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 very
10 * experimental code generating micro computation tasks
15 #include "msg_private.h"
16 #include "xbt/sysdep.h"
18 #include "simgrid/platf.h"
19 #include "simgrid/host.h"
21 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
22 "Cloud-oriented parts of the MSG API");
25 /* **** ******** GENERAL ********* **** */
27 /** \ingroup m_vm_management
28 * \brief Returns the value of a given vm property
31 * \param name a property name
32 * \return value of a property (or NULL if property not set)
35 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
37 return MSG_host_get_property_value(vm, name);
40 /** \ingroup m_vm_management
41 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
44 * \return a dict containing the properties
46 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
48 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
50 return (simcall_host_get_properties(vm));
53 /** \ingroup m_host_management
54 * \brief Change the value of a given host property
57 * \param name a property name
58 * \param value what to change the property to
59 * \param free_ctn the freeing function to use to kill the value on need
61 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
63 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
66 /** \ingroup msg_vm_management
67 * \brief Finds a msg_vm_t using its name.
69 * This is a name directory service
70 * \param name the name of a vm.
71 * \return the corresponding vm
73 * Please note that a VM is a specific host. Hence, you should give a different name
77 msg_vm_t MSG_vm_get_by_name(const char *name)
79 return MSG_host_by_name(name);
82 /** \ingroup m_vm_management
84 * \brief Return the name of the #msg_host_t.
86 * This functions checks whether \a host is a valid pointer or not and return
89 const char *MSG_vm_get_name(msg_vm_t vm)
91 return MSG_host_get_name(vm);
95 /* **** Check state of a VM **** */
96 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
98 return simcall_vm_get_state(vm) == state;
101 /** @brief Returns whether the given VM has just created, not running.
104 int MSG_vm_is_created(msg_vm_t vm)
106 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
109 /** @brief Returns whether the given VM is currently running
112 int MSG_vm_is_running(msg_vm_t vm)
114 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
117 /** @brief Returns whether the given VM is currently migrating
120 int MSG_vm_is_migrating(msg_vm_t vm)
122 msg_host_priv_t priv = sg_host_msg(vm);
123 return priv->is_migrating;
126 /** @brief Returns whether the given VM is currently suspended, not running.
129 int MSG_vm_is_suspended(msg_vm_t vm)
131 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
134 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
137 int MSG_vm_is_saving(msg_vm_t vm)
139 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
142 /** @brief Returns whether the given VM has been saved, not running.
145 int MSG_vm_is_saved(msg_vm_t vm)
147 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
150 /** @brief Returns whether the given VM is being restored, not running.
153 int MSG_vm_is_restoring(msg_vm_t vm)
155 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
160 /* ------------------------------------------------------------------------- */
161 /* ------------------------------------------------------------------------- */
163 /* **** ******** MSG vm actions ********* **** */
165 /** @brief Create a new VM with specified parameters.
167 * All parameters are in MBytes
170 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name,
171 int ncpus, int ramsize,
172 int net_cap, char *disk_path, int disksize,
173 int mig_netspeed, int dp_intensity)
175 /* For the moment, intensity_rate is the percentage against the migration
177 double host_speed = MSG_get_host_speed(ind_pm);
178 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
180 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
181 s_vm_params_t params;
182 memset(¶ms, 0, sizeof(params));
183 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
184 //params.overcommit = 0;
186 params.skip_stage2 = 0;
187 params.max_downtime = 0.03;
188 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
189 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
190 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
192 //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate, params.mig_speed, dp_intensity, update_speed, host_speed);
193 simcall_host_set_params(vm, ¶ms);
199 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
202 * A VM is treated as a host. The name of the VM must be unique among all hosts.
204 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
206 /* make sure the VM of the same name does not exit */
208 xbt_dictelm_t ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
209 if (ind_host_tmp && sg_host_simix(ind_host_tmp) != NULL) {
210 XBT_ERROR("host %s already exits", name);
215 /* Note: ind_vm and vm_workstation point to the same elm object. */
216 msg_vm_t ind_vm = NULL;
217 void *ind_vm_workstation = NULL;
219 /* Ask the SIMIX layer to create the surf vm resource */
220 ind_vm_workstation = simcall_vm_create(name, ind_pm);
221 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
223 XBT_DEBUG("A new VM (%s) has been created", name);
225 TRACE_msg_vm_create(name, ind_pm);
230 /** @brief Destroy a VM. Destroy the VM object from the simulation.
233 void MSG_vm_destroy(msg_vm_t vm)
235 if (MSG_vm_is_migrating(vm))
236 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
238 /* First, terminate all processes on the VM if necessary */
239 if (MSG_vm_is_running(vm))
240 simcall_vm_shutdown(vm);
242 if (!MSG_vm_is_created(vm)) {
243 XBT_CRITICAL("shutdown the given VM before destroying it");
247 /* Then, destroy the VM object */
248 simcall_vm_destroy(vm);
250 __MSG_host_destroy(vm);
252 TRACE_msg_vm_end(vm);
256 /** @brief Start a vm (i.e., boot the guest operating system)
259 * If the VM cannot be started, an exception is generated.
262 void MSG_vm_start(msg_vm_t vm)
264 simcall_vm_start(vm);
266 TRACE_msg_vm_start(vm);
271 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
274 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
275 * use a #MSG_process_sleep() or something. I'm not quite sure.
277 void MSG_vm_shutdown(msg_vm_t vm)
279 /* msg_vm_t equals to msg_host_t */
280 simcall_vm_shutdown(vm);
282 // TRACE_msg_vm_(vm);
287 /* We have two mailboxes. mbox is used to transfer migration data between
288 * source and destination PMs. mbox_ctl is used to detect the completion of a
289 * migration. The names of these mailboxes must not conflict with others. */
290 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
292 char *vm_name = sg_host_name(vm);
293 char *src_pm_name = sg_host_name(src_pm);
294 char *dst_pm_name = sg_host_name(dst_pm);
296 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
299 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
301 char *vm_name = sg_host_name(vm);
302 char *src_pm_name = sg_host_name(src_pm);
303 char *dst_pm_name = sg_host_name(dst_pm);
305 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
308 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
310 char *vm_name = sg_host_name(vm);
311 char *src_pm_name = sg_host_name(src_pm);
312 char *dst_pm_name = sg_host_name(dst_pm);
314 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
317 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
319 char *vm_name = sg_host_name(vm);
320 char *src_pm_name = sg_host_name(src_pm);
321 char *dst_pm_name = sg_host_name(dst_pm);
323 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
326 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
328 char *vm_name = sg_host_name(vm);
329 char *src_pm_name = sg_host_name(src_pm);
330 char *dst_pm_name = sg_host_name(dst_pm);
332 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
336 struct migration_session {
341 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
342 * know the completion of the migration. */
344 /* The migration_rx and migration_tx processes use mbox to transfer migration
350 static int migration_rx_fun(int argc, char *argv[])
352 XBT_DEBUG("mig: rx_start");
354 // The structure has been created in the do_migration function and should only be freed in the same place ;)
355 struct migration_session *ms = MSG_process_get_data(MSG_process_self());
357 s_vm_params_t params;
358 simcall_host_get_params(ms->vm, ¶ms);
362 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
366 msg_task_t task = NULL;
367 ret = MSG_task_recv(&task, ms->mbox);
370 // An error occured, clean the code and return
371 // The owner did not change, hence the task should be only destroyed on the other side
372 xbt_free(finalize_task_name);
377 if (strcmp(task->name, finalize_task_name) == 0)
380 MSG_task_destroy(task);
386 // Here Stage 1, 2 and 3 have been performed.
387 // Hence complete the migration
389 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
390 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms structure at the begining and free it at the end of each function)
391 msg_vm_t vm = ms->vm;
392 msg_host_t src_pm = ms->src_pm;
393 msg_host_t dst_pm = ms-> dst_pm;
394 msg_host_priv_t priv = sg_host_msg(vm);
396 // // TODO: we have an issue, if the DST node is turning off during the three next calls, then the VM is in an inconsistent state
397 // // I should check with Takahiro in order to make this portion of code atomic
398 // /* deinstall the current affinity setting for the CPU */
399 // simcall_vm_set_affinity(vm, src_pm, 0);
401 // /* Update the vm location */
402 // simcall_vm_migrate(vm, dst_pm);
404 // /* Resume the VM */
405 // simcall_vm_resume(vm);
407 simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
409 /* install the affinity setting of the VM on the destination pm */
412 unsigned long affinity_mask = (unsigned long)(uintptr_t) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *)dst_pm, sizeof(msg_host_t));
413 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
414 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
419 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
420 msg_host_priv_t priv = sg_host_msg(vm);
421 priv->is_migrating = 0;
422 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->key, ms->src_pm->key, ms->dst_pm->key);
423 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
425 // Inform the SRC that the migration has been correctly performed
427 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
428 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
429 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
430 // xbt_assert(ret == MSG_OK);
431 if(ret == MSG_HOST_FAILURE){
432 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM has been correctly migrated on the DST node
433 // TODO What does it mean ? What should we do ?
434 MSG_task_destroy(task);
435 } else if(ret == MSG_TRANSFER_FAILURE){
436 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
437 MSG_task_destroy(task);
444 xbt_free(finalize_task_name);
446 XBT_DEBUG("mig: rx_done");
451 static void reset_dirty_pages(msg_vm_t vm)
453 msg_host_priv_t priv = sg_host_msg(vm);
456 xbt_dict_cursor_t cursor = NULL;
457 dirty_page_t dp = NULL;
458 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
459 double remaining = MSG_task_get_flops_amount(dp->task);
460 dp->prev_clock = MSG_get_clock();
461 dp->prev_remaining = remaining;
463 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
467 static void start_dirty_page_tracking(msg_vm_t vm)
469 msg_host_priv_t priv = sg_host_msg(vm);
470 priv->dp_enabled = 1;
472 reset_dirty_pages(vm);
475 static void stop_dirty_page_tracking(msg_vm_t vm)
477 msg_host_priv_t priv = sg_host_msg(vm);
478 priv->dp_enabled = 0;
482 /* It might be natural that we define dp_rate for each task. But, we will also
483 * have to care about how each task behavior affects the memory update behavior
484 * at the operating system level. It may not be easy to model it with a simple algorithm. */
485 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
487 double computed = dp->prev_remaining - remaining;
488 double duration = clock - dp->prev_clock;
489 double updated = dp->task->dp_rate * computed;
491 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
492 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
493 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
494 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
500 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
502 double computed = dp->prev_remaining - remaining;
503 double duration = clock - dp->prev_clock;
505 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
506 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
511 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
513 msg_host_priv_t priv = sg_host_msg(vm);
517 xbt_dict_cursor_t cursor = NULL;
518 dirty_page_t dp = NULL;
519 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
520 double remaining = MSG_task_get_flops_amount(dp->task);
522 double clock = MSG_get_clock();
524 // total += calc_updated_pages(key, vm, dp, remaining, clock);
525 total += get_computed(key, vm, dp, remaining, clock);
527 dp->prev_remaining = remaining;
528 dp->prev_clock = clock;
531 total += priv->dp_updated_by_deleted_tasks;
533 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
534 stage_for_fancy_debug,
535 stage2_round_for_fancy_debug,
536 total, priv->dp_updated_by_deleted_tasks);
540 priv->dp_updated_by_deleted_tasks = 0;
546 // TODO Is this code redundant with the information provided by
547 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
548 void MSG_host_add_task(msg_host_t host, msg_task_t task)
550 msg_host_priv_t priv = sg_host_msg(host);
551 double remaining = MSG_task_get_flops_amount(task);
552 char *key = bprintf("%s-%p", task->name, task);
554 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
557 /* It should be okay that we add a task onto a migrating VM. */
558 if (priv->dp_enabled) {
559 dp->prev_clock = MSG_get_clock();
560 dp->prev_remaining = remaining;
563 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
564 xbt_dict_set(priv->dp_objs, key, dp, NULL);
565 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
570 void MSG_host_del_task(msg_host_t host, msg_task_t task)
572 msg_host_priv_t priv = sg_host_msg(host);
574 char *key = bprintf("%s-%p", task->name, task);
576 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
577 xbt_assert(dp->task == task);
579 /* If we are in the middle of dirty page tracking, we record how much
580 * computation has been done until now, and keep the information for the
581 * lookup_() function that will called soon. */
582 if (priv->dp_enabled) {
583 double remaining = MSG_task_get_flops_amount(task);
584 double clock = MSG_get_clock();
585 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
586 double updated = get_computed(key, host, dp, remaining, clock);
588 priv->dp_updated_by_deleted_tasks += updated;
591 xbt_dict_remove(priv->dp_objs, key);
594 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
602 static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm,
603 sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double timeout)
606 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
607 msg_task_t task = MSG_task_create(task_name, 0, (double)size, NULL);
611 double clock_sta = MSG_get_clock();
615 ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
617 ret = MSG_task_send(task, mbox);
623 } else if (ret == MSG_TIMEOUT) {
624 sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
625 sent = size - remaining;
626 XBT_INFO("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu",
627 timeout, remaining, size);
630 /* FIXME: why try-and-catch is used here? */
631 if(ret == MSG_HOST_FAILURE){
632 //XBT_INFO("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
633 MSG_task_destroy(task);
634 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
635 }else if(ret == MSG_TRANSFER_FAILURE){
636 //XBT_INFO("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
637 MSG_task_destroy(task);
638 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
641 double clock_end = MSG_get_clock();
642 double duration = clock_end - clock_sta;
643 double actual_speed = size / duration;
646 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration, actual_speed, mig_speed);
648 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed, mig_speed);
653 static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap)
655 double updated_size = computed * dp_rate;
656 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
657 if (updated_size > dp_cap) {
658 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
659 updated_size = dp_cap;
662 return (sg_size_t) updated_size;
665 static double send_stage1(struct migration_session *ms,
666 sg_size_t ramsize, double mig_speed, double dp_rate, double dp_cap)
669 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
670 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
671 sg_size_t remaining = ramsize;
672 double computed_total = 0;
674 while (remaining > 0) {
675 sg_size_t datasize = chunksize;
676 if (remaining < chunksize)
677 datasize = remaining;
679 remaining -= datasize;
680 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1);
681 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
682 computed_total += computed;
685 return computed_total;
690 static double get_threshold_value(double bandwidth, double max_downtime)
692 return max_downtime * bandwidth;
695 static int migration_tx_fun(int argc, char *argv[])
697 XBT_DEBUG("mig: tx_start");
699 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
700 struct migration_session *ms = MSG_process_get_data(MSG_process_self());
702 s_vm_params_t params;
703 simcall_host_get_params(ms->vm, ¶ms);
704 const sg_size_t ramsize = params.ramsize;
705 const sg_size_t devsize = params.devsize;
706 const int skip_stage1 = params.skip_stage1;
707 int skip_stage2 = params.skip_stage2;
708 const double dp_rate = params.dp_rate;
709 const double dp_cap = params.dp_cap;
710 const double mig_speed = params.mig_speed;
711 double max_downtime = params.max_downtime;
713 /* hard code it temporally. Fix Me */
714 #define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0
715 double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME;
717 double remaining_size = (double) (ramsize + devsize);
718 double threshold = 0.0;
720 /* check parameters */
722 XBT_WARN("migrate a VM, but ramsize is zero");
724 if (max_downtime == 0) {
725 XBT_WARN("use the default max_downtime value 30ms");
729 /* Stage1: send all memory pages to the destination. */
730 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
731 start_dirty_page_tracking(ms->vm);
733 double computed_during_stage1 = 0;
735 double clock_prev_send = MSG_get_clock();
738 /* At stage 1, we do not need timeout. We have to send all the memory
739 * pages even though the duration of this tranfer exceeds the timeout
741 XBT_INFO("Stage 1: Gonna send %llu", ramsize);
742 sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1);
743 remaining_size -= sent;
744 computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0);
746 if (sent < ramsize) {
747 XBT_INFO("mig-stage1: timeout, force moving to stage 3");
749 } else if (sent > ramsize)
753 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
754 // Stop the dirty page tracking an return (there is no memory space to release)
755 stop_dirty_page_tracking(ms->vm);
759 double clock_post_send = MSG_get_clock();
760 mig_timeout -= (clock_post_send - clock_prev_send);
761 if (mig_timeout < 0) {
762 XBT_INFO("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2",
763 (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME);
767 /* estimate bandwidth */
768 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
769 threshold = get_threshold_value(bandwidth, max_downtime);
770 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
774 /* Stage2: send update pages iteratively until the size of remaining states
775 * becomes smaller than the threshold value. */
780 int stage2_round = 0;
783 sg_size_t updated_size = 0;
784 if (stage2_round == 0) {
785 /* just after stage1, nothing has been updated. But, we have to send the
786 * data updated during stage1 */
787 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
789 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
790 updated_size = get_updated_size(computed, dp_rate, dp_cap);
793 XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f",
794 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
797 /* Check whether the remaining size is below the threshold value. If so,
798 * move to stage 3. */
799 remaining_size += updated_size;
800 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
801 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
802 if (remaining_size < threshold)
807 double clock_prev_send = MSG_get_clock();
809 XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
810 sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed, mig_timeout);
812 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
813 // Stop the dirty page tracking an return (there is no memory space to release)
814 stop_dirty_page_tracking(ms->vm);
817 double clock_post_send = MSG_get_clock();
819 if (sent == updated_size) {
820 /* timeout did not happen */
821 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
822 threshold = get_threshold_value(bandwidth, max_downtime);
823 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
824 remaining_size -= sent;
826 mig_timeout -= (clock_post_send - clock_prev_send);
827 xbt_assert(mig_timeout > 0);
829 } else if (sent < updated_size) {
830 /* When timeout happens, we move to stage 3. The size of memory pages
831 * updated before timeout must be added to the remaining size. */
832 XBT_INFO("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf",
833 stage2_round, sent, updated_size, (clock_post_send - clock_prev_send));
834 remaining_size -= sent;
836 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
837 updated_size = get_updated_size(computed, dp_rate, dp_cap);
838 remaining_size += updated_size;
847 /* Stage3: stop the VM and copy the rest of states. */
848 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
849 simcall_vm_suspend(ms->vm);
850 stop_dirty_page_tracking(ms->vm);
853 XBT_DEBUG("Stage 3: Gonna send %f", remaining_size);
854 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1);
856 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
857 // Stop the dirty page tracking an return (there is no memory space to release)
858 simcall_vm_resume(ms->vm);
862 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate effectively the VM on the DST node.
863 XBT_DEBUG("mig: tx_done");
870 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
872 struct migration_session *ms = xbt_new(struct migration_session, 1);
876 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
877 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
880 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
881 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
883 // msg_process_t tx_process, rx_process;
884 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
885 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
888 char **argv = xbt_new(char *, 2);
889 argv[0] = pr_rx_name;
891 /*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
894 char **argv = xbt_new(char *, 2);
895 argv[0] = pr_tx_name;
897 /* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
901 /* wait until the migration have finished or on error has occured */
903 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
904 msg_task_t task = NULL;
905 msg_error_t ret = MSG_TIMEOUT;
906 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
907 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4);
909 xbt_free(ms->mbox_ctl);
913 //xbt_assert(ret == MSG_OK);
914 if(ret == MSG_HOST_FAILURE){
915 // Note that since the communication failed, the owner did not change and the task should be destroyed on the other side.
916 // Hence, just throw the execption
917 XBT_INFO("SRC crashes, throw an exception (m-control)");
918 //MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is required or not
921 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){ // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
922 XBT_INFO("DST crashes, throw an exception (m-control)");
927 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
928 xbt_assert(strcmp(task->name, expected_task_name) == 0);
929 xbt_free(expected_task_name);
930 MSG_task_destroy(task);
938 /** @brief Migrate the VM to the given host.
941 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
942 * MSG_task_send() before or after, depending on whether you want to do cold or hot
945 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
948 * - One approach is ...
949 * We first create a new VM (i.e., destination VM) on the destination
950 * physical host. The destination VM will receive the state of the source
951 * VM over network. We will finally destroy the source VM.
952 * - This behavior is similar to the way of migration in the real world.
953 * Even before a migration is completed, we will see a destination VM,
954 * consuming resources.
955 * - We have to relocate all processes. The existing process migraion code
956 * will work for this?
957 * - The name of the VM is a somewhat unique ID in the code. It is tricky
958 * for the destination VM?
960 * - Another one is ...
961 * We update the information of the given VM to place it to the destination
964 * The second one would be easier.
968 msg_host_t old_pm = simcall_vm_get_pm(vm);
970 if(MSG_host_is_off(old_pm))
971 THROWF(vm_error, 0, "SRC host(%s) seems off, cannot start a migration", sg_host_name(old_pm));
973 if(MSG_host_is_off(new_pm))
974 THROWF(vm_error, 0, "DST host(%s) seems off, cannot start a migration", sg_host_name(new_pm));
976 if (!MSG_vm_is_running(vm))
977 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
979 if (MSG_vm_is_migrating(vm))
980 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_name(vm));
982 msg_host_priv_t priv = sg_host_msg(vm);
983 priv->is_migrating = 1;
987 int ret = do_migration(vm, old_pm, new_pm);
989 priv->is_migrating = 0;
990 THROWF(host_error, 0, "SRC host failed during migration");
993 priv->is_migrating = 0;
994 THROWF(host_error, 0, "DST host failed during migration");
998 // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration process
999 // In that case, the VM has been already assigned to the DST node.
1000 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1001 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1005 /** @brief Immediately suspend the execution of all processes within the given VM.
1008 * This function stops the execution of the VM. All the processes on this VM
1009 * will pause. The state of the VM is preserved. We can later resume it again.
1011 * No suspension cost occurs.
1013 void MSG_vm_suspend(msg_vm_t vm)
1015 if (MSG_vm_is_migrating(vm))
1016 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
1018 simcall_vm_suspend(vm);
1020 XBT_DEBUG("vm_suspend done");
1022 TRACE_msg_vm_suspend(vm);
1026 /** @brief Resume the execution of the VM. All processes on the VM run again.
1029 * No resume cost occurs.
1031 void MSG_vm_resume(msg_vm_t vm)
1033 simcall_vm_resume(vm);
1035 TRACE_msg_vm_resume(vm);
1039 /** @brief Immediately save the execution of all processes within the given VM.
1042 * This function stops the execution of the VM. All the processes on this VM
1043 * will pause. The state of the VM is preserved. We can later resume it again.
1045 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1046 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1047 * of VM save to you.
1049 void MSG_vm_save(msg_vm_t vm)
1051 if (MSG_vm_is_migrating(vm))
1052 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
1054 simcall_vm_save(vm);
1055 TRACE_msg_vm_save(vm);
1058 /** @brief Restore the execution of the VM. All processes on the VM run again.
1061 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1062 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1063 * of VM restore to you.
1065 void MSG_vm_restore(msg_vm_t vm)
1067 simcall_vm_restore(vm);
1069 TRACE_msg_vm_restore(vm);
1073 /** @brief Get the physical host of a given VM.
1076 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1078 return simcall_vm_get_pm(vm);
1082 /** @brief Set a CPU bound for a given VM.
1086 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1089 * On PM0, there are Task1 and VM0.
1090 * On VM0, there is Task2.
1091 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1093 * Task1\@PM0 gets 50%.
1094 * Task2\@VM0 gets 25%.
1095 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1097 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1098 * task on the VM does not affect the bound of the dummy CPU action. The bound
1099 * of the dummy CPU action is unlimited.
1101 * There are some solutions for this problem. One option is to update the bound
1102 * of the dummy CPU action automatically. It should be the sum of all tasks on
1103 * the VM. But, this solution might be costly, because we have to scan all tasks
1104 * on the VM in share_resource() or we have to trap both the start and end of
1107 * The current solution is to use MSG_vm_set_bound(), which allows us to
1108 * directly set the bound of the dummy CPU action.
1112 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1113 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1114 * exceeds the capacity of a CPU core.
1116 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1118 simcall_vm_set_bound(vm, bound);
1122 /** @brief Set the CPU affinity of a given VM.
1125 * This function changes the CPU affinity of a given VM. Usage is the same as
1126 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1128 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1130 msg_host_priv_t priv = sg_host_msg(vm);
1133 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1135 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *)(uintptr_t) mask, NULL);
1137 msg_host_t pm_now = MSG_vm_get_pm(vm);
1139 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1140 simcall_vm_set_affinity(vm, pm, mask);
1142 XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));