1 /* Copyright (c) 2012. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
7 // 1./ check how and where a new VM is added to the list of the hosts
8 // 2./ Diff between SIMIX_Actions and SURF_Actions
9 // => SIMIX_actions : point synchro entre processus de niveau (theoretically speaking I do not have to create such SIMIX_ACTION
13 // MSG_TRACE can be revisited in order to use the host
14 // To implement a mixed model between workstation and vm_workstation,
15 // please give a look at surf_model_private_t model_private at SURF Level and to the share resource functions
16 // double (*share_resources) (double now);
17 // For the action into the vm workstation model, we should be able to leverage the usual one (and if needed, look at
18 // the workstation model.
20 #include "msg_private.h"
21 #include "xbt/sysdep.h"
24 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
25 "Cloud-oriented parts of the MSG API");
28 /* **** ******** GENERAL ********* **** */
30 /** \ingroup m_vm_management
31 * \brief Returns the value of a given vm property
34 * \param name a property name
35 * \return value of a property (or NULL if property not set)
38 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
40 return MSG_host_get_property_value(vm, name);
43 /** \ingroup m_vm_management
44 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
47 * \return a dict containing the properties
49 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
51 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
53 return (simcall_host_get_properties(vm));
56 /** \ingroup m_host_management
57 * \brief Change the value of a given host property
60 * \param name a property name
61 * \param value what to change the property to
62 * \param free_ctn the freeing function to use to kill the value on need
64 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
66 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
69 /** \ingroup msg_vm_management
70 * \brief Finds a msg_vm_t using its name.
72 * This is a name directory service
73 * \param name the name of a vm.
74 * \return the corresponding vm
76 * Please note that a VM is a specific host. Hence, you should give a different name
80 msg_vm_t MSG_vm_get_by_name(const char *name)
82 return MSG_get_host_by_name(name);
85 /** \ingroup m_vm_management
87 * \brief Return the name of the #msg_host_t.
89 * This functions checks whether \a host is a valid pointer or not and return
92 const char *MSG_vm_get_name(msg_vm_t vm)
94 return MSG_host_get_name(vm);
98 /* **** Check state of a VM **** */
99 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
101 return simcall_vm_get_state(vm) == state;
104 /** @brief Returns whether the given VM has just reated, not running.
107 int MSG_vm_is_created(msg_vm_t vm)
109 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
112 /** @brief Returns whether the given VM is currently running
115 int MSG_vm_is_running(msg_vm_t vm)
117 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
120 /** @brief Returns whether the given VM is currently migrating
123 int MSG_vm_is_migrating(msg_vm_t vm)
125 return __MSG_vm_is_state(vm, SURF_VM_STATE_MIGRATING);
128 /** @brief Returns whether the given VM is currently suspended, not running.
131 int MSG_vm_is_suspended(msg_vm_t vm)
133 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
136 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
139 int MSG_vm_is_saving(msg_vm_t vm)
141 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
144 /** @brief Returns whether the given VM has been saved, not running.
147 int MSG_vm_is_saved(msg_vm_t vm)
149 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
152 /** @brief Returns whether the given VM is being restored, not running.
155 int MSG_vm_is_restoring(msg_vm_t vm)
157 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
162 /* ------------------------------------------------------------------------- */
163 /* ------------------------------------------------------------------------- */
165 /* **** ******** MSG vm actions ********* **** */
167 /** @brief Create a new VM with specified parameters.
169 * All parameters are in MBytes
172 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name, int ncpus, long ramsize,
173 long net_cap, char *disk_path, long disksize,
174 long dp_rate, long mig_netspeed)
176 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
177 s_ws_params_t params;
178 memset(¶ms, 0, sizeof(params));
179 params.ramsize = 1L * 1024 * 1024 * ramsize;
180 //params.overcommit = 0;
182 params.skip_stage2 = 0;
183 params.max_downtime = 0.03;
184 params.dp_rate = 1L * 1024 * 1024 * dp_rate;
185 params.dp_cap = params.ramsize / 0.9; // working set memory is 90%
186 params.mig_speed = 1L * 1024 * 1024 * mig_netspeed; // mig_speed
188 simcall_host_set_params(vm, ¶ms);
194 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
197 * A VM is treated as a host. The name of the VM must be unique among all hosts.
199 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
201 /* make sure the VM of the same name does not exit */
203 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
205 XBT_ERROR("host %s already exits", name);
210 /* Note: ind_vm and vm_workstation point to the same elm object. */
211 msg_vm_t ind_vm = NULL;
212 void *ind_vm_workstation = NULL;
214 /* Ask the SIMIX layer to create the surf vm resource */
215 ind_vm_workstation = simcall_vm_create(name, ind_pm);
216 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
218 XBT_DEBUG("A new VM (%s) has been created", name);
221 TRACE_msg_vm_create(name, ind_pm);
227 /** @brief Destroy a VM. Destroy the VM object from the simulation.
230 void MSG_vm_destroy(msg_vm_t vm)
232 /* First, terminate all processes on the VM if necessary */
233 if (MSG_vm_is_running(vm))
234 simcall_vm_shutdown(vm);
236 if (!MSG_vm_is_created(vm)) {
237 XBT_CRITICAL("shutdown the given VM before destroying it");
241 /* Then, destroy the VM object */
242 simcall_vm_destroy(vm);
244 __MSG_host_destroy(vm);
247 TRACE_msg_vm_end(vm);
252 /** @brief Start a vm (i.e., boot the guest operating system)
255 * If the VM cannot be started, an exception is generated.
258 void MSG_vm_start(msg_vm_t vm)
260 simcall_vm_start(vm);
263 TRACE_msg_vm_start(vm);
269 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
272 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
273 * use a #MSG_process_sleep() or something. I'm not quite sure.
275 void MSG_vm_shutdown(msg_vm_t vm)
277 /* msg_vm_t equals to msg_host_t */
278 simcall_vm_shutdown(vm);
280 // #ifdef HAVE_TRACING
281 // TRACE_msg_vm_(vm);
287 /* We have two mailboxes. mbox is used to transfer migration data between
288 * source and destiantion 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(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
292 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
295 static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
297 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
300 static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
302 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
305 static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
307 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
310 static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
312 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
315 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio);
317 static int migration_rx_fun(int argc, char *argv[])
319 XBT_DEBUG("mig: rx_start");
321 xbt_assert(argc == 4);
322 const char *vm_name = argv[1];
323 const char *src_pm_name = argv[2];
324 const char *dst_pm_name = argv[3];
325 msg_vm_t vm = MSG_get_host_by_name(vm_name);
326 msg_vm_t dst_pm = MSG_get_host_by_name(dst_pm_name);
329 s_ws_params_t params;
330 simcall_host_get_params(vm, ¶ms);
331 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
336 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
337 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
338 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
341 msg_task_t task = NULL;
342 MSG_task_recv(&task, mbox);
344 double received = MSG_task_get_data_size(task);
346 // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
347 launch_deferred_exec_process(vm, received * xfer_cpu_overhead, 1);
350 if (strcmp(task->name, finalize_task_name) == 0)
353 MSG_task_destroy(task);
360 simcall_vm_migrate(vm, dst_pm);
361 simcall_vm_resume(vm);
364 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
366 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
367 msg_error_t ret = MSG_task_send(task, mbox_ctl);
368 xbt_assert(ret == MSG_OK);
376 xbt_free(finalize_task_name);
378 XBT_DEBUG("mig: rx_done");
383 static void reset_dirty_pages(msg_vm_t vm)
385 msg_host_priv_t priv = msg_host_resource_priv(vm);
388 xbt_dict_cursor_t cursor = NULL;
389 dirty_page_t dp = NULL;
390 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
391 double remaining = MSG_task_get_remaining_computation(dp->task);
392 dp->prev_clock = MSG_get_clock();
393 dp->prev_remaining = remaining;
395 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
399 static void start_dirty_page_tracking(msg_vm_t vm)
401 msg_host_priv_t priv = msg_host_resource_priv(vm);
402 priv->dp_enabled = 1;
404 reset_dirty_pages(vm);
407 static void stop_dirty_page_tracking(msg_vm_t vm)
409 msg_host_priv_t priv = msg_host_resource_priv(vm);
410 priv->dp_enabled = 0;
414 /* It might be natural that we define dp_rate for each task. But, we will also
415 * have to care about how each task behavior affects the memory update behavior
416 * at the operating system level. It may not be easy to model it with a simple algorithm. */
417 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
419 double computed = dp->prev_remaining - remaining;
420 double duration = clock - dp->prev_clock;
421 double updated = dp->task->dp_rate * computed;
423 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
424 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
425 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
426 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
432 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
434 double computed = dp->prev_remaining - remaining;
435 double duration = clock - dp->prev_clock;
437 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
438 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
443 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
445 msg_host_priv_t priv = msg_host_resource_priv(vm);
449 xbt_dict_cursor_t cursor = NULL;
450 dirty_page_t dp = NULL;
451 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
452 double remaining = MSG_task_get_remaining_computation(dp->task);
453 double clock = MSG_get_clock();
455 // total += calc_updated_pages(key, vm, dp, remaining, clock);
456 total += get_computed(key, vm, dp, remaining, clock);
458 dp->prev_remaining = remaining;
459 dp->prev_clock = clock;
462 total += priv->dp_updated_by_deleted_tasks;
464 XBT_INFO("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
465 stage_for_fancy_debug,
466 stage2_round_for_fancy_debug,
467 total, priv->dp_updated_by_deleted_tasks);
471 priv->dp_updated_by_deleted_tasks = 0;
477 // TODO Is this code redundant with the information provided by
478 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
479 void MSG_host_add_task(msg_host_t host, msg_task_t task)
481 msg_host_priv_t priv = msg_host_resource_priv(host);
482 double remaining = MSG_task_get_remaining_computation(task);
483 char *key = bprintf("%s-%lld", task->name, task->counter);
485 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
488 /* It should be okay that we add a task onto a migrating VM. */
489 if (priv->dp_enabled) {
490 dp->prev_clock = MSG_get_clock();
491 dp->prev_remaining = remaining;
494 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
495 xbt_dict_set(priv->dp_objs, key, dp, NULL);
496 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
501 void MSG_host_del_task(msg_host_t host, msg_task_t task)
503 msg_host_priv_t priv = msg_host_resource_priv(host);
505 char *key = bprintf("%s-%lld", task->name, task->counter);
507 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
508 xbt_assert(dp->task == task);
510 /* If we are in the middle of dirty page tracking, we record how much
511 * computaion has been done until now, and keep the information for the
512 * lookup_() function that will called soon. */
513 if (priv->dp_enabled) {
514 double remaining = MSG_task_get_remaining_computation(task);
515 double clock = MSG_get_clock();
516 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
517 double updated = get_computed(key, host, dp, remaining, clock);
519 priv->dp_updated_by_deleted_tasks += updated;
522 xbt_dict_remove(priv->dp_objs, key);
525 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
531 static int deferred_exec_fun(int argc, char *argv[])
533 xbt_assert(argc == 3);
534 const char *comp_str = argv[1];
535 double computaion = atof(comp_str);
536 const char *prio_str = argv[2];
537 double prio = atof(prio_str);
539 msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
540 // XBT_INFO("exec deferred %f", computaion);
542 /* dpt is the results of the VM activity */
543 MSG_task_set_priority(task, prio);
544 MSG_task_execute(task);
548 MSG_task_destroy(task);
553 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
555 char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
558 char **argv = xbt_new(char *, nargvs);
559 argv[0] = xbt_strdup(pr_name);
560 argv[1] = bprintf("%lf", computation);
561 argv[2] = bprintf("%lf", prio);
564 MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv);
570 static int task_tx_overhead_fun(int argc, char *argv[])
572 xbt_assert(argc == 2);
573 const char *mbox = argv[1];
577 // XBT_INFO("start %s", mbox);
580 msg_task_t task = NULL;
581 MSG_task_recv(&task, mbox);
583 // XBT_INFO("task->name %s", task->name);
585 if (strcmp(task->name, "finalize_making_overhead") == 0)
589 // MSG_task_set_priority(task, 1000000);
590 MSG_task_execute(task);
591 MSG_task_destroy(task);
602 static void start_overhead_process(msg_task_t comm_task)
604 char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task));
605 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
608 char **argv = xbt_new(char *, nargvs);
609 argv[0] = xbt_strdup(pr_name);
610 argv[1] = xbt_strdup(mbox);
613 // XBT_INFO("micro start: mbox %s", mbox);
614 MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv);
620 static void shutdown_overhead_process(msg_task_t comm_task)
622 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
624 msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
626 // XBT_INFO("micro shutdown: mbox %s", mbox);
627 msg_error_t ret = MSG_task_send(task, mbox);
628 xbt_assert(ret == MSG_OK);
631 // XBT_INFO("shutdown done");
634 static void request_overhead(msg_task_t comm_task, double computation)
636 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
638 msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
640 // XBT_INFO("req overhead");
641 msg_error_t ret = MSG_task_send(task, mbox);
642 xbt_assert(ret == MSG_OK);
647 /* alpha is (floating_operations / bytes).
649 * When actual migration traffic was 32 mbytes/s, we observed the CPU
650 * utilization of the main thread of the Qemu process was 10 %.
651 * alpha = 0.1 * C / (32 * 1024 * 1024)
652 * where the CPU capacity of the PM is C flops/s.
655 static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
657 const double chunk_size = 1024 * 1024 * 10;
658 double remaining = MSG_task_get_data_size(comm_task);
660 start_overhead_process(comm_task);
663 while (remaining > 0) {
664 double data_size = chunk_size;
665 if (remaining < chunk_size)
666 data_size = remaining;
668 remaining -= data_size;
670 // XBT_INFO("remaining %f bytes", remaining);
673 double clock_sta = MSG_get_clock();
675 /* create a micro task */
677 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
678 msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
680 request_overhead(comm_task, data_size * alpha);
682 msg_error_t ret = MSG_task_send(mtask, mbox);
683 xbt_assert(ret == MSG_OK);
685 xbt_free(mtask_name);
690 /* In the real world, sending data involves small CPU computation. */
691 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
692 msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
693 MSG_task_execute(mtask);
694 MSG_task_destroy(mtask);
695 xbt_free(mtask_name);
701 double clock_end = MSG_get_clock();
706 * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
709 * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
711 * If time_to_sleep is smaller than zero, the elapsed time was too big. We
712 * do not need a micro sleep.
714 double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
715 if (time_to_sleep > 0)
716 MSG_process_sleep(time_to_sleep);
719 //XBT_INFO("duration %f", clock_end - clock_sta);
720 //XBT_INFO("time_to_sleep %f", time_to_sleep);
724 // XBT_INFO("%s", MSG_task_get_name(comm_task));
725 shutdown_overhead_process(comm_task);
731 static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
733 double prev_remaining = init_comm_size;
736 double remaining = MSG_task_get_remaining_communication(comm_task);
740 double sent = prev_remaining - remaining;
741 double comp_size = sent * overhead;
744 char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
745 msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
746 MSG_task_execute(comp_task);
747 MSG_task_destroy(comp_task);
752 prev_remaining = remaining;
756 xbt_free(comp_task_name);
760 #define USE_MICRO_TASK 1
763 // const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
764 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
765 // const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
766 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
767 // const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
768 //const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
769 // const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
770 ////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
771 // const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
772 // const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
773 // const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
775 /* CPU 22% when 80Mbyte/s */
776 const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
780 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
781 double size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
783 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
784 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
788 double clock_sta = MSG_get_clock();
790 #ifdef USE_MICRO_TASK
792 task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
797 ret = MSG_task_send_bounded(task, mbox, mig_speed);
799 ret = MSG_task_send(task, mbox);
800 xbt_assert(ret == MSG_OK);
803 double clock_end = MSG_get_clock();
804 double duration = clock_end - clock_sta;
805 double actual_speed = size / duration;
806 #ifdef USE_MICRO_TASK
807 double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
809 double cpu_utilization = 0;
816 XBT_INFO("mig-stage%d.%d: sent %f duration %f actual_speed %f (target %f) cpu %f", stage, stage2_round, size, duration, actual_speed, mig_speed, cpu_utilization);
818 XBT_INFO("mig-stage%d: sent %f duration %f actual_speed %f (target %f) cpu %f", stage, size, duration, actual_speed, mig_speed, cpu_utilization);
824 #ifdef USE_MICRO_TASK
825 /* The name of a micro task starts with __micro, which does not match the
826 * special name that finalizes the receiver loop. Thus, we send the special task.
830 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
831 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
832 msg_error_t ret = MSG_task_send(task, mbox);
833 xbt_assert(ret == MSG_OK);
840 static double get_updated_size(double computed, double dp_rate, double dp_cap)
842 double updated_size = computed * dp_rate;
843 XBT_INFO("updated_size %f dp_rate %f", updated_size, dp_rate);
844 if (updated_size > dp_cap) {
845 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
846 updated_size = dp_cap;
852 static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *dst_pm_name,
853 long ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead)
855 const char *vm_name = MSG_host_get_name(vm);
856 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
858 const long chunksize = 1024 * 1024 * 100;
859 long remaining = ramsize;
860 double computed_total = 0;
862 while (remaining > 0) {
863 long datasize = chunksize;
864 if (remaining < chunksize)
865 datasize = remaining;
867 remaining -= datasize;
869 send_migration_data(vm_name, src_pm_name, dst_pm_name, datasize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
871 double computed = lookup_computed_flop_counts(vm, 1, 0);
872 computed_total += computed;
875 double updated_size = get_updated_size(computed, dp_rate, dp_cap);
877 double overhead = dpt_cpu_overhead * updated_size;
878 launch_deferred_exec_process(vm, overhead, 10000);
882 return computed_total;
888 static int migration_tx_fun(int argc, char *argv[])
890 XBT_DEBUG("mig: tx_start");
892 xbt_assert(argc == 4);
893 const char *vm_name = argv[1];
894 const char *src_pm_name = argv[2];
895 const char *dst_pm_name = argv[3];
896 msg_vm_t vm = MSG_get_host_by_name(vm_name);
899 s_ws_params_t params;
900 simcall_host_get_params(vm, ¶ms);
901 const long ramsize = params.ramsize;
902 const long devsize = params.devsize;
903 const int skip_stage1 = params.skip_stage1;
904 const int skip_stage2 = params.skip_stage2;
905 const double dp_rate = params.dp_rate;
906 const double dp_cap = params.dp_cap;
907 const double mig_speed = params.mig_speed;
908 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
909 const double dpt_cpu_overhead = params.dpt_cpu_overhead;
911 double remaining_size = ramsize + devsize;
913 double max_downtime = params.max_downtime;
914 if (max_downtime == 0) {
915 XBT_WARN("use the default max_downtime value 30ms");
919 /* This value assumes the network link is 1Gbps. */
920 double threshold = max_downtime * 125 * 1024 * 1024;
922 /* setting up parameters has done */
926 XBT_WARN("migrate a VM, but ramsize is zero");
928 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
930 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
932 /* Stage1: send all memory pages to the destination. */
933 start_dirty_page_tracking(vm);
935 double computed_during_stage1 = 0;
937 // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
939 /* send ramsize, but split it */
940 computed_during_stage1 = send_stage1(vm, src_pm_name, dst_pm_name, ramsize, mig_speed, xfer_cpu_overhead, dp_rate, dp_cap, dpt_cpu_overhead);
941 remaining_size -= ramsize;
945 /* Stage2: send update pages iteratively until the size of remaining states
946 * becomes smaller than the threshold value. */
949 if (max_downtime == 0) {
950 XBT_WARN("no max_downtime parameter, skip stage2");
955 int stage2_round = 0;
958 double updated_size = 0;
959 if (stage2_round == 0) {
960 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
961 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
963 double computed = lookup_computed_flop_counts(vm, 2, stage2_round);
964 updated_size = get_updated_size(computed, dp_rate, dp_cap);
967 XBT_INFO("%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
968 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
971 if (stage2_round != 0) {
972 /* during stage1, we have already created overhead tasks */
973 double overhead = dpt_cpu_overhead * updated_size;
974 XBT_INFO("updated %f overhead %f", updated_size, overhead);
975 launch_deferred_exec_process(vm, overhead, 10000);
980 remaining_size += updated_size;
982 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
983 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
985 if (remaining_size < threshold)
990 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
992 remaining_size -= updated_size;
998 /* Stage3: stop the VM and copy the rest of states. */
999 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
1000 simcall_vm_suspend(vm);
1001 stop_dirty_page_tracking(vm);
1003 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead);
1007 XBT_DEBUG("mig: tx_done");
1014 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
1016 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1019 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1021 char **argv = xbt_new(char *, nargvs);
1022 argv[0] = xbt_strdup(pr_name);
1023 argv[1] = xbt_strdup(sg_host_name(vm));
1024 argv[2] = xbt_strdup(sg_host_name(src_pm));
1025 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1028 MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
1034 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1036 char **argv = xbt_new(char *, nargvs);
1037 argv[0] = xbt_strdup(pr_name);
1038 argv[1] = xbt_strdup(sg_host_name(vm));
1039 argv[2] = xbt_strdup(sg_host_name(src_pm));
1040 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1042 MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
1047 /* wait until the migration have finished */
1049 msg_task_t task = NULL;
1050 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
1051 xbt_assert(ret == MSG_OK);
1053 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
1054 xbt_assert(strcmp(task->name, expected_task_name) == 0);
1055 xbt_free(expected_task_name);
1062 /** @brief Migrate the VM to the given host.
1065 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
1066 * MSG_task_send() before or after, depending on whether you want to do cold or hot
1069 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
1072 * - One approach is ...
1073 * We first create a new VM (i.e., destination VM) on the destination
1074 * physical host. The destination VM will receive the state of the source
1075 * VM over network. We will finally destroy the source VM.
1076 * - This behavior is similar to the way of migration in the real world.
1077 * Even before a migration is completed, we will see a destination VM,
1078 * consuming resources.
1079 * - We have to relocate all processes. The existing process migraion code
1080 * will work for this?
1081 * - The name of the VM is a somewhat unique ID in the code. It is tricky
1082 * for the destination VM?
1084 * - Another one is ...
1085 * We update the information of the given VM to place it to the destination
1088 * The second one would be easier.
1092 msg_host_t old_pm = simcall_vm_get_pm(vm);
1094 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
1095 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
1097 do_migration(vm, old_pm, new_pm);
1101 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1104 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1109 /** @brief Immediately suspend the execution of all processes within the given VM.
1112 * This function stops the exection of the VM. All the processes on this VM
1113 * will pause. The state of the VM is perserved. We can later resume it again.
1115 * No suspension cost occurs.
1117 void MSG_vm_suspend(msg_vm_t vm)
1119 simcall_vm_suspend(vm);
1121 XBT_DEBUG("vm_suspend done");
1124 TRACE_msg_vm_suspend(vm);
1129 /** @brief Resume the execution of the VM. All processes on the VM run again.
1132 * No resume cost occurs.
1134 void MSG_vm_resume(msg_vm_t vm)
1136 simcall_vm_resume(vm);
1139 TRACE_msg_vm_resume(vm);
1144 /** @brief Immediately save the execution of all processes within the given VM.
1147 * This function stops the exection of the VM. All the processes on this VM
1148 * will pause. The state of the VM is perserved. We can later resume it again.
1150 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1151 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1152 * of VM save to you.
1154 void MSG_vm_save(msg_vm_t vm)
1156 simcall_vm_save(vm);
1158 TRACE_msg_vm_save(vm);
1162 /** @brief Restore the execution of the VM. All processes on the VM run again.
1165 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1166 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1167 * of VM restore to you.
1169 void MSG_vm_restore(msg_vm_t vm)
1171 simcall_vm_restore(vm);
1174 TRACE_msg_vm_restore(vm);
1179 /** @brief Get the physical host of a given VM.
1182 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1184 return simcall_vm_get_pm(vm);
1188 /** @brief Set a CPU bound for a given VM.
1192 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1195 * On PM0, there are Task1 and VM0.
1196 * On VM0, there is Task2.
1197 * Now we bound 75% to Task1@PM0 and bound 25% to Task2@VM0.
1199 * Task1@PM0 gets 50%.
1200 * Task2@VM0 gets 25%.
1201 * This is NOT 75% for Task1@PM0 and 25% for Task2@VM0, respectively.
1203 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1204 * task on the VM does not affect the bound of the dummy CPU action. The bound
1205 * of the dummy CPU action is unlimited.
1207 * There are some solutions for this problem. One option is to update the bound
1208 * of the dummy CPU action automatically. It should be the sum of all tasks on
1209 * the VM. But, this solution might be costy, because we have to scan all tasks
1210 * on the VM in share_resource() or we have to trap both the start and end of
1213 * The current solution is to use MSG_vm_set_bound(), which allows us to
1214 * directly set the bound of the dummy CPU action.
1218 * Note that bound == 0 means no bound (i.e., unlimited).
1220 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1222 return simcall_vm_set_bound(vm, bound);