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. add the support of trace
9 * 2. use parallel tasks to simulate CPU overhead and remove the very
10 * experimental code generating micro computation tasks
15 #include "msg_private.h"
16 #include "xbt/sysdep.h"
18 #include "simgrid/platf.h"
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
21 "Cloud-oriented parts of the MSG API");
24 /* **** ******** GENERAL ********* **** */
26 /** \ingroup m_vm_management
27 * \brief Returns the value of a given vm property
30 * \param name a property name
31 * \return value of a property (or NULL if property not set)
34 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
36 return MSG_host_get_property_value(vm, name);
39 /** \ingroup m_vm_management
40 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
43 * \return a dict containing the properties
45 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
47 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
49 return (simcall_host_get_properties(vm));
52 /** \ingroup m_host_management
53 * \brief Change the value of a given host property
56 * \param name a property name
57 * \param value what to change the property to
58 * \param free_ctn the freeing function to use to kill the value on need
60 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
62 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
65 /** \ingroup msg_vm_management
66 * \brief Finds a msg_vm_t using its name.
68 * This is a name directory service
69 * \param name the name of a vm.
70 * \return the corresponding vm
72 * Please note that a VM is a specific host. Hence, you should give a different name
76 msg_vm_t MSG_vm_get_by_name(const char *name)
78 return MSG_get_host_by_name(name);
81 /** \ingroup m_vm_management
83 * \brief Return the name of the #msg_host_t.
85 * This functions checks whether \a host is a valid pointer or not and return
88 const char *MSG_vm_get_name(msg_vm_t vm)
90 return MSG_host_get_name(vm);
94 /* **** Check state of a VM **** */
95 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
97 return simcall_vm_get_state(vm) == state;
100 /** @brief Returns whether the given VM has just created, not running.
103 int MSG_vm_is_created(msg_vm_t vm)
105 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
108 /** @brief Returns whether the given VM is currently running
111 int MSG_vm_is_running(msg_vm_t vm)
113 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
116 /** @brief Returns whether the given VM is currently migrating
119 int MSG_vm_is_migrating(msg_vm_t vm)
121 msg_host_priv_t priv = msg_host_resource_priv(vm);
122 return priv->is_migrating;
125 /** @brief Returns whether the given VM is currently suspended, not running.
128 int MSG_vm_is_suspended(msg_vm_t vm)
130 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
133 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
136 int MSG_vm_is_saving(msg_vm_t vm)
138 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
141 /** @brief Returns whether the given VM has been saved, not running.
144 int MSG_vm_is_saved(msg_vm_t vm)
146 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
149 /** @brief Returns whether the given VM is being restored, not running.
152 int MSG_vm_is_restoring(msg_vm_t vm)
154 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
159 /* ------------------------------------------------------------------------- */
160 /* ------------------------------------------------------------------------- */
162 /* **** ******** MSG vm actions ********* **** */
164 /** @brief Create a new VM with specified parameters.
166 * All parameters are in MBytes
169 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name,
170 int ncpus, int ramsize,
171 int net_cap, char *disk_path, int disksize,
172 int mig_netspeed, int dp_intensity)
174 /* For the moment, intensity_rate is the percentage against the migration
176 double host_speed = MSG_get_host_speed(ind_pm);
177 double update_speed = ((double)dp_intensity/100) * mig_netspeed;
179 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
180 s_ws_params_t params;
181 memset(¶ms, 0, sizeof(params));
182 params.ramsize = (sg_size_t)ramsize * 1024 * 1024;
183 //params.overcommit = 0;
185 params.skip_stage2 = 0;
186 params.max_downtime = 0.03;
187 params.dp_rate = (update_speed * 1024 * 1024) / host_speed;
188 params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize
189 params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed
191 //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);
192 simcall_host_set_params(vm, ¶ms);
198 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
201 * A VM is treated as a host. The name of the VM must be unique among all hosts.
203 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
205 /* make sure the VM of the same name does not exit */
207 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
209 XBT_ERROR("host %s already exits", name);
214 /* Note: ind_vm and vm_workstation point to the same elm object. */
215 msg_vm_t ind_vm = NULL;
216 void *ind_vm_workstation = NULL;
218 /* Ask the SIMIX layer to create the surf vm resource */
219 ind_vm_workstation = simcall_vm_create(name, ind_pm);
220 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
222 XBT_DEBUG("A new VM (%s) has been created", name);
225 TRACE_msg_vm_create(name, ind_pm);
231 /** @brief Destroy a VM. Destroy the VM object from the simulation.
234 void MSG_vm_destroy(msg_vm_t vm)
236 if (MSG_vm_is_migrating(vm))
237 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
239 /* First, terminate all processes on the VM if necessary */
240 if (MSG_vm_is_running(vm))
241 simcall_vm_shutdown(vm);
243 if (!MSG_vm_is_created(vm)) {
244 XBT_CRITICAL("shutdown the given VM before destroying it");
248 /* Then, destroy the VM object */
249 simcall_vm_destroy(vm);
251 __MSG_host_destroy(vm);
254 TRACE_msg_vm_end(vm);
256 xbt_lib_remove(host_lib, sg_host_name(vm));
260 /** @brief Start a vm (i.e., boot the guest operating system)
263 * If the VM cannot be started, an exception is generated.
266 void MSG_vm_start(msg_vm_t vm)
268 simcall_vm_start(vm);
271 TRACE_msg_vm_start(vm);
277 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
280 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
281 * use a #MSG_process_sleep() or something. I'm not quite sure.
283 void MSG_vm_shutdown(msg_vm_t vm)
285 /* msg_vm_t equals to msg_host_t */
286 simcall_vm_shutdown(vm);
288 // #ifdef HAVE_TRACING
289 // TRACE_msg_vm_(vm);
295 /* We have two mailboxes. mbox is used to transfer migration data between
296 * source and destination PMs. mbox_ctl is used to detect the completion of a
297 * migration. The names of these mailboxes must not conflict with others. */
298 static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
300 char *vm_name = sg_host_name(vm);
301 char *src_pm_name = sg_host_name(src_pm);
302 char *dst_pm_name = sg_host_name(dst_pm);
304 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
307 static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
309 char *vm_name = sg_host_name(vm);
310 char *src_pm_name = sg_host_name(src_pm);
311 char *dst_pm_name = sg_host_name(dst_pm);
313 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
316 static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
318 char *vm_name = sg_host_name(vm);
319 char *src_pm_name = sg_host_name(src_pm);
320 char *dst_pm_name = sg_host_name(dst_pm);
322 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
325 static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
327 char *vm_name = sg_host_name(vm);
328 char *src_pm_name = sg_host_name(src_pm);
329 char *dst_pm_name = sg_host_name(dst_pm);
331 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
334 static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
336 char *vm_name = sg_host_name(vm);
337 char *src_pm_name = sg_host_name(src_pm);
338 char *dst_pm_name = sg_host_name(dst_pm);
340 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
343 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio);
346 struct migration_session {
351 /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
352 * know the completion of the migration. */
354 /* The migration_rx and migration_tx processes use mbox to transfer migration
360 static int migration_rx_fun(int argc, char *argv[])
362 XBT_DEBUG("mig: rx_start");
364 // The structure has been created in the do_migration function and should only be freed in the same place ;)
365 struct migration_session *ms = MSG_process_get_data(MSG_process_self());
367 s_ws_params_t params;
368 simcall_host_get_params(ms->vm, ¶ms);
369 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
373 char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
377 msg_task_t task = NULL;
378 ret = MSG_task_recv(&task, ms->mbox);
382 received = MSG_task_get_data_size(task);
384 // An error occured, clean the code and return
385 // The owner did not change, hence the task should be only destroyed on the other side
386 xbt_free(finalize_task_name);
390 // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
391 launch_deferred_exec_process(ms->vm, received * xfer_cpu_overhead, 1);
394 if (strcmp(task->name, finalize_task_name) == 0)
397 MSG_task_destroy(task);
403 // Here Stage 1, 2 and 3 have been performed.
404 // Hence complete the migration
406 // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
407 // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms structure at the begining and free it at the end of each function)
408 msg_vm_t vm = ms->vm;
409 msg_host_t src_pm = ms->src_pm;
410 msg_host_t dst_pm = ms-> dst_pm;
411 msg_host_priv_t priv = msg_host_resource_priv(vm);
413 // TODO: we have an issue, if the DST node is turning off during the three next calls, then the VM is in an inconsistent state
414 // I should check with Takahiro in order to make this portion of code atomic
415 /* deinstall the current affinity setting for the CPU */
416 simcall_vm_set_affinity(vm, src_pm, 0);
418 /* Update the vm location */
419 simcall_vm_migrate(vm, dst_pm);
422 simcall_vm_resume(vm);
424 /* install the affinity setting of the VM on the destination pm */
427 unsigned long affinity_mask = (unsigned long) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *)dst_pm, sizeof(msg_host_t));
428 simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
429 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
434 // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
435 msg_host_priv_t priv = msg_host_resource_priv(vm);
436 priv->is_migrating = 0;
437 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->key, ms->src_pm->key, ms->dst_pm->key);
439 TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
443 // Inform the SRC that the migration has been correctly performed
445 char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
446 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
447 msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
448 // xbt_assert(ret == MSG_OK);
449 if(ret == MSG_HOST_FAILURE){
450 // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM has been correctly migrated on the DST node
451 // TODO What does it mean ? What should we do ?
452 MSG_task_destroy(task);
453 } else if(ret == MSG_TRANSFER_FAILURE){
454 // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
455 MSG_task_destroy(task);
462 xbt_free(finalize_task_name);
464 XBT_DEBUG("mig: rx_done");
469 static void reset_dirty_pages(msg_vm_t vm)
471 msg_host_priv_t priv = msg_host_resource_priv(vm);
474 xbt_dict_cursor_t cursor = NULL;
475 dirty_page_t dp = NULL;
476 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
477 double remaining = MSG_task_get_remaining_computation(dp->task);
478 dp->prev_clock = MSG_get_clock();
479 dp->prev_remaining = remaining;
481 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
485 static void start_dirty_page_tracking(msg_vm_t vm)
487 msg_host_priv_t priv = msg_host_resource_priv(vm);
488 priv->dp_enabled = 1;
490 reset_dirty_pages(vm);
493 static void stop_dirty_page_tracking(msg_vm_t vm)
495 msg_host_priv_t priv = msg_host_resource_priv(vm);
496 priv->dp_enabled = 0;
500 /* It might be natural that we define dp_rate for each task. But, we will also
501 * have to care about how each task behavior affects the memory update behavior
502 * at the operating system level. It may not be easy to model it with a simple algorithm. */
503 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
505 double computed = dp->prev_remaining - remaining;
506 double duration = clock - dp->prev_clock;
507 double updated = dp->task->dp_rate * computed;
509 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
510 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
511 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
512 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
518 static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
520 double computed = dp->prev_remaining - remaining;
521 double duration = clock - dp->prev_clock;
523 XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
524 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
529 static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
531 msg_host_priv_t priv = msg_host_resource_priv(vm);
535 xbt_dict_cursor_t cursor = NULL;
536 dirty_page_t dp = NULL;
537 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
538 double remaining = MSG_task_get_remaining_computation(dp->task);
540 double clock = MSG_get_clock();
542 // total += calc_updated_pages(key, vm, dp, remaining, clock);
543 total += get_computed(key, vm, dp, remaining, clock);
545 dp->prev_remaining = remaining;
546 dp->prev_clock = clock;
549 total += priv->dp_updated_by_deleted_tasks;
551 XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
552 stage_for_fancy_debug,
553 stage2_round_for_fancy_debug,
554 total, priv->dp_updated_by_deleted_tasks);
558 priv->dp_updated_by_deleted_tasks = 0;
564 // TODO Is this code redundant with the information provided by
565 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
566 void MSG_host_add_task(msg_host_t host, msg_task_t task)
568 msg_host_priv_t priv = msg_host_resource_priv(host);
569 double remaining = MSG_task_get_remaining_computation(task);
570 char *key = bprintf("%s-%p", task->name, task);
572 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
575 /* It should be okay that we add a task onto a migrating VM. */
576 if (priv->dp_enabled) {
577 dp->prev_clock = MSG_get_clock();
578 dp->prev_remaining = remaining;
581 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
582 xbt_dict_set(priv->dp_objs, key, dp, NULL);
583 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
588 void MSG_host_del_task(msg_host_t host, msg_task_t task)
590 msg_host_priv_t priv = msg_host_resource_priv(host);
592 char *key = bprintf("%s-%p", task->name, task);
594 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
595 xbt_assert(dp->task == task);
597 /* If we are in the middle of dirty page tracking, we record how much
598 * computation has been done until now, and keep the information for the
599 * lookup_() function that will called soon. */
600 if (priv->dp_enabled) {
601 double remaining = MSG_task_get_remaining_computation(task);
602 double clock = MSG_get_clock();
603 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
604 double updated = get_computed(key, host, dp, remaining, clock);
606 priv->dp_updated_by_deleted_tasks += updated;
609 xbt_dict_remove(priv->dp_objs, key);
612 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
618 static int deferred_exec_fun(int argc, char *argv[])
620 xbt_assert(argc == 3);
621 const char *comp_str = argv[1];
622 double computaion = atof(comp_str);
623 const char *prio_str = argv[2];
624 double prio = atof(prio_str);
626 msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL);
627 // XBT_INFO("exec deferred %f", computation);
629 /* dpt is the results of the VM activity */
630 MSG_task_set_priority(task, prio);
631 MSG_task_execute(task);
635 MSG_task_destroy(task);
640 static void launch_deferred_exec_process(msg_host_t host, double computation, double prio)
642 char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host));
645 char **argv = xbt_new(char *, nargvs);
647 argv[1] = bprintf("%f", computation);
648 argv[2] = bprintf("%f", prio);
651 MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv);
655 static int task_tx_overhead_fun(int argc, char *argv[])
657 xbt_assert(argc == 2);
658 const char *mbox = argv[1];
662 // XBT_INFO("start %s", mbox);
665 msg_task_t task = NULL;
666 MSG_task_recv(&task, mbox);
668 // XBT_INFO("task->name %s", task->name);
670 if (strcmp(task->name, "finalize_making_overhead") == 0)
674 // MSG_task_set_priority(task, 1000000);
675 MSG_task_execute(task);
676 MSG_task_destroy(task);
687 static void start_overhead_process(msg_task_t comm_task)
689 char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task));
690 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
693 char **argv = xbt_new(char *, nargvs);
698 // XBT_INFO("micro start: mbox %s", mbox);
699 MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv);
702 static void shutdown_overhead_process(msg_task_t comm_task)
704 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
706 msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL);
708 // XBT_INFO("micro shutdown: mbox %s", mbox);
709 msg_error_t ret = MSG_task_send(task, mbox);
711 xbt_die("shutdown error - task not sent");
714 // XBT_INFO("shutdown done");
717 static void request_overhead(msg_task_t comm_task, double computation)
719 char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task));
721 msg_task_t task = MSG_task_create("micro", computation, 0, NULL);
723 // XBT_INFO("req overhead");
724 msg_error_t ret = MSG_task_send(task, mbox);
726 xbt_die("req overhead error - task not sent");
731 /* alpha is (floating_operations / bytes).
733 * When actual migration traffic was 32 mbytes/s, we observed the CPU
734 * utilization of the main thread of the Qemu process was 10 %.
735 * alpha = 0.1 * C / (32 * 1024 * 1024)
736 * where the CPU capacity of the PM is C flops/s.
739 static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha)
741 const double chunk_size = 1024 * 1024 * 10;
742 double remaining = MSG_task_get_data_size(comm_task);
744 start_overhead_process(comm_task);
747 while (remaining > 0) {
748 double data_size = chunk_size;
749 if (remaining < chunk_size)
750 data_size = remaining;
752 remaining -= data_size;
754 // XBT_INFO("remaining %f bytes", remaining);
757 double clock_sta = MSG_get_clock();
759 /* create a micro task */
761 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
762 msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL);
764 request_overhead(comm_task, data_size * alpha);
766 msg_error_t ret = MSG_task_send(mtask, mbox);
768 xbt_die("migration error - task not sent");
770 xbt_free(mtask_name);
775 /* In the real world, sending data involves small CPU computation. */
776 char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task));
777 msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL);
778 MSG_task_execute(mtask);
779 MSG_task_destroy(mtask);
780 xbt_free(mtask_name);
786 double clock_end = MSG_get_clock();
791 * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep)
794 * time_to_sleep > data_size / (max bandwidth) - (elapsed time)
796 * If time_to_sleep is smaller than zero, the elapsed time was too big. We
797 * do not need a micro sleep.
799 double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta);
800 if (time_to_sleep > 0)
801 MSG_process_sleep(time_to_sleep);
804 //XBT_INFO("duration %f", clock_end - clock_sta);
805 //XBT_INFO("time_to_sleep %f", time_to_sleep);
809 // XBT_INFO("%s", MSG_task_get_name(comm_task));
810 shutdown_overhead_process(comm_task);
816 static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size)
818 double prev_remaining = init_comm_size;
821 double remaining = MSG_task_get_remaining_communication(comm_task);
825 double sent = prev_remaining - remaining;
826 double comp_size = sent * overhead;
829 char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task));
830 msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL);
831 MSG_task_execute(comp_task);
832 MSG_task_destroy(comp_task);
837 prev_remaining = remaining;
841 xbt_free(comp_task_name);
845 // #define USE_MICRO_TASK 1
848 // const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024);
849 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
850 // const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024);
851 // const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024);
852 // const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s
853 //const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024);
854 // const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024);
855 ////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
856 // const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024);
857 // const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024);
858 // const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024);
860 /* CPU 22% when 80Mbyte/s */
861 const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024);
865 static void send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm,
866 sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead)
868 char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
869 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
873 double clock_sta = MSG_get_clock();
875 #ifdef USE_MICRO_TASK
877 task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead);
882 ret = MSG_task_send_bounded(task, mbox, mig_speed);
884 ret = MSG_task_send(task, mbox);
886 // xbt_assert(ret == MSG_OK);
888 if(ret == MSG_HOST_FAILURE){
889 //XBT_INFO("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
890 MSG_task_destroy(task);
891 THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
892 }else if(ret == MSG_TRANSFER_FAILURE){
893 //XBT_INFO("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
894 MSG_task_destroy(task);
895 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
898 // XBT_INFO("Ret != FAILURE !!!!");
901 double clock_end = MSG_get_clock();
902 double duration = clock_end - clock_sta;
903 double actual_speed = size / duration;
904 #ifdef USE_MICRO_TASK
905 double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8;
907 double cpu_utilization = 0;
911 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);}
913 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);
917 #ifdef USE_MICRO_TASK
918 /* The name of a micro task starts with __micro, which does not match the
919 * special name that finalizes the receiver loop. Thus, we send the special task.
923 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
924 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
925 msg_error_t ret = MSG_task_send(task, mbox);
926 // xbt_assert(ret == MSG_OK);
928 if(ret == MSG_HOST_FAILURE){
929 //XBT_INFO("SRC host failed during migration of %s (stage 3)", sg_host_name(vm));
930 MSG_task_destroy(task);
931 THROWF(host_error, 0, "SRC host failed during migration of VM %s (stage 3)", sg_host_name(vm));
932 // The owner of the task did not change so destroy the task
934 }else if(ret == MSG_TRANSFER_FAILURE){
935 //XBT_INFO("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
936 MSG_task_destroy(task);
937 THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
947 static double get_updated_size(double computed, double dp_rate, double dp_cap)
949 double updated_size = computed * dp_rate;
950 XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate);
951 if (updated_size > dp_cap) {
952 // XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f", stage2_round, updated_size, dp_cap);
953 updated_size = dp_cap;
959 static double send_stage1(struct migration_session *ms,
960 sg_size_t ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead)
963 // const long chunksize = (sg_size_t)1024 * 1024 * 100;
964 const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000;
965 sg_size_t remaining = ramsize;
966 double computed_total = 0;
968 while (remaining > 0) {
969 sg_size_t datasize = chunksize;
970 if (remaining < chunksize)
971 datasize = remaining;
973 remaining -= datasize;
974 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, xfer_cpu_overhead);
975 double computed = lookup_computed_flop_counts(ms->vm, 1, 0);
976 computed_total += computed;
979 // double updated_size = get_updated_size(computed, dp_rate, dp_cap);
981 // double overhead = dpt_cpu_overhead * updated_size;
982 // launch_deferred_exec_process(vm, overhead, 10000);
986 return computed_total;
991 static double get_threshold_value(double bandwidth, double max_downtime)
993 /* This value assumes the network link is 1Gbps. */
994 // double threshold = max_downtime * 125 * 1024 * 1024;
995 double threshold = max_downtime * bandwidth;
1000 static int migration_tx_fun(int argc, char *argv[])
1002 XBT_DEBUG("mig: tx_start");
1004 // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;)
1005 struct migration_session *ms = MSG_process_get_data(MSG_process_self());
1007 s_ws_params_t params;
1008 simcall_host_get_params(ms->vm, ¶ms);
1009 const sg_size_t ramsize = params.ramsize;
1010 const sg_size_t devsize = params.devsize;
1011 const int skip_stage1 = params.skip_stage1;
1012 const int skip_stage2 = params.skip_stage2;
1013 const double dp_rate = params.dp_rate;
1014 const double dp_cap = params.dp_cap;
1015 const double mig_speed = params.mig_speed;
1016 const double xfer_cpu_overhead = params.xfer_cpu_overhead;
1017 const double dpt_cpu_overhead = params.dpt_cpu_overhead;
1021 double remaining_size = ramsize + devsize;
1023 double max_downtime = params.max_downtime;
1024 if (max_downtime == 0) {
1025 XBT_WARN("use the default max_downtime value 30ms");
1026 max_downtime = 0.03;
1029 double threshold = 0.00001; /* TODO: cleanup */
1031 /* setting up parameters has done */
1035 XBT_WARN("migrate a VM, but ramsize is zero");
1038 XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size);
1040 /* Stage1: send all memory pages to the destination. */
1041 start_dirty_page_tracking(vm);
1043 double computed_during_stage1 = 0;
1045 // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead);
1047 /* send ramsize, but split it */
1048 double clock_prev_send = MSG_get_clock();
1051 computed_during_stage1 = send_stage1(ms, ramsize, mig_speed, xfer_cpu_overhead, dp_rate, dp_cap, dpt_cpu_overhead);
1053 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
1054 // Stop the dirty page tracking an return (there is no memory space to release)
1055 stop_dirty_page_tracking(vm);
1058 remaining_size -= ramsize;
1060 double clock_post_send = MSG_get_clock();
1061 double bandwidth = ramsize / (clock_post_send - clock_prev_send);
1062 threshold = get_threshold_value(bandwidth, max_downtime);
1063 XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold);
1067 /* Stage2: send update pages iteratively until the size of remaining states
1068 * becomes smaller than the threshold value. */
1071 if (max_downtime == 0) {
1072 XBT_WARN("no max_downtime parameter, skip stage2");
1077 int stage2_round = 0;
1080 double updated_size = 0;
1081 if (stage2_round == 0) {
1082 /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */
1083 updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap);
1085 double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
1086 updated_size = get_updated_size(computed, dp_rate, dp_cap);
1089 XBT_DEBUG("mig-stage 2:%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f",
1090 stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap);
1093 // if (stage2_round != 0) {
1094 // /* during stage1, we have already created overhead tasks */
1095 // double overhead = dpt_cpu_overhead * updated_size;
1096 // XBT_DEBUG("updated %f overhead %f", updated_size, overhead);
1097 // launch_deferred_exec_process(vm, overhead, 10000);
1102 remaining_size += updated_size;
1104 XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
1105 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
1107 if (remaining_size < threshold)
1111 double clock_prev_send = MSG_get_clock();
1113 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead);
1115 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
1116 // Stop the dirty page tracking an return (there is no memory space to release)
1117 stop_dirty_page_tracking(vm);
1120 double clock_post_send = MSG_get_clock();
1122 double bandwidth = updated_size / (clock_post_send - clock_prev_send);
1123 threshold = get_threshold_value(bandwidth, max_downtime);
1124 XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold);
1127 remaining_size -= updated_size;
1133 /* Stage3: stop the VM and copy the rest of states. */
1134 XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size);
1135 simcall_vm_suspend(vm);
1136 stop_dirty_page_tracking(vm);
1139 send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, remaining_size, ms->mbox, 3, 0, mig_speed, xfer_cpu_overhead);
1141 //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code)
1142 // Stop the dirty page tracking an return (there is no memory space to release)
1143 simcall_vm_resume(vm);
1147 // At that point the Migration is considered valid for the SRC node but remind that the DST side should relocate effectively the VM on the DST node.
1149 XBT_DEBUG("mig: tx_done");
1156 static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
1158 struct migration_session *ms = xbt_new(struct migration_session, 1);
1160 ms->src_pm = src_pm;
1161 ms->dst_pm = dst_pm;
1162 ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm);
1163 ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm);
1166 char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
1167 char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
1169 // MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm);
1170 // MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm);
1173 char **argv = xbt_new(char *, 2);
1174 argv[0] = pr_rx_name;
1176 MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv);
1179 char **argv = xbt_new(char *, 2);
1180 argv[0] = pr_tx_name;
1182 MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv);
1186 /* wait until the migration have finished or on error has occured */
1188 XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed");
1189 msg_task_t task = NULL;
1190 msg_error_t ret = MSG_TIMEOUT;
1191 while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o
1192 ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 10);
1194 xbt_free(ms->mbox_ctl);
1198 //xbt_assert(ret == MSG_OK);
1199 if(ret == MSG_HOST_FAILURE){
1200 // Note that since the communication failed, the owner did not change and the task should be destroyed on the other side.
1201 // Hence, just throw the execption
1202 //XBT_INFO("SRC crashes, throw an exception (m-control)");
1205 else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){ // MSG_TIMEOUT here means that MSG_host_is_avail() returned false.
1206 //XBT_INFO("DST crashes, throw an exception (m-control)");
1211 char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4);
1212 xbt_assert(strcmp(task->name, expected_task_name) == 0);
1213 xbt_free(expected_task_name);
1214 MSG_task_destroy(task);
1222 /** @brief Migrate the VM to the given host.
1225 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
1226 * MSG_task_send() before or after, depending on whether you want to do cold or hot
1229 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
1232 * - One approach is ...
1233 * We first create a new VM (i.e., destination VM) on the destination
1234 * physical host. The destination VM will receive the state of the source
1235 * VM over network. We will finally destroy the source VM.
1236 * - This behavior is similar to the way of migration in the real world.
1237 * Even before a migration is completed, we will see a destination VM,
1238 * consuming resources.
1239 * - We have to relocate all processes. The existing process migraion code
1240 * will work for this?
1241 * - The name of the VM is a somewhat unique ID in the code. It is tricky
1242 * for the destination VM?
1244 * - Another one is ...
1245 * We update the information of the given VM to place it to the destination
1248 * The second one would be easier.
1252 msg_host_t old_pm = simcall_vm_get_pm(vm);
1254 if (!MSG_vm_is_running(vm))
1255 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
1257 if (MSG_vm_is_migrating(vm))
1258 THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_name(vm));
1260 msg_host_priv_t priv = msg_host_resource_priv(vm);
1261 priv->is_migrating = 1;
1265 int ret = do_migration(vm, old_pm, new_pm);
1267 priv->is_migrating = 0;
1268 THROWF(host_error, 0, "SRC host failed during migration");
1271 priv->is_migrating = 0;
1272 THROWF(host_error, 0, "DST host failed during migration");
1276 // This part is done in the RX code, to handle the corner case where SRC can crash just at the end of the migration process
1277 // In that case, the VM has been already assigned to the DST node.
1278 //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
1279 //#ifdef HAVE_TRACING
1280 //TRACE_msg_vm_change_host(vm, old_pm, new_pm);
1285 /** @brief Immediately suspend the execution of all processes within the given VM.
1288 * This function stops the execution of the VM. All the processes on this VM
1289 * will pause. The state of the VM is preserved. We can later resume it again.
1291 * No suspension cost occurs.
1293 void MSG_vm_suspend(msg_vm_t vm)
1295 if (MSG_vm_is_migrating(vm))
1296 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
1298 simcall_vm_suspend(vm);
1300 XBT_DEBUG("vm_suspend done");
1303 TRACE_msg_vm_suspend(vm);
1308 /** @brief Resume the execution of the VM. All processes on the VM run again.
1311 * No resume cost occurs.
1313 void MSG_vm_resume(msg_vm_t vm)
1315 simcall_vm_resume(vm);
1318 TRACE_msg_vm_resume(vm);
1323 /** @brief Immediately save the execution of all processes within the given VM.
1326 * This function stops the execution of the VM. All the processes on this VM
1327 * will pause. The state of the VM is preserved. We can later resume it again.
1329 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
1330 * use a \ref MSG_file_write() before or after, depending on the exact semantic
1331 * of VM save to you.
1333 void MSG_vm_save(msg_vm_t vm)
1335 if (MSG_vm_is_migrating(vm))
1336 THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm));
1338 simcall_vm_save(vm);
1340 TRACE_msg_vm_save(vm);
1344 /** @brief Restore the execution of the VM. All processes on the VM run again.
1347 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
1348 * use a \ref MSG_file_read() before or after, depending on the exact semantic
1349 * of VM restore to you.
1351 void MSG_vm_restore(msg_vm_t vm)
1353 simcall_vm_restore(vm);
1356 TRACE_msg_vm_restore(vm);
1361 /** @brief Get the physical host of a given VM.
1364 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
1366 return simcall_vm_get_pm(vm);
1370 /** @brief Set a CPU bound for a given VM.
1374 * Note that in some cases MSG_task_set_bound() may not intuitively work for VMs.
1377 * On PM0, there are Task1 and VM0.
1378 * On VM0, there is Task2.
1379 * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0.
1381 * Task1\@PM0 gets 50%.
1382 * Task2\@VM0 gets 25%.
1383 * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively.
1385 * This is because a VM has the dummy CPU action in the PM layer. Putting a
1386 * task on the VM does not affect the bound of the dummy CPU action. The bound
1387 * of the dummy CPU action is unlimited.
1389 * There are some solutions for this problem. One option is to update the bound
1390 * of the dummy CPU action automatically. It should be the sum of all tasks on
1391 * the VM. But, this solution might be costly, because we have to scan all tasks
1392 * on the VM in share_resource() or we have to trap both the start and end of
1395 * The current solution is to use MSG_vm_set_bound(), which allows us to
1396 * directly set the bound of the dummy CPU action.
1400 * Note that bound == 0 means no bound (i.e., unlimited). But, if a host has
1401 * multiple CPU cores, the CPU share of a computation task (or a VM) never
1402 * exceeds the capacity of a CPU core.
1404 void MSG_vm_set_bound(msg_vm_t vm, double bound)
1406 return simcall_vm_set_bound(vm, bound);
1410 /** @brief Set the CPU affinity of a given VM.
1413 * This function changes the CPU affinity of a given VM. Usage is the same as
1414 * MSG_task_set_affinity(). See the MSG_task_set_affinity() for details.
1416 void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask)
1418 msg_host_priv_t priv = msg_host_resource_priv(vm);
1421 xbt_dict_remove_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm));
1423 xbt_dict_set_ext(priv->affinity_mask_db, (char *) pm, sizeof(pm), (void *) mask, NULL);
1425 msg_host_t pm_now = MSG_vm_get_pm(vm);
1427 XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));
1428 simcall_vm_set_affinity(vm, pm, mask);
1430 XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm));