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"
25 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
26 "Cloud-oriented parts of the MSG API");
29 /* **** ******** GENERAL ********* **** */
31 /** \ingroup m_vm_management
32 * \brief Returns the value of a given vm property
35 * \param name a property name
36 * \return value of a property (or NULL if property not set)
39 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
41 return MSG_host_get_property_value(vm, name);
44 /** \ingroup m_vm_management
45 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
48 * \return a dict containing the properties
50 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
52 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
54 return (simcall_host_get_properties(vm));
57 /** \ingroup m_host_management
58 * \brief Change the value of a given host property
61 * \param name a property name
62 * \param value what to change the property to
63 * \param free_ctn the freeing function to use to kill the value on need
65 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
67 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
70 /** \ingroup msg_vm_management
71 * \brief Finds a msg_vm_t using its name.
73 * This is a name directory service
74 * \param name the name of a vm.
75 * \return the corresponding vm
77 * Please note that a VM is a specific host. Hence, you should give a different name
81 msg_vm_t MSG_vm_get_by_name(const char *name)
83 return MSG_get_host_by_name(name);
86 /** \ingroup m_vm_management
88 * \brief Return the name of the #msg_host_t.
90 * This functions checks whether \a host is a valid pointer or not and return
93 const char *MSG_vm_get_name(msg_vm_t vm)
95 return MSG_host_get_name(vm);
99 /* **** Check state of a VM **** */
100 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
102 return simcall_vm_get_state(vm) == state;
105 /** @brief Returns whether the given VM has just reated, not running.
108 int MSG_vm_is_created(msg_vm_t vm)
110 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
113 /** @brief Returns whether the given VM is currently running
116 int MSG_vm_is_running(msg_vm_t vm)
118 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
121 /** @brief Returns whether the given VM is currently migrating
124 int MSG_vm_is_migrating(msg_vm_t vm)
126 return __MSG_vm_is_state(vm, SURF_VM_STATE_MIGRATING);
129 /** @brief Returns whether the given VM is currently suspended, not running.
132 int MSG_vm_is_suspended(msg_vm_t vm)
134 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
137 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
140 int MSG_vm_is_saving(msg_vm_t vm)
142 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
145 /** @brief Returns whether the given VM has been saved, not running.
148 int MSG_vm_is_saved(msg_vm_t vm)
150 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
153 /** @brief Returns whether the given VM is being restored, not running.
156 int MSG_vm_is_restoring(msg_vm_t vm)
158 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
163 /* ------------------------------------------------------------------------- */
164 /* ------------------------------------------------------------------------- */
166 /* **** ******** MSG vm actions ********* **** */
168 /** @brief Create a new VM with specified parameters.
170 * All parameters are in MBytes
173 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name, int ncpus, int ramsize,
174 int net_cap, char *disk_path, int disksize,
175 int mig_netspeed, int dp_intensity)
177 /* For the moment, intensity_rate is the percentage against the migration bandwidth */
178 double host_speed = MSG_get_host_speed(ind_pm);
179 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
181 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
182 s_ws_params_t params;
183 memset(¶ms, 0, sizeof(params));
184 params.ramsize = 1L * 1024 * 1024 * ramsize;
185 //params.overcommit = 0;
187 params.skip_stage2 = 0;
188 params.max_downtime = 0.03;
189 params.dp_rate = (update_speed * 1L * 1024 * 1024 ) / host_speed;
190 params.dp_cap = params.ramsize / 0.9; // working set memory is 90%
191 params.mig_speed = 1L * 1024 * 1024 * mig_netspeed; // mig_speed
193 //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);
194 simcall_host_set_params(vm, ¶ms);
200 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
203 * A VM is treated as a host. The name of the VM must be unique among all hosts.
205 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
207 /* make sure the VM of the same name does not exit */
209 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
211 XBT_ERROR("host %s already exits", name);
216 /* Note: ind_vm and vm_workstation point to the same elm object. */
217 msg_vm_t ind_vm = NULL;
218 void *ind_vm_workstation = NULL;
220 /* Ask the SIMIX layer to create the surf vm resource */
221 ind_vm_workstation = simcall_vm_create(name, ind_pm);
222 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
224 XBT_DEBUG("A new VM (%s) has been created", name);
227 TRACE_msg_vm_create(name, ind_pm);
233 /** @brief Destroy a VM. Destroy the VM object from the simulation.
236 void MSG_vm_destroy(msg_vm_t vm)
238 /* First, terminate all processes on the VM if necessary */
239 if (MSG_vm_is_running(vm))
240 simcall_vm_shutdown(vm);
242 if (!MSG_vm_is_created(vm)) {
243 XBT_CRITICAL("shutdown the given VM before destroying it");
247 /* Then, destroy the VM object */
248 simcall_vm_destroy(vm);
250 __MSG_host_destroy(vm);
253 TRACE_msg_vm_end(vm);
258 /** @brief Start a vm (i.e., boot the guest operating system)
261 * If the VM cannot be started, an exception is generated.
264 void MSG_vm_start(msg_vm_t vm)
266 simcall_vm_start(vm);
269 TRACE_msg_vm_start(vm);
275 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
278 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
279 * use a #MSG_process_sleep() or something. I'm not quite sure.
281 void MSG_vm_shutdown(msg_vm_t vm)
283 /* msg_vm_t equals to msg_host_t */
284 simcall_vm_shutdown(vm);
286 // #ifdef HAVE_TRACING
287 // TRACE_msg_vm_(vm);
293 /* We have two mailboxes. mbox is used to transfer migration data between
294 * source and destiantion PMs. mbox_ctl is used to detect the completion of a
295 * migration. The names of these mailboxes must not conflict with others. */
296 static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
298 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
301 static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
303 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
306 static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
308 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
311 static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
313 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
316 static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
318 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
321 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio);
323 static int migration_rx_fun(int argc, char *argv[])
325 XBT_DEBUG("mig: rx_start");
327 xbt_assert(argc == 4);
328 const char *vm_name = argv[1];
329 const char *src_pm_name = argv[2];
330 const char *dst_pm_name = argv[3];
331 msg_vm_t vm = MSG_get_host_by_name(vm_name);
332 msg_host_t src_pm = MSG_get_host_by_name(src_pm_name);
333 msg_host_t dst_pm = MSG_get_host_by_name(dst_pm_name);
336 s_ws_params_t params;
337 simcall_host_get_params(vm, ¶ms);
338 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
343 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
344 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
345 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
348 msg_task_t task = NULL;
349 MSG_task_recv(&task, mbox);
351 double received = MSG_task_get_data_size(task);
353 // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
354 launch_deferred_exec_process(vm, received * xfer_cpu_overhead, 1);
357 if (strcmp(task->name, finalize_task_name) == 0)
360 MSG_task_destroy(task);
367 /* deinstall the current affinity setting */
368 simcall_vm_set_affinity(vm, src_pm, 0);
370 simcall_vm_migrate(vm, dst_pm);
371 simcall_vm_resume(vm);
373 /* install the affinity setting of the VM on the destination pm */
375 msg_host_priv_t priv = msg_host_resource_priv(vm);
377 unsigned long affinity_mask = (unsigned long) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *) dst_pm, sizeof(msg_host_t));
378 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
379 XBT_INFO("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
383 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
385 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
386 msg_error_t ret = MSG_task_send(task, mbox_ctl);
387 xbt_assert(ret == MSG_OK);
395 xbt_free(finalize_task_name);
397 XBT_DEBUG("mig: rx_done");
402 static void reset_dirty_pages(msg_vm_t vm)
404 msg_host_priv_t priv = msg_host_resource_priv(vm);
407 xbt_dict_cursor_t cursor = NULL;
408 dirty_page_t dp = NULL;
409 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
410 double remaining = MSG_task_get_remaining_computation(dp->task);
411 dp->prev_clock = MSG_get_clock();
412 dp->prev_remaining = remaining;
414 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
418 static void start_dirty_page_tracking(msg_vm_t vm)
420 msg_host_priv_t priv = msg_host_resource_priv(vm);
421 priv->dp_enabled = 1;
423 reset_dirty_pages(vm);
426 static void stop_dirty_page_tracking(msg_vm_t vm)
428 msg_host_priv_t priv = msg_host_resource_priv(vm);
429 priv->dp_enabled = 0;
433 /* It might be natural that we define dp_rate for each task. But, we will also
434 * have to care about how each task behavior affects the memory update behavior
435 * at the operating system level. It may not be easy to model it with a simple algorithm. */
436 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
438 double computed = dp->prev_remaining - remaining;
439 double duration = clock - dp->prev_clock;
440 double updated = dp->task->dp_rate * computed;
442 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
443 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
444 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
445 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
451 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
453 double computed = dp->prev_remaining - remaining;
454 double duration = clock - dp->prev_clock;
456 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
457 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
462 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
464 msg_host_priv_t priv = msg_host_resource_priv(vm);
468 xbt_dict_cursor_t cursor = NULL;
469 dirty_page_t dp = NULL;
470 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
471 double remaining = MSG_task_get_remaining_computation(dp->task);
473 double clock = MSG_get_clock();
475 // total += calc_updated_pages(key, vm, dp, remaining, clock);
476 total += get_computed(key, vm, dp, remaining, clock);
478 dp->prev_remaining = remaining;
479 dp->prev_clock = clock;
482 total += priv->dp_updated_by_deleted_tasks;
484 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
485 stage_for_fancy_debug,
486 stage2_round_for_fancy_debug,
487 total, priv->dp_updated_by_deleted_tasks);
491 priv->dp_updated_by_deleted_tasks = 0;
497 // TODO Is this code redundant with the information provided by
498 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
499 void MSG_host_add_task(msg_host_t host, msg_task_t task)
501 msg_host_priv_t priv = msg_host_resource_priv(host);
502 double remaining = MSG_task_get_remaining_computation(task);
503 char *key = bprintf("%s-%lld", task->name, task->counter);
505 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
508 /* It should be okay that we add a task onto a migrating VM. */
509 if (priv->dp_enabled) {
510 dp->prev_clock = MSG_get_clock();
511 dp->prev_remaining = remaining;
514 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
515 xbt_dict_set(priv->dp_objs, key, dp, NULL);
516 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
521 void MSG_host_del_task(msg_host_t host, msg_task_t task)
523 msg_host_priv_t priv = msg_host_resource_priv(host);
525 char *key = bprintf("%s-%lld", task->name, task->counter);
527 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
528 xbt_assert(dp->task == task);
530 /* If we are in the middle of dirty page tracking, we record how much
531 * computaion has been done until now, and keep the information for the
532 * lookup_() function that will called soon. */
533 if (priv->dp_enabled) {
534 double remaining = MSG_task_get_remaining_computation(task);
535 double clock = MSG_get_clock();
536 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
537 double updated = get_computed(key, host, dp, remaining, clock);
539 priv->dp_updated_by_deleted_tasks += updated;
542 xbt_dict_remove(priv->dp_objs, key);
545 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
551 static int deferred_exec_fun(int argc, char *argv[])
553 xbt_assert(argc == 3);
554 const char *comp_str = argv[1];
555 double computaion = atof(comp_str);
556 const char *prio_str = argv[2];
557 double prio = atof(prio_str);
559 msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
560 // XBT_INFO("exec deferred %f", computaion);
562 /* dpt is the results of the VM activity */
563 MSG_task_set_priority(task, prio);
564 MSG_task_execute(task);
568 MSG_task_destroy(task);
573 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
575 char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
578 char **argv = xbt_new(char *, nargvs);
579 argv[0] = xbt_strdup(pr_name);
580 argv[1] = bprintf("%lf", computation);
581 argv[2] = bprintf("%lf", prio);
584 MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv);
590 static int task_tx_overhead_fun(int argc, char *argv[])
592 xbt_assert(argc == 2);
593 const char *mbox = argv[1];
597 // XBT_INFO("start %s", mbox);
600 msg_task_t task = NULL;
601 MSG_task_recv(&task, mbox);
603 // XBT_INFO("task->name %s", task->name);
605 if (strcmp(task->name, "finalize_making_overhead") == 0)
609 // MSG_task_set_priority(task, 1000000);
610 MSG_task_execute(task);
611 MSG_task_destroy(task);
622 static void start_overhead_process(msg_task_t comm_task)
624 char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task));
625 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
628 char **argv = xbt_new(char *, nargvs);
629 argv[0] = xbt_strdup(pr_name);
630 argv[1] = xbt_strdup(mbox);
633 // XBT_INFO("micro start: mbox %s", mbox);
634 MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv);
640 static void shutdown_overhead_process(msg_task_t comm_task)
642 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
644 msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
646 // XBT_INFO("micro shutdown: mbox %s", mbox);
647 msg_error_t ret = MSG_task_send(task, mbox);
648 xbt_assert(ret == MSG_OK);
651 // XBT_INFO("shutdown done");
654 static void request_overhead(msg_task_t comm_task, double computation)
656 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
658 msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
660 // XBT_INFO("req overhead");
661 msg_error_t ret = MSG_task_send(task, mbox);
662 xbt_assert(ret == MSG_OK);
667 /* alpha is (floating_operations / bytes).
669 * When actual migration traffic was 32 mbytes/s, we observed the CPU
670 * utilization of the main thread of the Qemu process was 10 %.
671 * alpha = 0.1 * C / (32 * 1024 * 1024)
672 * where the CPU capacity of the PM is C flops/s.
675 static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
677 const double chunk_size = 1024 * 1024 * 10;
678 double remaining = MSG_task_get_data_size(comm_task);
680 start_overhead_process(comm_task);
683 while (remaining > 0) {
684 double data_size = chunk_size;
685 if (remaining < chunk_size)
686 data_size = remaining;
688 remaining -= data_size;
690 // XBT_INFO("remaining %f bytes", remaining);
693 double clock_sta = MSG_get_clock();
695 /* create a micro task */
697 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
698 msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
700 request_overhead(comm_task, data_size * alpha);
702 msg_error_t ret = MSG_task_send(mtask, mbox);
703 xbt_assert(ret == MSG_OK);
705 xbt_free(mtask_name);
710 /* In the real world, sending data involves small CPU computation. */
711 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
712 msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
713 MSG_task_execute(mtask);
714 MSG_task_destroy(mtask);
715 xbt_free(mtask_name);
721 double clock_end = MSG_get_clock();
726 * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
729 * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
731 * If time_to_sleep is smaller than zero, the elapsed time was too big. We
732 * do not need a micro sleep.
734 double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
735 if (time_to_sleep > 0)
736 MSG_process_sleep(time_to_sleep);
739 //XBT_INFO("duration %f", clock_end - clock_sta);
740 //XBT_INFO("time_to_sleep %f", time_to_sleep);
744 // XBT_INFO("%s", MSG_task_get_name(comm_task));
745 shutdown_overhead_process(comm_task);
751 static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
753 double prev_remaining = init_comm_size;
756 double remaining = MSG_task_get_remaining_communication(comm_task);
760 double sent = prev_remaining - remaining;
761 double comp_size = sent * overhead;
764 char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
765 msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
766 MSG_task_execute(comp_task);
767 MSG_task_destroy(comp_task);
772 prev_remaining = remaining;
776 xbt_free(comp_task_name);
780 // #define USE_MICRO_TASK 1
783 // const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
784 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
785 // const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
786 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
787 // const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
788 //const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
789 // const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
790 ////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
791 // const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
792 // const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
793 // const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
795 /* CPU 22% when 80Mbyte/s */
796 const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
800 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
801 double size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
803 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
804 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
808 double clock_sta = MSG_get_clock();
810 #ifdef USE_MICRO_TASK
812 task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
817 ret = MSG_task_send_bounded(task, mbox, mig_speed);
819 ret = MSG_task_send(task, mbox);
820 xbt_assert(ret == MSG_OK);
823 double clock_end = MSG_get_clock();
824 double duration = clock_end - clock_sta;
825 double actual_speed = size / duration;
826 #ifdef USE_MICRO_TASK
827 double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
829 double cpu_utilization = 0;
836 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);}
838 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);
845 #ifdef USE_MICRO_TASK
846 /* The name of a micro task starts with __micro, which does not match the
847 * special name that finalizes the receiver loop. Thus, we send the special task.
851 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
852 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
853 msg_error_t ret = MSG_task_send(task, mbox);
854 xbt_assert(ret == MSG_OK);
861 static double get_updated_size(double computed, double dp_rate, double dp_cap)
863 double updated_size = computed * dp_rate;
864 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
865 if (updated_size > dp_cap) {
866 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
867 updated_size = dp_cap;
873 static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *dst_pm_name,
874 long ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead)
876 const char *vm_name = MSG_host_get_name(vm);
877 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
879 // const long chunksize = 1024 * 1024 * 100;
880 const long chunksize = 1024L * 1024 * 100000;
881 long remaining = ramsize;
882 double computed_total = 0;
884 while (remaining > 0) {
885 long datasize = chunksize;
886 if (remaining < chunksize)
887 datasize = remaining;
889 remaining -= datasize;
891 send_migration_data(vm_name, src_pm_name, dst_pm_name, datasize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
893 double computed = lookup_computed_flop_counts(vm, 1, 0);
894 computed_total += computed;
897 // double updated_size = get_updated_size(computed, dp_rate, dp_cap);
899 // double overhead = dpt_cpu_overhead * updated_size;
900 // launch_deferred_exec_process(vm, overhead, 10000);
904 return computed_total;
909 static double get_threshold_value(double bandwidth, double max_downtime)
911 /* This value assumes the network link is 1Gbps. */
912 // double threshold = max_downtime * 125 * 1024 * 1024;
913 double threshold = max_downtime * bandwidth;
918 static int migration_tx_fun(int argc, char *argv[])
920 XBT_DEBUG("mig: tx_start");
922 xbt_assert(argc == 4);
923 const char *vm_name = argv[1];
924 const char *src_pm_name = argv[2];
925 const char *dst_pm_name = argv[3];
926 msg_vm_t vm = MSG_get_host_by_name(vm_name);
929 s_ws_params_t params;
930 simcall_host_get_params(vm, ¶ms);
931 const long ramsize = params.ramsize;
932 const long devsize = params.devsize;
933 const int skip_stage1 = params.skip_stage1;
934 const int skip_stage2 = params.skip_stage2;
935 const double dp_rate = params.dp_rate;
936 const double dp_cap = params.dp_cap;
937 const double mig_speed = params.mig_speed;
938 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
939 const double dpt_cpu_overhead = params.dpt_cpu_overhead;
941 double remaining_size = ramsize + devsize;
943 double max_downtime = params.max_downtime;
944 if (max_downtime == 0) {
945 XBT_WARN("use the default max_downtime value 30ms");
949 double threshold = 0.00001; /* TODO: cleanup */
951 /* setting up parameters has done */
955 XBT_WARN("migrate a VM, but ramsize is zero");
957 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
959 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
961 /* Stage1: send all memory pages to the destination. */
962 start_dirty_page_tracking(vm);
964 double computed_during_stage1 = 0;
966 // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
968 /* send ramsize, but split it */
969 double clock_prev_send = MSG_get_clock();
971 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);
972 remaining_size -= ramsize;
974 double clock_post_send = MSG_get_clock();
975 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
976 threshold = get_threshold_value(bandwidth, max_downtime);
977 XBT_INFO("actual banwdidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
981 /* Stage2: send update pages iteratively until the size of remaining states
982 * becomes smaller than the threshold value. */
985 if (max_downtime == 0) {
986 XBT_WARN("no max_downtime parameter, skip stage2");
991 int stage2_round = 0;
994 double updated_size = 0;
995 if (stage2_round == 0) {
996 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
997 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
999 double computed = lookup_computed_flop_counts(vm, 2, stage2_round);
1000 updated_size = get_updated_size(computed, dp_rate, dp_cap);
1003 XBT_INFO("mig-stage 2:%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
1004 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
1007 // if (stage2_round != 0) {
1008 // /* during stage1, we have already created overhead tasks */
1009 // double overhead = dpt_cpu_overhead * updated_size;
1010 // XBT_DEBUG("updated %f overhead %f", updated_size, overhead);
1011 // launch_deferred_exec_process(vm, overhead, 10000);
1016 remaining_size += updated_size;
1018 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
1019 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
1021 if (remaining_size < threshold)
1025 double clock_prev_send = MSG_get_clock();
1027 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
1029 double clock_post_send = MSG_get_clock();
1031 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
1032 threshold = get_threshold_value(bandwidth, max_downtime);
1033 XBT_INFO("actual banwdidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
1041 remaining_size -= updated_size;
1047 /* Stage3: stop the VM and copy the rest of states. */
1048 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
1049 simcall_vm_suspend(vm);
1050 stop_dirty_page_tracking(vm);
1052 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead);
1056 XBT_DEBUG("mig: tx_done");
1063 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
1065 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1068 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1070 char **argv = xbt_new(char *, nargvs);
1071 argv[0] = xbt_strdup(pr_name);
1072 argv[1] = xbt_strdup(sg_host_name(vm));
1073 argv[2] = xbt_strdup(sg_host_name(src_pm));
1074 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1077 MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
1083 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
1085 char **argv = xbt_new(char *, nargvs);
1086 argv[0] = xbt_strdup(pr_name);
1087 argv[1] = xbt_strdup(sg_host_name(vm));
1088 argv[2] = xbt_strdup(sg_host_name(src_pm));
1089 argv[3] = xbt_strdup(sg_host_name(dst_pm));
1091 MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
1096 /* wait until the migration have finished */
1098 msg_task_t task = NULL;
1099 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
1100 xbt_assert(ret == MSG_OK);
1102 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
1103 xbt_assert(strcmp(task->name, expected_task_name) == 0);
1104 xbt_free(expected_task_name);
1111 /** @brief Migrate the VM to the given host.
1114 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
1115 * MSG_task_send() before or after, depending on whether you want to do cold or hot
1118 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
1121 * - One approach is ...
1122 * We first create a new VM (i.e., destination VM) on the destination
1123 * physical host. The destination VM will receive the state of the source
1124 * VM over network. We will finally destroy the source VM.
1125 * - This behavior is similar to the way of migration in the real world.
1126 * Even before a migration is completed, we will see a destination VM,
1127 * consuming resources.
1128 * - We have to relocate all processes. The existing process migraion code
1129 * will work for this?
1130 * - The name of the VM is a somewhat unique ID in the code. It is tricky
1131 * for the destination VM?
1133 * - Another one is ...
1134 * We update the information of the given VM to place it to the destination
1137 * The second one would be easier.
1141 msg_host_t old_pm = simcall_vm_get_pm(vm);
1143 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
1144 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
1146 do_migration(vm, old_pm, new_pm);
1150 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1153 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1158 /** @brief Immediately suspend the execution of all processes within the given VM.
1161 * This function stops the exection of the VM. All the processes on this VM
1162 * will pause. The state of the VM is perserved. We can later resume it again.
1164 * No suspension cost occurs.
1166 void MSG_vm_suspend(msg_vm_t vm)
1168 simcall_vm_suspend(vm);
1170 XBT_DEBUG("vm_suspend done");
1173 TRACE_msg_vm_suspend(vm);
1178 /** @brief Resume the execution of the VM. All processes on the VM run again.
1181 * No resume cost occurs.
1183 void MSG_vm_resume(msg_vm_t vm)
1185 simcall_vm_resume(vm);
1188 TRACE_msg_vm_resume(vm);
1193 /** @brief Immediately save the execution of all processes within the given VM.
1196 * This function stops the exection of the VM. All the processes on this VM
1197 * will pause. The state of the VM is perserved. We can later resume it again.
1199 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1200 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1201 * of VM save to you.
1203 void MSG_vm_save(msg_vm_t vm)
1205 simcall_vm_save(vm);
1207 TRACE_msg_vm_save(vm);
1211 /** @brief Restore the execution of the VM. All processes on the VM run again.
1214 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1215 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1216 * of VM restore to you.
1218 void MSG_vm_restore(msg_vm_t vm)
1220 simcall_vm_restore(vm);
1223 TRACE_msg_vm_restore(vm);
1228 /** @brief Get the physical host of a given VM.
1231 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1233 return simcall_vm_get_pm(vm);
1237 /** @brief Set a CPU bound for a given VM.
1241 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1244 * On PM0, there are Task1 and VM0.
1245 * On VM0, there is Task2.
1246 * Now we bound 75% to Task1@PM0 and bound 25% to Task2@VM0.
1248 * Task1@PM0 gets 50%.
1249 * Task2@VM0 gets 25%.
1250 * This is NOT 75% for Task1@PM0 and 25% for Task2@VM0, respectively.
1252 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1253 * task on the VM does not affect the bound of the dummy CPU action. The bound
1254 * of the dummy CPU action is unlimited.
1256 * There are some solutions for this problem. One option is to update the bound
1257 * of the dummy CPU action automatically. It should be the sum of all tasks on
1258 * the VM. But, this solution might be costy, because we have to scan all tasks
1259 * on the VM in share_resource() or we have to trap both the start and end of
1262 * The current solution is to use MSG_vm_set_bound(), which allows us to
1263 * directly set the bound of the dummy CPU action.
1267 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1268 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1269 * exceeds the capacity of a CPU core.
1271 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1273 return simcall_vm_set_bound(vm, bound);
1277 /** @brief Set the CPU affinity of a given VM.
1280 * This function changes the CPU affinity of a given VM. Usage is the same as
1281 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1283 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1285 msg_host_priv_t priv = msg_host_resource_priv(vm);
1288 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1290 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *) mask, NULL);
1292 msg_host_t pm_now = MSG_vm_get_pm(vm);
1294 XBT_INFO("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1295 simcall_vm_set_affinity(vm, pm, mask);
1297 XBT_INFO("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));