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 const char *pr_name = MSG_process_get_name(MSG_process_self());
316 const char *host_name = MSG_host_get_name(MSG_host_self());
318 XBT_DEBUG("mig: rx_start");
320 xbt_assert(argc == 4);
321 const char *vm_name = argv[1];
322 const char *src_pm_name = argv[2];
323 const char *dst_pm_name = argv[3];
324 msg_vm_t vm = MSG_get_host_by_name(vm_name);
325 msg_vm_t dst_pm = MSG_get_host_by_name(dst_pm_name);
328 s_ws_params_t params;
329 simcall_host_get_params(vm, ¶ms);
330 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
335 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
336 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
337 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
340 msg_task_t task = NULL;
341 MSG_task_recv(&task, mbox);
343 double received = MSG_task_get_data_size(task);
345 // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
346 launch_deferred_exec_process(vm, received * xfer_cpu_overhead, 1);
349 if (strcmp(task->name, finalize_task_name) == 0)
352 MSG_task_destroy(task);
359 simcall_vm_migrate(vm, dst_pm);
360 simcall_vm_resume(vm);
363 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
365 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
366 msg_error_t ret = MSG_task_send(task, mbox_ctl);
367 xbt_assert(ret == MSG_OK);
375 xbt_free(finalize_task_name);
377 XBT_DEBUG("mig: rx_done");
383 typedef struct dirty_page {
385 double prev_remaining;
387 } s_dirty_page, *dirty_page_t;
390 static void reset_dirty_pages(msg_vm_t vm)
392 msg_host_priv_t priv = msg_host_resource_priv(vm);
395 xbt_dict_cursor_t cursor = NULL;
396 dirty_page_t dp = NULL;
397 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
398 double remaining = MSG_task_get_remaining_computation(dp->task);
399 dp->prev_clock = MSG_get_clock();
400 dp->prev_remaining = remaining;
402 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
406 static void start_dirty_page_tracking(msg_vm_t vm)
408 msg_host_priv_t priv = msg_host_resource_priv(vm);
409 priv->dp_enabled = 1;
411 reset_dirty_pages(vm);
414 static void stop_dirty_page_tracking(msg_vm_t vm)
416 msg_host_priv_t priv = msg_host_resource_priv(vm);
417 priv->dp_enabled = 0;
421 /* It might be natural that we define dp_rate for each task. But, we will also
422 * have to care about how each task behavior affects the memory update behavior
423 * at the operating system level. It may not be easy to model it with a simple algorithm. */
424 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
426 double computed = dp->prev_remaining - remaining;
427 double duration = clock - dp->prev_clock;
428 double updated = dp->task->dp_rate * computed;
430 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
431 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
432 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
433 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
439 double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
441 double computed = dp->prev_remaining - remaining;
442 double duration = clock - dp->prev_clock;
444 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
445 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
450 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
452 msg_host_priv_t priv = msg_host_resource_priv(vm);
456 xbt_dict_cursor_t cursor = NULL;
457 dirty_page_t dp = NULL;
458 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
459 double remaining = MSG_task_get_remaining_computation(dp->task);
460 double clock = MSG_get_clock();
462 // total += calc_updated_pages(key, vm, dp, remaining, clock);
463 total += get_computed(key, vm, dp, remaining, clock);
465 dp->prev_remaining = remaining;
466 dp->prev_clock = clock;
469 total += priv->dp_updated_by_deleted_tasks;
471 XBT_INFO("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
472 stage_for_fancy_debug,
473 stage2_round_for_fancy_debug,
474 total, priv->dp_updated_by_deleted_tasks);
478 priv->dp_updated_by_deleted_tasks = 0;
484 // TODO Is this code redundant with the information provided by
485 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
486 void MSG_host_add_task(msg_host_t host, msg_task_t task)
488 msg_host_priv_t priv = msg_host_resource_priv(host);
489 double remaining = MSG_task_get_remaining_computation(task);
490 char *key = bprintf("%s-%lld", task->name, task->counter);
492 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
495 /* It should be okay that we add a task onto a migrating VM. */
496 if (priv->dp_enabled) {
497 dp->prev_clock = MSG_get_clock();
498 dp->prev_remaining = remaining;
501 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
502 xbt_dict_set(priv->dp_objs, key, dp, NULL);
503 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
508 void MSG_host_del_task(msg_host_t host, msg_task_t task)
510 msg_host_priv_t priv = msg_host_resource_priv(host);
512 char *key = bprintf("%s-%lld", task->name, task->counter);
514 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
515 xbt_assert(dp->task == task);
517 /* If we are in the middle of dirty page tracking, we record how much
518 * computaion has been done until now, and keep the information for the
519 * lookup_() function that will called soon. */
520 if (priv->dp_enabled) {
521 double remaining = MSG_task_get_remaining_computation(task);
522 double clock = MSG_get_clock();
523 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
524 double updated = get_computed(key, host, dp, remaining, clock);
526 priv->dp_updated_by_deleted_tasks += updated;
529 xbt_dict_remove(priv->dp_objs, key);
532 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
538 static int deferred_exec_fun(int argc, char *argv[])
540 xbt_assert(argc == 3);
541 const char *comp_str = argv[1];
542 double computaion = atof(comp_str);
543 const char *prio_str = argv[2];
544 double prio = atof(prio_str);
546 msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
547 // XBT_INFO("exec deferred %f", computaion);
549 /* dpt is the results of the VM activity */
550 MSG_task_set_priority(task, prio);
551 MSG_task_execute(task);
555 MSG_task_destroy(task);
560 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
562 char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
565 char **argv = xbt_new(char *, nargvs);
566 argv[0] = xbt_strdup(pr_name);
567 argv[1] = bprintf("%lf", computation);
568 argv[2] = bprintf("%lf", prio);
571 msg_process_t pr = MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv);
577 static int task_tx_overhead_fun(int argc, char *argv[])
579 xbt_assert(argc == 2);
580 const char *mbox = argv[1];
584 // XBT_INFO("start %s", mbox);
587 msg_task_t task = NULL;
588 MSG_task_recv(&task, mbox);
590 // XBT_INFO("task->name %s", task->name);
592 if (strcmp(task->name, "finalize_making_overhead") == 0)
596 // MSG_task_set_priority(task, 1000000);
597 MSG_task_execute(task);
598 MSG_task_destroy(task);
609 static void start_overhead_process(msg_task_t comm_task)
611 char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task));
612 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
615 char **argv = xbt_new(char *, nargvs);
616 argv[0] = xbt_strdup(pr_name);
617 argv[1] = xbt_strdup(mbox);
620 // XBT_INFO("micro start: mbox %s", mbox);
621 msg_process_t pr = MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv);
627 static void shutdown_overhead_process(msg_task_t comm_task)
629 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
631 msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
633 // XBT_INFO("micro shutdown: mbox %s", mbox);
634 msg_error_t ret = MSG_task_send(task, mbox);
635 xbt_assert(ret == MSG_OK);
638 // XBT_INFO("shutdown done");
641 static void request_overhead(msg_task_t comm_task, double computation)
643 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
645 msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
647 // XBT_INFO("req overhead");
648 msg_error_t ret = MSG_task_send(task, mbox);
649 xbt_assert(ret == MSG_OK);
654 /* alpha is (floating_operations / bytes).
656 * When actual migration traffic was 32 mbytes/s, we observed the CPU
657 * utilization of the main thread of the Qemu process was 10 %.
658 * alpha = 0.1 * C / (32 * 1024 * 1024)
659 * where the CPU capacity of the PM is C flops/s.
662 static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
664 const double chunk_size = 1024 * 1024 * 10;
665 double remaining = MSG_task_get_data_size(comm_task);
667 start_overhead_process(comm_task);
670 while (remaining > 0) {
671 double data_size = chunk_size;
672 if (remaining < chunk_size)
673 data_size = remaining;
675 remaining -= data_size;
677 // XBT_INFO("remaining %f bytes", remaining);
680 double clock_sta = MSG_get_clock();
682 /* create a micro task */
684 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
685 msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
687 request_overhead(comm_task, data_size * alpha);
689 msg_error_t ret = MSG_task_send(mtask, mbox);
690 xbt_assert(ret == MSG_OK);
692 xbt_free(mtask_name);
697 /* In the real world, sending data involves small CPU computation. */
698 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
699 msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
700 MSG_task_execute(mtask);
701 MSG_task_destroy(mtask);
702 xbt_free(mtask_name);
708 double clock_end = MSG_get_clock();
713 * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
716 * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
718 * If time_to_sleep is smaller than zero, the elapsed time was too big. We
719 * do not need a micro sleep.
721 double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
722 if (time_to_sleep > 0)
723 MSG_process_sleep(time_to_sleep);
726 //XBT_INFO("duration %f", clock_end - clock_sta);
727 //XBT_INFO("time_to_sleep %f", time_to_sleep);
731 // XBT_INFO("%s", MSG_task_get_name(comm_task));
732 shutdown_overhead_process(comm_task);
738 static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
740 double prev_remaining = init_comm_size;
743 double remaining = MSG_task_get_remaining_communication(comm_task);
747 double sent = prev_remaining - remaining;
748 double comp_size = sent * overhead;
751 char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
752 msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
753 MSG_task_execute(comp_task);
754 MSG_task_destroy(comp_task);
759 prev_remaining = remaining;
763 xbt_free(comp_task_name);
767 #define USE_MICRO_TASK 1
770 // const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
771 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
772 // const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
773 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
774 // const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
775 //const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
776 // const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
777 ////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
778 // const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
779 // const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
780 // const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
782 /* CPU 22% when 80Mbyte/s */
783 const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
787 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
788 double size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
790 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
791 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
795 double clock_sta = MSG_get_clock();
797 #ifdef USE_MICRO_TASK
799 task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
804 ret = MSG_task_send_bounded(task, mbox, mig_speed);
806 ret = MSG_task_send(task, mbox);
807 xbt_assert(ret == MSG_OK);
810 double clock_end = MSG_get_clock();
811 double duration = clock_end - clock_sta;
812 double actual_speed = size / duration;
813 #ifdef USE_MICRO_TASK
814 double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
816 double cpu_utilization = 0;
823 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);
825 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);
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 double get_updated_size(double computed, double dp_rate, double dp_cap)
849 double updated_size = computed * dp_rate;
850 XBT_INFO("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 const char *pr_name = MSG_process_get_name(MSG_process_self());
898 const char *host_name = MSG_host_get_name(MSG_host_self());
900 XBT_DEBUG("mig: tx_start");
902 xbt_assert(argc == 4);
903 const char *vm_name = argv[1];
904 const char *src_pm_name = argv[2];
905 const char *dst_pm_name = argv[3];
906 msg_vm_t vm = MSG_get_host_by_name(vm_name);
909 s_ws_params_t params;
910 simcall_host_get_params(vm, ¶ms);
911 const long ramsize = params.ramsize;
912 const long devsize = params.devsize;
913 const int skip_stage1 = params.skip_stage1;
914 const int skip_stage2 = params.skip_stage2;
915 const double dp_rate = params.dp_rate;
916 const double dp_cap = params.dp_cap;
917 const double mig_speed = params.mig_speed;
918 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
919 const double dpt_cpu_overhead = params.dpt_cpu_overhead;
921 double remaining_size = ramsize + devsize;
923 double max_downtime = params.max_downtime;
924 if (max_downtime == 0) {
925 XBT_WARN("use the default max_downtime value 30ms");
929 /* This value assumes the network link is 1Gbps. */
930 double threshold = max_downtime * 125 * 1024 * 1024;
932 /* setting up parameters has done */
936 XBT_WARN("migrate a VM, but ramsize is zero");
938 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
940 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
942 /* Stage1: send all memory pages to the destination. */
943 start_dirty_page_tracking(vm);
945 double computed_during_stage1 = 0;
947 // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
949 /* send ramsize, but split it */
950 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);
951 remaining_size -= ramsize;
955 /* Stage2: send update pages iteratively until the size of remaining states
956 * becomes smaller than the threshold value. */
959 if (max_downtime == 0) {
960 XBT_WARN("no max_downtime parameter, skip stage2");
965 int stage2_round = 0;
968 double updated_size = 0;
969 if (stage2_round == 0) {
970 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
971 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
973 double computed = lookup_computed_flop_counts(vm, 2, stage2_round);
974 updated_size = get_updated_size(computed, dp_rate, dp_cap);
977 XBT_INFO("%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
978 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
981 if (stage2_round != 0) {
982 /* during stage1, we have already created overhead tasks */
983 double overhead = dpt_cpu_overhead * updated_size;
984 XBT_INFO("updated %f overhead %f", updated_size, overhead);
985 launch_deferred_exec_process(vm, overhead, 10000);
990 remaining_size += updated_size;
992 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
993 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
995 if (remaining_size < threshold)
1000 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
1002 remaining_size -= updated_size;
1008 /* Stage3: stop the VM and copy the rest of states. */
1009 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
1010 simcall_vm_suspend(vm);
1011 stop_dirty_page_tracking(vm);
1013 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead);
1017 XBT_DEBUG("mig: tx_done");
1024 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
1026 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1029 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1031 char **argv = xbt_new(char *, nargvs);
1032 argv[0] = xbt_strdup(pr_name);
1033 argv[1] = xbt_strdup(sg_host_name(vm));
1034 argv[2] = xbt_strdup(sg_host_name(src_pm));
1035 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1038 msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
1044 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1046 char **argv = xbt_new(char *, nargvs);
1047 argv[0] = xbt_strdup(pr_name);
1048 argv[1] = xbt_strdup(sg_host_name(vm));
1049 argv[2] = xbt_strdup(sg_host_name(src_pm));
1050 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1052 msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
1057 /* wait until the migration have finished */
1059 msg_task_t task = NULL;
1060 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
1061 xbt_assert(ret == MSG_OK);
1063 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
1064 xbt_assert(strcmp(task->name, expected_task_name) == 0);
1065 xbt_free(expected_task_name);
1072 /** @brief Migrate the VM to the given host.
1075 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
1076 * MSG_task_send() before or after, depending on whether you want to do cold or hot
1079 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
1082 * - One approach is ...
1083 * We first create a new VM (i.e., destination VM) on the destination
1084 * physical host. The destination VM will receive the state of the source
1085 * VM over network. We will finally destroy the source VM.
1086 * - This behavior is similar to the way of migration in the real world.
1087 * Even before a migration is completed, we will see a destination VM,
1088 * consuming resources.
1089 * - We have to relocate all processes. The existing process migraion code
1090 * will work for this?
1091 * - The name of the VM is a somewhat unique ID in the code. It is tricky
1092 * for the destination VM?
1094 * - Another one is ...
1095 * We update the information of the given VM to place it to the destination
1098 * The second one would be easier.
1102 msg_host_t old_pm = simcall_vm_get_pm(vm);
1104 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
1105 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
1107 do_migration(vm, old_pm, new_pm);
1111 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1114 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1119 /** @brief Immediately suspend the execution of all processes within the given VM.
1122 * This function stops the exection of the VM. All the processes on this VM
1123 * will pause. The state of the VM is perserved. We can later resume it again.
1125 * No suspension cost occurs.
1127 void MSG_vm_suspend(msg_vm_t vm)
1129 simcall_vm_suspend(vm);
1131 XBT_DEBUG("vm_suspend done");
1134 TRACE_msg_vm_suspend(vm);
1139 /** @brief Resume the execution of the VM. All processes on the VM run again.
1142 * No resume cost occurs.
1144 void MSG_vm_resume(msg_vm_t vm)
1146 simcall_vm_resume(vm);
1149 TRACE_msg_vm_resume(vm);
1154 /** @brief Immediately save the execution of all processes within the given VM.
1157 * This function stops the exection of the VM. All the processes on this VM
1158 * will pause. The state of the VM is perserved. We can later resume it again.
1160 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1161 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1162 * of VM save to you.
1164 void MSG_vm_save(msg_vm_t vm)
1166 simcall_vm_save(vm);
1168 TRACE_msg_vm_save(vm);
1172 /** @brief Restore the execution of the VM. All processes on the VM run again.
1175 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1176 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1177 * of VM restore to you.
1179 void MSG_vm_restore(msg_vm_t vm)
1181 simcall_vm_restore(vm);
1184 TRACE_msg_vm_restore(vm);
1191 /** @brief Get the physical host of a given VM.
1194 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1196 return simcall_vm_get_pm(vm);