1 /* Copyright (c) 2012-2014. 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,
175 int ncpus, int ramsize,
176 int net_cap, char *disk_path, int disksize,
177 int mig_netspeed, int dp_intensity)
179 /* For the moment, intensity_rate is the percentage against the migration
181 double host_speed = MSG_get_host_speed(ind_pm);
182 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
184 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
185 s_ws_params_t params;
186 memset(¶ms, 0, sizeof(params));
187 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
188 //params.overcommit = 0;
190 params.skip_stage2 = 0;
191 params.max_downtime = 0.03;
192 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
193 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
194 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
196 //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);
197 simcall_host_set_params(vm, ¶ms);
203 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
206 * A VM is treated as a host. The name of the VM must be unique among all hosts.
208 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
210 /* make sure the VM of the same name does not exit */
212 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
214 XBT_ERROR("host %s already exits", name);
219 /* Note: ind_vm and vm_workstation point to the same elm object. */
220 msg_vm_t ind_vm = NULL;
221 void *ind_vm_workstation = NULL;
223 /* Ask the SIMIX layer to create the surf vm resource */
224 ind_vm_workstation = simcall_vm_create(name, ind_pm);
225 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
227 XBT_DEBUG("A new VM (%s) has been created", name);
230 TRACE_msg_vm_create(name, ind_pm);
236 /** @brief Destroy a VM. Destroy the VM object from the simulation.
239 void MSG_vm_destroy(msg_vm_t vm)
241 /* First, terminate all processes on the VM if necessary */
242 if (MSG_vm_is_running(vm))
243 simcall_vm_shutdown(vm);
245 if (!MSG_vm_is_created(vm)) {
246 XBT_CRITICAL("shutdown the given VM before destroying it");
250 /* Then, destroy the VM object */
251 simcall_vm_destroy(vm);
253 __MSG_host_destroy(vm);
256 TRACE_msg_vm_end(vm);
261 /** @brief Start a vm (i.e., boot the guest operating system)
264 * If the VM cannot be started, an exception is generated.
267 void MSG_vm_start(msg_vm_t vm)
269 simcall_vm_start(vm);
272 TRACE_msg_vm_start(vm);
278 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
281 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
282 * use a #MSG_process_sleep() or something. I'm not quite sure.
284 void MSG_vm_shutdown(msg_vm_t vm)
286 /* msg_vm_t equals to msg_host_t */
287 simcall_vm_shutdown(vm);
289 // #ifdef HAVE_TRACING
290 // TRACE_msg_vm_(vm);
296 /* We have two mailboxes. mbox is used to transfer migration data between
297 * source and destiantion PMs. mbox_ctl is used to detect the completion of a
298 * migration. The names of these mailboxes must not conflict with others. */
299 static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
301 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
304 static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
306 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
309 static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
311 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
314 static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
316 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
319 static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
321 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
324 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio);
326 static int migration_rx_fun(int argc, char *argv[])
328 XBT_DEBUG("mig: rx_start");
330 xbt_assert(argc == 4);
331 const char *vm_name = argv[1];
332 const char *src_pm_name = argv[2];
333 const char *dst_pm_name = argv[3];
334 msg_vm_t vm = MSG_get_host_by_name(vm_name);
335 msg_host_t src_pm = MSG_get_host_by_name(src_pm_name);
336 msg_host_t dst_pm = MSG_get_host_by_name(dst_pm_name);
339 s_ws_params_t params;
340 simcall_host_get_params(vm, ¶ms);
341 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
346 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
347 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
348 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
351 msg_task_t task = NULL;
352 MSG_task_recv(&task, mbox);
354 double received = MSG_task_get_data_size(task);
356 // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
357 launch_deferred_exec_process(vm, received * xfer_cpu_overhead, 1);
360 if (strcmp(task->name, finalize_task_name) == 0)
363 MSG_task_destroy(task);
370 /* deinstall the current affinity setting */
371 simcall_vm_set_affinity(vm, src_pm, 0);
373 simcall_vm_migrate(vm, dst_pm);
374 simcall_vm_resume(vm);
376 /* install the affinity setting of the VM on the destination pm */
378 msg_host_priv_t priv = msg_host_resource_priv(vm);
380 unsigned long affinity_mask = (unsigned long) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *) dst_pm, sizeof(msg_host_t));
381 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
382 XBT_INFO("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
386 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
388 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
389 msg_error_t ret = MSG_task_send(task, mbox_ctl);
390 xbt_assert(ret == MSG_OK);
398 xbt_free(finalize_task_name);
400 XBT_DEBUG("mig: rx_done");
405 static void reset_dirty_pages(msg_vm_t vm)
407 msg_host_priv_t priv = msg_host_resource_priv(vm);
410 xbt_dict_cursor_t cursor = NULL;
411 dirty_page_t dp = NULL;
412 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
413 double remaining = MSG_task_get_remaining_computation(dp->task);
414 dp->prev_clock = MSG_get_clock();
415 dp->prev_remaining = remaining;
417 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
421 static void start_dirty_page_tracking(msg_vm_t vm)
423 msg_host_priv_t priv = msg_host_resource_priv(vm);
424 priv->dp_enabled = 1;
426 reset_dirty_pages(vm);
429 static void stop_dirty_page_tracking(msg_vm_t vm)
431 msg_host_priv_t priv = msg_host_resource_priv(vm);
432 priv->dp_enabled = 0;
436 /* It might be natural that we define dp_rate for each task. But, we will also
437 * have to care about how each task behavior affects the memory update behavior
438 * at the operating system level. It may not be easy to model it with a simple algorithm. */
439 double calc_updated_pages(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;
443 double updated = dp->task->dp_rate * computed;
445 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
446 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
447 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
448 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
454 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
456 double computed = dp->prev_remaining - remaining;
457 double duration = clock - dp->prev_clock;
459 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
460 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
465 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
467 msg_host_priv_t priv = msg_host_resource_priv(vm);
471 xbt_dict_cursor_t cursor = NULL;
472 dirty_page_t dp = NULL;
473 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
474 double remaining = MSG_task_get_remaining_computation(dp->task);
476 double clock = MSG_get_clock();
478 // total += calc_updated_pages(key, vm, dp, remaining, clock);
479 total += get_computed(key, vm, dp, remaining, clock);
481 dp->prev_remaining = remaining;
482 dp->prev_clock = clock;
485 total += priv->dp_updated_by_deleted_tasks;
487 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
488 stage_for_fancy_debug,
489 stage2_round_for_fancy_debug,
490 total, priv->dp_updated_by_deleted_tasks);
494 priv->dp_updated_by_deleted_tasks = 0;
500 // TODO Is this code redundant with the information provided by
501 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
502 void MSG_host_add_task(msg_host_t host, msg_task_t task)
504 msg_host_priv_t priv = msg_host_resource_priv(host);
505 double remaining = MSG_task_get_remaining_computation(task);
506 char *key = bprintf("%s-%p", task->name, task);
508 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
511 /* It should be okay that we add a task onto a migrating VM. */
512 if (priv->dp_enabled) {
513 dp->prev_clock = MSG_get_clock();
514 dp->prev_remaining = remaining;
517 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
518 xbt_dict_set(priv->dp_objs, key, dp, NULL);
519 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
524 void MSG_host_del_task(msg_host_t host, msg_task_t task)
526 msg_host_priv_t priv = msg_host_resource_priv(host);
528 char *key = bprintf("%s-%p", task->name, task);
530 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
531 xbt_assert(dp->task == task);
533 /* If we are in the middle of dirty page tracking, we record how much
534 * computaion has been done until now, and keep the information for the
535 * lookup_() function that will called soon. */
536 if (priv->dp_enabled) {
537 double remaining = MSG_task_get_remaining_computation(task);
538 double clock = MSG_get_clock();
539 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
540 double updated = get_computed(key, host, dp, remaining, clock);
542 priv->dp_updated_by_deleted_tasks += updated;
545 xbt_dict_remove(priv->dp_objs, key);
548 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
554 static int deferred_exec_fun(int argc, char *argv[])
556 xbt_assert(argc == 3);
557 const char *comp_str = argv[1];
558 double computaion = atof(comp_str);
559 const char *prio_str = argv[2];
560 double prio = atof(prio_str);
562 msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
563 // XBT_INFO("exec deferred %f", computaion);
565 /* dpt is the results of the VM activity */
566 MSG_task_set_priority(task, prio);
567 MSG_task_execute(task);
571 MSG_task_destroy(task);
576 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
578 char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
581 char **argv = xbt_new(char *, nargvs);
583 argv[1] = bprintf("%f", computation);
584 argv[2] = bprintf("%f", prio);
587 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);
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);
638 static void shutdown_overhead_process(msg_task_t comm_task)
640 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
642 msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
644 // XBT_INFO("micro shutdown: mbox %s", mbox);
645 msg_error_t ret = MSG_task_send(task, mbox);
646 xbt_assert(ret == MSG_OK);
649 // XBT_INFO("shutdown done");
652 static void request_overhead(msg_task_t comm_task, double computation)
654 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
656 msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
658 // XBT_INFO("req overhead");
659 msg_error_t ret = MSG_task_send(task, mbox);
660 xbt_assert(ret == MSG_OK);
665 /* alpha is (floating_operations / bytes).
667 * When actual migration traffic was 32 mbytes/s, we observed the CPU
668 * utilization of the main thread of the Qemu process was 10 %.
669 * alpha = 0.1 * C / (32 * 1024 * 1024)
670 * where the CPU capacity of the PM is C flops/s.
673 static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
675 const double chunk_size = 1024 * 1024 * 10;
676 double remaining = MSG_task_get_data_size(comm_task);
678 start_overhead_process(comm_task);
681 while (remaining > 0) {
682 double data_size = chunk_size;
683 if (remaining < chunk_size)
684 data_size = remaining;
686 remaining -= data_size;
688 // XBT_INFO("remaining %f bytes", remaining);
691 double clock_sta = MSG_get_clock();
693 /* create a micro task */
695 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
696 msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
698 request_overhead(comm_task, data_size * alpha);
700 msg_error_t ret = MSG_task_send(mtask, mbox);
701 xbt_assert(ret == MSG_OK);
703 xbt_free(mtask_name);
708 /* In the real world, sending data involves small CPU computation. */
709 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
710 msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
711 MSG_task_execute(mtask);
712 MSG_task_destroy(mtask);
713 xbt_free(mtask_name);
719 double clock_end = MSG_get_clock();
724 * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
727 * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
729 * If time_to_sleep is smaller than zero, the elapsed time was too big. We
730 * do not need a micro sleep.
732 double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
733 if (time_to_sleep > 0)
734 MSG_process_sleep(time_to_sleep);
737 //XBT_INFO("duration %f", clock_end - clock_sta);
738 //XBT_INFO("time_to_sleep %f", time_to_sleep);
742 // XBT_INFO("%s", MSG_task_get_name(comm_task));
743 shutdown_overhead_process(comm_task);
749 static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
751 double prev_remaining = init_comm_size;
754 double remaining = MSG_task_get_remaining_communication(comm_task);
758 double sent = prev_remaining - remaining;
759 double comp_size = sent * overhead;
762 char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
763 msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
764 MSG_task_execute(comp_task);
765 MSG_task_destroy(comp_task);
770 prev_remaining = remaining;
774 xbt_free(comp_task_name);
778 // #define USE_MICRO_TASK 1
781 // const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
782 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
783 // const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
784 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
785 // const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
786 //const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
787 // const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
788 ////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
789 // const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
790 // const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
791 // const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
793 /* CPU 22% when 80Mbyte/s */
794 const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
798 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
799 sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
801 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
802 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
806 double clock_sta = MSG_get_clock();
808 #ifdef USE_MICRO_TASK
810 task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
815 ret = MSG_task_send_bounded(task, mbox, mig_speed);
817 ret = MSG_task_send(task, mbox);
818 xbt_assert(ret == MSG_OK);
821 double clock_end = MSG_get_clock();
822 double duration = clock_end - clock_sta;
823 double actual_speed = size / duration;
824 #ifdef USE_MICRO_TASK
825 double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
827 double cpu_utilization = 0;
834 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);}
836 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);
843 #ifdef USE_MICRO_TASK
844 /* The name of a micro task starts with __micro, which does not match the
845 * special name that finalizes the receiver loop. Thus, we send the special task.
849 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
850 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
851 msg_error_t ret = MSG_task_send(task, mbox);
852 xbt_assert(ret == MSG_OK);
859 static double get_updated_size(double computed, double dp_rate, double dp_cap)
861 double updated_size = computed * dp_rate;
862 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
863 if (updated_size > dp_cap) {
864 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
865 updated_size = dp_cap;
871 static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *dst_pm_name,
872 sg_size_t ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead)
874 const char *vm_name = MSG_host_get_name(vm);
875 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
877 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
878 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
879 sg_size_t remaining = ramsize;
880 double computed_total = 0;
882 while (remaining > 0) {
883 sg_size_t datasize = chunksize;
884 if (remaining < chunksize)
885 datasize = remaining;
887 remaining -= datasize;
889 send_migration_data(vm_name, src_pm_name, dst_pm_name, datasize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
890 double computed = lookup_computed_flop_counts(vm, 1, 0);
891 computed_total += computed;
894 // double updated_size = get_updated_size(computed, dp_rate, dp_cap);
896 // double overhead = dpt_cpu_overhead * updated_size;
897 // launch_deferred_exec_process(vm, overhead, 10000);
901 return computed_total;
906 static double get_threshold_value(double bandwidth, double max_downtime)
908 /* This value assumes the network link is 1Gbps. */
909 // double threshold = max_downtime * 125 * 1024 * 1024;
910 double threshold = max_downtime * bandwidth;
915 static int migration_tx_fun(int argc, char *argv[])
917 XBT_DEBUG("mig: tx_start");
919 xbt_assert(argc == 4);
920 const char *vm_name = argv[1];
921 const char *src_pm_name = argv[2];
922 const char *dst_pm_name = argv[3];
923 msg_vm_t vm = MSG_get_host_by_name(vm_name);
926 s_ws_params_t params;
927 simcall_host_get_params(vm, ¶ms);
928 const sg_size_t ramsize = params.ramsize;
929 const sg_size_t devsize = params.devsize;
930 const int skip_stage1 = params.skip_stage1;
931 const int skip_stage2 = params.skip_stage2;
932 const double dp_rate = params.dp_rate;
933 const double dp_cap = params.dp_cap;
934 const double mig_speed = params.mig_speed;
935 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
936 const double dpt_cpu_overhead = params.dpt_cpu_overhead;
938 double remaining_size = ramsize + devsize;
940 double max_downtime = params.max_downtime;
941 if (max_downtime == 0) {
942 XBT_WARN("use the default max_downtime value 30ms");
946 double threshold = 0.00001; /* TODO: cleanup */
948 /* setting up parameters has done */
952 XBT_WARN("migrate a VM, but ramsize is zero");
954 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
956 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
958 /* Stage1: send all memory pages to the destination. */
959 start_dirty_page_tracking(vm);
961 double computed_during_stage1 = 0;
963 // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
965 /* send ramsize, but split it */
966 double clock_prev_send = MSG_get_clock();
968 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);
969 remaining_size -= ramsize;
971 double clock_post_send = MSG_get_clock();
972 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
973 threshold = get_threshold_value(bandwidth, max_downtime);
974 XBT_INFO("actual banwdidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
978 /* Stage2: send update pages iteratively until the size of remaining states
979 * becomes smaller than the threshold value. */
982 if (max_downtime == 0) {
983 XBT_WARN("no max_downtime parameter, skip stage2");
988 int stage2_round = 0;
991 double updated_size = 0;
992 if (stage2_round == 0) {
993 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
994 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
996 double computed = lookup_computed_flop_counts(vm, 2, stage2_round);
997 updated_size = get_updated_size(computed, dp_rate, dp_cap);
1000 XBT_INFO("mig-stage 2:%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
1001 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
1004 // if (stage2_round != 0) {
1005 // /* during stage1, we have already created overhead tasks */
1006 // double overhead = dpt_cpu_overhead * updated_size;
1007 // XBT_DEBUG("updated %f overhead %f", updated_size, overhead);
1008 // launch_deferred_exec_process(vm, overhead, 10000);
1013 remaining_size += updated_size;
1015 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
1016 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
1018 if (remaining_size < threshold)
1022 double clock_prev_send = MSG_get_clock();
1024 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
1026 double clock_post_send = MSG_get_clock();
1028 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
1029 threshold = get_threshold_value(bandwidth, max_downtime);
1030 XBT_INFO("actual banwdidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
1038 remaining_size -= updated_size;
1044 /* Stage3: stop the VM and copy the rest of states. */
1045 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
1046 simcall_vm_suspend(vm);
1047 stop_dirty_page_tracking(vm);
1049 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead);
1053 XBT_DEBUG("mig: tx_done");
1060 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
1062 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1065 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1067 char **argv = xbt_new(char *, nargvs);
1069 argv[1] = xbt_strdup(sg_host_name(vm));
1070 argv[2] = xbt_strdup(sg_host_name(src_pm));
1071 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1074 MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
1078 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1080 char **argv = xbt_new(char *, nargvs);
1082 argv[1] = xbt_strdup(sg_host_name(vm));
1083 argv[2] = xbt_strdup(sg_host_name(src_pm));
1084 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1086 MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
1089 /* wait until the migration have finished */
1091 msg_task_t task = NULL;
1092 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
1094 xbt_assert(ret == MSG_OK);
1096 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
1097 xbt_assert(strcmp(task->name, expected_task_name) == 0);
1098 xbt_free(expected_task_name);
1099 MSG_task_destroy(task);
1106 /** @brief Migrate the VM to the given host.
1109 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
1110 * MSG_task_send() before or after, depending on whether you want to do cold or hot
1113 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
1116 * - One approach is ...
1117 * We first create a new VM (i.e., destination VM) on the destination
1118 * physical host. The destination VM will receive the state of the source
1119 * VM over network. We will finally destroy the source VM.
1120 * - This behavior is similar to the way of migration in the real world.
1121 * Even before a migration is completed, we will see a destination VM,
1122 * consuming resources.
1123 * - We have to relocate all processes. The existing process migraion code
1124 * will work for this?
1125 * - The name of the VM is a somewhat unique ID in the code. It is tricky
1126 * for the destination VM?
1128 * - Another one is ...
1129 * We update the information of the given VM to place it to the destination
1132 * The second one would be easier.
1136 msg_host_t old_pm = simcall_vm_get_pm(vm);
1138 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
1139 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
1141 do_migration(vm, old_pm, new_pm);
1145 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1148 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1153 /** @brief Immediately suspend the execution of all processes within the given VM.
1156 * This function stops the exection of the VM. All the processes on this VM
1157 * will pause. The state of the VM is perserved. We can later resume it again.
1159 * No suspension cost occurs.
1161 void MSG_vm_suspend(msg_vm_t vm)
1163 simcall_vm_suspend(vm);
1165 XBT_DEBUG("vm_suspend done");
1168 TRACE_msg_vm_suspend(vm);
1173 /** @brief Resume the execution of the VM. All processes on the VM run again.
1176 * No resume cost occurs.
1178 void MSG_vm_resume(msg_vm_t vm)
1180 simcall_vm_resume(vm);
1183 TRACE_msg_vm_resume(vm);
1188 /** @brief Immediately save the execution of all processes within the given VM.
1191 * This function stops the exection of the VM. All the processes on this VM
1192 * will pause. The state of the VM is perserved. We can later resume it again.
1194 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1195 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1196 * of VM save to you.
1198 void MSG_vm_save(msg_vm_t vm)
1200 simcall_vm_save(vm);
1202 TRACE_msg_vm_save(vm);
1206 /** @brief Restore the execution of the VM. All processes on the VM run again.
1209 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1210 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1211 * of VM restore to you.
1213 void MSG_vm_restore(msg_vm_t vm)
1215 simcall_vm_restore(vm);
1218 TRACE_msg_vm_restore(vm);
1223 /** @brief Get the physical host of a given VM.
1226 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1228 return simcall_vm_get_pm(vm);
1232 /** @brief Set a CPU bound for a given VM.
1236 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1239 * On PM0, there are Task1 and VM0.
1240 * On VM0, there is Task2.
1241 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1243 * Task1\@PM0 gets 50%.
1244 * Task2\@VM0 gets 25%.
1245 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1247 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1248 * task on the VM does not affect the bound of the dummy CPU action. The bound
1249 * of the dummy CPU action is unlimited.
1251 * There are some solutions for this problem. One option is to update the bound
1252 * of the dummy CPU action automatically. It should be the sum of all tasks on
1253 * the VM. But, this solution might be costy, because we have to scan all tasks
1254 * on the VM in share_resource() or we have to trap both the start and end of
1257 * The current solution is to use MSG_vm_set_bound(), which allows us to
1258 * directly set the bound of the dummy CPU action.
1262 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1263 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1264 * exceeds the capacity of a CPU core.
1266 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1268 return simcall_vm_set_bound(vm, bound);
1272 /** @brief Set the CPU affinity of a given VM.
1275 * This function changes the CPU affinity of a given VM. Usage is the same as
1276 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1278 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1280 msg_host_priv_t priv = msg_host_resource_priv(vm);
1283 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1285 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *) mask, NULL);
1287 msg_host_t pm_now = MSG_vm_get_pm(vm);
1289 XBT_INFO("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1290 simcall_vm_set_affinity(vm, pm, mask);
1292 XBT_INFO("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));