1 /* Copyright (c) 2012. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
7 // 1./ check how and where a new VM is added to the list of the hosts
8 // 2./ Diff between SIMIX_Actions and SURF_Actions
9 // => SIMIX_actions : point synchro entre processus de niveau (theoretically speaking I do not have to create such SIMIX_ACTION
13 // MSG_TRACE can be revisited in order to use the host
14 // To implement a mixed model between workstation and vm_workstation,
15 // please give a look at surf_model_private_t model_private at SURF Level and to the share resource functions
16 // double (*share_resources) (double now);
17 // For the action into the vm workstation model, we should be able to leverage the usual one (and if needed, look at
18 // the workstation model.
20 #include "msg_private.h"
21 #include "xbt/sysdep.h"
24 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
25 "Cloud-oriented parts of the MSG API");
28 /* **** ******** GENERAL ********* **** */
30 /** \ingroup m_vm_management
31 * \brief Returns the value of a given vm property
34 * \param name a property name
35 * \return value of a property (or NULL if property not set)
38 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
40 return MSG_host_get_property_value(vm, name);
43 /** \ingroup m_vm_management
44 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
47 * \return a dict containing the properties
49 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
51 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
53 return (simcall_host_get_properties(vm));
56 /** \ingroup m_host_management
57 * \brief Change the value of a given host property
60 * \param name a property name
61 * \param value what to change the property to
62 * \param free_ctn the freeing function to use to kill the value on need
64 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
66 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
69 /** \ingroup msg_vm_management
70 * \brief Finds a msg_vm_t using its name.
72 * This is a name directory service
73 * \param name the name of a vm.
74 * \return the corresponding vm
76 * Please note that a VM is a specific host. Hence, you should give a different name
80 msg_vm_t MSG_vm_get_by_name(const char *name)
82 return MSG_get_host_by_name(name);
85 /** \ingroup m_vm_management
87 * \brief Return the name of the #msg_host_t.
89 * This functions checks whether \a host is a valid pointer or not and return
92 const char *MSG_vm_get_name(msg_vm_t vm)
94 return MSG_host_get_name(vm);
98 /* **** Check state of a VM **** */
99 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
101 return simcall_vm_get_state(vm) == state;
104 /** @brief Returns whether the given VM has just reated, not running.
107 int MSG_vm_is_created(msg_vm_t vm)
109 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
112 /** @brief Returns whether the given VM is currently running
115 int MSG_vm_is_running(msg_vm_t vm)
117 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
120 /** @brief Returns whether the given VM is currently migrating
123 int MSG_vm_is_migrating(msg_vm_t vm)
125 return __MSG_vm_is_state(vm, SURF_VM_STATE_MIGRATING);
128 /** @brief Returns whether the given VM is currently suspended, not running.
131 int MSG_vm_is_suspended(msg_vm_t vm)
133 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
136 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
139 int MSG_vm_is_saving(msg_vm_t vm)
141 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
144 /** @brief Returns whether the given VM has been saved, not running.
147 int MSG_vm_is_saved(msg_vm_t vm)
149 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
152 /** @brief Returns whether the given VM is being restored, not running.
155 int MSG_vm_is_restoring(msg_vm_t vm)
157 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
162 /* ------------------------------------------------------------------------- */
163 /* ------------------------------------------------------------------------- */
165 /* **** ******** MSG vm actions ********* **** */
167 /** @brief Create a new VM with specified parameters.
169 * All parameters are in MBytes
172 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name, int ncpus, int ramsize,
173 int net_cap, char *disk_path, int disksize,
174 int mig_netspeed, int dp_intensity)
176 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
177 double host_speed = MSG_get_host_speed(ind_pm);
178 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
180 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
181 s_ws_params_t params;
182 memset(¶ms, 0, sizeof(params));
183 params.ramsize = 1L * 1024 * 1024 * ramsize;
184 //params.overcommit = 0;
186 params.skip_stage2 = 0;
187 params.max_downtime = 0.03;
188 params.dp_rate = (update_speed * 1L * 1024 * 1024 ) / host_speed;
189 params.dp_cap = params.ramsize / 0.9; // working set memory is 90%
190 params.mig_speed = 1L * 1024 * 1024 * mig_netspeed; // mig_speed
192 //XBT_INFO("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f",params.dp_rate, params.mig_speed, dp_intensity, update_speed, host_speed);
193 simcall_host_set_params(vm, ¶ms);
199 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
202 * A VM is treated as a host. The name of the VM must be unique among all hosts.
204 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
206 /* make sure the VM of the same name does not exit */
208 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
210 XBT_ERROR("host %s already exits", name);
215 /* Note: ind_vm and vm_workstation point to the same elm object. */
216 msg_vm_t ind_vm = NULL;
217 void *ind_vm_workstation = NULL;
219 /* Ask the SIMIX layer to create the surf vm resource */
220 ind_vm_workstation = simcall_vm_create(name, ind_pm);
221 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
223 XBT_DEBUG("A new VM (%s) has been created", name);
226 TRACE_msg_vm_create(name, ind_pm);
232 /** @brief Destroy a VM. Destroy the VM object from the simulation.
235 void MSG_vm_destroy(msg_vm_t vm)
237 /* First, terminate all processes on the VM if necessary */
238 if (MSG_vm_is_running(vm))
239 simcall_vm_shutdown(vm);
241 if (!MSG_vm_is_created(vm)) {
242 XBT_CRITICAL("shutdown the given VM before destroying it");
246 /* Then, destroy the VM object */
247 simcall_vm_destroy(vm);
249 __MSG_host_destroy(vm);
252 TRACE_msg_vm_end(vm);
257 /** @brief Start a vm (i.e., boot the guest operating system)
260 * If the VM cannot be started, an exception is generated.
263 void MSG_vm_start(msg_vm_t vm)
265 simcall_vm_start(vm);
268 TRACE_msg_vm_start(vm);
274 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
277 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
278 * use a #MSG_process_sleep() or something. I'm not quite sure.
280 void MSG_vm_shutdown(msg_vm_t vm)
282 /* msg_vm_t equals to msg_host_t */
283 simcall_vm_shutdown(vm);
285 // #ifdef HAVE_TRACING
286 // TRACE_msg_vm_(vm);
292 /* We have two mailboxes. mbox is used to transfer migration data between
293 * source and destiantion PMs. mbox_ctl is used to detect the completion of a
294 * migration. The names of these mailboxes must not conflict with others. */
295 static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
297 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
300 static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
302 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
305 static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
307 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
310 static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
312 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
315 static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
317 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
320 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio);
322 static int migration_rx_fun(int argc, char *argv[])
324 XBT_DEBUG("mig: rx_start");
326 xbt_assert(argc == 4);
327 const char *vm_name = argv[1];
328 const char *src_pm_name = argv[2];
329 const char *dst_pm_name = argv[3];
330 msg_vm_t vm = MSG_get_host_by_name(vm_name);
331 msg_host_t src_pm = MSG_get_host_by_name(src_pm_name);
332 msg_host_t dst_pm = MSG_get_host_by_name(dst_pm_name);
335 s_ws_params_t params;
336 simcall_host_get_params(vm, ¶ms);
337 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
342 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
343 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
344 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
347 msg_task_t task = NULL;
348 MSG_task_recv(&task, mbox);
350 double received = MSG_task_get_data_size(task);
352 // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
353 launch_deferred_exec_process(vm, received * xfer_cpu_overhead, 1);
356 if (strcmp(task->name, finalize_task_name) == 0)
359 MSG_task_destroy(task);
366 /* deinstall the current affinity setting */
367 simcall_vm_set_affinity(vm, src_pm, 0);
369 simcall_vm_migrate(vm, dst_pm);
370 simcall_vm_resume(vm);
372 /* install the affinity setting of the VM on the destination pm */
374 msg_host_priv_t priv = msg_host_resource_priv(vm);
376 unsigned long affinity_mask = (unsigned long) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *) dst_pm, sizeof(msg_host_t));
377 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
378 XBT_INFO("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
382 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
384 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
385 msg_error_t ret = MSG_task_send(task, mbox_ctl);
386 xbt_assert(ret == MSG_OK);
394 xbt_free(finalize_task_name);
396 XBT_DEBUG("mig: rx_done");
401 static void reset_dirty_pages(msg_vm_t vm)
403 msg_host_priv_t priv = msg_host_resource_priv(vm);
406 xbt_dict_cursor_t cursor = NULL;
407 dirty_page_t dp = NULL;
408 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
409 double remaining = MSG_task_get_remaining_computation(dp->task);
410 dp->prev_clock = MSG_get_clock();
411 dp->prev_remaining = remaining;
413 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
417 static void start_dirty_page_tracking(msg_vm_t vm)
419 msg_host_priv_t priv = msg_host_resource_priv(vm);
420 priv->dp_enabled = 1;
422 reset_dirty_pages(vm);
425 static void stop_dirty_page_tracking(msg_vm_t vm)
427 msg_host_priv_t priv = msg_host_resource_priv(vm);
428 priv->dp_enabled = 0;
432 /* It might be natural that we define dp_rate for each task. But, we will also
433 * have to care about how each task behavior affects the memory update behavior
434 * at the operating system level. It may not be easy to model it with a simple algorithm. */
435 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
437 double computed = dp->prev_remaining - remaining;
438 double duration = clock - dp->prev_clock;
439 double updated = dp->task->dp_rate * computed;
441 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
442 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
443 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
444 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
450 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
452 double computed = dp->prev_remaining - remaining;
453 double duration = clock - dp->prev_clock;
455 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
456 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
461 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
463 msg_host_priv_t priv = msg_host_resource_priv(vm);
467 xbt_dict_cursor_t cursor = NULL;
468 dirty_page_t dp = NULL;
469 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
470 double remaining = MSG_task_get_remaining_computation(dp->task);
472 double clock = MSG_get_clock();
474 // total += calc_updated_pages(key, vm, dp, remaining, clock);
475 total += get_computed(key, vm, dp, remaining, clock);
477 dp->prev_remaining = remaining;
478 dp->prev_clock = clock;
481 total += priv->dp_updated_by_deleted_tasks;
483 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
484 stage_for_fancy_debug,
485 stage2_round_for_fancy_debug,
486 total, priv->dp_updated_by_deleted_tasks);
490 priv->dp_updated_by_deleted_tasks = 0;
496 // TODO Is this code redundant with the information provided by
497 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
498 void MSG_host_add_task(msg_host_t host, msg_task_t task)
500 msg_host_priv_t priv = msg_host_resource_priv(host);
501 double remaining = MSG_task_get_remaining_computation(task);
502 char *key = bprintf("%s-%lld", task->name, task->counter);
504 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
507 /* It should be okay that we add a task onto a migrating VM. */
508 if (priv->dp_enabled) {
509 dp->prev_clock = MSG_get_clock();
510 dp->prev_remaining = remaining;
513 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
514 xbt_dict_set(priv->dp_objs, key, dp, NULL);
515 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
520 void MSG_host_del_task(msg_host_t host, msg_task_t task)
522 msg_host_priv_t priv = msg_host_resource_priv(host);
524 char *key = bprintf("%s-%lld", task->name, task->counter);
526 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
527 xbt_assert(dp->task == task);
529 /* If we are in the middle of dirty page tracking, we record how much
530 * computaion has been done until now, and keep the information for the
531 * lookup_() function that will called soon. */
532 if (priv->dp_enabled) {
533 double remaining = MSG_task_get_remaining_computation(task);
534 double clock = MSG_get_clock();
535 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
536 double updated = get_computed(key, host, dp, remaining, clock);
538 priv->dp_updated_by_deleted_tasks += updated;
541 xbt_dict_remove(priv->dp_objs, key);
544 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
550 static int deferred_exec_fun(int argc, char *argv[])
552 xbt_assert(argc == 3);
553 const char *comp_str = argv[1];
554 double computaion = atof(comp_str);
555 const char *prio_str = argv[2];
556 double prio = atof(prio_str);
558 msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
559 // XBT_INFO("exec deferred %f", computaion);
561 /* dpt is the results of the VM activity */
562 MSG_task_set_priority(task, prio);
563 MSG_task_execute(task);
567 MSG_task_destroy(task);
572 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
574 char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
577 char **argv = xbt_new(char *, nargvs);
578 argv[0] = xbt_strdup(pr_name);
579 argv[1] = bprintf("%lf", computation);
580 argv[2] = bprintf("%lf", prio);
583 MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv);
589 static int task_tx_overhead_fun(int argc, char *argv[])
591 xbt_assert(argc == 2);
592 const char *mbox = argv[1];
596 // XBT_INFO("start %s", mbox);
599 msg_task_t task = NULL;
600 MSG_task_recv(&task, mbox);
602 // XBT_INFO("task->name %s", task->name);
604 if (strcmp(task->name, "finalize_making_overhead") == 0)
608 // MSG_task_set_priority(task, 1000000);
609 MSG_task_execute(task);
610 MSG_task_destroy(task);
621 static void start_overhead_process(msg_task_t comm_task)
623 char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task));
624 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
627 char **argv = xbt_new(char *, nargvs);
628 argv[0] = xbt_strdup(pr_name);
629 argv[1] = xbt_strdup(mbox);
632 // XBT_INFO("micro start: mbox %s", mbox);
633 MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv);
639 static void shutdown_overhead_process(msg_task_t comm_task)
641 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
643 msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
645 // XBT_INFO("micro shutdown: mbox %s", mbox);
646 msg_error_t ret = MSG_task_send(task, mbox);
647 xbt_assert(ret == MSG_OK);
650 // XBT_INFO("shutdown done");
653 static void request_overhead(msg_task_t comm_task, double computation)
655 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
657 msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
659 // XBT_INFO("req overhead");
660 msg_error_t ret = MSG_task_send(task, mbox);
661 xbt_assert(ret == MSG_OK);
666 /* alpha is (floating_operations / bytes).
668 * When actual migration traffic was 32 mbytes/s, we observed the CPU
669 * utilization of the main thread of the Qemu process was 10 %.
670 * alpha = 0.1 * C / (32 * 1024 * 1024)
671 * where the CPU capacity of the PM is C flops/s.
674 static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
676 const double chunk_size = 1024 * 1024 * 10;
677 double remaining = MSG_task_get_data_size(comm_task);
679 start_overhead_process(comm_task);
682 while (remaining > 0) {
683 double data_size = chunk_size;
684 if (remaining < chunk_size)
685 data_size = remaining;
687 remaining -= data_size;
689 // XBT_INFO("remaining %f bytes", remaining);
692 double clock_sta = MSG_get_clock();
694 /* create a micro task */
696 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
697 msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
699 request_overhead(comm_task, data_size * alpha);
701 msg_error_t ret = MSG_task_send(mtask, mbox);
702 xbt_assert(ret == MSG_OK);
704 xbt_free(mtask_name);
709 /* In the real world, sending data involves small CPU computation. */
710 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
711 msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
712 MSG_task_execute(mtask);
713 MSG_task_destroy(mtask);
714 xbt_free(mtask_name);
720 double clock_end = MSG_get_clock();
725 * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
728 * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
730 * If time_to_sleep is smaller than zero, the elapsed time was too big. We
731 * do not need a micro sleep.
733 double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
734 if (time_to_sleep > 0)
735 MSG_process_sleep(time_to_sleep);
738 //XBT_INFO("duration %f", clock_end - clock_sta);
739 //XBT_INFO("time_to_sleep %f", time_to_sleep);
743 // XBT_INFO("%s", MSG_task_get_name(comm_task));
744 shutdown_overhead_process(comm_task);
750 static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
752 double prev_remaining = init_comm_size;
755 double remaining = MSG_task_get_remaining_communication(comm_task);
759 double sent = prev_remaining - remaining;
760 double comp_size = sent * overhead;
763 char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
764 msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
765 MSG_task_execute(comp_task);
766 MSG_task_destroy(comp_task);
771 prev_remaining = remaining;
775 xbt_free(comp_task_name);
779 // #define USE_MICRO_TASK 1
782 // const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
783 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
784 // const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
785 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
786 // const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
787 //const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
788 // const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
789 ////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
790 // const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
791 // const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
792 // const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
794 /* CPU 22% when 80Mbyte/s */
795 const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
799 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
800 double size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
802 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
803 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
807 double clock_sta = MSG_get_clock();
809 #ifdef USE_MICRO_TASK
811 task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
816 ret = MSG_task_send_bounded(task, mbox, mig_speed);
818 ret = MSG_task_send(task, mbox);
819 xbt_assert(ret == MSG_OK);
822 double clock_end = MSG_get_clock();
823 double duration = clock_end - clock_sta;
824 double actual_speed = size / duration;
825 #ifdef USE_MICRO_TASK
826 double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
828 double cpu_utilization = 0;
835 XBT_DEBUG("mig-stage%d.%d: sent %f duration %f actual_speed %f (target %f) cpu %f", stage, stage2_round, size, duration, actual_speed, mig_speed, cpu_utilization);}
837 XBT_DEBUG("mig-stage%d: sent %f duration %f actual_speed %f (target %f) cpu %f", stage, size, duration, actual_speed, mig_speed, cpu_utilization);
844 #ifdef USE_MICRO_TASK
845 /* The name of a micro task starts with __micro, which does not match the
846 * special name that finalizes the receiver loop. Thus, we send the special task.
850 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
851 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
852 msg_error_t ret = MSG_task_send(task, mbox);
853 xbt_assert(ret == MSG_OK);
860 static double get_updated_size(double computed, double dp_rate, double dp_cap)
862 double updated_size = computed * dp_rate;
863 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
864 if (updated_size > dp_cap) {
865 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
866 updated_size = dp_cap;
872 static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *dst_pm_name,
873 long ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead)
875 const char *vm_name = MSG_host_get_name(vm);
876 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
878 // const long chunksize = 1024 * 1024 * 100;
879 const long chunksize = 1024L * 1024 * 100000;
880 long remaining = ramsize;
881 double computed_total = 0;
883 while (remaining > 0) {
884 long datasize = chunksize;
885 if (remaining < chunksize)
886 datasize = remaining;
888 remaining -= datasize;
890 send_migration_data(vm_name, src_pm_name, dst_pm_name, datasize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
892 double computed = lookup_computed_flop_counts(vm, 1, 0);
893 computed_total += computed;
896 // double updated_size = get_updated_size(computed, dp_rate, dp_cap);
898 // double overhead = dpt_cpu_overhead * updated_size;
899 // launch_deferred_exec_process(vm, overhead, 10000);
903 return computed_total;
908 static double get_threshold_value(double bandwidth, double max_downtime)
910 /* This value assumes the network link is 1Gbps. */
911 // double threshold = max_downtime * 125 * 1024 * 1024;
912 double threshold = max_downtime * bandwidth;
917 static int migration_tx_fun(int argc, char *argv[])
919 XBT_DEBUG("mig: tx_start");
921 xbt_assert(argc == 4);
922 const char *vm_name = argv[1];
923 const char *src_pm_name = argv[2];
924 const char *dst_pm_name = argv[3];
925 msg_vm_t vm = MSG_get_host_by_name(vm_name);
928 s_ws_params_t params;
929 simcall_host_get_params(vm, ¶ms);
930 const long ramsize = params.ramsize;
931 const long devsize = params.devsize;
932 const int skip_stage1 = params.skip_stage1;
933 const int skip_stage2 = params.skip_stage2;
934 const double dp_rate = params.dp_rate;
935 const double dp_cap = params.dp_cap;
936 const double mig_speed = params.mig_speed;
937 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
938 const double dpt_cpu_overhead = params.dpt_cpu_overhead;
940 double remaining_size = ramsize + devsize;
942 double max_downtime = params.max_downtime;
943 if (max_downtime == 0) {
944 XBT_WARN("use the default max_downtime value 30ms");
948 double threshold = 0.00001; /* TODO: cleanup */
950 /* setting up parameters has done */
954 XBT_WARN("migrate a VM, but ramsize is zero");
956 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
958 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
960 /* Stage1: send all memory pages to the destination. */
961 start_dirty_page_tracking(vm);
963 double computed_during_stage1 = 0;
965 // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
967 /* send ramsize, but split it */
968 double clock_prev_send = MSG_get_clock();
970 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);
971 remaining_size -= ramsize;
973 double clock_post_send = MSG_get_clock();
974 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
975 threshold = get_threshold_value(bandwidth, max_downtime);
976 XBT_INFO("actual banwdidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
980 /* Stage2: send update pages iteratively until the size of remaining states
981 * becomes smaller than the threshold value. */
984 if (max_downtime == 0) {
985 XBT_WARN("no max_downtime parameter, skip stage2");
990 int stage2_round = 0;
993 double updated_size = 0;
994 if (stage2_round == 0) {
995 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
996 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
998 double computed = lookup_computed_flop_counts(vm, 2, stage2_round);
999 updated_size = get_updated_size(computed, dp_rate, dp_cap);
1002 XBT_INFO("mig-stage 2:%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
1003 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
1006 // if (stage2_round != 0) {
1007 // /* during stage1, we have already created overhead tasks */
1008 // double overhead = dpt_cpu_overhead * updated_size;
1009 // XBT_DEBUG("updated %f overhead %f", updated_size, overhead);
1010 // launch_deferred_exec_process(vm, overhead, 10000);
1015 remaining_size += updated_size;
1017 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
1018 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
1020 if (remaining_size < threshold)
1024 double clock_prev_send = MSG_get_clock();
1026 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
1028 double clock_post_send = MSG_get_clock();
1030 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
1031 threshold = get_threshold_value(bandwidth, max_downtime);
1032 XBT_INFO("actual banwdidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
1040 remaining_size -= updated_size;
1046 /* Stage3: stop the VM and copy the rest of states. */
1047 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
1048 simcall_vm_suspend(vm);
1049 stop_dirty_page_tracking(vm);
1051 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead);
1055 XBT_DEBUG("mig: tx_done");
1062 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
1064 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1067 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1069 char **argv = xbt_new(char *, nargvs);
1070 argv[0] = xbt_strdup(pr_name);
1071 argv[1] = xbt_strdup(sg_host_name(vm));
1072 argv[2] = xbt_strdup(sg_host_name(src_pm));
1073 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1076 MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
1082 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1084 char **argv = xbt_new(char *, nargvs);
1085 argv[0] = xbt_strdup(pr_name);
1086 argv[1] = xbt_strdup(sg_host_name(vm));
1087 argv[2] = xbt_strdup(sg_host_name(src_pm));
1088 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1090 MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
1095 /* wait until the migration have finished */
1097 msg_task_t task = NULL;
1098 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
1099 xbt_assert(ret == MSG_OK);
1101 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
1102 xbt_assert(strcmp(task->name, expected_task_name) == 0);
1103 xbt_free(expected_task_name);
1110 /** @brief Migrate the VM to the given host.
1113 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
1114 * MSG_task_send() before or after, depending on whether you want to do cold or hot
1117 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
1120 * - One approach is ...
1121 * We first create a new VM (i.e., destination VM) on the destination
1122 * physical host. The destination VM will receive the state of the source
1123 * VM over network. We will finally destroy the source VM.
1124 * - This behavior is similar to the way of migration in the real world.
1125 * Even before a migration is completed, we will see a destination VM,
1126 * consuming resources.
1127 * - We have to relocate all processes. The existing process migraion code
1128 * will work for this?
1129 * - The name of the VM is a somewhat unique ID in the code. It is tricky
1130 * for the destination VM?
1132 * - Another one is ...
1133 * We update the information of the given VM to place it to the destination
1136 * The second one would be easier.
1140 msg_host_t old_pm = simcall_vm_get_pm(vm);
1142 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
1143 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
1145 do_migration(vm, old_pm, new_pm);
1149 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1152 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1157 /** @brief Immediately suspend the execution of all processes within the given VM.
1160 * This function stops the exection of the VM. All the processes on this VM
1161 * will pause. The state of the VM is perserved. We can later resume it again.
1163 * No suspension cost occurs.
1165 void MSG_vm_suspend(msg_vm_t vm)
1167 simcall_vm_suspend(vm);
1169 XBT_DEBUG("vm_suspend done");
1172 TRACE_msg_vm_suspend(vm);
1177 /** @brief Resume the execution of the VM. All processes on the VM run again.
1180 * No resume cost occurs.
1182 void MSG_vm_resume(msg_vm_t vm)
1184 simcall_vm_resume(vm);
1187 TRACE_msg_vm_resume(vm);
1192 /** @brief Immediately save the execution of all processes within the given VM.
1195 * This function stops the exection of the VM. All the processes on this VM
1196 * will pause. The state of the VM is perserved. We can later resume it again.
1198 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1199 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1200 * of VM save to you.
1202 void MSG_vm_save(msg_vm_t vm)
1204 simcall_vm_save(vm);
1206 TRACE_msg_vm_save(vm);
1210 /** @brief Restore the execution of the VM. All processes on the VM run again.
1213 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1214 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1215 * of VM restore to you.
1217 void MSG_vm_restore(msg_vm_t vm)
1219 simcall_vm_restore(vm);
1222 TRACE_msg_vm_restore(vm);
1227 /** @brief Get the physical host of a given VM.
1230 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1232 return simcall_vm_get_pm(vm);
1236 /** @brief Set a CPU bound for a given VM.
1240 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1243 * On PM0, there are Task1 and VM0.
1244 * On VM0, there is Task2.
1245 * Now we bound 75% to Task1@PM0 and bound 25% to Task2@VM0.
1247 * Task1@PM0 gets 50%.
1248 * Task2@VM0 gets 25%.
1249 * This is NOT 75% for Task1@PM0 and 25% for Task2@VM0, respectively.
1251 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1252 * task on the VM does not affect the bound of the dummy CPU action. The bound
1253 * of the dummy CPU action is unlimited.
1255 * There are some solutions for this problem. One option is to update the bound
1256 * of the dummy CPU action automatically. It should be the sum of all tasks on
1257 * the VM. But, this solution might be costy, because we have to scan all tasks
1258 * on the VM in share_resource() or we have to trap both the start and end of
1261 * The current solution is to use MSG_vm_set_bound(), which allows us to
1262 * directly set the bound of the dummy CPU action.
1266 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1267 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1268 * exceeds the capacity of a CPU core.
1270 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1272 return simcall_vm_set_bound(vm, bound);
1276 /** @brief Set the CPU affinity of a given VM.
1279 * This function changes the CPU affinity of a given VM. Usage is the same as
1280 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1282 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1284 msg_host_priv_t priv = msg_host_resource_priv(vm);
1287 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1289 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *) mask, NULL);
1291 msg_host_t pm_now = MSG_vm_get_pm(vm);
1293 XBT_INFO("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1294 simcall_vm_set_affinity(vm, pm, mask);
1296 XBT_INFO("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));