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);
458 double clock = MSG_get_clock();
460 // total += calc_updated_pages(key, vm, dp, remaining, clock);
461 total += get_computed(key, vm, dp, remaining, clock);
463 dp->prev_remaining = remaining;
464 dp->prev_clock = clock;
467 total += priv->dp_updated_by_deleted_tasks;
469 XBT_INFO("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
470 stage_for_fancy_debug,
471 stage2_round_for_fancy_debug,
472 total, priv->dp_updated_by_deleted_tasks);
476 priv->dp_updated_by_deleted_tasks = 0;
482 // TODO Is this code redundant with the information provided by
483 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
484 void MSG_host_add_task(msg_host_t host, msg_task_t task)
486 msg_host_priv_t priv = msg_host_resource_priv(host);
487 double remaining = MSG_task_get_remaining_computation(task);
488 char *key = bprintf("%s-%lld", task->name, task->counter);
490 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
493 /* It should be okay that we add a task onto a migrating VM. */
494 if (priv->dp_enabled) {
495 dp->prev_clock = MSG_get_clock();
496 dp->prev_remaining = remaining;
499 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
500 xbt_dict_set(priv->dp_objs, key, dp, NULL);
501 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
506 void MSG_host_del_task(msg_host_t host, msg_task_t task)
508 msg_host_priv_t priv = msg_host_resource_priv(host);
510 char *key = bprintf("%s-%lld", task->name, task->counter);
512 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
513 xbt_assert(dp->task == task);
515 /* If we are in the middle of dirty page tracking, we record how much
516 * computaion has been done until now, and keep the information for the
517 * lookup_() function that will called soon. */
518 if (priv->dp_enabled) {
519 double remaining = MSG_task_get_remaining_computation(task);
520 double clock = MSG_get_clock();
521 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
522 double updated = get_computed(key, host, dp, remaining, clock);
524 priv->dp_updated_by_deleted_tasks += updated;
527 xbt_dict_remove(priv->dp_objs, key);
530 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
536 static int deferred_exec_fun(int argc, char *argv[])
538 xbt_assert(argc == 3);
539 const char *comp_str = argv[1];
540 double computaion = atof(comp_str);
541 const char *prio_str = argv[2];
542 double prio = atof(prio_str);
544 msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
545 // XBT_INFO("exec deferred %f", computaion);
547 /* dpt is the results of the VM activity */
548 MSG_task_set_priority(task, prio);
549 MSG_task_execute(task);
553 MSG_task_destroy(task);
558 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
560 char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
563 char **argv = xbt_new(char *, nargvs);
564 argv[0] = xbt_strdup(pr_name);
565 argv[1] = bprintf("%lf", computation);
566 argv[2] = bprintf("%lf", prio);
569 MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv);
575 static int task_tx_overhead_fun(int argc, char *argv[])
577 xbt_assert(argc == 2);
578 const char *mbox = argv[1];
582 // XBT_INFO("start %s", mbox);
585 msg_task_t task = NULL;
586 MSG_task_recv(&task, mbox);
588 // XBT_INFO("task->name %s", task->name);
590 if (strcmp(task->name, "finalize_making_overhead") == 0)
594 // MSG_task_set_priority(task, 1000000);
595 MSG_task_execute(task);
596 MSG_task_destroy(task);
607 static void start_overhead_process(msg_task_t comm_task)
609 char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task));
610 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
613 char **argv = xbt_new(char *, nargvs);
614 argv[0] = xbt_strdup(pr_name);
615 argv[1] = xbt_strdup(mbox);
618 // XBT_INFO("micro start: mbox %s", mbox);
619 MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv);
625 static void shutdown_overhead_process(msg_task_t comm_task)
627 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
629 msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
631 // XBT_INFO("micro shutdown: mbox %s", mbox);
632 msg_error_t ret = MSG_task_send(task, mbox);
633 xbt_assert(ret == MSG_OK);
636 // XBT_INFO("shutdown done");
639 static void request_overhead(msg_task_t comm_task, double computation)
641 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
643 msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
645 // XBT_INFO("req overhead");
646 msg_error_t ret = MSG_task_send(task, mbox);
647 xbt_assert(ret == MSG_OK);
652 /* alpha is (floating_operations / bytes).
654 * When actual migration traffic was 32 mbytes/s, we observed the CPU
655 * utilization of the main thread of the Qemu process was 10 %.
656 * alpha = 0.1 * C / (32 * 1024 * 1024)
657 * where the CPU capacity of the PM is C flops/s.
660 static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
662 const double chunk_size = 1024 * 1024 * 10;
663 double remaining = MSG_task_get_data_size(comm_task);
665 start_overhead_process(comm_task);
668 while (remaining > 0) {
669 double data_size = chunk_size;
670 if (remaining < chunk_size)
671 data_size = remaining;
673 remaining -= data_size;
675 // XBT_INFO("remaining %f bytes", remaining);
678 double clock_sta = MSG_get_clock();
680 /* create a micro task */
682 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
683 msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
685 request_overhead(comm_task, data_size * alpha);
687 msg_error_t ret = MSG_task_send(mtask, mbox);
688 xbt_assert(ret == MSG_OK);
690 xbt_free(mtask_name);
695 /* In the real world, sending data involves small CPU computation. */
696 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
697 msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
698 MSG_task_execute(mtask);
699 MSG_task_destroy(mtask);
700 xbt_free(mtask_name);
706 double clock_end = MSG_get_clock();
711 * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
714 * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
716 * If time_to_sleep is smaller than zero, the elapsed time was too big. We
717 * do not need a micro sleep.
719 double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
720 if (time_to_sleep > 0)
721 MSG_process_sleep(time_to_sleep);
724 //XBT_INFO("duration %f", clock_end - clock_sta);
725 //XBT_INFO("time_to_sleep %f", time_to_sleep);
729 // XBT_INFO("%s", MSG_task_get_name(comm_task));
730 shutdown_overhead_process(comm_task);
736 static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
738 double prev_remaining = init_comm_size;
741 double remaining = MSG_task_get_remaining_communication(comm_task);
745 double sent = prev_remaining - remaining;
746 double comp_size = sent * overhead;
749 char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
750 msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
751 MSG_task_execute(comp_task);
752 MSG_task_destroy(comp_task);
757 prev_remaining = remaining;
761 xbt_free(comp_task_name);
765 #define USE_MICRO_TASK 1
768 // const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
769 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
770 // const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
771 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
772 // const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
773 //const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
774 // const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
775 ////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
776 // const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
777 // const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
778 // const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
780 /* CPU 22% when 80Mbyte/s */
781 const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
785 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
786 double size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
788 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
789 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
793 double clock_sta = MSG_get_clock();
795 #ifdef USE_MICRO_TASK
797 task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
802 ret = MSG_task_send_bounded(task, mbox, mig_speed);
804 ret = MSG_task_send(task, mbox);
805 xbt_assert(ret == MSG_OK);
808 double clock_end = MSG_get_clock();
809 double duration = clock_end - clock_sta;
810 double actual_speed = size / duration;
811 #ifdef USE_MICRO_TASK
812 double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
814 double cpu_utilization = 0;
821 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);
823 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);
829 #ifdef USE_MICRO_TASK
830 /* The name of a micro task starts with __micro, which does not match the
831 * special name that finalizes the receiver loop. Thus, we send the special task.
835 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
836 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
837 msg_error_t ret = MSG_task_send(task, mbox);
838 xbt_assert(ret == MSG_OK);
845 static double get_updated_size(double computed, double dp_rate, double dp_cap)
847 double updated_size = computed * dp_rate;
848 XBT_INFO("updated_size %f dp_rate %f", updated_size, dp_rate);
849 if (updated_size > dp_cap) {
850 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
851 updated_size = dp_cap;
857 static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *dst_pm_name,
858 long ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead)
860 const char *vm_name = MSG_host_get_name(vm);
861 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
863 const long chunksize = 1024 * 1024 * 100;
864 long remaining = ramsize;
865 double computed_total = 0;
867 while (remaining > 0) {
868 long datasize = chunksize;
869 if (remaining < chunksize)
870 datasize = remaining;
872 remaining -= datasize;
874 send_migration_data(vm_name, src_pm_name, dst_pm_name, datasize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
876 double computed = lookup_computed_flop_counts(vm, 1, 0);
877 computed_total += computed;
880 double updated_size = get_updated_size(computed, dp_rate, dp_cap);
882 double overhead = dpt_cpu_overhead * updated_size;
883 launch_deferred_exec_process(vm, overhead, 10000);
887 return computed_total;
893 static int migration_tx_fun(int argc, char *argv[])
895 XBT_DEBUG("mig: tx_start");
897 xbt_assert(argc == 4);
898 const char *vm_name = argv[1];
899 const char *src_pm_name = argv[2];
900 const char *dst_pm_name = argv[3];
901 msg_vm_t vm = MSG_get_host_by_name(vm_name);
904 s_ws_params_t params;
905 simcall_host_get_params(vm, ¶ms);
906 const long ramsize = params.ramsize;
907 const long devsize = params.devsize;
908 const int skip_stage1 = params.skip_stage1;
909 const int skip_stage2 = params.skip_stage2;
910 const double dp_rate = params.dp_rate;
911 const double dp_cap = params.dp_cap;
912 const double mig_speed = params.mig_speed;
913 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
914 const double dpt_cpu_overhead = params.dpt_cpu_overhead;
916 double remaining_size = ramsize + devsize;
918 double max_downtime = params.max_downtime;
919 if (max_downtime == 0) {
920 XBT_WARN("use the default max_downtime value 30ms");
924 /* This value assumes the network link is 1Gbps. */
925 double threshold = max_downtime * 125 * 1024 * 1024;
927 /* setting up parameters has done */
931 XBT_WARN("migrate a VM, but ramsize is zero");
933 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
935 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
937 /* Stage1: send all memory pages to the destination. */
938 start_dirty_page_tracking(vm);
940 double computed_during_stage1 = 0;
942 // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
944 /* send ramsize, but split it */
945 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);
946 remaining_size -= ramsize;
950 /* Stage2: send update pages iteratively until the size of remaining states
951 * becomes smaller than the threshold value. */
954 if (max_downtime == 0) {
955 XBT_WARN("no max_downtime parameter, skip stage2");
960 int stage2_round = 0;
963 double updated_size = 0;
964 if (stage2_round == 0) {
965 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
966 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
968 double computed = lookup_computed_flop_counts(vm, 2, stage2_round);
969 updated_size = get_updated_size(computed, dp_rate, dp_cap);
972 XBT_INFO("%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
973 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
976 if (stage2_round != 0) {
977 /* during stage1, we have already created overhead tasks */
978 double overhead = dpt_cpu_overhead * updated_size;
979 XBT_INFO("updated %f overhead %f", updated_size, overhead);
980 launch_deferred_exec_process(vm, overhead, 10000);
985 remaining_size += updated_size;
987 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
988 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
990 if (remaining_size < threshold)
995 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
997 remaining_size -= updated_size;
1003 /* Stage3: stop the VM and copy the rest of states. */
1004 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
1005 simcall_vm_suspend(vm);
1006 stop_dirty_page_tracking(vm);
1008 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead);
1012 XBT_DEBUG("mig: tx_done");
1019 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
1021 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1024 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1026 char **argv = xbt_new(char *, nargvs);
1027 argv[0] = xbt_strdup(pr_name);
1028 argv[1] = xbt_strdup(sg_host_name(vm));
1029 argv[2] = xbt_strdup(sg_host_name(src_pm));
1030 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1033 MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
1039 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1041 char **argv = xbt_new(char *, nargvs);
1042 argv[0] = xbt_strdup(pr_name);
1043 argv[1] = xbt_strdup(sg_host_name(vm));
1044 argv[2] = xbt_strdup(sg_host_name(src_pm));
1045 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1047 MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
1052 /* wait until the migration have finished */
1054 msg_task_t task = NULL;
1055 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
1056 xbt_assert(ret == MSG_OK);
1058 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
1059 xbt_assert(strcmp(task->name, expected_task_name) == 0);
1060 xbt_free(expected_task_name);
1067 /** @brief Migrate the VM to the given host.
1070 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
1071 * MSG_task_send() before or after, depending on whether you want to do cold or hot
1074 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
1077 * - One approach is ...
1078 * We first create a new VM (i.e., destination VM) on the destination
1079 * physical host. The destination VM will receive the state of the source
1080 * VM over network. We will finally destroy the source VM.
1081 * - This behavior is similar to the way of migration in the real world.
1082 * Even before a migration is completed, we will see a destination VM,
1083 * consuming resources.
1084 * - We have to relocate all processes. The existing process migraion code
1085 * will work for this?
1086 * - The name of the VM is a somewhat unique ID in the code. It is tricky
1087 * for the destination VM?
1089 * - Another one is ...
1090 * We update the information of the given VM to place it to the destination
1093 * The second one would be easier.
1097 msg_host_t old_pm = simcall_vm_get_pm(vm);
1099 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
1100 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
1102 do_migration(vm, old_pm, new_pm);
1106 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1109 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1114 /** @brief Immediately suspend the execution of all processes within the given VM.
1117 * This function stops the exection of the VM. All the processes on this VM
1118 * will pause. The state of the VM is perserved. We can later resume it again.
1120 * No suspension cost occurs.
1122 void MSG_vm_suspend(msg_vm_t vm)
1124 simcall_vm_suspend(vm);
1126 XBT_DEBUG("vm_suspend done");
1129 TRACE_msg_vm_suspend(vm);
1134 /** @brief Resume the execution of the VM. All processes on the VM run again.
1137 * No resume cost occurs.
1139 void MSG_vm_resume(msg_vm_t vm)
1141 simcall_vm_resume(vm);
1144 TRACE_msg_vm_resume(vm);
1149 /** @brief Immediately save the execution of all processes within the given VM.
1152 * This function stops the exection of the VM. All the processes on this VM
1153 * will pause. The state of the VM is perserved. We can later resume it again.
1155 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1156 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1157 * of VM save to you.
1159 void MSG_vm_save(msg_vm_t vm)
1161 simcall_vm_save(vm);
1163 TRACE_msg_vm_save(vm);
1167 /** @brief Restore the execution of the VM. All processes on the VM run again.
1170 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1171 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1172 * of VM restore to you.
1174 void MSG_vm_restore(msg_vm_t vm)
1176 simcall_vm_restore(vm);
1179 TRACE_msg_vm_restore(vm);
1184 /** @brief Get the physical host of a given VM.
1187 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1189 return simcall_vm_get_pm(vm);
1193 /** @brief Set a CPU bound for a given VM.
1197 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1200 * On PM0, there are Task1 and VM0.
1201 * On VM0, there is Task2.
1202 * Now we bound 75% to Task1@PM0 and bound 25% to Task2@VM0.
1204 * Task1@PM0 gets 50%.
1205 * Task2@VM0 gets 25%.
1206 * This is NOT 75% for Task1@PM0 and 25% for Task2@VM0, respectively.
1208 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1209 * task on the VM does not affect the bound of the dummy CPU action. The bound
1210 * of the dummy CPU action is unlimited.
1212 * There are some solutions for this problem. One option is to update the bound
1213 * of the dummy CPU action automatically. It should be the sum of all tasks on
1214 * the VM. But, this solution might be costy, because we have to scan all tasks
1215 * on the VM in share_resource() or we have to trap both the start and end of
1218 * The current solution is to use MSG_vm_set_bound(), which allows us to
1219 * directly set the bound of the dummy CPU action.
1223 * Note that bound == 0 means no bound (i.e., unlimited).
1225 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1227 return simcall_vm_set_bound(vm, bound);