1 /* Copyright (c) 2012-2013. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
8 // 1./ check how and where a new VM is added to the list of the hosts
9 // 2./ Diff between SIMIX_Actions and SURF_Actions
10 // => SIMIX_actions : point synchro entre processus de niveau (theoretically speaking I do not have to create such SIMIX_ACTION
14 // MSG_TRACE can be revisited in order to use the host
15 // To implement a mixed model between workstation and vm_workstation,
16 // please give a look at surf_model_private_t model_private at SURF Level and to the share resource functions
17 // double (*share_resources) (double now);
18 // For the action into the vm workstation model, we should be able to leverage the usual one (and if needed, look at
19 // the workstation model.
21 #include "msg_private.h"
22 #include "xbt/sysdep.h"
24 #include "simgrid/platf.h"
26 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
27 "Cloud-oriented parts of the MSG API");
30 /* **** ******** GENERAL ********* **** */
32 /** \ingroup m_vm_management
33 * \brief Returns the value of a given vm property
36 * \param name a property name
37 * \return value of a property (or NULL if property not set)
40 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
42 return MSG_host_get_property_value(vm, name);
45 /** \ingroup m_vm_management
46 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
49 * \return a dict containing the properties
51 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
53 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
55 return (simcall_host_get_properties(vm));
58 /** \ingroup m_host_management
59 * \brief Change the value of a given host property
62 * \param name a property name
63 * \param value what to change the property to
64 * \param free_ctn the freeing function to use to kill the value on need
66 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
68 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
71 /** \ingroup msg_vm_management
72 * \brief Finds a msg_vm_t using its name.
74 * This is a name directory service
75 * \param name the name of a vm.
76 * \return the corresponding vm
78 * Please note that a VM is a specific host. Hence, you should give a different name
82 msg_vm_t MSG_vm_get_by_name(const char *name)
84 return MSG_get_host_by_name(name);
87 /** \ingroup m_vm_management
89 * \brief Return the name of the #msg_host_t.
91 * This functions checks whether \a host is a valid pointer or not and return
94 const char *MSG_vm_get_name(msg_vm_t vm)
96 return MSG_host_get_name(vm);
100 /* **** Check state of a VM **** */
101 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
103 return simcall_vm_get_state(vm) == state;
106 /** @brief Returns whether the given VM has just reated, not running.
109 int MSG_vm_is_created(msg_vm_t vm)
111 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
114 /** @brief Returns whether the given VM is currently running
117 int MSG_vm_is_running(msg_vm_t vm)
119 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
122 /** @brief Returns whether the given VM is currently migrating
125 int MSG_vm_is_migrating(msg_vm_t vm)
127 return __MSG_vm_is_state(vm, SURF_VM_STATE_MIGRATING);
130 /** @brief Returns whether the given VM is currently suspended, not running.
133 int MSG_vm_is_suspended(msg_vm_t vm)
135 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
138 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
141 int MSG_vm_is_saving(msg_vm_t vm)
143 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
146 /** @brief Returns whether the given VM has been saved, not running.
149 int MSG_vm_is_saved(msg_vm_t vm)
151 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
154 /** @brief Returns whether the given VM is being restored, not running.
157 int MSG_vm_is_restoring(msg_vm_t vm)
159 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
164 /* ------------------------------------------------------------------------- */
165 /* ------------------------------------------------------------------------- */
167 /* **** ******** MSG vm actions ********* **** */
169 /** @brief Create a new VM with specified parameters.
171 * All parameters are in MBytes
174 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name, int ncpus, int ramsize,
175 int net_cap, char *disk_path, int disksize,
176 int mig_netspeed, int dp_intensity)
178 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
179 double host_speed = MSG_get_host_speed(ind_pm);
180 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
182 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
183 s_ws_params_t params;
184 memset(¶ms, 0, sizeof(params));
185 params.ramsize = 1L * 1024 * 1024 * ramsize;
186 //params.overcommit = 0;
188 params.skip_stage2 = 0;
189 params.max_downtime = 0.03;
190 params.dp_rate = (update_speed * 1L * 1024 * 1024 ) / host_speed;
191 params.dp_cap = params.ramsize / 0.9; // working set memory is 90%
192 params.mig_speed = 1L * 1024 * 1024 * mig_netspeed; // mig_speed
194 //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);
195 simcall_host_set_params(vm, ¶ms);
201 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
204 * A VM is treated as a host. The name of the VM must be unique among all hosts.
206 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
208 /* make sure the VM of the same name does not exit */
210 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
212 XBT_ERROR("host %s already exits", name);
217 /* Note: ind_vm and vm_workstation point to the same elm object. */
218 msg_vm_t ind_vm = NULL;
219 void *ind_vm_workstation = NULL;
221 /* Ask the SIMIX layer to create the surf vm resource */
222 ind_vm_workstation = simcall_vm_create(name, ind_pm);
223 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
225 XBT_DEBUG("A new VM (%s) has been created", name);
228 TRACE_msg_vm_create(name, ind_pm);
234 /** @brief Destroy a VM. Destroy the VM object from the simulation.
237 void MSG_vm_destroy(msg_vm_t vm)
239 /* First, terminate all processes on the VM if necessary */
240 if (MSG_vm_is_running(vm))
241 simcall_vm_shutdown(vm);
243 if (!MSG_vm_is_created(vm)) {
244 XBT_CRITICAL("shutdown the given VM before destroying it");
248 /* Then, destroy the VM object */
249 simcall_vm_destroy(vm);
251 __MSG_host_destroy(vm);
254 TRACE_msg_vm_end(vm);
259 /** @brief Start a vm (i.e., boot the guest operating system)
262 * If the VM cannot be started, an exception is generated.
265 void MSG_vm_start(msg_vm_t vm)
267 simcall_vm_start(vm);
270 TRACE_msg_vm_start(vm);
276 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
279 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
280 * use a #MSG_process_sleep() or something. I'm not quite sure.
282 void MSG_vm_shutdown(msg_vm_t vm)
284 /* msg_vm_t equals to msg_host_t */
285 simcall_vm_shutdown(vm);
287 // #ifdef HAVE_TRACING
288 // TRACE_msg_vm_(vm);
294 /* We have two mailboxes. mbox is used to transfer migration data between
295 * source and destiantion PMs. mbox_ctl is used to detect the completion of a
296 * migration. The names of these mailboxes must not conflict with others. */
297 static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
299 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
302 static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
304 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
307 static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
309 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
312 static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
314 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
317 static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
319 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
322 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio);
324 static int migration_rx_fun(int argc, char *argv[])
326 XBT_DEBUG("mig: rx_start");
328 xbt_assert(argc == 4);
329 const char *vm_name = argv[1];
330 const char *src_pm_name = argv[2];
331 const char *dst_pm_name = argv[3];
332 msg_vm_t vm = MSG_get_host_by_name(vm_name);
333 msg_host_t src_pm = MSG_get_host_by_name(src_pm_name);
334 msg_host_t dst_pm = MSG_get_host_by_name(dst_pm_name);
337 s_ws_params_t params;
338 simcall_host_get_params(vm, ¶ms);
339 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
344 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
345 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
346 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
349 msg_task_t task = NULL;
350 MSG_task_recv(&task, mbox);
352 double received = MSG_task_get_data_size(task);
354 // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
355 launch_deferred_exec_process(vm, received * xfer_cpu_overhead, 1);
358 if (strcmp(task->name, finalize_task_name) == 0)
361 MSG_task_destroy(task);
368 /* deinstall the current affinity setting */
369 simcall_vm_set_affinity(vm, src_pm, 0);
371 simcall_vm_migrate(vm, dst_pm);
372 simcall_vm_resume(vm);
374 /* install the affinity setting of the VM on the destination pm */
376 msg_host_priv_t priv = msg_host_resource_priv(vm);
378 unsigned long affinity_mask = (unsigned long) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *) dst_pm, sizeof(msg_host_t));
379 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
380 XBT_INFO("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
384 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
386 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
387 msg_error_t ret = MSG_task_send(task, mbox_ctl);
388 xbt_assert(ret == MSG_OK);
396 xbt_free(finalize_task_name);
398 XBT_DEBUG("mig: rx_done");
403 static void reset_dirty_pages(msg_vm_t vm)
405 msg_host_priv_t priv = msg_host_resource_priv(vm);
408 xbt_dict_cursor_t cursor = NULL;
409 dirty_page_t dp = NULL;
410 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
411 double remaining = MSG_task_get_remaining_computation(dp->task);
412 dp->prev_clock = MSG_get_clock();
413 dp->prev_remaining = remaining;
415 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
419 static void start_dirty_page_tracking(msg_vm_t vm)
421 msg_host_priv_t priv = msg_host_resource_priv(vm);
422 priv->dp_enabled = 1;
424 reset_dirty_pages(vm);
427 static void stop_dirty_page_tracking(msg_vm_t vm)
429 msg_host_priv_t priv = msg_host_resource_priv(vm);
430 priv->dp_enabled = 0;
434 /* It might be natural that we define dp_rate for each task. But, we will also
435 * have to care about how each task behavior affects the memory update behavior
436 * at the operating system level. It may not be easy to model it with a simple algorithm. */
437 double calc_updated_pages(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;
441 double updated = dp->task->dp_rate * computed;
443 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
444 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
445 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
446 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
452 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
454 double computed = dp->prev_remaining - remaining;
455 double duration = clock - dp->prev_clock;
457 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
458 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
463 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
465 msg_host_priv_t priv = msg_host_resource_priv(vm);
469 xbt_dict_cursor_t cursor = NULL;
470 dirty_page_t dp = NULL;
471 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
472 double remaining = MSG_task_get_remaining_computation(dp->task);
474 double clock = MSG_get_clock();
476 // total += calc_updated_pages(key, vm, dp, remaining, clock);
477 total += get_computed(key, vm, dp, remaining, clock);
479 dp->prev_remaining = remaining;
480 dp->prev_clock = clock;
483 total += priv->dp_updated_by_deleted_tasks;
485 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
486 stage_for_fancy_debug,
487 stage2_round_for_fancy_debug,
488 total, priv->dp_updated_by_deleted_tasks);
492 priv->dp_updated_by_deleted_tasks = 0;
498 // TODO Is this code redundant with the information provided by
499 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
500 void MSG_host_add_task(msg_host_t host, msg_task_t task)
502 msg_host_priv_t priv = msg_host_resource_priv(host);
503 double remaining = MSG_task_get_remaining_computation(task);
504 char *key = bprintf("%s-%lld", task->name, task->counter);
506 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
509 /* It should be okay that we add a task onto a migrating VM. */
510 if (priv->dp_enabled) {
511 dp->prev_clock = MSG_get_clock();
512 dp->prev_remaining = remaining;
515 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
516 xbt_dict_set(priv->dp_objs, key, dp, NULL);
517 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
522 void MSG_host_del_task(msg_host_t host, msg_task_t task)
524 msg_host_priv_t priv = msg_host_resource_priv(host);
526 char *key = bprintf("%s-%lld", task->name, task->counter);
528 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
529 xbt_assert(dp->task == task);
531 /* If we are in the middle of dirty page tracking, we record how much
532 * computaion has been done until now, and keep the information for the
533 * lookup_() function that will called soon. */
534 if (priv->dp_enabled) {
535 double remaining = MSG_task_get_remaining_computation(task);
536 double clock = MSG_get_clock();
537 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
538 double updated = get_computed(key, host, dp, remaining, clock);
540 priv->dp_updated_by_deleted_tasks += updated;
543 xbt_dict_remove(priv->dp_objs, key);
546 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
552 static int deferred_exec_fun(int argc, char *argv[])
554 xbt_assert(argc == 3);
555 const char *comp_str = argv[1];
556 double computaion = atof(comp_str);
557 const char *prio_str = argv[2];
558 double prio = atof(prio_str);
560 msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
561 // XBT_INFO("exec deferred %f", computaion);
563 /* dpt is the results of the VM activity */
564 MSG_task_set_priority(task, prio);
565 MSG_task_execute(task);
569 MSG_task_destroy(task);
574 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
576 char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
579 char **argv = xbt_new(char *, nargvs);
580 argv[0] = xbt_strdup(pr_name);
581 argv[1] = bprintf("%lf", computation);
582 argv[2] = bprintf("%lf", prio);
585 MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv);
591 static int task_tx_overhead_fun(int argc, char *argv[])
593 xbt_assert(argc == 2);
594 const char *mbox = argv[1];
598 // XBT_INFO("start %s", mbox);
601 msg_task_t task = NULL;
602 MSG_task_recv(&task, mbox);
604 // XBT_INFO("task->name %s", task->name);
606 if (strcmp(task->name, "finalize_making_overhead") == 0)
610 // MSG_task_set_priority(task, 1000000);
611 MSG_task_execute(task);
612 MSG_task_destroy(task);
623 static void start_overhead_process(msg_task_t comm_task)
625 char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task));
626 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
629 char **argv = xbt_new(char *, nargvs);
630 argv[0] = xbt_strdup(pr_name);
631 argv[1] = xbt_strdup(mbox);
634 // XBT_INFO("micro start: mbox %s", mbox);
635 MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv);
641 static void shutdown_overhead_process(msg_task_t comm_task)
643 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
645 msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
647 // XBT_INFO("micro shutdown: mbox %s", mbox);
648 msg_error_t ret = MSG_task_send(task, mbox);
649 xbt_assert(ret == MSG_OK);
652 // XBT_INFO("shutdown done");
655 static void request_overhead(msg_task_t comm_task, double computation)
657 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
659 msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
661 // XBT_INFO("req overhead");
662 msg_error_t ret = MSG_task_send(task, mbox);
663 xbt_assert(ret == MSG_OK);
668 /* alpha is (floating_operations / bytes).
670 * When actual migration traffic was 32 mbytes/s, we observed the CPU
671 * utilization of the main thread of the Qemu process was 10 %.
672 * alpha = 0.1 * C / (32 * 1024 * 1024)
673 * where the CPU capacity of the PM is C flops/s.
676 static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
678 const double chunk_size = 1024 * 1024 * 10;
679 double remaining = MSG_task_get_data_size(comm_task);
681 start_overhead_process(comm_task);
684 while (remaining > 0) {
685 double data_size = chunk_size;
686 if (remaining < chunk_size)
687 data_size = remaining;
689 remaining -= data_size;
691 // XBT_INFO("remaining %f bytes", remaining);
694 double clock_sta = MSG_get_clock();
696 /* create a micro task */
698 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
699 msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
701 request_overhead(comm_task, data_size * alpha);
703 msg_error_t ret = MSG_task_send(mtask, mbox);
704 xbt_assert(ret == MSG_OK);
706 xbt_free(mtask_name);
711 /* In the real world, sending data involves small CPU computation. */
712 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
713 msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
714 MSG_task_execute(mtask);
715 MSG_task_destroy(mtask);
716 xbt_free(mtask_name);
722 double clock_end = MSG_get_clock();
727 * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
730 * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
732 * If time_to_sleep is smaller than zero, the elapsed time was too big. We
733 * do not need a micro sleep.
735 double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
736 if (time_to_sleep > 0)
737 MSG_process_sleep(time_to_sleep);
740 //XBT_INFO("duration %f", clock_end - clock_sta);
741 //XBT_INFO("time_to_sleep %f", time_to_sleep);
745 // XBT_INFO("%s", MSG_task_get_name(comm_task));
746 shutdown_overhead_process(comm_task);
752 static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
754 double prev_remaining = init_comm_size;
757 double remaining = MSG_task_get_remaining_communication(comm_task);
761 double sent = prev_remaining - remaining;
762 double comp_size = sent * overhead;
765 char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
766 msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
767 MSG_task_execute(comp_task);
768 MSG_task_destroy(comp_task);
773 prev_remaining = remaining;
777 xbt_free(comp_task_name);
781 // #define USE_MICRO_TASK 1
784 // const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
785 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
786 // const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
787 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
788 // const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
789 //const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
790 // const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
791 ////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
792 // const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
793 // const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
794 // const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
796 /* CPU 22% when 80Mbyte/s */
797 const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
801 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
802 sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
804 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
805 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
809 double clock_sta = MSG_get_clock();
811 #ifdef USE_MICRO_TASK
813 task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
818 ret = MSG_task_send_bounded(task, mbox, mig_speed);
820 ret = MSG_task_send(task, mbox);
821 xbt_assert(ret == MSG_OK);
824 double clock_end = MSG_get_clock();
825 double duration = clock_end - clock_sta;
826 double actual_speed = size / duration;
827 #ifdef USE_MICRO_TASK
828 double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
830 double cpu_utilization = 0;
837 XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f) cpu %f", stage, stage2_round, size, duration, actual_speed, mig_speed, cpu_utilization);}
839 XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f) cpu %f", stage, size, duration, actual_speed, mig_speed, cpu_utilization);
846 #ifdef USE_MICRO_TASK
847 /* The name of a micro task starts with __micro, which does not match the
848 * special name that finalizes the receiver loop. Thus, we send the special task.
852 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
853 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
854 msg_error_t ret = MSG_task_send(task, mbox);
855 xbt_assert(ret == MSG_OK);
862 static double get_updated_size(double computed, double dp_rate, double dp_cap)
864 double updated_size = computed * dp_rate;
865 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
866 if (updated_size > dp_cap) {
867 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
868 updated_size = dp_cap;
874 static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *dst_pm_name,
875 sg_size_t ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead)
877 const char *vm_name = MSG_host_get_name(vm);
878 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
880 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
881 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
882 sg_size_t remaining = ramsize;
883 double computed_total = 0;
885 while (remaining > 0) {
886 sg_size_t datasize = chunksize;
887 if (remaining < chunksize)
888 datasize = remaining;
890 remaining -= datasize;
892 send_migration_data(vm_name, src_pm_name, dst_pm_name, datasize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
894 double computed = lookup_computed_flop_counts(vm, 1, 0);
895 computed_total += computed;
898 // double updated_size = get_updated_size(computed, dp_rate, dp_cap);
900 // double overhead = dpt_cpu_overhead * updated_size;
901 // launch_deferred_exec_process(vm, overhead, 10000);
905 return computed_total;
910 static double get_threshold_value(double bandwidth, double max_downtime)
912 /* This value assumes the network link is 1Gbps. */
913 // double threshold = max_downtime * 125 * 1024 * 1024;
914 double threshold = max_downtime * bandwidth;
919 static int migration_tx_fun(int argc, char *argv[])
921 XBT_DEBUG("mig: tx_start");
923 xbt_assert(argc == 4);
924 const char *vm_name = argv[1];
925 const char *src_pm_name = argv[2];
926 const char *dst_pm_name = argv[3];
927 msg_vm_t vm = MSG_get_host_by_name(vm_name);
930 s_ws_params_t params;
931 simcall_host_get_params(vm, ¶ms);
932 const sg_size_t ramsize = params.ramsize;
933 const long devsize = params.devsize;
934 const int skip_stage1 = params.skip_stage1;
935 const int skip_stage2 = params.skip_stage2;
936 const double dp_rate = params.dp_rate;
937 const double dp_cap = params.dp_cap;
938 const double mig_speed = params.mig_speed;
939 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
940 const double dpt_cpu_overhead = params.dpt_cpu_overhead;
942 double remaining_size = ramsize + devsize;
944 double max_downtime = params.max_downtime;
945 if (max_downtime == 0) {
946 XBT_WARN("use the default max_downtime value 30ms");
950 double threshold = 0.00001; /* TODO: cleanup */
952 /* setting up parameters has done */
956 XBT_WARN("migrate a VM, but ramsize is zero");
958 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
960 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
962 /* Stage1: send all memory pages to the destination. */
963 start_dirty_page_tracking(vm);
965 double computed_during_stage1 = 0;
967 // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
969 /* send ramsize, but split it */
970 double clock_prev_send = MSG_get_clock();
972 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);
973 remaining_size -= ramsize;
975 double clock_post_send = MSG_get_clock();
976 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
977 threshold = get_threshold_value(bandwidth, max_downtime);
978 XBT_INFO("actual banwdidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
982 /* Stage2: send update pages iteratively until the size of remaining states
983 * becomes smaller than the threshold value. */
986 if (max_downtime == 0) {
987 XBT_WARN("no max_downtime parameter, skip stage2");
992 int stage2_round = 0;
995 double updated_size = 0;
996 if (stage2_round == 0) {
997 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
998 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
1000 double computed = lookup_computed_flop_counts(vm, 2, stage2_round);
1001 updated_size = get_updated_size(computed, dp_rate, dp_cap);
1004 XBT_INFO("mig-stage 2:%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
1005 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
1008 // if (stage2_round != 0) {
1009 // /* during stage1, we have already created overhead tasks */
1010 // double overhead = dpt_cpu_overhead * updated_size;
1011 // XBT_DEBUG("updated %f overhead %f", updated_size, overhead);
1012 // launch_deferred_exec_process(vm, overhead, 10000);
1017 remaining_size += updated_size;
1019 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
1020 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
1022 if (remaining_size < threshold)
1026 double clock_prev_send = MSG_get_clock();
1028 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
1030 double clock_post_send = MSG_get_clock();
1032 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
1033 threshold = get_threshold_value(bandwidth, max_downtime);
1034 XBT_INFO("actual banwdidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
1042 remaining_size -= updated_size;
1048 /* Stage3: stop the VM and copy the rest of states. */
1049 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
1050 simcall_vm_suspend(vm);
1051 stop_dirty_page_tracking(vm);
1053 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead);
1057 XBT_DEBUG("mig: tx_done");
1064 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
1066 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1069 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1071 char **argv = xbt_new(char *, nargvs);
1072 argv[0] = xbt_strdup(pr_name);
1073 argv[1] = xbt_strdup(sg_host_name(vm));
1074 argv[2] = xbt_strdup(sg_host_name(src_pm));
1075 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1078 MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
1084 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1086 char **argv = xbt_new(char *, nargvs);
1087 argv[0] = xbt_strdup(pr_name);
1088 argv[1] = xbt_strdup(sg_host_name(vm));
1089 argv[2] = xbt_strdup(sg_host_name(src_pm));
1090 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1092 MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
1097 /* wait until the migration have finished */
1099 msg_task_t task = NULL;
1100 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
1101 xbt_assert(ret == MSG_OK);
1103 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
1104 xbt_assert(strcmp(task->name, expected_task_name) == 0);
1105 xbt_free(expected_task_name);
1112 /** @brief Migrate the VM to the given host.
1115 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
1116 * MSG_task_send() before or after, depending on whether you want to do cold or hot
1119 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
1122 * - One approach is ...
1123 * We first create a new VM (i.e., destination VM) on the destination
1124 * physical host. The destination VM will receive the state of the source
1125 * VM over network. We will finally destroy the source VM.
1126 * - This behavior is similar to the way of migration in the real world.
1127 * Even before a migration is completed, we will see a destination VM,
1128 * consuming resources.
1129 * - We have to relocate all processes. The existing process migraion code
1130 * will work for this?
1131 * - The name of the VM is a somewhat unique ID in the code. It is tricky
1132 * for the destination VM?
1134 * - Another one is ...
1135 * We update the information of the given VM to place it to the destination
1138 * The second one would be easier.
1142 msg_host_t old_pm = simcall_vm_get_pm(vm);
1144 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
1145 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
1147 do_migration(vm, old_pm, new_pm);
1151 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1154 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1159 /** @brief Immediately suspend the execution of all processes within the given VM.
1162 * This function stops the exection of the VM. All the processes on this VM
1163 * will pause. The state of the VM is perserved. We can later resume it again.
1165 * No suspension cost occurs.
1167 void MSG_vm_suspend(msg_vm_t vm)
1169 simcall_vm_suspend(vm);
1171 XBT_DEBUG("vm_suspend done");
1174 TRACE_msg_vm_suspend(vm);
1179 /** @brief Resume the execution of the VM. All processes on the VM run again.
1182 * No resume cost occurs.
1184 void MSG_vm_resume(msg_vm_t vm)
1186 simcall_vm_resume(vm);
1189 TRACE_msg_vm_resume(vm);
1194 /** @brief Immediately save the execution of all processes within the given VM.
1197 * This function stops the exection of the VM. All the processes on this VM
1198 * will pause. The state of the VM is perserved. We can later resume it again.
1200 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1201 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1202 * of VM save to you.
1204 void MSG_vm_save(msg_vm_t vm)
1206 simcall_vm_save(vm);
1208 TRACE_msg_vm_save(vm);
1212 /** @brief Restore the execution of the VM. All processes on the VM run again.
1215 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1216 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1217 * of VM restore to you.
1219 void MSG_vm_restore(msg_vm_t vm)
1221 simcall_vm_restore(vm);
1224 TRACE_msg_vm_restore(vm);
1229 /** @brief Get the physical host of a given VM.
1232 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1234 return simcall_vm_get_pm(vm);
1238 /** @brief Set a CPU bound for a given VM.
1242 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1245 * On PM0, there are Task1 and VM0.
1246 * On VM0, there is Task2.
1247 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1249 * Task1\@PM0 gets 50%.
1250 * Task2\@VM0 gets 25%.
1251 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1253 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1254 * task on the VM does not affect the bound of the dummy CPU action. The bound
1255 * of the dummy CPU action is unlimited.
1257 * There are some solutions for this problem. One option is to update the bound
1258 * of the dummy CPU action automatically. It should be the sum of all tasks on
1259 * the VM. But, this solution might be costy, because we have to scan all tasks
1260 * on the VM in share_resource() or we have to trap both the start and end of
1263 * The current solution is to use MSG_vm_set_bound(), which allows us to
1264 * directly set the bound of the dummy CPU action.
1268 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1269 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1270 * exceeds the capacity of a CPU core.
1272 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1274 return simcall_vm_set_bound(vm, bound);
1278 /** @brief Set the CPU affinity of a given VM.
1281 * This function changes the CPU affinity of a given VM. Usage is the same as
1282 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1284 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1286 msg_host_priv_t priv = msg_host_resource_priv(vm);
1289 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1291 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *) mask, NULL);
1293 msg_host_t pm_now = MSG_vm_get_pm(vm);
1295 XBT_INFO("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1296 simcall_vm_set_affinity(vm, pm, mask);
1298 XBT_INFO("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));