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.
171 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name,
172 int ncpus, long ramsize, long net_cap, char *disk_path, long disksize)
174 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 = ramsize;
180 //params.overcommit = 0;
181 simcall_host_set_params(vm, ¶ms);
184 /* TODO: Limit net capability, take into account disk considerations. */
190 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
193 * A VM is treated as a host. The name of the VM must be unique among all hosts.
195 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
197 /* make sure the VM of the same name does not exit */
199 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
201 XBT_ERROR("host %s already exits", name);
206 /* Note: ind_vm and vm_workstation point to the same elm object. */
207 msg_vm_t ind_vm = NULL;
208 void *ind_vm_workstation = NULL;
210 /* Ask the SIMIX layer to create the surf vm resource */
211 ind_vm_workstation = simcall_vm_create(name, ind_pm);
212 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
214 XBT_DEBUG("A new VM (%s) has been created", name);
217 TRACE_msg_vm_create(name, ind_pm);
223 /** @brief Destroy a VM. Destroy the VM object from the simulation.
226 void MSG_vm_destroy(msg_vm_t vm)
228 /* First, terminate all processes on the VM if necessary */
229 if (MSG_vm_is_running(vm))
230 simcall_vm_shutdown(vm);
232 if (!MSG_vm_is_created(vm)) {
233 XBT_CRITICAL("shutdown the given VM before destroying it");
237 /* Then, destroy the VM object */
238 simcall_vm_destroy(vm);
240 __MSG_host_destroy(vm);
243 TRACE_msg_vm_end(vm);
248 /** @brief Start a vm (i.e., boot the guest operating system)
251 * If the VM cannot be started, an exception is generated.
254 void MSG_vm_start(msg_vm_t vm)
256 simcall_vm_start(vm);
259 TRACE_msg_vm_start(vm);
265 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
268 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
269 * use a #MSG_process_sleep() or something. I'm not quite sure.
271 void MSG_vm_shutdown(msg_vm_t vm)
273 /* msg_vm_t equals to msg_host_t */
274 simcall_vm_shutdown(vm);
276 // #ifdef HAVE_TRACING
277 // TRACE_msg_vm_(vm);
283 /* We have two mailboxes. mbox is used to transfer migration data between
284 * source and destiantion PMs. mbox_ctl is used to detect the completion of a
285 * migration. The names of these mailboxes must not conflict with others. */
286 static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
288 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
291 static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
293 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
296 static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
298 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
301 static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
303 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
306 static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
308 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
311 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio);
313 static int migration_rx_fun(int argc, char *argv[])
315 XBT_DEBUG("mig: rx_start");
317 xbt_assert(argc == 4);
318 const char *vm_name = argv[1];
319 const char *src_pm_name = argv[2];
320 const char *dst_pm_name = argv[3];
321 msg_vm_t vm = MSG_get_host_by_name(vm_name);
322 msg_vm_t dst_pm = MSG_get_host_by_name(dst_pm_name);
325 s_ws_params_t params;
326 simcall_host_get_params(vm, ¶ms);
327 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
332 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
333 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
334 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
337 msg_task_t task = NULL;
338 MSG_task_recv(&task, mbox);
340 double received = MSG_task_get_data_size(task);
342 // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
343 launch_deferred_exec_process(vm, received * xfer_cpu_overhead, 1);
346 if (strcmp(task->name, finalize_task_name) == 0)
349 MSG_task_destroy(task);
356 simcall_vm_migrate(vm, dst_pm);
357 simcall_vm_resume(vm);
360 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
362 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
363 msg_error_t ret = MSG_task_send(task, mbox_ctl);
364 xbt_assert(ret == MSG_OK);
372 xbt_free(finalize_task_name);
374 XBT_DEBUG("mig: rx_done");
379 static void reset_dirty_pages(msg_vm_t vm)
381 msg_host_priv_t priv = msg_host_resource_priv(vm);
384 xbt_dict_cursor_t cursor = NULL;
385 dirty_page_t dp = NULL;
386 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
387 double remaining = MSG_task_get_remaining_computation(dp->task);
388 dp->prev_clock = MSG_get_clock();
389 dp->prev_remaining = remaining;
391 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
395 static void start_dirty_page_tracking(msg_vm_t vm)
397 msg_host_priv_t priv = msg_host_resource_priv(vm);
398 priv->dp_enabled = 1;
400 reset_dirty_pages(vm);
403 static void stop_dirty_page_tracking(msg_vm_t vm)
405 msg_host_priv_t priv = msg_host_resource_priv(vm);
406 priv->dp_enabled = 0;
410 /* It might be natural that we define dp_rate for each task. But, we will also
411 * have to care about how each task behavior affects the memory update behavior
412 * at the operating system level. It may not be easy to model it with a simple algorithm. */
413 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
415 double computed = dp->prev_remaining - remaining;
416 double duration = clock - dp->prev_clock;
417 double updated = dp->task->dp_rate * computed;
419 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
420 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
421 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
422 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
428 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
430 double computed = dp->prev_remaining - remaining;
431 double duration = clock - dp->prev_clock;
433 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
434 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
439 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
441 msg_host_priv_t priv = msg_host_resource_priv(vm);
445 xbt_dict_cursor_t cursor = NULL;
446 dirty_page_t dp = NULL;
447 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
448 double remaining = MSG_task_get_remaining_computation(dp->task);
449 double clock = MSG_get_clock();
451 // total += calc_updated_pages(key, vm, dp, remaining, clock);
452 total += get_computed(key, vm, dp, remaining, clock);
454 dp->prev_remaining = remaining;
455 dp->prev_clock = clock;
458 total += priv->dp_updated_by_deleted_tasks;
460 XBT_INFO("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
461 stage_for_fancy_debug,
462 stage2_round_for_fancy_debug,
463 total, priv->dp_updated_by_deleted_tasks);
467 priv->dp_updated_by_deleted_tasks = 0;
473 // TODO Is this code redundant with the information provided by
474 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
475 void MSG_host_add_task(msg_host_t host, msg_task_t task)
477 msg_host_priv_t priv = msg_host_resource_priv(host);
478 double remaining = MSG_task_get_remaining_computation(task);
479 char *key = bprintf("%s-%lld", task->name, task->counter);
481 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
484 /* It should be okay that we add a task onto a migrating VM. */
485 if (priv->dp_enabled) {
486 dp->prev_clock = MSG_get_clock();
487 dp->prev_remaining = remaining;
490 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
491 xbt_dict_set(priv->dp_objs, key, dp, NULL);
492 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
497 void MSG_host_del_task(msg_host_t host, msg_task_t task)
499 msg_host_priv_t priv = msg_host_resource_priv(host);
501 char *key = bprintf("%s-%lld", task->name, task->counter);
503 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
504 xbt_assert(dp->task == task);
506 /* If we are in the middle of dirty page tracking, we record how much
507 * computaion has been done until now, and keep the information for the
508 * lookup_() function that will called soon. */
509 if (priv->dp_enabled) {
510 double remaining = MSG_task_get_remaining_computation(task);
511 double clock = MSG_get_clock();
512 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
513 double updated = get_computed(key, host, dp, remaining, clock);
515 priv->dp_updated_by_deleted_tasks += updated;
518 xbt_dict_remove(priv->dp_objs, key);
521 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
527 static int deferred_exec_fun(int argc, char *argv[])
529 xbt_assert(argc == 3);
530 const char *comp_str = argv[1];
531 double computaion = atof(comp_str);
532 const char *prio_str = argv[2];
533 double prio = atof(prio_str);
535 msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
536 // XBT_INFO("exec deferred %f", computaion);
538 /* dpt is the results of the VM activity */
539 MSG_task_set_priority(task, prio);
540 MSG_task_execute(task);
544 MSG_task_destroy(task);
549 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
551 char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
554 char **argv = xbt_new(char *, nargvs);
555 argv[0] = xbt_strdup(pr_name);
556 argv[1] = bprintf("%lf", computation);
557 argv[2] = bprintf("%lf", prio);
560 MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv);
566 static int task_tx_overhead_fun(int argc, char *argv[])
568 xbt_assert(argc == 2);
569 const char *mbox = argv[1];
573 // XBT_INFO("start %s", mbox);
576 msg_task_t task = NULL;
577 MSG_task_recv(&task, mbox);
579 // XBT_INFO("task->name %s", task->name);
581 if (strcmp(task->name, "finalize_making_overhead") == 0)
585 // MSG_task_set_priority(task, 1000000);
586 MSG_task_execute(task);
587 MSG_task_destroy(task);
598 static void start_overhead_process(msg_task_t comm_task)
600 char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task));
601 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
604 char **argv = xbt_new(char *, nargvs);
605 argv[0] = xbt_strdup(pr_name);
606 argv[1] = xbt_strdup(mbox);
609 // XBT_INFO("micro start: mbox %s", mbox);
610 MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv);
616 static void shutdown_overhead_process(msg_task_t comm_task)
618 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
620 msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
622 // XBT_INFO("micro shutdown: mbox %s", mbox);
623 msg_error_t ret = MSG_task_send(task, mbox);
624 xbt_assert(ret == MSG_OK);
627 // XBT_INFO("shutdown done");
630 static void request_overhead(msg_task_t comm_task, double computation)
632 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
634 msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
636 // XBT_INFO("req overhead");
637 msg_error_t ret = MSG_task_send(task, mbox);
638 xbt_assert(ret == MSG_OK);
643 /* alpha is (floating_operations / bytes).
645 * When actual migration traffic was 32 mbytes/s, we observed the CPU
646 * utilization of the main thread of the Qemu process was 10 %.
647 * alpha = 0.1 * C / (32 * 1024 * 1024)
648 * where the CPU capacity of the PM is C flops/s.
651 static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
653 const double chunk_size = 1024 * 1024 * 10;
654 double remaining = MSG_task_get_data_size(comm_task);
656 start_overhead_process(comm_task);
659 while (remaining > 0) {
660 double data_size = chunk_size;
661 if (remaining < chunk_size)
662 data_size = remaining;
664 remaining -= data_size;
666 // XBT_INFO("remaining %f bytes", remaining);
669 double clock_sta = MSG_get_clock();
671 /* create a micro task */
673 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
674 msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
676 request_overhead(comm_task, data_size * alpha);
678 msg_error_t ret = MSG_task_send(mtask, mbox);
679 xbt_assert(ret == MSG_OK);
681 xbt_free(mtask_name);
686 /* In the real world, sending data involves small CPU computation. */
687 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
688 msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
689 MSG_task_execute(mtask);
690 MSG_task_destroy(mtask);
691 xbt_free(mtask_name);
697 double clock_end = MSG_get_clock();
702 * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
705 * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
707 * If time_to_sleep is smaller than zero, the elapsed time was too big. We
708 * do not need a micro sleep.
710 double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
711 if (time_to_sleep > 0)
712 MSG_process_sleep(time_to_sleep);
715 //XBT_INFO("duration %f", clock_end - clock_sta);
716 //XBT_INFO("time_to_sleep %f", time_to_sleep);
720 // XBT_INFO("%s", MSG_task_get_name(comm_task));
721 shutdown_overhead_process(comm_task);
727 static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
729 double prev_remaining = init_comm_size;
732 double remaining = MSG_task_get_remaining_communication(comm_task);
736 double sent = prev_remaining - remaining;
737 double comp_size = sent * overhead;
740 char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
741 msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
742 MSG_task_execute(comp_task);
743 MSG_task_destroy(comp_task);
748 prev_remaining = remaining;
752 xbt_free(comp_task_name);
756 #define USE_MICRO_TASK 1
759 // const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
760 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
761 // const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
762 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
763 // const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
764 //const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
765 // const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
766 ////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
767 // const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
768 // const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
769 // const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
771 /* CPU 22% when 80Mbyte/s */
772 const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
776 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
777 double size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
779 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
780 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
784 double clock_sta = MSG_get_clock();
786 #ifdef USE_MICRO_TASK
788 task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
793 ret = MSG_task_send_bounded(task, mbox, mig_speed);
795 ret = MSG_task_send(task, mbox);
796 xbt_assert(ret == MSG_OK);
799 double clock_end = MSG_get_clock();
800 double duration = clock_end - clock_sta;
801 double actual_speed = size / duration;
802 #ifdef USE_MICRO_TASK
803 double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
805 double cpu_utilization = 0;
812 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);
814 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);
820 #ifdef USE_MICRO_TASK
821 /* The name of a micro task starts with __micro, which does not match the
822 * special name that finalizes the receiver loop. Thus, we send the special task.
826 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
827 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
828 msg_error_t ret = MSG_task_send(task, mbox);
829 xbt_assert(ret == MSG_OK);
836 static double get_updated_size(double computed, double dp_rate, double dp_cap)
838 double updated_size = computed * dp_rate;
839 XBT_INFO("updated_size %f dp_rate %f", updated_size, dp_rate);
840 if (updated_size > dp_cap) {
841 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
842 updated_size = dp_cap;
848 static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *dst_pm_name,
849 long ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead)
851 const char *vm_name = MSG_host_get_name(vm);
852 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
854 const long chunksize = 1024 * 1024 * 100;
855 long remaining = ramsize;
856 double computed_total = 0;
858 while (remaining > 0) {
859 long datasize = chunksize;
860 if (remaining < chunksize)
861 datasize = remaining;
863 remaining -= datasize;
865 send_migration_data(vm_name, src_pm_name, dst_pm_name, datasize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
867 double computed = lookup_computed_flop_counts(vm, 1, 0);
868 computed_total += computed;
871 double updated_size = get_updated_size(computed, dp_rate, dp_cap);
873 double overhead = dpt_cpu_overhead * updated_size;
874 launch_deferred_exec_process(vm, overhead, 10000);
878 return computed_total;
884 static int migration_tx_fun(int argc, char *argv[])
886 XBT_DEBUG("mig: tx_start");
888 xbt_assert(argc == 4);
889 const char *vm_name = argv[1];
890 const char *src_pm_name = argv[2];
891 const char *dst_pm_name = argv[3];
892 msg_vm_t vm = MSG_get_host_by_name(vm_name);
895 s_ws_params_t params;
896 simcall_host_get_params(vm, ¶ms);
897 const long ramsize = params.ramsize;
898 const long devsize = params.devsize;
899 const int skip_stage1 = params.skip_stage1;
900 const int skip_stage2 = params.skip_stage2;
901 const double dp_rate = params.dp_rate;
902 const double dp_cap = params.dp_cap;
903 const double mig_speed = params.mig_speed;
904 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
905 const double dpt_cpu_overhead = params.dpt_cpu_overhead;
907 double remaining_size = ramsize + devsize;
909 double max_downtime = params.max_downtime;
910 if (max_downtime == 0) {
911 XBT_WARN("use the default max_downtime value 30ms");
915 /* This value assumes the network link is 1Gbps. */
916 double threshold = max_downtime * 125 * 1024 * 1024;
918 /* setting up parameters has done */
922 XBT_WARN("migrate a VM, but ramsize is zero");
924 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
926 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
928 /* Stage1: send all memory pages to the destination. */
929 start_dirty_page_tracking(vm);
931 double computed_during_stage1 = 0;
933 // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
935 /* send ramsize, but split it */
936 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);
937 remaining_size -= ramsize;
941 /* Stage2: send update pages iteratively until the size of remaining states
942 * becomes smaller than the threshold value. */
945 if (max_downtime == 0) {
946 XBT_WARN("no max_downtime parameter, skip stage2");
951 int stage2_round = 0;
954 double updated_size = 0;
955 if (stage2_round == 0) {
956 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
957 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
959 double computed = lookup_computed_flop_counts(vm, 2, stage2_round);
960 updated_size = get_updated_size(computed, dp_rate, dp_cap);
963 XBT_INFO("%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
964 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
967 if (stage2_round != 0) {
968 /* during stage1, we have already created overhead tasks */
969 double overhead = dpt_cpu_overhead * updated_size;
970 XBT_INFO("updated %f overhead %f", updated_size, overhead);
971 launch_deferred_exec_process(vm, overhead, 10000);
976 remaining_size += updated_size;
978 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
979 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
981 if (remaining_size < threshold)
986 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
988 remaining_size -= updated_size;
994 /* Stage3: stop the VM and copy the rest of states. */
995 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
996 simcall_vm_suspend(vm);
997 stop_dirty_page_tracking(vm);
999 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead);
1003 XBT_DEBUG("mig: tx_done");
1010 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
1012 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1015 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1017 char **argv = xbt_new(char *, nargvs);
1018 argv[0] = xbt_strdup(pr_name);
1019 argv[1] = xbt_strdup(sg_host_name(vm));
1020 argv[2] = xbt_strdup(sg_host_name(src_pm));
1021 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1024 MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
1030 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1032 char **argv = xbt_new(char *, nargvs);
1033 argv[0] = xbt_strdup(pr_name);
1034 argv[1] = xbt_strdup(sg_host_name(vm));
1035 argv[2] = xbt_strdup(sg_host_name(src_pm));
1036 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1038 MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
1043 /* wait until the migration have finished */
1045 msg_task_t task = NULL;
1046 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
1047 xbt_assert(ret == MSG_OK);
1049 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
1050 xbt_assert(strcmp(task->name, expected_task_name) == 0);
1051 xbt_free(expected_task_name);
1058 /** @brief Migrate the VM to the given host.
1061 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
1062 * MSG_task_send() before or after, depending on whether you want to do cold or hot
1065 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
1068 * - One approach is ...
1069 * We first create a new VM (i.e., destination VM) on the destination
1070 * physical host. The destination VM will receive the state of the source
1071 * VM over network. We will finally destroy the source VM.
1072 * - This behavior is similar to the way of migration in the real world.
1073 * Even before a migration is completed, we will see a destination VM,
1074 * consuming resources.
1075 * - We have to relocate all processes. The existing process migraion code
1076 * will work for this?
1077 * - The name of the VM is a somewhat unique ID in the code. It is tricky
1078 * for the destination VM?
1080 * - Another one is ...
1081 * We update the information of the given VM to place it to the destination
1084 * The second one would be easier.
1088 msg_host_t old_pm = simcall_vm_get_pm(vm);
1090 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
1091 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
1093 do_migration(vm, old_pm, new_pm);
1097 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1100 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1105 /** @brief Immediately suspend the execution of all processes within the given VM.
1108 * This function stops the exection of the VM. All the processes on this VM
1109 * will pause. The state of the VM is perserved. We can later resume it again.
1111 * No suspension cost occurs.
1113 void MSG_vm_suspend(msg_vm_t vm)
1115 simcall_vm_suspend(vm);
1117 XBT_DEBUG("vm_suspend done");
1120 TRACE_msg_vm_suspend(vm);
1125 /** @brief Resume the execution of the VM. All processes on the VM run again.
1128 * No resume cost occurs.
1130 void MSG_vm_resume(msg_vm_t vm)
1132 simcall_vm_resume(vm);
1135 TRACE_msg_vm_resume(vm);
1140 /** @brief Immediately save the execution of all processes within the given VM.
1143 * This function stops the exection of the VM. All the processes on this VM
1144 * will pause. The state of the VM is perserved. We can later resume it again.
1146 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1147 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1148 * of VM save to you.
1150 void MSG_vm_save(msg_vm_t vm)
1152 simcall_vm_save(vm);
1154 TRACE_msg_vm_save(vm);
1158 /** @brief Restore the execution of the VM. All processes on the VM run again.
1161 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1162 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1163 * of VM restore to you.
1165 void MSG_vm_restore(msg_vm_t vm)
1167 simcall_vm_restore(vm);
1170 TRACE_msg_vm_restore(vm);
1175 /** @brief Get the physical host of a given VM.
1178 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1180 return simcall_vm_get_pm(vm);
1184 /** @brief Set a CPU bound for a given VM.
1187 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1190 * On PM0, there are Task1 and VM0.
1191 * On VM0, there is Task2.
1192 * Now we bound 75% to Task1@PM0 and bound 25% to Task2@VM0.
1194 * Task1@PM0 gets 50%.
1195 * Task2@VM0 gets 25%.
1196 * This is NOT 75% for Task1@PM0 and 25% for Task2@VM0, respectively.
1198 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1199 * task on the VM does not affect the bound of the dummy CPU action. The bound
1200 * of the dummy CPU action is unlimited.
1202 * There are some solutions for this problem. One option is to update the bound
1203 * of the dummy CPU action automatically. It should be the sum of all tasks on
1204 * the VM. But, this solution might be costy, because we have to scan all tasks
1205 * on the VM in share_resource() or we have to trap both the start and end of
1208 * The current solution is to use MSG_vm_set_bound(), which allows us to
1209 * directly set the bound of the dummy CPU action.
1211 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1213 return simcall_vm_set_bound(vm, bound);