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, int ramsize,
173 int net_cap, char *disk_path, int disksize,
174 int mig_netspeed, int dp_intensity)
176 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
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_ws_params_t params;
182 memset(¶ms, 0, sizeof(params));
183 params.ramsize = 1L * 1024 * 1024 * ramsize;
184 //params.overcommit = 0;
186 params.skip_stage2 = 0;
187 params.max_downtime = 0.03;
188 params.dp_rate = (update_speed * 1L * 1024 * 1024 ) / host_speed;
189 params.dp_cap = params.ramsize / 0.9; // working set memory is 90%
190 params.mig_speed = 1L * 1024 * 1024 * mig_netspeed; // 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 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
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);
226 TRACE_msg_vm_create(name, ind_pm);
232 /** @brief Destroy a VM. Destroy the VM object from the simulation.
235 void MSG_vm_destroy(msg_vm_t vm)
237 /* First, terminate all processes on the VM if necessary */
238 if (MSG_vm_is_running(vm))
239 simcall_vm_shutdown(vm);
241 if (!MSG_vm_is_created(vm)) {
242 XBT_CRITICAL("shutdown the given VM before destroying it");
246 /* Then, destroy the VM object */
247 simcall_vm_destroy(vm);
249 __MSG_host_destroy(vm);
252 TRACE_msg_vm_end(vm);
257 /** @brief Start a vm (i.e., boot the guest operating system)
260 * If the VM cannot be started, an exception is generated.
263 void MSG_vm_start(msg_vm_t vm)
265 simcall_vm_start(vm);
268 TRACE_msg_vm_start(vm);
274 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
277 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
278 * use a #MSG_process_sleep() or something. I'm not quite sure.
280 void MSG_vm_shutdown(msg_vm_t vm)
282 /* msg_vm_t equals to msg_host_t */
283 simcall_vm_shutdown(vm);
285 // #ifdef HAVE_TRACING
286 // TRACE_msg_vm_(vm);
292 /* We have two mailboxes. mbox is used to transfer migration data between
293 * source and destiantion PMs. mbox_ctl is used to detect the completion of a
294 * migration. The names of these mailboxes must not conflict with others. */
295 static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
297 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
300 static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
302 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
305 static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
307 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
310 static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
312 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
315 static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
317 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
320 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio);
322 static int migration_rx_fun(int argc, char *argv[])
324 XBT_DEBUG("mig: rx_start");
326 xbt_assert(argc == 4);
327 const char *vm_name = argv[1];
328 const char *src_pm_name = argv[2];
329 const char *dst_pm_name = argv[3];
330 msg_vm_t vm = MSG_get_host_by_name(vm_name);
331 msg_vm_t dst_pm = MSG_get_host_by_name(dst_pm_name);
334 s_ws_params_t params;
335 simcall_host_get_params(vm, ¶ms);
336 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
341 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
342 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
343 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
346 msg_task_t task = NULL;
347 MSG_task_recv(&task, mbox);
349 double received = MSG_task_get_data_size(task);
351 // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
352 launch_deferred_exec_process(vm, received * xfer_cpu_overhead, 1);
355 if (strcmp(task->name, finalize_task_name) == 0)
358 MSG_task_destroy(task);
365 simcall_vm_migrate(vm, dst_pm);
366 simcall_vm_resume(vm);
369 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
371 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
372 msg_error_t ret = MSG_task_send(task, mbox_ctl);
373 xbt_assert(ret == MSG_OK);
381 xbt_free(finalize_task_name);
383 XBT_DEBUG("mig: rx_done");
388 static void reset_dirty_pages(msg_vm_t vm)
390 msg_host_priv_t priv = msg_host_resource_priv(vm);
393 xbt_dict_cursor_t cursor = NULL;
394 dirty_page_t dp = NULL;
395 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
396 double remaining = MSG_task_get_remaining_computation(dp->task);
397 dp->prev_clock = MSG_get_clock();
398 dp->prev_remaining = remaining;
400 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
404 static void start_dirty_page_tracking(msg_vm_t vm)
406 msg_host_priv_t priv = msg_host_resource_priv(vm);
407 priv->dp_enabled = 1;
409 reset_dirty_pages(vm);
412 static void stop_dirty_page_tracking(msg_vm_t vm)
414 msg_host_priv_t priv = msg_host_resource_priv(vm);
415 priv->dp_enabled = 0;
419 /* It might be natural that we define dp_rate for each task. But, we will also
420 * have to care about how each task behavior affects the memory update behavior
421 * at the operating system level. It may not be easy to model it with a simple algorithm. */
422 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
424 double computed = dp->prev_remaining - remaining;
425 double duration = clock - dp->prev_clock;
426 double updated = dp->task->dp_rate * computed;
428 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
429 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
430 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
431 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
437 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
439 double computed = dp->prev_remaining - remaining;
440 double duration = clock - dp->prev_clock;
442 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
443 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
448 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
450 msg_host_priv_t priv = msg_host_resource_priv(vm);
454 xbt_dict_cursor_t cursor = NULL;
455 dirty_page_t dp = NULL;
456 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
457 double remaining = MSG_task_get_remaining_computation(dp->task);
459 double clock = MSG_get_clock();
461 // total += calc_updated_pages(key, vm, dp, remaining, clock);
462 total += get_computed(key, vm, dp, remaining, clock);
464 dp->prev_remaining = remaining;
465 dp->prev_clock = clock;
468 total += priv->dp_updated_by_deleted_tasks;
470 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
471 stage_for_fancy_debug,
472 stage2_round_for_fancy_debug,
473 total, priv->dp_updated_by_deleted_tasks);
477 priv->dp_updated_by_deleted_tasks = 0;
483 // TODO Is this code redundant with the information provided by
484 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
485 void MSG_host_add_task(msg_host_t host, msg_task_t task)
487 msg_host_priv_t priv = msg_host_resource_priv(host);
488 double remaining = MSG_task_get_remaining_computation(task);
489 char *key = bprintf("%s-%lld", task->name, task->counter);
491 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
494 /* It should be okay that we add a task onto a migrating VM. */
495 if (priv->dp_enabled) {
496 dp->prev_clock = MSG_get_clock();
497 dp->prev_remaining = remaining;
500 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
501 xbt_dict_set(priv->dp_objs, key, dp, NULL);
502 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
507 void MSG_host_del_task(msg_host_t host, msg_task_t task)
509 msg_host_priv_t priv = msg_host_resource_priv(host);
511 char *key = bprintf("%s-%lld", task->name, task->counter);
513 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
514 xbt_assert(dp->task == task);
516 /* If we are in the middle of dirty page tracking, we record how much
517 * computaion has been done until now, and keep the information for the
518 * lookup_() function that will called soon. */
519 if (priv->dp_enabled) {
520 double remaining = MSG_task_get_remaining_computation(task);
521 double clock = MSG_get_clock();
522 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
523 double updated = get_computed(key, host, dp, remaining, clock);
525 priv->dp_updated_by_deleted_tasks += updated;
528 xbt_dict_remove(priv->dp_objs, key);
531 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
537 static int deferred_exec_fun(int argc, char *argv[])
539 xbt_assert(argc == 3);
540 const char *comp_str = argv[1];
541 double computaion = atof(comp_str);
542 const char *prio_str = argv[2];
543 double prio = atof(prio_str);
545 msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
546 // XBT_INFO("exec deferred %f", computaion);
548 /* dpt is the results of the VM activity */
549 MSG_task_set_priority(task, prio);
550 MSG_task_execute(task);
554 MSG_task_destroy(task);
559 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
561 char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
564 char **argv = xbt_new(char *, nargvs);
565 argv[0] = xbt_strdup(pr_name);
566 argv[1] = bprintf("%lf", computation);
567 argv[2] = bprintf("%lf", prio);
570 MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv);
576 static int task_tx_overhead_fun(int argc, char *argv[])
578 xbt_assert(argc == 2);
579 const char *mbox = argv[1];
583 // XBT_INFO("start %s", mbox);
586 msg_task_t task = NULL;
587 MSG_task_recv(&task, mbox);
589 // XBT_INFO("task->name %s", task->name);
591 if (strcmp(task->name, "finalize_making_overhead") == 0)
595 // MSG_task_set_priority(task, 1000000);
596 MSG_task_execute(task);
597 MSG_task_destroy(task);
608 static void start_overhead_process(msg_task_t comm_task)
610 char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task));
611 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
614 char **argv = xbt_new(char *, nargvs);
615 argv[0] = xbt_strdup(pr_name);
616 argv[1] = xbt_strdup(mbox);
619 // XBT_INFO("micro start: mbox %s", mbox);
620 MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv);
626 static void shutdown_overhead_process(msg_task_t comm_task)
628 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
630 msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
632 // XBT_INFO("micro shutdown: mbox %s", mbox);
633 msg_error_t ret = MSG_task_send(task, mbox);
634 xbt_assert(ret == MSG_OK);
637 // XBT_INFO("shutdown done");
640 static void request_overhead(msg_task_t comm_task, double computation)
642 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
644 msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
646 // XBT_INFO("req overhead");
647 msg_error_t ret = MSG_task_send(task, mbox);
648 xbt_assert(ret == MSG_OK);
653 /* alpha is (floating_operations / bytes).
655 * When actual migration traffic was 32 mbytes/s, we observed the CPU
656 * utilization of the main thread of the Qemu process was 10 %.
657 * alpha = 0.1 * C / (32 * 1024 * 1024)
658 * where the CPU capacity of the PM is C flops/s.
661 static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
663 const double chunk_size = 1024 * 1024 * 10;
664 double remaining = MSG_task_get_data_size(comm_task);
666 start_overhead_process(comm_task);
669 while (remaining > 0) {
670 double data_size = chunk_size;
671 if (remaining < chunk_size)
672 data_size = remaining;
674 remaining -= data_size;
676 // XBT_INFO("remaining %f bytes", remaining);
679 double clock_sta = MSG_get_clock();
681 /* create a micro task */
683 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
684 msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
686 request_overhead(comm_task, data_size * alpha);
688 msg_error_t ret = MSG_task_send(mtask, mbox);
689 xbt_assert(ret == MSG_OK);
691 xbt_free(mtask_name);
696 /* In the real world, sending data involves small CPU computation. */
697 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
698 msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
699 MSG_task_execute(mtask);
700 MSG_task_destroy(mtask);
701 xbt_free(mtask_name);
707 double clock_end = MSG_get_clock();
712 * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
715 * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
717 * If time_to_sleep is smaller than zero, the elapsed time was too big. We
718 * do not need a micro sleep.
720 double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
721 if (time_to_sleep > 0)
722 MSG_process_sleep(time_to_sleep);
725 //XBT_INFO("duration %f", clock_end - clock_sta);
726 //XBT_INFO("time_to_sleep %f", time_to_sleep);
730 // XBT_INFO("%s", MSG_task_get_name(comm_task));
731 shutdown_overhead_process(comm_task);
737 static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
739 double prev_remaining = init_comm_size;
742 double remaining = MSG_task_get_remaining_communication(comm_task);
746 double sent = prev_remaining - remaining;
747 double comp_size = sent * overhead;
750 char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
751 msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
752 MSG_task_execute(comp_task);
753 MSG_task_destroy(comp_task);
758 prev_remaining = remaining;
762 xbt_free(comp_task_name);
766 #define USE_MICRO_TASK 1
769 // const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
770 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
771 // const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
772 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
773 // const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
774 //const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
775 // const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
776 ////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
777 // const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
778 // const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
779 // const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
781 /* CPU 22% when 80Mbyte/s */
782 const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
786 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
787 double size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
789 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
790 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
794 double clock_sta = MSG_get_clock();
796 #ifdef USE_MICRO_TASK
798 task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
803 ret = MSG_task_send_bounded(task, mbox, mig_speed);
805 ret = MSG_task_send(task, mbox);
806 xbt_assert(ret == MSG_OK);
809 double clock_end = MSG_get_clock();
810 double duration = clock_end - clock_sta;
811 double actual_speed = size / duration;
812 #ifdef USE_MICRO_TASK
813 double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
815 double cpu_utilization = 0;
822 XBT_DEBUG("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);}
824 XBT_DEBUG("mig-stage%d: sent %f duration %f actual_speed %f (target %f) cpu %f", stage, size, duration, actual_speed, mig_speed, cpu_utilization);
831 #ifdef USE_MICRO_TASK
832 /* The name of a micro task starts with __micro, which does not match the
833 * special name that finalizes the receiver loop. Thus, we send the special task.
837 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
838 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
839 msg_error_t ret = MSG_task_send(task, mbox);
840 xbt_assert(ret == MSG_OK);
847 static double get_updated_size(double computed, double dp_rate, double dp_cap)
849 double updated_size = computed * dp_rate;
850 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
851 if (updated_size > dp_cap) {
852 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
853 updated_size = dp_cap;
859 static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *dst_pm_name,
860 long ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead)
862 const char *vm_name = MSG_host_get_name(vm);
863 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
865 const long chunksize = 1024 * 1024 * 100;
866 long remaining = ramsize;
867 double computed_total = 0;
869 while (remaining > 0) {
870 long datasize = chunksize;
871 if (remaining < chunksize)
872 datasize = remaining;
874 remaining -= datasize;
876 send_migration_data(vm_name, src_pm_name, dst_pm_name, datasize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
878 double computed = lookup_computed_flop_counts(vm, 1, 0);
879 computed_total += computed;
882 double updated_size = get_updated_size(computed, dp_rate, dp_cap);
884 double overhead = dpt_cpu_overhead * updated_size;
885 launch_deferred_exec_process(vm, overhead, 10000);
889 return computed_total;
895 static int migration_tx_fun(int argc, char *argv[])
897 XBT_DEBUG("mig: tx_start");
899 xbt_assert(argc == 4);
900 const char *vm_name = argv[1];
901 const char *src_pm_name = argv[2];
902 const char *dst_pm_name = argv[3];
903 msg_vm_t vm = MSG_get_host_by_name(vm_name);
906 s_ws_params_t params;
907 simcall_host_get_params(vm, ¶ms);
908 const long ramsize = params.ramsize;
909 const long devsize = params.devsize;
910 const int skip_stage1 = params.skip_stage1;
911 const int skip_stage2 = params.skip_stage2;
912 const double dp_rate = params.dp_rate;
913 const double dp_cap = params.dp_cap;
914 const double mig_speed = params.mig_speed;
915 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
916 const double dpt_cpu_overhead = params.dpt_cpu_overhead;
918 double remaining_size = ramsize + devsize;
920 double max_downtime = params.max_downtime;
921 if (max_downtime == 0) {
922 XBT_WARN("use the default max_downtime value 30ms");
926 /* This value assumes the network link is 1Gbps. */
927 double threshold = max_downtime * 125 * 1024 * 1024;
929 /* setting up parameters has done */
933 XBT_WARN("migrate a VM, but ramsize is zero");
935 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
937 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
939 /* Stage1: send all memory pages to the destination. */
940 start_dirty_page_tracking(vm);
942 double computed_during_stage1 = 0;
944 // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
946 /* send ramsize, but split it */
947 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);
948 remaining_size -= ramsize;
952 /* Stage2: send update pages iteratively until the size of remaining states
953 * becomes smaller than the threshold value. */
956 if (max_downtime == 0) {
957 XBT_WARN("no max_downtime parameter, skip stage2");
962 int stage2_round = 0;
965 double updated_size = 0;
966 if (stage2_round == 0) {
967 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
968 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
970 double computed = lookup_computed_flop_counts(vm, 2, stage2_round);
971 updated_size = get_updated_size(computed, dp_rate, dp_cap);
974 XBT_INFO("mig-stage 2:%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
975 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
978 if (stage2_round != 0) {
979 /* during stage1, we have already created overhead tasks */
980 double overhead = dpt_cpu_overhead * updated_size;
981 XBT_DEBUG("updated %f overhead %f", updated_size, overhead);
982 launch_deferred_exec_process(vm, overhead, 10000);
987 remaining_size += updated_size;
989 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
990 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
992 if (remaining_size < threshold)
997 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
999 remaining_size -= updated_size;
1005 /* Stage3: stop the VM and copy the rest of states. */
1006 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
1007 simcall_vm_suspend(vm);
1008 stop_dirty_page_tracking(vm);
1010 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead);
1014 XBT_DEBUG("mig: tx_done");
1021 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
1023 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1026 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1028 char **argv = xbt_new(char *, nargvs);
1029 argv[0] = xbt_strdup(pr_name);
1030 argv[1] = xbt_strdup(sg_host_name(vm));
1031 argv[2] = xbt_strdup(sg_host_name(src_pm));
1032 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1035 MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
1041 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1043 char **argv = xbt_new(char *, nargvs);
1044 argv[0] = xbt_strdup(pr_name);
1045 argv[1] = xbt_strdup(sg_host_name(vm));
1046 argv[2] = xbt_strdup(sg_host_name(src_pm));
1047 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1049 MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
1054 /* wait until the migration have finished */
1056 msg_task_t task = NULL;
1057 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
1058 xbt_assert(ret == MSG_OK);
1060 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
1061 xbt_assert(strcmp(task->name, expected_task_name) == 0);
1062 xbt_free(expected_task_name);
1069 /** @brief Migrate the VM to the given host.
1072 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
1073 * MSG_task_send() before or after, depending on whether you want to do cold or hot
1076 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
1079 * - One approach is ...
1080 * We first create a new VM (i.e., destination VM) on the destination
1081 * physical host. The destination VM will receive the state of the source
1082 * VM over network. We will finally destroy the source VM.
1083 * - This behavior is similar to the way of migration in the real world.
1084 * Even before a migration is completed, we will see a destination VM,
1085 * consuming resources.
1086 * - We have to relocate all processes. The existing process migraion code
1087 * will work for this?
1088 * - The name of the VM is a somewhat unique ID in the code. It is tricky
1089 * for the destination VM?
1091 * - Another one is ...
1092 * We update the information of the given VM to place it to the destination
1095 * The second one would be easier.
1099 msg_host_t old_pm = simcall_vm_get_pm(vm);
1101 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
1102 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
1104 do_migration(vm, old_pm, new_pm);
1108 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1111 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1116 /** @brief Immediately suspend the execution of all processes within the given VM.
1119 * This function stops the exection of the VM. All the processes on this VM
1120 * will pause. The state of the VM is perserved. We can later resume it again.
1122 * No suspension cost occurs.
1124 void MSG_vm_suspend(msg_vm_t vm)
1126 simcall_vm_suspend(vm);
1128 XBT_DEBUG("vm_suspend done");
1131 TRACE_msg_vm_suspend(vm);
1136 /** @brief Resume the execution of the VM. All processes on the VM run again.
1139 * No resume cost occurs.
1141 void MSG_vm_resume(msg_vm_t vm)
1143 simcall_vm_resume(vm);
1146 TRACE_msg_vm_resume(vm);
1151 /** @brief Immediately save the execution of all processes within the given VM.
1154 * This function stops the exection of the VM. All the processes on this VM
1155 * will pause. The state of the VM is perserved. We can later resume it again.
1157 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1158 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1159 * of VM save to you.
1161 void MSG_vm_save(msg_vm_t vm)
1163 simcall_vm_save(vm);
1165 TRACE_msg_vm_save(vm);
1169 /** @brief Restore the execution of the VM. All processes on the VM run again.
1172 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1173 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1174 * of VM restore to you.
1176 void MSG_vm_restore(msg_vm_t vm)
1178 simcall_vm_restore(vm);
1181 TRACE_msg_vm_restore(vm);
1186 /** @brief Get the physical host of a given VM.
1189 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1191 return simcall_vm_get_pm(vm);
1195 /** @brief Set a CPU bound for a given VM.
1199 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1202 * On PM0, there are Task1 and VM0.
1203 * On VM0, there is Task2.
1204 * Now we bound 75% to Task1@PM0 and bound 25% to Task2@VM0.
1206 * Task1@PM0 gets 50%.
1207 * Task2@VM0 gets 25%.
1208 * This is NOT 75% for Task1@PM0 and 25% for Task2@VM0, respectively.
1210 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1211 * task on the VM does not affect the bound of the dummy CPU action. The bound
1212 * of the dummy CPU action is unlimited.
1214 * There are some solutions for this problem. One option is to update the bound
1215 * of the dummy CPU action automatically. It should be the sum of all tasks on
1216 * the VM. But, this solution might be costy, because we have to scan all tasks
1217 * on the VM in share_resource() or we have to trap both the start and end of
1220 * The current solution is to use MSG_vm_set_bound(), which allows us to
1221 * directly set the bound of the dummy CPU action.
1225 * Note that bound == 0 means no bound (i.e., unlimited).
1227 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1229 return simcall_vm_set_bound(vm, bound);