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);
364 XBT_DEBUG("mig: rx_done");
368 static void reset_dirty_pages(msg_vm_t vm)
370 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
373 xbt_dict_cursor_t cursor = nullptr;
374 dirty_page_t dp = nullptr;
377 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
378 double remaining = MSG_task_get_flops_amount(dp->task);
379 dp->prev_clock = MSG_get_clock();
380 dp->prev_remaining = remaining;
382 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
386 static void start_dirty_page_tracking(msg_vm_t vm)
388 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 1;
390 reset_dirty_pages(vm);
393 static void stop_dirty_page_tracking(msg_vm_t vm)
395 static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->dp_enabled = 0;
398 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
400 double computed = dp->prev_remaining - remaining;
401 double duration = clock - dp->prev_clock;
403 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
404 key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
409 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
411 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
415 xbt_dict_cursor_t cursor = nullptr;
416 dirty_page_t dp = nullptr;
417 xbt_dict_foreach (pimpl->dp_objs, cursor, key, dp) {
418 double remaining = MSG_task_get_flops_amount(dp->task);
420 double clock = MSG_get_clock();
422 // total += calc_updated_pages(key, vm, dp, remaining, clock);
423 total += get_computed(key, vm, dp, remaining, clock);
425 dp->prev_remaining = remaining;
426 dp->prev_clock = clock;
429 total += pimpl->dp_updated_by_deleted_tasks;
431 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
432 stage2_round_for_fancy_debug, total, pimpl->dp_updated_by_deleted_tasks);
434 pimpl->dp_updated_by_deleted_tasks = 0;
439 // TODO Is this code redundant with the information provided by
440 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
441 /** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
442 void MSG_host_add_task(msg_host_t host, msg_task_t task)
444 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
447 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
449 double remaining = MSG_task_get_flops_amount(task);
450 char *key = bprintf("%s-%p", task->name, task);
452 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
454 if (pimpl->dp_enabled) {
455 dp->prev_clock = MSG_get_clock();
456 dp->prev_remaining = remaining;
459 pimpl->dp_objs = xbt_dict_new();
460 xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr);
461 xbt_dict_set(pimpl->dp_objs, key, dp, nullptr);
462 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, pimpl->dp_enabled);
467 void MSG_host_del_task(msg_host_t host, msg_task_t task)
469 simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
472 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
474 char *key = bprintf("%s-%p", task->name, task);
475 dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL);
476 xbt_assert(dp->task == task);
478 /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep
479 * the information for the lookup_() function that will called soon. */
480 if (pimpl->dp_enabled) {
481 double remaining = MSG_task_get_flops_amount(task);
482 double clock = MSG_get_clock();
483 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
484 double updated = get_computed(key, host, dp, remaining, clock);
486 pimpl->dp_updated_by_deleted_tasks += updated;
489 xbt_dict_remove(pimpl->dp_objs, key);
492 XBT_DEBUG("del %s on %s", key, sg_host_get_name(host));
496 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,
497 int stage, int stage2_round, double mig_speed, double timeout)
500 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
501 msg_task_t task = MSG_task_create(task_name, 0, (double)size, nullptr);
505 double clock_sta = MSG_get_clock();
509 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
511 ret = MSG_task_send(task, mbox);
517 } else if (ret == MSG_TIMEOUT) {
518 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
519 sent = size - remaining;
520 XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
523 /* FIXME: why try-and-catch is used here? */
524 if(ret == MSG_HOST_FAILURE){
525 //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
526 MSG_task_destroy(task);
527 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
528 }else if(ret == MSG_TRANSFER_FAILURE){
529 //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
530 MSG_task_destroy(task);
531 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage);
534 double clock_end = MSG_get_clock();
535 double duration = clock_end - clock_sta;
536 double actual_speed = size / duration;
539 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration,
540 actual_speed, mig_speed);
542 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed,
548 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
550 double updated_size = computed * dp_rate;
551 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
552 if (updated_size > dp_cap) {
553 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size,
555 updated_size = dp_cap;
558 return (sg_size_t) updated_size;
561 static double send_stage1(struct migration_session* ms, sg_size_t ramsize, double mig_speed, double dp_rate,
564 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
565 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
566 sg_size_t remaining = ramsize;
567 double computed_total = 0;
569 while (remaining > 0) {
570 sg_size_t datasize = chunksize;
571 if (remaining < chunksize)
572 datasize = remaining;
574 remaining -= datasize;
575 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
576 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
577 computed_total += computed;
580 return computed_total;
583 static double get_threshold_value(double bandwidth, double max_downtime)
585 return max_downtime * bandwidth;
588 static int migration_tx_fun(int argc, char *argv[])
590 XBT_DEBUG("mig: tx_start");
592 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
593 migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self());
595 s_vm_params_t params;
596 static_cast<simgrid::s4u::VirtualMachine*>(ms->vm)->parameters(¶ms);
597 const sg_size_t ramsize = params.ramsize;
598 const sg_size_t devsize = params.devsize;
599 const int skip_stage1 = params.skip_stage1;
600 int skip_stage2 = params.skip_stage2;
601 const double dp_rate = params.dp_rate;
602 const double dp_cap = params.dp_cap;
603 const double mig_speed = params.mig_speed;
604 double max_downtime = params.max_downtime;
606 /* hard code it temporally. Fix Me */
607 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
608 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
610 double remaining_size = (double) (ramsize + devsize);
611 double threshold = 0.0;
613 /* check parameters */
615 XBT_WARN("migrate a VM, but ramsize is zero");
617 if (max_downtime <= 0) {
618 XBT_WARN("use the default max_downtime value 30ms");
622 /* Stage1: send all memory pages to the destination. */
623 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
624 start_dirty_page_tracking(ms->vm);
626 double computed_during_stage1 = 0;
628 double clock_prev_send = MSG_get_clock();
631 /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this
632 * transfer exceeds the timeout value. */
633 XBT_VERB("Stage 1: Gonna send %llu", ramsize);
634 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
635 remaining_size -= sent;
636 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
638 if (sent < ramsize) {
639 XBT_VERB("mig-stage1: timeout, force moving to stage 3");
641 } else if (sent > ramsize)
646 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
647 // Stop the dirty page tracking an return (there is no memory space to release)
648 stop_dirty_page_tracking(ms->vm);
652 double clock_post_send = MSG_get_clock();
653 mig_timeout -= (clock_post_send - clock_prev_send);
654 if (mig_timeout < 0) {
655 XBT_VERB("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
656 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
660 /* estimate bandwidth */
661 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
662 threshold = get_threshold_value(bandwidth, max_downtime);
663 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
667 /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */
670 int stage2_round = 0;
673 sg_size_t updated_size = 0;
674 if (stage2_round == 0) {
675 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
676 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
678 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
679 updated_size = get_updated_size(computed, dp_rate, dp_cap);
682 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round,
683 updated_size, computed_during_stage1, dp_rate, dp_cap);
685 /* Check whether the remaining size is below the threshold value. If so, move to stage 3. */
686 remaining_size += updated_size;
687 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, remaining_size,
688 (remaining_size < threshold) ? "<" : ">", threshold);
689 if (remaining_size < threshold)
693 double clock_prev_send = MSG_get_clock();
695 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
696 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed,
698 } catch (xbt_ex& e) {
699 // hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data
701 // Stop the dirty page tracking an return (there is no memory space to release)
702 stop_dirty_page_tracking(ms->vm);
705 double clock_post_send = MSG_get_clock();
707 if (sent == updated_size) {
708 /* timeout did not happen */
709 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
710 threshold = get_threshold_value(bandwidth, max_downtime);
711 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
712 remaining_size -= sent;
714 mig_timeout -= (clock_post_send - clock_prev_send);
715 xbt_assert(mig_timeout > 0);
717 } else if (sent < updated_size) {
718 /* When timeout happens, we move to stage 3. The size of memory pages
719 * updated before timeout must be added to the remaining size. */
720 XBT_VERB("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", stage2_round, sent,
721 updated_size, (clock_post_send - clock_prev_send));
722 remaining_size -= sent;
724 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
725 updated_size = get_updated_size(computed, dp_rate, dp_cap);
726 remaining_size += updated_size;
733 /* Stage3: stop the VM and copy the rest of states. */
734 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
735 simcall_vm_suspend(ms->vm);
736 stop_dirty_page_tracking(ms->vm);
739 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
740 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
743 //hostfailure (if you want to know whether this is the SRC or the DST check directly in send_migration_data code)
744 // Stop the dirty page tracking an return (there is no memory space to release)
745 simcall_vm_resume(ms->vm);
749 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate
750 // effectively the VM on the DST node.
751 XBT_DEBUG("mig: tx_done");
756 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
758 struct migration_session *ms = xbt_new(struct migration_session, 1);
762 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
763 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
765 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
766 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
768 char** argv = xbt_new(char*, 2);
769 argv[0] = pr_rx_name;
771 MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
773 argv = xbt_new(char*, 2);
774 argv[0] = pr_tx_name;
776 MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
778 /* wait until the migration have finished or on error has occurred */
779 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
780 msg_task_t task = nullptr;
781 msg_error_t ret = MSG_TIMEOUT;
782 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) // Wait while you receive the message o
783 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
785 xbt_free(ms->mbox_ctl);
789 // xbt_assert(ret == MSG_OK);
790 if (ret == MSG_HOST_FAILURE) {
791 // Note that since the communication failed, the owner did not change and the task should be destroyed on the
792 // other side. Hence, just throw the execption
793 XBT_ERROR("SRC crashes, throw an exception (m-control)");
794 // MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is
797 } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) {
798 // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
799 XBT_ERROR("DST crashes, throw an exception (m-control)");
803 char* expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
804 xbt_assert(strcmp(task->name, expected_task_name) == 0);
805 xbt_free(expected_task_name);
806 MSG_task_destroy(task);
810 /** @brief Migrate the VM to the given host.
813 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after,
814 * depending on whether you want to do cold or hot migration.
816 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
819 * - One approach is ...
820 * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will
821 * receive the state of the source
822 * VM over network. We will finally destroy the source VM.
823 * - This behavior is similar to the way of migration in the real world.
824 * Even before a migration is completed, we will see a destination VM, consuming resources.
825 * - We have to relocate all processes. The existing process migration code will work for this?
826 * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM?
828 * - Another one is ...
829 * We update the information of the given VM to place it to the destination physical host.
831 * The second one would be easier.
834 simgrid::surf::VirtualMachineImpl* pimpl = static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_;
835 msg_host_t old_pm = pimpl->getPm();
838 THROWF(vm_error, 0, "Cannot start a migration from host '%s', which is offline.", sg_host_get_name(old_pm));
841 THROWF(vm_error, 0, "Cannot start a migration to host '%s', which is offline.", sg_host_get_name(new_pm));
843 if (!MSG_vm_is_running(vm))
844 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm));
846 if (MSG_vm_is_migrating(vm))
847 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm));
849 pimpl->isMigrating = 1;
851 int ret = do_migration(vm, old_pm, new_pm);
853 pimpl->isMigrating = 0;
854 THROWF(host_error, 0, "SRC host failed during migration");
855 } else if (ret == -2) {
856 pimpl->isMigrating = 0;
857 THROWF(host_error, 0, "DST host failed during migration");
861 /** @brief Immediately suspend the execution of all processes within the given VM.
864 * This function stops the execution of the VM. All the processes on this VM
865 * will pause. The state of the VM is preserved. We can later resume it again.
867 * No suspension cost occurs.
869 void MSG_vm_suspend(msg_vm_t vm)
871 if (MSG_vm_is_migrating(vm))
872 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
874 simcall_vm_suspend(vm);
876 XBT_DEBUG("vm_suspend done");
878 if (TRACE_msg_vm_is_enabled()) {
879 container_t vm_container = PJ_container_get(vm->name().c_str());
880 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
881 val_t value = PJ_value_get_or_new("suspend", "1 0 0", type); // suspend is red
882 new_pajePushState(MSG_get_clock(), vm_container, type, value);
886 /** @brief Resume the execution of the VM. All processes on the VM run again.
889 * No resume cost occurs.
891 void MSG_vm_resume(msg_vm_t vm)
893 simcall_vm_resume(vm);
895 if (TRACE_msg_vm_is_enabled()) {
896 container_t vm_container = PJ_container_get(vm->name().c_str());
897 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
898 new_pajePopState(MSG_get_clock(), vm_container, type);
903 /** @brief Immediately save the execution of all processes within the given VM.
906 * This function stops the execution of the VM. All the processes on this VM
907 * will pause. The state of the VM is preserved. We can later resume it again.
909 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_write() before
910 * or after, depending on the exact semantic of VM save to you.
912 void MSG_vm_save(msg_vm_t vm)
914 if (MSG_vm_is_migrating(vm))
915 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm));
919 if (TRACE_msg_vm_is_enabled()) {
920 container_t vm_container = PJ_container_get(vm->name().c_str());
921 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
922 val_t value = PJ_value_get_or_new("save", "0 1 0", type); // save is green
923 new_pajePushState(MSG_get_clock(), vm_container, type, value);
927 /** @brief Restore the execution of the VM. All processes on the VM run again.
930 * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or
931 * after, depending on the exact semantic of VM restore to you.
933 void MSG_vm_restore(msg_vm_t vm)
935 simcall_vm_restore(vm);
937 if (TRACE_msg_vm_is_enabled()) {
938 container_t vm_container = PJ_container_get(vm->name().c_str());
939 type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type);
940 new_pajePopState(MSG_get_clock(), vm_container, type);
944 /** @brief Get the physical host of a given VM.
947 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
949 return static_cast<simgrid::s4u::VirtualMachine*>(vm)->pimpl_vm_->getPm();
952 /** @brief Set a CPU bound for a given VM.
955 * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
958 * On PM0, there are Task1 and VM0.
959 * On VM0, there is Task2.
960 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
962 * Task1\@PM0 gets 50%.
963 * Task2\@VM0 gets 25%.
964 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
966 * 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
967 * the dummy CPU action. The bound of the dummy CPU action is unlimited.
969 * There are some solutions for this problem. One option is to update the bound of the dummy CPU action automatically.
970 * It should be the sum of all tasks on the VM. But, this solution might be costly, because we have to scan all tasks
971 * on the VM in share_resource() or we have to trap both the start and end of task execution.
973 * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action.
975 * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a
976 * computation task (or a VM) never exceeds the capacity of a CPU core.
978 void MSG_vm_set_bound(msg_vm_t vm, double bound)
980 simcall_vm_set_bound(vm, bound);