X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/eb5b6b0c3c4d9a38bb205b2c8bc9aeeba8674a25..8b858c848c9b85707b853622d3bf88d9392199f0:/src/msg/msg_vm.c diff --git a/src/msg/msg_vm.c b/src/msg/msg_vm.c index 7972de267b..4d83a7d308 100644 --- a/src/msg/msg_vm.c +++ b/src/msg/msg_vm.c @@ -4,19 +4,13 @@ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ -// QUESTIONS: -// 1./ check how and where a new VM is added to the list of the hosts -// 2./ Diff between SIMIX_Actions and SURF_Actions -// => SIMIX_actions : point synchro entre processus de niveau (theoretically speaking I do not have to create such SIMIX_ACTION -// => Surf_Actions - -// TODO -// MSG_TRACE can be revisited in order to use the host -// To implement a mixed model between workstation and vm_workstation, -// please give a look at surf_model_private_t model_private at SURF Level and to the share resource functions -// double (*share_resources) (double now); -// For the action into the vm workstation model, we should be able to leverage the usual one (and if needed, look at -// the workstation model. +/* TODO: + * 1. add the support of trace + * 2. use parallel tasks to simulate CPU overhead and remove the very + * experimental code generating micro computation tasks + */ + + #include "msg_private.h" #include "xbt/sysdep.h" @@ -81,7 +75,7 @@ void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_ msg_vm_t MSG_vm_get_by_name(const char *name) { - return MSG_get_host_by_name(name); + return MSG_get_host_by_name(name); } /** \ingroup m_vm_management @@ -103,7 +97,7 @@ static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state) return simcall_vm_get_state(vm) == state; } -/** @brief Returns whether the given VM has just reated, not running. +/** @brief Returns whether the given VM has just created, not running. * @ingroup msg_VMs */ int MSG_vm_is_created(msg_vm_t vm) @@ -124,7 +118,8 @@ int MSG_vm_is_running(msg_vm_t vm) */ int MSG_vm_is_migrating(msg_vm_t vm) { - return __MSG_vm_is_state(vm, SURF_VM_STATE_MIGRATING); + msg_host_priv_t priv = msg_host_resource_priv(vm); + return priv->is_migrating; } /** @brief Returns whether the given VM is currently suspended, not running. @@ -180,7 +175,7 @@ msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name, * bandwidth */ double host_speed = MSG_get_host_speed(ind_pm); double update_speed = ((double)dp_intensity/100) * mig_netspeed; - + msg_vm_t vm = MSG_vm_create_core(ind_pm, name); s_ws_params_t params; memset(¶ms, 0, sizeof(params)); @@ -209,8 +204,8 @@ msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name) { /* make sure the VM of the same name does not exit */ { - void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name); - if (ind_host_tmp) { + xbt_dictelm_t ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name); + if (ind_host_tmp && xbt_lib_get_level(ind_host_tmp, SIMIX_HOST_LEVEL) != NULL) { XBT_ERROR("host %s already exits", name); return NULL; } @@ -238,6 +233,9 @@ msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name) */ void MSG_vm_destroy(msg_vm_t vm) { + if (MSG_vm_is_migrating(vm)) + THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm)); + /* First, terminate all processes on the VM if necessary */ if (MSG_vm_is_running(vm)) simcall_vm_shutdown(vm); @@ -294,67 +292,93 @@ void MSG_vm_shutdown(msg_vm_t vm) /* We have two mailboxes. mbox is used to transfer migration data between - * source and destiantion PMs. mbox_ctl is used to detect the completion of a + * source and destination PMs. mbox_ctl is used to detect the completion of a * migration. The names of these mailboxes must not conflict with others. */ -static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name) +static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) { + char *vm_name = sg_host_name(vm); + char *src_pm_name = sg_host_name(src_pm); + char *dst_pm_name = sg_host_name(dst_pm); + return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name); } -static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name) +static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) { + char *vm_name = sg_host_name(vm); + char *src_pm_name = sg_host_name(src_pm); + char *dst_pm_name = sg_host_name(dst_pm); + return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name); } -static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name) +static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) { + char *vm_name = sg_host_name(vm); + char *src_pm_name = sg_host_name(src_pm); + char *dst_pm_name = sg_host_name(dst_pm); + return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name); } -static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name) +static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) { + char *vm_name = sg_host_name(vm); + char *src_pm_name = sg_host_name(src_pm); + char *dst_pm_name = sg_host_name(dst_pm); + return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name); } -static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage) +static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage) { + char *vm_name = sg_host_name(vm); + char *src_pm_name = sg_host_name(src_pm); + char *dst_pm_name = sg_host_name(dst_pm); + return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name); } -static void launch_deferred_exec_process(msg_host_t host, double computation, double prio); + +struct migration_session { + msg_vm_t vm; + msg_host_t src_pm; + msg_host_t dst_pm; + + /* The miration_rx process uses mbox_ctl to let the caller of do_migration() + * know the completion of the migration. */ + char *mbox_ctl; + /* The migration_rx and migration_tx processes use mbox to transfer migration + * data. */ + char *mbox; +}; + static int migration_rx_fun(int argc, char *argv[]) { XBT_DEBUG("mig: rx_start"); - xbt_assert(argc == 4); - const char *vm_name = argv[1]; - const char *src_pm_name = argv[2]; - const char *dst_pm_name = argv[3]; - msg_vm_t vm = MSG_get_host_by_name(vm_name); - msg_host_t src_pm = MSG_get_host_by_name(src_pm_name); - msg_host_t dst_pm = MSG_get_host_by_name(dst_pm_name); - + // The structure has been created in the do_migration function and should only be freed in the same place ;) + struct migration_session *ms = MSG_process_get_data(MSG_process_self()); s_ws_params_t params; - simcall_host_get_params(vm, ¶ms); - const double xfer_cpu_overhead = params.xfer_cpu_overhead; - + simcall_host_get_params(ms->vm, ¶ms); int need_exit = 0; - char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name); - char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name); - char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3); + char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3); + int ret = 0; for (;;) { msg_task_t task = NULL; - MSG_task_recv(&task, mbox); + ret = MSG_task_recv(&task, ms->mbox); { - double received = MSG_task_get_data_size(task); - /* TODO: clean up */ - // const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024); - launch_deferred_exec_process(vm, received * xfer_cpu_overhead, 1); + if (ret != MSG_OK) { + // An error occured, clean the code and return + // The owner did not change, hence the task should be only destroyed on the other side + xbt_free(finalize_task_name); + return 0; + } } if (strcmp(task->name, finalize_task_name) == 0) @@ -366,35 +390,65 @@ static int migration_rx_fun(int argc, char *argv[]) break; } + // Here Stage 1, 2 and 3 have been performed. + // Hence complete the migration + + // Copy the reference to the vm (if SRC crashes now, do_migration will free ms) + // 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) + msg_vm_t vm = ms->vm; + msg_host_t src_pm = ms->src_pm; + msg_host_t dst_pm = ms-> dst_pm; + msg_host_priv_t priv = msg_host_resource_priv(vm); - /* deinstall the current affinity setting */ +// 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 +// I should check with Takahiro in order to make this portion of code atomic + /* deinstall the current affinity setting for the CPU */ simcall_vm_set_affinity(vm, src_pm, 0); + /* Update the vm location */ simcall_vm_migrate(vm, dst_pm); + + /* Resume the VM */ simcall_vm_resume(vm); /* install the affinity setting of the VM on the destination pm */ { - msg_host_priv_t priv = msg_host_resource_priv(vm); - unsigned long affinity_mask = (unsigned long) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *) dst_pm, sizeof(msg_host_t)); + unsigned long affinity_mask = (unsigned long) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *)dst_pm, sizeof(msg_host_t)); simcall_vm_set_affinity(vm, dst_pm, affinity_mask); - XBT_INFO("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm)); + XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm)); } { - char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4); + // Now the VM is running on the new host (the migration is completed) (even if the SRC crash) + msg_host_priv_t priv = msg_host_resource_priv(vm); + priv->is_migrating = 0; + XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->key, ms->src_pm->key, ms->dst_pm->key); + #ifdef HAVE_TRACING + TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm); + #endif + + } + // Inform the SRC that the migration has been correctly performed + { + char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4); msg_task_t task = MSG_task_create(task_name, 0, 0, NULL); - msg_error_t ret = MSG_task_send(task, mbox_ctl); - xbt_assert(ret == MSG_OK); + msg_error_t ret = MSG_task_send(task, ms->mbox_ctl); + // xbt_assert(ret == MSG_OK); + if(ret == MSG_HOST_FAILURE){ + // 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 + // TODO What does it mean ? What should we do ? + MSG_task_destroy(task); + } else if(ret == MSG_TRANSFER_FAILURE){ + // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node + MSG_task_destroy(task); + } xbt_free(task_name); } - xbt_free(mbox); - xbt_free(mbox_ctl); xbt_free(finalize_task_name); XBT_DEBUG("mig: rx_done"); @@ -456,7 +510,7 @@ static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remai double computed = dp->prev_remaining - remaining; double duration = clock - dp->prev_clock; - XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)", + XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock); return computed; @@ -472,8 +526,8 @@ static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug dirty_page_t dp = NULL; xbt_dict_foreach(priv->dp_objs, cursor, key, dp) { double remaining = MSG_task_get_remaining_computation(dp->task); - - double clock = MSG_get_clock(); + + double clock = MSG_get_clock(); // total += calc_updated_pages(key, vm, dp, remaining, clock); total += get_computed(key, vm, dp, remaining, clock); @@ -531,7 +585,7 @@ void MSG_host_del_task(msg_host_t host, msg_task_t task) xbt_assert(dp->task == task); /* If we are in the middle of dirty page tracking, we record how much - * computaion has been done until now, and keep the information for the + * computation has been done until now, and keep the information for the * lookup_() function that will called soon. */ if (priv->dp_enabled) { double remaining = MSG_task_get_remaining_computation(task); @@ -551,312 +605,60 @@ void MSG_host_del_task(msg_host_t host, msg_task_t task) } -static int deferred_exec_fun(int argc, char *argv[]) -{ - xbt_assert(argc == 3); - const char *comp_str = argv[1]; - double computaion = atof(comp_str); - const char *prio_str = argv[2]; - double prio = atof(prio_str); - - msg_task_t task = MSG_task_create("__task_deferred", computaion, 0, NULL); - // XBT_INFO("exec deferred %f", computaion); - - /* dpt is the results of the VM activity */ - MSG_task_set_priority(task, prio); - MSG_task_execute(task); - - - - MSG_task_destroy(task); - - return 0; -} - -static void launch_deferred_exec_process(msg_host_t host, double computation, double prio) -{ - char *pr_name = bprintf("__pr_deferred_exec_%s", MSG_host_get_name(host)); - - int nargvs = 4; - char **argv = xbt_new(char *, nargvs); - argv[0] = pr_name; - argv[1] = bprintf("%f", computation); - argv[2] = bprintf("%f", prio); - argv[3] = NULL; - MSG_process_create_with_arguments(pr_name, deferred_exec_fun, NULL, host, nargvs - 1, argv); -} - -static int task_tx_overhead_fun(int argc, char *argv[]) +static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, + sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double timeout) { - xbt_assert(argc == 2); - const char *mbox = argv[1]; - - int need_exit = 0; - - // XBT_INFO("start %s", mbox); - - for (;;) { - msg_task_t task = NULL; - MSG_task_recv(&task, mbox); - - // XBT_INFO("task->name %s", task->name); - - if (strcmp(task->name, "finalize_making_overhead") == 0) - need_exit = 1; - - // XBT_INFO("exec"); - // MSG_task_set_priority(task, 1000000); - MSG_task_execute(task); - MSG_task_destroy(task); - - if (need_exit) - break; - } - - // XBT_INFO("bye"); - - return 0; -} - -static void start_overhead_process(msg_task_t comm_task) -{ - char *pr_name = bprintf("__pr_task_tx_overhead_%s", MSG_task_get_name(comm_task)); - char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task)); - - int nargvs = 3; - char **argv = xbt_new(char *, nargvs); - argv[0] = pr_name; - argv[1] = mbox; - argv[2] = NULL; - - // XBT_INFO("micro start: mbox %s", mbox); - MSG_process_create_with_arguments(pr_name, task_tx_overhead_fun, NULL, MSG_host_self(), nargvs - 1, argv); -} - -static void shutdown_overhead_process(msg_task_t comm_task) -{ - char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task)); - - msg_task_t task = MSG_task_create("finalize_making_overhead", 0, 0, NULL); - - // XBT_INFO("micro shutdown: mbox %s", mbox); - msg_error_t ret = MSG_task_send(task, mbox); - xbt_assert(ret == MSG_OK); - - xbt_free(mbox); - // XBT_INFO("shutdown done"); -} - -static void request_overhead(msg_task_t comm_task, double computation) -{ - char *mbox = bprintf("__mb_task_tx_overhead_%s", MSG_task_get_name(comm_task)); - - msg_task_t task = MSG_task_create("micro", computation, 0, NULL); - - // XBT_INFO("req overhead"); - msg_error_t ret = MSG_task_send(task, mbox); - xbt_assert(ret == MSG_OK); - - xbt_free(mbox); -} - -/* alpha is (floating_operations / bytes). - * - * When actual migration traffic was 32 mbytes/s, we observed the CPU - * utilization of the main thread of the Qemu process was 10 %. - * alpha = 0.1 * C / (32 * 1024 * 1024) - * where the CPU capacity of the PM is C flops/s. - * - * */ -static void task_send_bounded_with_cpu_overhead(msg_task_t comm_task, char *mbox, double mig_speed, double alpha) -{ - const double chunk_size = 1024 * 1024 * 10; - double remaining = MSG_task_get_data_size(comm_task); - - start_overhead_process(comm_task); - - - while (remaining > 0) { - double data_size = chunk_size; - if (remaining < chunk_size) - data_size = remaining; - - remaining -= data_size; - - // XBT_INFO("remaining %f bytes", remaining); - - - double clock_sta = MSG_get_clock(); - - /* create a micro task */ - { - char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task)); - msg_task_t mtask = MSG_task_create(mtask_name, 0, data_size, NULL); - - request_overhead(comm_task, data_size * alpha); - - msg_error_t ret = MSG_task_send(mtask, mbox); - xbt_assert(ret == MSG_OK); - - xbt_free(mtask_name); - } - -#if 0 - { - /* In the real world, sending data involves small CPU computation. */ - char *mtask_name = bprintf("__micro_%s", MSG_task_get_name(comm_task)); - msg_task_t mtask = MSG_task_create(mtask_name, data_size * alpha, data_size, NULL); - MSG_task_execute(mtask); - MSG_task_destroy(mtask); - xbt_free(mtask_name); - } -#endif - - /* TODO */ - - double clock_end = MSG_get_clock(); - - - if (mig_speed > 0) { - /* - * (max bandwidth) > data_size / ((elapsed time) + time_to_sleep) - * - * Thus, we get - * time_to_sleep > data_size / (max bandwidth) - (elapsed time) - * - * If time_to_sleep is smaller than zero, the elapsed time was too big. We - * do not need a micro sleep. - **/ - double time_to_sleep = data_size / mig_speed - (clock_end - clock_sta); - if (time_to_sleep > 0) - MSG_process_sleep(time_to_sleep); - - - //XBT_INFO("duration %f", clock_end - clock_sta); - //XBT_INFO("time_to_sleep %f", time_to_sleep); - } - } - - // XBT_INFO("%s", MSG_task_get_name(comm_task)); - shutdown_overhead_process(comm_task); - -} - - -#if 0 -static void make_cpu_overhead_of_data_transfer(msg_task_t comm_task, double init_comm_size) -{ - double prev_remaining = init_comm_size; - - for (;;) { - double remaining = MSG_task_get_remaining_communication(comm_task); - if (remaining == 0) - need_exit = 1; - - double sent = prev_remaining - remaining; - double comp_size = sent * overhead; - - - char *comp_task_name = bprintf("__sender_overhead%s", MSG_task_get_name(comm_task)); - msg_task_t comp_task = MSG_task_create(comp_task_name, comp_size, 0, NULL); - MSG_task_execute(comp_task); - MSG_task_destroy(comp_task); - - if (need_exit) - break; - - prev_remaining = remaining; - - } - - xbt_free(comp_task_name); -} -#endif - -// #define USE_MICRO_TASK 1 - -#if 0 -// const double alpha = 0.1L * 1.0E8 / (32L * 1024 * 1024); -// const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024); -// const double alpha = 0.20L * 1.0E8 / (85L * 1024 * 1024); -// const double alpha = 0.25L * 1.0E8 / (85L * 1024 * 1024); -// const double alpha = 0.32L * 1.0E8 / (24L * 1024 * 1024); // makes super good values for 32 mbytes/s -//const double alpha = 0.32L * 1.0E8 / (32L * 1024 * 1024); -// const double alpha = 0.56L * 1.0E8 / (80L * 1024 * 1024); -////const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024); -// const double alpha = 0.56L * 1.0E8 / (90L * 1024 * 1024); -// const double alpha = 0.66L * 1.0E8 / (90L * 1024 * 1024); -// const double alpha = 0.20L * 1.0E8 / (80L * 1024 * 1024); - -/* CPU 22% when 80Mbyte/s */ -const double alpha = 0.22L * 1.0E8 / (80L * 1024 * 1024); -#endif - - -static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, - sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double xfer_cpu_overhead) -{ - char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage); + sg_size_t sent = 0; + char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage); msg_task_t task = MSG_task_create(task_name, 0, size, NULL); /* TODO: clean up */ double clock_sta = MSG_get_clock(); -#ifdef USE_MICRO_TASK - - task_send_bounded_with_cpu_overhead(task, mbox, mig_speed, xfer_cpu_overhead); - -#else msg_error_t ret; if (mig_speed > 0) - ret = MSG_task_send_bounded(task, mbox, mig_speed); + ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed); else ret = MSG_task_send(task, mbox); - xbt_assert(ret == MSG_OK); -#endif - - double clock_end = MSG_get_clock(); - double duration = clock_end - clock_sta; - double actual_speed = size / duration; -#ifdef USE_MICRO_TASK - double cpu_utilization = size * xfer_cpu_overhead / duration / 1.0E8; -#else - double cpu_utilization = 0; -#endif - - + xbt_free(task_name); - if (stage == 2){ - 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);} - else{ - 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); + if (ret == MSG_OK) { + sent = size; + } else if (ret == MSG_TIMEOUT) { + sg_size_t remaining = MSG_task_get_remaining_communication(task); + sent = size - remaining; + XBT_INFO("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", + timeout, remaining, size); } - xbt_free(task_name); + /* FIXME: why try-and-catch is used here? */ + if(ret == MSG_HOST_FAILURE){ + //XBT_INFO("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage); + MSG_task_destroy(task); + THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage); + }else if(ret == MSG_TRANSFER_FAILURE){ + //XBT_INFO("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage); + MSG_task_destroy(task); + THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage); + } + double clock_end = MSG_get_clock(); + double duration = clock_end - clock_sta; + double actual_speed = size / duration; + if (stage == 2) + XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration, actual_speed, mig_speed); + else + XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed, mig_speed); -#ifdef USE_MICRO_TASK - /* The name of a micro task starts with __micro, which does not match the - * special name that finalizes the receiver loop. Thus, we send the special task. - **/ - { - if (stage == 3) { - char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage); - msg_task_t task = MSG_task_create(task_name, 0, 0, NULL); - msg_error_t ret = MSG_task_send(task, mbox); - xbt_assert(ret == MSG_OK); - xbt_free(task_name); - } - } -#endif + return sent; } -static double get_updated_size(double computed, double dp_rate, double dp_cap) +static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap) { double updated_size = computed * dp_rate; XBT_DEBUG("updated_size %f dp_rate %f", updated_size, dp_rate); @@ -865,14 +667,12 @@ static double get_updated_size(double computed, double dp_rate, double dp_cap) updated_size = dp_cap; } - return updated_size; + return (sg_size_t) updated_size; } -static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *dst_pm_name, - sg_size_t ramsize, double mig_speed, double xfer_cpu_overhead, double dp_rate, double dp_cap, double dpt_cpu_overhead) +static double send_stage1(struct migration_session *ms, + sg_size_t ramsize, double mig_speed, double dp_rate, double dp_cap) { - const char *vm_name = MSG_host_get_name(vm); - char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name); // const long chunksize = (sg_size_t)1024 * 1024 * 100; const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000; @@ -885,19 +685,11 @@ static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *ds datasize = remaining; remaining -= datasize; - - send_migration_data(vm_name, src_pm_name, dst_pm_name, datasize, mbox, 1, 0, mig_speed, xfer_cpu_overhead); - double computed = lookup_computed_flop_counts(vm, 1, 0); + send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, datasize, ms->mbox, 1, 0, mig_speed, -1); + double computed = lookup_computed_flop_counts(ms->vm, 1, 0); computed_total += computed; - - // { - // double updated_size = get_updated_size(computed, dp_rate, dp_cap); - - // double overhead = dpt_cpu_overhead * updated_size; - // launch_deferred_exec_process(vm, overhead, 10000); - // } } - xbt_free(mbox); + return computed_total; } @@ -905,73 +697,85 @@ static double send_stage1(msg_host_t vm, const char *src_pm_name, const char *ds static double get_threshold_value(double bandwidth, double max_downtime) { - /* This value assumes the network link is 1Gbps. */ - // double threshold = max_downtime * 125 * 1024 * 1024; - double threshold = max_downtime * bandwidth; - - return threshold; + return max_downtime * bandwidth; } static int migration_tx_fun(int argc, char *argv[]) { XBT_DEBUG("mig: tx_start"); - xbt_assert(argc == 4); - const char *vm_name = argv[1]; - const char *src_pm_name = argv[2]; - const char *dst_pm_name = argv[3]; - msg_vm_t vm = MSG_get_host_by_name(vm_name); - + // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;) + struct migration_session *ms = MSG_process_get_data(MSG_process_self()); s_ws_params_t params; - simcall_host_get_params(vm, ¶ms); + simcall_host_get_params(ms->vm, ¶ms); const sg_size_t ramsize = params.ramsize; const sg_size_t devsize = params.devsize; const int skip_stage1 = params.skip_stage1; - const int skip_stage2 = params.skip_stage2; + int skip_stage2 = params.skip_stage2; const double dp_rate = params.dp_rate; const double dp_cap = params.dp_cap; const double mig_speed = params.mig_speed; - const double xfer_cpu_overhead = params.xfer_cpu_overhead; - const double dpt_cpu_overhead = params.dpt_cpu_overhead; + double max_downtime = params.max_downtime; + + /* hard code it temporally. Fix Me */ +#define MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME 10000000.0 + double mig_timeout = MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME; double remaining_size = ramsize + devsize; + double threshold = 0.0; + + /* check parameters */ + if (ramsize == 0) + XBT_WARN("migrate a VM, but ramsize is zero"); - double max_downtime = params.max_downtime; if (max_downtime == 0) { XBT_WARN("use the default max_downtime value 30ms"); max_downtime = 0.03; } - double threshold = 0.00001; /* TODO: cleanup */ - - /* setting up parameters has done */ - - - if (ramsize == 0) - XBT_WARN("migrate a VM, but ramsize is zero"); - - char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name); - - XBT_INFO("mig-stage1: remaining_size %f", remaining_size); - /* Stage1: send all memory pages to the destination. */ - start_dirty_page_tracking(vm); + XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size); + start_dirty_page_tracking(ms->vm); double computed_during_stage1 = 0; if (!skip_stage1) { - // send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed, xfer_cpu_overhead); - - /* send ramsize, but split it */ double clock_prev_send = MSG_get_clock(); - 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); - remaining_size -= ramsize; + TRY { + /* At stage 1, we do not need timeout. We have to send all the memory + * pages even though the duration of this tranfer exceeds the timeout + * value. */ + XBT_INFO("Stage 1: Gonna send %llu", ramsize); + sg_size_t sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, ramsize, ms->mbox, 1, 0, mig_speed, -1); + remaining_size -= sent; + computed_during_stage1 = lookup_computed_flop_counts(ms->vm, 1, 0); + + if (sent < ramsize) { + XBT_INFO("mig-stage1: timeout, force moving to stage 3"); + skip_stage2 = 1; + } else if (sent > ramsize) + XBT_CRITICAL("bug"); + + } CATCH_ANONYMOUS { + //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code) + // Stop the dirty page tracking an return (there is no memory space to release) + stop_dirty_page_tracking(ms->vm); + return 0; + } double clock_post_send = MSG_get_clock(); + mig_timeout -= (clock_post_send - clock_prev_send); + if (mig_timeout < 0) { + XBT_INFO("The duration of stage 1 exceeds the timeout value (%lf > %lf), skip stage 2", + (clock_post_send - clock_prev_send), MIGRATION_TIMEOUT_DO_NOT_HARDCODE_ME); + skip_stage2 = 1; + } + + /* estimate bandwidth */ double bandwidth = ramsize / (clock_post_send - clock_prev_send); threshold = get_threshold_value(bandwidth, max_downtime); - XBT_INFO("actual banwdidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold); + XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold); } @@ -979,130 +783,166 @@ static int migration_tx_fun(int argc, char *argv[]) * becomes smaller than the threshold value. */ if (skip_stage2) goto stage3; - if (max_downtime == 0) { - XBT_WARN("no max_downtime parameter, skip stage2"); - goto stage3; - } int stage2_round = 0; for (;;) { - double updated_size = 0; - if (stage2_round == 0) { - /* just after stage1, nothing has been updated. But, we have to send the data updated during stage1 */ + sg_size_t updated_size = 0; + if (stage2_round == 0) { + /* just after stage1, nothing has been updated. But, we have to send the + * data updated during stage1 */ updated_size = get_updated_size(computed_during_stage1, dp_rate, dp_cap); } else { - double computed = lookup_computed_flop_counts(vm, 2, stage2_round); + double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round); updated_size = get_updated_size(computed, dp_rate, dp_cap); } - XBT_INFO("mig-stage 2:%d updated_size %f computed_during_stage1 %f dp_rate %f dp_cap %f", + XBT_DEBUG("mig-stage 2:%d updated_size %llu computed_during_stage1 %f dp_rate %f dp_cap %f", stage2_round, updated_size, computed_during_stage1, dp_rate, dp_cap); - // if (stage2_round != 0) { - // /* during stage1, we have already created overhead tasks */ - // double overhead = dpt_cpu_overhead * updated_size; - // XBT_DEBUG("updated %f overhead %f", updated_size, overhead); - // launch_deferred_exec_process(vm, overhead, 10000); - // } - - - { - remaining_size += updated_size; - - XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, - remaining_size, (remaining_size < threshold) ? "<" : ">", threshold); + /* Check whether the remaining size is below the threshold value. If so, + * move to stage 3. */ + remaining_size += updated_size; + XBT_DEBUG("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round, + remaining_size, (remaining_size < threshold) ? "<" : ">", threshold); + if (remaining_size < threshold) + break; - if (remaining_size < threshold) - break; - } + sg_size_t sent = 0; double clock_prev_send = MSG_get_clock(); - - send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed, xfer_cpu_overhead); - + TRY { + XBT_INFO("Stage 2, gonna send %llu", updated_size); + sent = send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, updated_size, ms->mbox, 2, stage2_round, mig_speed, mig_timeout); + } CATCH_ANONYMOUS { + //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code) + // Stop the dirty page tracking an return (there is no memory space to release) + stop_dirty_page_tracking(ms->vm); + return 0; + } double clock_post_send = MSG_get_clock(); - double bandwidth = updated_size / (clock_post_send - clock_prev_send); - threshold = get_threshold_value(bandwidth, max_downtime); - XBT_INFO("actual banwdidth %f, threshold %f", bandwidth / 1024 / 1024, threshold); - - - - - - + if (sent == updated_size) { + /* timeout did not happen */ + double bandwidth = updated_size / (clock_post_send - clock_prev_send); + threshold = get_threshold_value(bandwidth, max_downtime); + XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold); + remaining_size -= sent; + stage2_round += 1; + mig_timeout -= (clock_post_send - clock_prev_send); + xbt_assert(mig_timeout > 0); + + } else if (sent < updated_size) { + /* When timeout happens, we move to stage 3. The size of memory pages + * updated before timeout must be added to the remaining size. */ + XBT_INFO("mig-stage2.%d: timeout, force moving to stage 3. sent %llu / %llu, eta %lf", + stage2_round, sent, updated_size, (clock_post_send - clock_prev_send)); + remaining_size -= sent; + + double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round); + updated_size = get_updated_size(computed, dp_rate, dp_cap); + remaining_size += updated_size; + break; - remaining_size -= updated_size; - stage2_round += 1; + } else + XBT_CRITICAL("bug"); } stage3: /* Stage3: stop the VM and copy the rest of states. */ - XBT_INFO("mig-stage3: remaining_size %f", remaining_size); - simcall_vm_suspend(vm); - stop_dirty_page_tracking(vm); - - send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed, xfer_cpu_overhead); - - xbt_free(mbox); + XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size); + simcall_vm_suspend(ms->vm); + stop_dirty_page_tracking(ms->vm); + + TRY { + XBT_INFO("Stage 3: Gonna send %f", remaining_size); + send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, remaining_size, ms->mbox, 3, 0, mig_speed, -1); + } CATCH_ANONYMOUS { + //hostfailure (if you want to know whether this is the SRC or the DST please check directly in send_migration_data code) + // Stop the dirty page tracking an return (there is no memory space to release) + simcall_vm_resume(ms->vm); + return 0; + } - XBT_DEBUG("mig: tx_done"); + // 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. + XBT_INFO("mig: tx_done"); return 0; } -static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) +static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) { - char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm)); - - { - char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm)); - int nargvs = 5; - char **argv = xbt_new(char *, nargvs); - argv[0] = pr_name; - argv[1] = xbt_strdup(sg_host_name(vm)); - argv[2] = xbt_strdup(sg_host_name(src_pm)); - argv[3] = xbt_strdup(sg_host_name(dst_pm)); - argv[4] = NULL; - - MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv); - } - - { - char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm)); - int nargvs = 5; - char **argv = xbt_new(char *, nargvs); - argv[0] = pr_name; - argv[1] = xbt_strdup(sg_host_name(vm)); - argv[2] = xbt_strdup(sg_host_name(src_pm)); - argv[3] = xbt_strdup(sg_host_name(dst_pm)); - argv[4] = NULL; - MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv); - } + struct migration_session *ms = xbt_new(struct migration_session, 1); + ms->vm = vm; + ms->src_pm = src_pm; + ms->dst_pm = dst_pm; + ms->mbox_ctl = get_mig_mbox_ctl(vm, src_pm, dst_pm); + ms->mbox = get_mig_mbox_src_dst(vm, src_pm, dst_pm); + + + char *pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm); + char *pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm); + +// msg_process_t tx_process, rx_process; +// MSG_process_create(pr_rx_name, migration_rx_fun, ms, dst_pm); +// MSG_process_create(pr_tx_name, migration_tx_fun, ms, src_pm); +#if 1 + { + char **argv = xbt_new(char *, 2); + argv[0] = pr_rx_name; + argv[1] = NULL; +/*rx_process = */ MSG_process_create_with_arguments(pr_rx_name, migration_rx_fun, ms, dst_pm, 1, argv); + } + { + char **argv = xbt_new(char *, 2); + argv[0] = pr_tx_name; + argv[1] = NULL; +/* tx_process = */MSG_process_create_with_arguments(pr_tx_name, migration_tx_fun, ms, src_pm, 1, argv); + } +#endif - /* wait until the migration have finished */ + /* wait until the migration have finished or on error has occured */ { + XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed"); msg_task_t task = NULL; - msg_error_t ret = MSG_task_recv(&task, mbox_ctl); + msg_error_t ret = MSG_TIMEOUT; + while (ret == MSG_TIMEOUT && MSG_host_is_on(dst_pm)) //Wait while you receive the message o + ret = MSG_task_receive_with_timeout(&task, ms->mbox_ctl, 4); - xbt_assert(ret == MSG_OK); + xbt_free(ms->mbox_ctl); + xbt_free(ms->mbox); + xbt_free(ms); + + //xbt_assert(ret == MSG_OK); + if(ret == MSG_HOST_FAILURE){ + // Note that since the communication failed, the owner did not change and the task should be destroyed on the other side. + // Hence, just throw the execption + XBT_INFO("SRC crashes, throw an exception (m-control)"); + //MSG_process_kill(tx_process); // Adrien, I made a merge on Nov 28th 2014, I'm not sure whether this line is required or not + return -1; + } + else if((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)){ // MSG_TIMEOUT here means that MSG_host_is_avail() returned false. + XBT_INFO("DST crashes, throw an exception (m-control)"); + return -2; + } - char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4); + + char *expected_task_name = get_mig_task_name(vm, src_pm, dst_pm, 4); xbt_assert(strcmp(task->name, expected_task_name) == 0); xbt_free(expected_task_name); MSG_task_destroy(task); + return 0; } - - xbt_free(mbox_ctl); } + + /** @brief Migrate the VM to the given host. * @ingroup msg_VMs * @@ -1130,36 +970,55 @@ void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm) * physical host. * * The second one would be easier. - * + * */ msg_host_t old_pm = simcall_vm_get_pm(vm); - if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING) + if (!MSG_vm_is_running(vm)) THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm)); - do_migration(vm, old_pm, new_pm); - + if (MSG_vm_is_migrating(vm)) + THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_name(vm)); + msg_host_priv_t priv = msg_host_resource_priv(vm); + priv->is_migrating = 1; - XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key); + { + + int ret = do_migration(vm, old_pm, new_pm); + if (ret == -1){ + priv->is_migrating = 0; + THROWF(host_error, 0, "SRC host failed during migration"); + } + else if(ret == -2){ + priv->is_migrating = 0; + THROWF(host_error, 0, "DST host failed during migration"); + } + } - #ifdef HAVE_TRACING - TRACE_msg_vm_change_host(vm, old_pm, new_pm); - #endif + // 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 + // In that case, the VM has been already assigned to the DST node. + //XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key); + //#ifdef HAVE_TRACING + //TRACE_msg_vm_change_host(vm, old_pm, new_pm); + //#endif } /** @brief Immediately suspend the execution of all processes within the given VM. * @ingroup msg_VMs * - * This function stops the exection of the VM. All the processes on this VM - * will pause. The state of the VM is perserved. We can later resume it again. + * This function stops the execution of the VM. All the processes on this VM + * will pause. The state of the VM is preserved. We can later resume it again. * * No suspension cost occurs. */ void MSG_vm_suspend(msg_vm_t vm) { + if (MSG_vm_is_migrating(vm)) + THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm)); + simcall_vm_suspend(vm); XBT_DEBUG("vm_suspend done"); @@ -1188,8 +1047,8 @@ void MSG_vm_resume(msg_vm_t vm) /** @brief Immediately save the execution of all processes within the given VM. * @ingroup msg_VMs * - * This function stops the exection of the VM. All the processes on this VM - * will pause. The state of the VM is perserved. We can later resume it again. + * This function stops the execution of the VM. All the processes on this VM + * will pause. The state of the VM is preserved. We can later resume it again. * * FIXME: No suspension cost occurs. If you want to simulate this too, you want to * use a \ref MSG_file_write() before or after, depending on the exact semantic @@ -1197,6 +1056,9 @@ void MSG_vm_resume(msg_vm_t vm) */ void MSG_vm_save(msg_vm_t vm) { + if (MSG_vm_is_migrating(vm)) + THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_name(vm)); + simcall_vm_save(vm); #ifdef HAVE_TRACING TRACE_msg_vm_save(vm); @@ -1239,7 +1101,7 @@ msg_host_t MSG_vm_get_pm(msg_vm_t vm) * On PM0, there are Task1 and VM0. * On VM0, there is Task2. * Now we bound 75% to Task1\@PM0 and bound 25% to Task2\@VM0. - * Then, + * Then, * Task1\@PM0 gets 50%. * Task2\@VM0 gets 25%. * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively. @@ -1250,7 +1112,7 @@ msg_host_t MSG_vm_get_pm(msg_vm_t vm) * * There are some solutions for this problem. One option is to update the bound * of the dummy CPU action automatically. It should be the sum of all tasks on - * the VM. But, this solution might be costy, because we have to scan all tasks + * the VM. But, this solution might be costly, because we have to scan all tasks * on the VM in share_resource() or we have to trap both the start and end of * task execution. * @@ -1265,7 +1127,7 @@ msg_host_t MSG_vm_get_pm(msg_vm_t vm) */ void MSG_vm_set_bound(msg_vm_t vm, double bound) { - return simcall_vm_set_bound(vm, bound); + return simcall_vm_set_bound(vm, bound); } @@ -1286,8 +1148,8 @@ void MSG_vm_set_affinity(msg_vm_t vm, msg_host_t pm, unsigned long mask) msg_host_t pm_now = MSG_vm_get_pm(vm); if (pm_now == pm) { - XBT_INFO("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm)); + XBT_DEBUG("set affinity(0x%04lx@%s) for %s", mask, MSG_host_get_name(pm), MSG_host_get_name(vm)); simcall_vm_set_affinity(vm, pm, mask); } else - XBT_INFO("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm)); + XBT_DEBUG("set affinity(0x%04lx@%s) for %s (not active now)", mask, MSG_host_get_name(pm), MSG_host_get_name(vm)); }