspend(vm);
863 XBT_DEBUG("vm_suspend done");
865 if (TRACE_msg_vm_is_enabled()) {
866 container_t vm_container = PJ_container_get(vm->cname());
867 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
868 val_t value = PJ_value_get_or_new("suspend", "1 0 0", type); // suspend is red
869 new_pajePushState(MSG_get_clock(), vm_container, type, value);
873 /** @brief Resume the execution of the VM. All processes on the VM run again.
876 * No resume cost occurs.
878 void MSG_vm_resume(msg_vm_t vm)
880 simcall_vm_resume(vm);
882 if (TRACE_msg_vm_is_enabled()) {
883 container_t vm_container = PJ_container_get(vm->cname());
884 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
885 new_pajePopState(MSG_get_clock(), vm_container, type);
890 /** @brief Immediately save the execution of all processes within the given VM.
893 * This function stops the execution of the VM. All the processes on this VM
894 * will pause. The state of the VM is preserved. We can later resume it again.
896 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
897 * or after, depending on the exact semantic of VM save to you.
899 void MSG_vm_save(msg_vm_t vm)
901 if (MSG_vm_is_migrating(vm))
902 THROWF(vm_error, 0, "Cannot save VM '%s', which is migrating.", vm->cname());
906 if (TRACE_msg_vm_is_enabled()) {
907 container_t vm_container = PJ_container_get(vm->cname());
908 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
909 val_t value = PJ_value_get_or_new("save", "0 1 0", type); // save is green
910 new_pajePushState(MSG_get_clock(), vm_container, type, value);
914 /** @brief Restore the execution of the VM. All processes on the VM run again.
917 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
918 * after, depending on the exact semantic of VM restore to you.
920 void MSG_vm_restore(msg_vm_t vm)
922 simcall_vm_restore(vm);
924 if (TRACE_msg_vm_is_enabled()) {
925 container_t vm_container = PJ_container_get(vm->cname());
926 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
927 new_pajePopState(MSG_get_clock(), vm_container, type);
931 /** @brief Get the physical host of a given VM.
934 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
936 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getPm();
939 /** @brief Set a CPU bound for a given VM.
942 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
945 * On PM0, there are Task1 and VM0.
946 * On VM0, there is Task2.
947 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
949 * Task1\@PM0 gets 50%.
950 * Task2\@VM0 gets 25%.
951 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
953 * 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
954 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
956 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
957 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
958 * on the VM in share_resource() or we have to trap both the start and end of task execution.
960 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
962 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
963 * computation task (or a VM) never exceeds the capacity of a CPU core.
965 void MSG_vm_set_bound(msg_vm_t vm, double bound)
967 simcall_vm_set_bound(vm, bound);