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 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
227 return bprintf("__pr_mig_tx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
230 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
232 return bprintf("__pr_mig_rx:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
235 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
237 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm->cname(), src_pm->cname(), dst_pm->cname());
240 struct migration_session {
245 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
246 * know the completion of the migration. */
248 /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
252 static int migration_rx_fun(int argc, char *argv[])
254 XBT_DEBUG("mig: rx_start");
256 // The structure has been created in the do_migration function and should only be freed in the same place ;)
257 struct migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
259 bool received_finalize = false;
261 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
262 while (!received_finalize) {
263 msg_task_t task = nullptr;
264 int ret = MSG_task_recv(&task, ms->mbox);
267 // An error occurred, clean the code and return
268 // The owner did not change, hence the task should be only destroyed on the other side
269 xbt_free(finalize_task_name);
273 if (strcmp(task->name, finalize_task_name) == 0)
274 received_finalize = 1;
276 MSG_task_destroy(task);
278 xbt_free(finalize_task_name);
280 // Here Stage 1, 2 and 3 have been performed.
281 // Hence complete the migration
283 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
284 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms
285 // structure at the beginning and free it at the end of each function)
286 simgrid::s4u::VirtualMachine* vm = static_cast<simgrid::s4u::VirtualMachine*>(ms->vm);
287 msg_host_t src_pm = ms->src_pm;
288 msg_host_t dst_pm = ms->dst_pm;
290 // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state
291 simgrid::simix::kernelImmediate([vm, src_pm, dst_pm]() {
292 /* Update the vm location */
293 /* precopy migration makes the VM temporally paused */
294 xbt_assert(static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED);
296 /* jump to vm_ws_xigrate(). this will update the vm location. */
297 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->migrate(dst_pm);
304 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
305 vm->pimpl_vm_->isMigrating = false;
306 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm),
307 sg_host_get_name(ms->dst_pm));
309 if (TRACE_msg_vm_is_enabled()) {
310 static long long int counter = 0;
311 char key[INSTR_DEFAULT_STR_SIZE];
312 snprintf(key, INSTR_DEFAULT_STR_SIZE, "%lld", counter++);
315 container_t msg = PJ_container_get(vm->name().c_str());
316 type_t type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
317 new_pajeStartLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
319 // destroy existing container of this vm
320 container_t existing_container = PJ_container_get(vm->name().c_str());
321 PJ_container_remove_from_parent(existing_container);
322 PJ_container_free(existing_container);
324 // create new container on the new_host location
325 PJ_container_new(vm->name().c_str(), INSTR_MSG_VM, PJ_container_get(sg_host_get_name(ms->dst_pm)));
328 msg = PJ_container_get(vm->name().c_str());
329 type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root());
330 new_pajeEndLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key);
333 // Inform the SRC that the migration has been correctly performed
335 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
336 msg_task_t task = MSG_task_create(task_name, 0, 0, nullptr);
337 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
338 // xbt_assert(ret == MSG_OK);
339 if(ret == MSG_HOST_FAILURE){
340 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM
341 // has been correctly migrated on the DST node
342 // TODO What does it mean ? What should we do ?
343 MSG_task_destroy(task);
344 } else if(ret == MSG_TRANSFER_FAILURE){
345 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
346 MSG_task_destroy(task);
351 XBT_DEBUG("mig: rx_done");
355 static void start_dirty_page_tracking(msg_vm_t vm)
357 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
359 pimpl->dp_enabled = 1;
364 xbt_dict_cursor_t cursor = nullptr;
365 dirty_page_t dp = nullptr;
366 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
367 double remaining = MSG_task_get_flops_amount(dp->task);
368 dp->prev_clock = MSG_get_clock();
369 dp->prev_remaining = remaining;
371 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
375 static void stop_dirty_page_tracking(msg_vm_t vm)
377 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
380 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
382 double computed = dp->prev_remaining - remaining;
383 double duration = clock - dp->prev_clock;
385 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
386 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
391 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
393 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
397 xbt_dict_cursor_t cursor = nullptr;
398 dirty_page_t dp = nullptr;
399 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
400 double remaining = MSG_task_get_flops_amount(dp->task);
402 double clock = MSG_get_clock();
404 // total += calc_updated_pages(key, vm, dp, remaining, clock);
405 total += get_computed(key, vm, dp, remaining, clock);
407 dp->prev_remaining = remaining;
408 dp->prev_clock = clock;
411 total += pimpl->dp_updated_by_deleted_tasks;
413 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
414 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
416 pimpl->dp_updated_by_deleted_tasks = 0;
421 // TODO Is this code redundant with the information provided by
422 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
423 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
424 void MSG_host_add_task(msg_host_t host, msg_task_t task)
426 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
429 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
431 double remaining = MSG_task_get_flops_amount(task);
432 char *key = bprintf("%s-%p", task->name, task);
434 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
436 if (pimpl->dp_enabled) {
437 dp->prev_clock = MSG_get_clock();
438 dp->prev_remaining = remaining;
441 pimpl->dp_objs = xbt_dict_new();
442 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
443 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
444 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, pimpl->dp_enabled);
449 void MSG_host_del_task(msg_host_t host, msg_task_t task)
451 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
454 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
456 char *key = bprintf("%s-%p", task->name, task);
457 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
458 xbt_assert(dp->task == task);
460 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
461 * the information for the lookup_() function that will called soon. */
462 if (pimpl->dp_enabled) {
463 double remaining = MSG_task_get_flops_amount(task);
464 double clock = MSG_get_clock();
465 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
466 double updated = get_computed(key, host, dp, remaining, clock);
468 pimpl->dp_updated_by_deleted_tasks += updated;
471 xbt_dict_remove(pimpl->dp_objs, key);
474 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
478 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,
479 int stage, int stage2_round, double mig_speed, double timeout)
482 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
483 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
487 double clock_sta = MSG_get_clock();
491 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
493 ret = MSG_task_send(task, mbox);
499 } else if (ret == MSG_TIMEOUT) {
500 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
501 sent = size - remaining;
502 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
505 /* FIXME: why try-and-catch is used here? */
506 if(ret == MSG_HOST_FAILURE){
507 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
508 MSG_task_destroy(task);
509 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
510 }else if(ret == MSG_TRANSFER_FAILURE){
511 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
512 MSG_task_destroy(task);
513 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
516 double clock_end = MSG_get_clock();
517 double duration = clock_end - clock_sta;
518 double actual_speed = size / duration;
521 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
522 actual_speed, mig_speed);
524 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
530 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
532 double updated_size = computed * dp_rate;
533 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
534 if (updated_size > dp_cap) {
535 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
537 updated_size = dp_cap;
540 return (sg_size_t) updated_size;
543 static double send_stage1(struct migration_session* ms, sg_size_t ramsize, double mig_speed, double dp_rate,
546 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
547 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
548 sg_size_t remaining = ramsize;
549 double computed_total = 0;
551 while (remaining > 0) {
552 sg_size_t datasize = chunksize;
553 if (remaining < chunksize)
554 datasize = remaining;
556 remaining -= datasize;
557 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
558 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
559 computed_total += computed;
562 return computed_total;
565 static double get_threshold_value(double bandwidth, double max_downtime)
567 return max_downtime * bandwidth;
570 static int migration_tx_fun(int argc, char *argv[])
572 XBT_DEBUG("mig: tx_start");
574 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
575 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
577 s_vm_params_t params;
578 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
579 const sg_size_t ramsize = params.ramsize;
580 const sg_size_t devsize = params.devsize;
581 const int skip_stage1 = params.skip_stage1;
582 int skip_stage2 = params.skip_stage2;
583 const double dp_rate = params.dp_rate;
584 const double dp_cap = params.dp_cap;
585 const double mig_speed = params.mig_speed;
586 double max_downtime = params.max_downtime;
588 /* hard code it temporally. Fix Me */
589 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
590 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
592 double remaining_size = (double) (ramsize + devsize);
593 double threshold = 0.0;
595 /* check parameters */
597 XBT_WARN("migrate a VM, but ramsize is zero");
599 if (max_downtime <= 0) {
600 XBT_WARN("use the default max_downtime value 30ms");
604 /* Stage1: send all memory pages to the destination. */
605 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
606 start_dirty_page_tracking(ms->vm);
608 double computed_during_stage1 = 0;
610 double clock_prev_send = MSG_get_clock();
613 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
614 * transfer exceeds the timeout value. */
615 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
616 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
617 remaining_size -= sent;
618 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
620 if (sent < ramsize) {
621 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
623 } else if (sent > ramsize)
628 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
629 // Stop the dirty page tracking an return (there is no memory space to release)
630 stop_dirty_page_tracking(ms->vm);
634 double clock_post_send = MSG_get_clock();
635 mig_timeout -= (clock_post_send - clock_prev_send);
636 if (mig_timeout < 0) {
637 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
638 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
642 /* estimate bandwidth */
643 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
644 threshold = get_threshold_value(bandwidth, max_downtime);
645 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
649 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
652 int stage2_round = 0;
655 sg_size_t updated_size = 0;
656 if (stage2_round == 0) {
657 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
658 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
660 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
661 updated_size = get_updated_size(computed, dp_rate, dp_cap);
664 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
665 updated_size, computed_during_stage1, dp_rate, dp_cap);
667 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
668 remaining_size += updated_size;
669 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
670 (remaining_size < threshold) ? "<" : ">", threshold);
671 if (remaining_size < threshold)
675 double clock_prev_send = MSG_get_clock();
677 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
678 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
680 } catch (xbt_ex& e) {
681 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
683 // Stop the dirty page tracking an return (there is no memory space to release)
684 stop_dirty_page_tracking(ms->vm);
687 double clock_post_send = MSG_get_clock();
689 if (sent == updated_size) {
690 /* timeout did not happen */
691 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
692 threshold = get_threshold_value(bandwidth, max_downtime);
693 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
694 remaining_size -= sent;
696 mig_timeout -= (clock_post_send - clock_prev_send);
697 xbt_assert(mig_timeout > 0);
699 } else if (sent < updated_size) {
700 /* When timeout happens, we move to stage 3. The size of memory pages
701 * updated before timeout must be added to the remaining size. */
702 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
703 updated_size, (clock_post_send - clock_prev_send));
704 remaining_size -= sent;
706 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
707 updated_size = get_updated_size(computed, dp_rate, dp_cap);
708 remaining_size += updated_size;
715 /* Stage3: stop the VM and copy the rest of states. */
716 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
717 simcall_vm_suspend(ms->vm);
718 stop_dirty_page_tracking(ms->vm);
721 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
722 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
725 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
726 // Stop the dirty page tracking an return (there is no memory space to release)
727 simcall_vm_resume(ms->vm);
731 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
732 // effectively the VM on the DST node.
733 XBT_DEBUG("mig: tx_done");
738 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
740 struct migration_session *ms = xbt_new(struct migration_session, 1);
745 /* We have two mailboxes. mbox is used to transfer migration data between source and destination PMs. mbox_ctl is used
746 * to detect the completion of a migration. The names of these mailboxes must not conflict with others. */
747 ms->mbox_ctl = bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
748 ms->mbox = bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname());
750 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
751 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
753 char** argv = xbt_new(char*, 2);
754 argv[0] = pr_rx_name;
756 MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
758 argv = xbt_new(char*, 2);
759 argv[0] = pr_tx_name;
761 MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
763 /* wait until the migration have finished or on error has occurred */
764 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
765 msg_task_t task = nullptr;
766 msg_error_t ret = MSG_TIMEOUT;
767 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) // Wait while you receive the message o
768 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
770 xbt_free(ms->mbox_ctl);
774 if (ret == MSG_HOST_FAILURE) {
775 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
776 // other side. Hence, just throw the execption
777 XBT_ERROR("SRC crashes, throw an exception (m-control)");
778 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
781 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
782 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
783 XBT_ERROR("DST crashes, throw an exception (m-control)");
787 char* expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
788 xbt_assert(strcmp(task->name, expected_task_name) == 0);
789 xbt_free(expected_task_name);
790 MSG_task_destroy(task);
794 /** @brief Migrate the VM to the given host.
797 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
798 * depending on whether you want to do cold or hot migration.
800 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
803 * - One approach is ...
804 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
805 * receive the state of the source
806 * VM over network. We will finally destroy the source VM.
807 * - This behavior is similar to the way of migration in the real world.
808 * Even before a migration is completed, we will see a destination VM, consuming resources.
809 * - We have to relocate all processes. The existing process migration code will work for this?
810 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
812 * - Another one is ...
813 * We update the information of the given VM to place it to the destination physical host.
815 * The second one would be easier.
818 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
819 msg_host_t old_pm = pimpl->getPm();
822 THROWF(vm_error, 0, "Cannot start a migration from host '%s', which is offline.", old_pm->cname());
825 THROWF(vm_error, 0, "Cannot start a migration to host '%s', which is offline.", new_pm->cname());
827 if (!MSG_vm_is_running(vm))
828 THROWF(vm_error, 0, "VM(%s) is not running", vm->cname());
830 if (MSG_vm_is_migrating(vm))
831 THROWF(vm_error, 0, "VM(%s) is already migrating", vm->cname());
833 pimpl->isMigrating = 1;
835 int ret = do_migration(vm, old_pm, new_pm);
837 pimpl->isMigrating = 0;
838 THROWF(host_error, 0, "SRC host failed during migration");
839 } else if (ret == -2) {
840 pimpl->isMigrating = 0;
841 THROWF(host_error, 0, "DST host failed during migration");
845 /** @brief Immediately suspend the execution of all processes within the given VM.
848 * This function stops the execution of the VM. All the processes on this VM
849 * will pause. The state of the VM is preserved. We can later resume it again.
851 * No suspension cost occurs.
853 void MSG_vm_suspend(msg_vm_t vm)
855 if (MSG_vm_is_migrating(vm))
856 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
858 simcall_vm_suspend(vm);
860 XBT_DEBUG("vm_suspend done");
862 if (TRACE_msg_vm_is_enabled()) {
863 container_t vm_container = PJ_container_get(vm->name().c_str());
864 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
865 val_t value = PJ_value_get_or_new("suspend", "1 0 0", type); // suspend is red
866 new_pajePushState(MSG_get_clock(), vm_container, type, value);
870 /** @brief Resume the execution of the VM. All processes on the VM run again.
873 * No resume cost occurs.
875 void MSG_vm_resume(msg_vm_t vm)
877 simcall_vm_resume(vm);
879 if (TRACE_msg_vm_is_enabled()) {
880 container_t vm_container = PJ_container_get(vm->name().c_str());
881 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
882 new_pajePopState(MSG_get_clock(), vm_container, type);
887 /** @brief Immediately save the execution of all processes within the given VM.
890 * This function stops the execution of the VM. All the processes on this VM
891 * will pause. The state of the VM is preserved. We can later resume it again.
893 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
894 * or after, depending on the exact semantic of VM save to you.
896 void MSG_vm_save(msg_vm_t vm)
898 if (MSG_vm_is_migrating(vm))
899 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
903 if (TRACE_msg_vm_is_enabled()) {
904 container_t vm_container = PJ_container_get(vm->name().c_str());
905 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
906 val_t value = PJ_value_get_or_new("save", "0 1 0", type); // save is green
907 new_pajePushState(MSG_get_clock(), vm_container, type, value);
911 /** @brief Restore the execution of the VM. All processes on the VM run again.
914 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
915 * after, depending on the exact semantic of VM restore to you.
917 void MSG_vm_restore(msg_vm_t vm)
919 simcall_vm_restore(vm);
921 if (TRACE_msg_vm_is_enabled()) {
922 container_t vm_container = PJ_container_get(vm->name().c_str());
923 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
924 new_pajePopState(MSG_get_clock(), vm_container, type);
928 /** @brief Get the physical host of a given VM.
931 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
933 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getPm();
936 /** @brief Set a CPU bound for a given VM.
939 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
942 * On PM0, there are Task1 and VM0.
943 * On VM0, there is Task2.
944 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
946 * Task1\@PM0 gets 50%.
947 * Task2\@VM0 gets 25%.
948 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
950 * 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
951 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
953 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
954 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
955 * on the VM in share_resource() or we have to trap both the start and end of task execution.
957 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
959 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
960 * computation task (or a VM) never exceeds the capacity of a CPU core.
962 void MSG_vm_set_bound(msg_vm_t vm, double bound)
964 simcall_vm_set_bound(vm, bound);