X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/397be753d4e5a7371b973245bb04ae657d0d7db0..f7623851a023484d8ba1d5c26134ee7850bac134:/src/msg/msg_vm.cpp diff --git a/src/msg/msg_vm.cpp b/src/msg/msg_vm.cpp index 998e8b4bc2..8d59a21358 100644 --- a/src/msg/msg_vm.cpp +++ b/src/msg/msg_vm.cpp @@ -11,20 +11,25 @@ #include +#include "src/instr/instr_private.h" +#include "src/msg/msg_private.h" #include "src/plugins/vm/VirtualMachineImpl.hpp" #include "src/plugins/vm/VmHostExt.hpp" -#include -#include -#include "msg_private.h" -#include "xbt/sysdep.h" -#include "xbt/log.h" #include "simgrid/host.h" +#include "simgrid/simix.hpp" -#include "src/simix/smx_host_private.h" /* don't ask me why the VM functions are in there (FIXME:KILLME) */ +SG_BEGIN_DECL() -XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API"); +struct dirty_page { + double prev_clock; + double prev_remaining; + msg_task_t task; +}; +typedef struct dirty_page s_dirty_page; +typedef struct dirty_page* dirty_page_t; +XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API"); /* **** ******** GENERAL ********* **** */ @@ -42,7 +47,7 @@ void MSG_vm_set_params(msg_vm_t vm, vm_params_t params) /** \ingroup m_vm_management * \brief Get the parameters of a given host * - * \param host a host + * \param vm the vm you are interested into * \param params a prameter object */ void MSG_vm_get_params(msg_vm_t vm, vm_params_t params) @@ -53,7 +58,8 @@ void MSG_vm_get_params(msg_vm_t vm, vm_params_t params) /* **** Check state of a VM **** */ static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state) { - return static_cast(vm)->pimpl_vm_->getState() == state; + simgrid::s4u::VirtualMachine* castedVm = static_cast(vm); + return castedVm->pimpl_vm_ != nullptr && castedVm->pimpl_vm_->getState() == state; } /** @brief Returns whether the given VM has just created, not running. @@ -88,72 +94,43 @@ int MSG_vm_is_suspended(msg_vm_t vm) return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED); } -/** @brief Returns whether the given VM is being saved (FIXME: live saving or not?). - * @ingroup msg_VMs - */ -int MSG_vm_is_saving(msg_vm_t vm) -{ - return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING); -} - -/** @brief Returns whether the given VM has been saved, not running. - * @ingroup msg_VMs - */ -int MSG_vm_is_saved(msg_vm_t vm) -{ - return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED); -} - -/** @brief Returns whether the given VM is being restored, not running. - * @ingroup msg_VMs - */ -int MSG_vm_is_restoring(msg_vm_t vm) -{ - return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING); -} - /* **** ******** MSG vm actions ********* **** */ /** @brief Create a new VM with specified parameters. * @ingroup msg_VMs* * @param pm Physical machine that will host the VM - * @param name [TODO] - * @param ncpus [TODO] + * @param name Must be unique + * @param coreAmount Must be >= 1 * @param ramsize [TODO] - * @param net_cap Maximal bandwidth that the VM can consume (in MByte/s) - * @param disk_path (unused) Path to the image that boots - * @param disksize (unused) will represent the size of the VM (will be used during migrations) * @param mig_netspeed Amount of Mbyte/s allocated to the migration (cannot be larger than net_cap). Use 0 if unsure. * @param dp_intensity Dirty page percentage according to migNetSpeed, [0-100]. Use 0 if unsure. */ -msg_vm_t MSG_vm_create(msg_host_t pm, const char* name, int ncpus, int ramsize, int net_cap, char* disk_path, - int disksize, int mig_netspeed, int dp_intensity) +msg_vm_t MSG_vm_create(msg_host_t pm, const char* name, int coreAmount, int ramsize, int mig_netspeed, int dp_intensity) { simgrid::vm::VmHostExt::ensureVmExtInstalled(); /* For the moment, intensity_rate is the percentage against the migration bandwidth */ double host_speed = MSG_host_get_speed(pm); - double update_speed = ((double)dp_intensity/100) * mig_netspeed; + double update_speed = (static_cast(dp_intensity)/100) * mig_netspeed; - msg_vm_t vm = MSG_vm_create_core(pm, name); + msg_vm_t vm = MSG_vm_create_multicore(pm, name, coreAmount); s_vm_params_t params; memset(¶ms, 0, sizeof(params)); - params.ramsize = (sg_size_t)ramsize * 1024 * 1024; - //params.overcommit = 0; + params.ramsize = static_cast(ramsize) * 1024 * 1024; params.devsize = 0; params.skip_stage2 = 0; params.max_downtime = 0.03; params.dp_rate = (update_speed * 1024 * 1024) / host_speed; params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize - params.mig_speed = (double)mig_netspeed * 1024 * 1024; // mig_speed + params.mig_speed = static_cast(mig_netspeed) * 1024 * 1024; // mig_speed - //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); + XBT_DEBUG("dp rate %f migspeed : %f intensity mem : %d, updatespeed %f, hostspeed %f", params.dp_rate, + params.mig_speed, dp_intensity, update_speed, host_speed); static_cast(vm)->setParameters(¶ms); return vm; } -/** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start(). +/** @brief Create a new VM object with the default parameters * @ingroup msg_VMs* * * A VM is treated as a host. The name of the VM must be unique among all hosts. @@ -163,7 +140,19 @@ msg_vm_t MSG_vm_create_core(msg_host_t pm, const char* name) xbt_assert(sg_host_by_name(name) == nullptr, "Cannot create a VM named %s: this name is already used by an host or a VM", name); - return new simgrid::s4u::VirtualMachine(name, pm); + return new simgrid::s4u::VirtualMachine(name, pm, 1); +} +/** @brief Create a new VM object with the default parameters, but with a specified amount of cores + * @ingroup msg_VMs* + * + * A VM is treated as a host. The name of the VM must be unique among all hosts. + */ +msg_vm_t MSG_vm_create_multicore(msg_host_t pm, const char* name, int coreAmount) +{ + xbt_assert(sg_host_by_name(name) == nullptr, + "Cannot create a VM named %s: this name is already used by an host or a VM", name); + + return new simgrid::s4u::VirtualMachine(name, pm, coreAmount); } /** @brief Destroy a VM. Destroy the VM object from the simulation. @@ -172,21 +161,19 @@ msg_vm_t MSG_vm_create_core(msg_host_t 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", vm->name().c_str()); + THROWF(vm_error, 0, "Cannot destroy VM '%s', which is migrating.", vm->cname()); /* First, terminate all processes on the VM if necessary */ if (MSG_vm_is_running(vm)) MSG_vm_shutdown(vm); - xbt_assert(MSG_vm_is_created(vm), "shutdown the given VM before destroying it"); - /* Then, destroy the VM object */ simgrid::simix::kernelImmediate([vm]() { vm->destroy(); }); if (TRACE_msg_vm_is_enabled()) { - container_t container = PJ_container_get(vm->name().c_str()); + container_t container = PJ_container_get(vm->cname()); PJ_container_remove_from_parent(container); PJ_container_free(container); } @@ -195,43 +182,64 @@ void MSG_vm_destroy(msg_vm_t vm) /** @brief Start a vm (i.e., boot the guest operating system) * @ingroup msg_VMs * - * If the VM cannot be started, an exception is generated. + * If the VM cannot be started (because of memory over-provisioning), an exception is generated. */ void MSG_vm_start(msg_vm_t vm) { - simcall_vm_start(vm); + simgrid::simix::kernelImmediate([vm]() { + simgrid::vm::VmHostExt::ensureVmExtInstalled(); + + simgrid::s4u::VirtualMachine* typedVM = static_cast(vm); + simgrid::s4u::Host* pm = typedVM->pimpl_vm_->getPm(); + if (pm->extension() == nullptr) + pm->extension_set(new simgrid::vm::VmHostExt()); + + long pm_ramsize = pm->extension()->ramsize; + int pm_overcommit = pm->extension()->overcommit; + long vm_ramsize = typedVM->getRamsize(); + + if (pm_ramsize && not pm_overcommit) { /* Only verify that we don't overcommit on need */ + /* Retrieve the memory occupied by the VMs on that host. Yep, we have to traverse all VMs of all hosts for that */ + long total_ramsize_of_vms = 0; + for (simgrid::s4u::VirtualMachine* ws_vm : simgrid::vm::VirtualMachineImpl::allVms_) + if (pm == ws_vm->pimpl_vm_->getPm()) + total_ramsize_of_vms += ws_vm->pimpl_vm_->getRamsize(); + + if (vm_ramsize > pm_ramsize - total_ramsize_of_vms) { + XBT_WARN("cannnot start %s@%s due to memory shortage: vm_ramsize %ld, free %ld, pm_ramsize %ld (bytes).", + vm->cname(), pm->cname(), vm_ramsize, pm_ramsize - total_ramsize_of_vms, pm_ramsize); + THROWF(vm_error, 0, "Memory shortage on host '%s', VM '%s' cannot be started", pm->cname(), vm->cname()); + } + } + + typedVM->pimpl_vm_->setState(SURF_VM_STATE_RUNNING); + }); if (TRACE_msg_vm_is_enabled()) { - container_t vm_container = PJ_container_get(vm->name().c_str()); + container_t vm_container = PJ_container_get(vm->cname()); type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type); val_t value = PJ_value_get_or_new("start", "0 0 1", type); // start is blue - new_pajePushState(MSG_get_clock(), vm_container, type, value); + new PushStateEvent(MSG_get_clock(), vm_container, type, value); } } -/** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked. +/** @brief Immediately kills all processes within the given VM. * @ingroup msg_VMs * - * FIXME: No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep() or something. - * I'm not quite sure. + * Any memory that they allocated will be leaked, unless you used #MSG_process_on_exit(). + * + * No extra delay occurs. If you want to simulate this too, you want to use a #MSG_process_sleep(). */ void MSG_vm_shutdown(msg_vm_t vm) { - simcall_vm_shutdown(vm); - MSG_process_sleep(0.); // Make sure that the processes in the VM are killed in this scheduling round before processing - // (eg with the VM destroy) -} - -/* We have two mailboxes. mbox is used to transfer migration data between 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(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) -{ - return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname()); -} + smx_actor_t issuer=SIMIX_process_self(); + simgrid::simix::kernelImmediate([vm,issuer]() { + static_cast(vm)->pimpl_vm_->shutdown(issuer); + }); -static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) -{ - return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname()); + // Make sure that the processes in the VM are killed in this scheduling round before processing + // (eg with the VM destroy) + MSG_process_sleep(0.); } static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) @@ -271,7 +279,7 @@ static int migration_rx_fun(int argc, char *argv[]) bool received_finalize = false; char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3); - while (!received_finalize) { + while (not received_finalize) { msg_task_t task = nullptr; int ret = MSG_task_recv(&task, ms->mbox); @@ -296,69 +304,64 @@ static int migration_rx_fun(int argc, char *argv[]) // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms // structure at the beginning and free it at the end of each function) simgrid::s4u::VirtualMachine* vm = static_cast(ms->vm); - msg_host_t src_pm = ms->src_pm; msg_host_t dst_pm = ms->dst_pm; // Make sure that we cannot get interrupted between the migrate and the resume to not end in an inconsistent state - simgrid::simix::kernelImmediate([vm, src_pm, dst_pm]() { + simgrid::simix::kernelImmediate([vm, dst_pm]() { /* Update the vm location */ /* precopy migration makes the VM temporally paused */ - xbt_assert(static_cast(vm)->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED); - - /* jump to vm_ws_xigrate(). this will update the vm location. */ - static_cast(vm)->pimpl_vm_->migrate(dst_pm); + xbt_assert(vm->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED); - /* Resume the VM */ - SIMIX_vm_resume(vm); + /* Update the vm location and resume it */ + vm->pimpl_vm_->setPm(dst_pm); + vm->pimpl_vm_->resume(); }); - { - // Now the VM is running on the new host (the migration is completed) (even if the SRC crash) - vm->pimpl_vm_->isMigrating = false; - XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", sg_host_get_name(ms->vm), sg_host_get_name(ms->src_pm), - sg_host_get_name(ms->dst_pm)); - - if (TRACE_msg_vm_is_enabled()) { - static long long int counter = 0; - char key[INSTR_DEFAULT_STR_SIZE]; - snprintf(key, INSTR_DEFAULT_STR_SIZE, "%lld", counter++); - - // start link - container_t msg = PJ_container_get(vm->name().c_str()); - type_t type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root()); - new_pajeStartLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key); - - // destroy existing container of this vm - container_t existing_container = PJ_container_get(vm->name().c_str()); - PJ_container_remove_from_parent(existing_container); - PJ_container_free(existing_container); - - // create new container on the new_host location - PJ_container_new(vm->name().c_str(), INSTR_MSG_VM, PJ_container_get(sg_host_get_name(ms->dst_pm))); - - // end link - msg = PJ_container_get(vm->name().c_str()); - type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root()); - new_pajeEndLink(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key); - } + + // Now the VM is running on the new host (the migration is completed) (even if the SRC crash) + vm->pimpl_vm_->isMigrating = false; + XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->cname(), ms->src_pm->cname(), ms->dst_pm->cname()); + + if (TRACE_msg_vm_is_enabled()) { + static long long int counter = 0; + char key[INSTR_DEFAULT_STR_SIZE]; + snprintf(key, INSTR_DEFAULT_STR_SIZE, "%lld", counter); + counter++; + + // start link + container_t msg = PJ_container_get(vm->cname()); + type_t type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root()); + new StartLinkEvent(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key); + + // destroy existing container of this vm + container_t existing_container = PJ_container_get(vm->cname()); + PJ_container_remove_from_parent(existing_container); + PJ_container_free(existing_container); + + // create new container on the new_host location + PJ_container_new(vm->cname(), INSTR_MSG_VM, PJ_container_get(ms->dst_pm->cname())); + + // end link + msg = PJ_container_get(vm->cname()); + type = PJ_type_get("MSG_VM_LINK", PJ_type_get_root()); + new EndLinkEvent(MSG_get_clock(), PJ_container_get_root(), type, msg, "M", key); } + // 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, nullptr); - 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); + 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, nullptr); + 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_DEBUG("mig: rx_done"); return 0; @@ -366,10 +369,10 @@ static int migration_rx_fun(int argc, char *argv[]) static void start_dirty_page_tracking(msg_vm_t vm) { - simgrid::surf::VirtualMachineImpl* pimpl = static_cast(vm)->pimpl_vm_; + simgrid::vm::VirtualMachineImpl* pimpl = static_cast(vm)->pimpl_vm_; pimpl->dp_enabled = 1; - if (!pimpl->dp_objs) + if (not pimpl->dp_objs) return; char *key = nullptr; @@ -394,15 +397,15 @@ 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: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", - key, sg_host_get_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock); + XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)", key, vm->cname(), computed, + dp->prev_remaining, remaining, duration, dp->prev_clock, clock); return computed; } static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug) { - simgrid::surf::VirtualMachineImpl* pimpl = static_cast(vm)->pimpl_vm_; + simgrid::vm::VirtualMachineImpl* pimpl = static_cast(vm)->pimpl_vm_; double total = 0; char *key = nullptr; @@ -438,7 +441,7 @@ void MSG_host_add_task(msg_host_t host, msg_task_t task) simgrid::s4u::VirtualMachine* vm = dynamic_cast(host); if (vm == nullptr) return; - simgrid::surf::VirtualMachineImpl* pimpl = static_cast(vm)->pimpl_vm_; + simgrid::vm::VirtualMachineImpl* pimpl = static_cast(vm)->pimpl_vm_; double remaining = MSG_task_get_flops_amount(task); char *key = bprintf("%s-%p", task->name, task); @@ -449,11 +452,11 @@ void MSG_host_add_task(msg_host_t host, msg_task_t task) dp->prev_clock = MSG_get_clock(); dp->prev_remaining = remaining; } - if (!pimpl->dp_objs) - pimpl->dp_objs = xbt_dict_new(); + if (not pimpl->dp_objs) + pimpl->dp_objs = xbt_dict_new_homogeneous(nullptr); xbt_assert(xbt_dict_get_or_null(pimpl->dp_objs, key) == nullptr); xbt_dict_set(pimpl->dp_objs, key, dp, nullptr); - XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_get_name(host), remaining, pimpl->dp_enabled); + XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, host->cname(), remaining, pimpl->dp_enabled); xbt_free(key); } @@ -463,7 +466,7 @@ void MSG_host_del_task(msg_host_t host, msg_task_t task) simgrid::s4u::VirtualMachine* vm = dynamic_cast(host); if (vm == nullptr) return; - simgrid::surf::VirtualMachineImpl* pimpl = static_cast(vm)->pimpl_vm_; + simgrid::vm::VirtualMachineImpl* pimpl = static_cast(vm)->pimpl_vm_; char *key = bprintf("%s-%p", task->name, task); dirty_page_t dp = (dirty_page_t)(pimpl->dp_objs ? xbt_dict_get_or_null(pimpl->dp_objs, key) : NULL); @@ -483,7 +486,7 @@ void MSG_host_del_task(msg_host_t host, msg_task_t task) xbt_dict_remove(pimpl->dp_objs, key); xbt_free(dp); - XBT_DEBUG("del %s on %s", key, sg_host_get_name(host)); + XBT_DEBUG("del %s on %s", key, host->cname()); xbt_free(key); } @@ -492,7 +495,7 @@ static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t { 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, (double)size, nullptr); + msg_task_t task = MSG_task_create(task_name, 0, static_cast(size), nullptr); /* TODO: clean up */ @@ -509,20 +512,20 @@ static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t if (ret == MSG_OK) { sent = size; } else if (ret == MSG_TIMEOUT) { - sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task); + sg_size_t remaining = static_cast(MSG_task_get_remaining_communication(task)); sent = size - remaining; XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size); } /* FIXME: why try-and-catch is used here? */ if(ret == MSG_HOST_FAILURE){ - //XBT_DEBUG("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage); + XBT_DEBUG("SRC host failed during migration of %s (stage %d)", vm->cname(), stage); MSG_task_destroy(task); - THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage); + THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", vm->cname(), stage); }else if(ret == MSG_TRANSFER_FAILURE){ - //XBT_DEBUG("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage); + XBT_DEBUG("DST host failed during migration of %s (stage %d)", vm->cname(), stage); MSG_task_destroy(task); - THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_get_name(vm), stage); + THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", vm->cname(), stage); } double clock_end = MSG_get_clock(); @@ -549,34 +552,7 @@ static sg_size_t get_updated_size(double computed, double dp_rate, double dp_cap updated_size = dp_cap; } - return (sg_size_t) updated_size; -} - -static double send_stage1(struct migration_session* ms, sg_size_t ramsize, double mig_speed, double dp_rate, - double dp_cap) -{ - // const long chunksize = (sg_size_t)1024 * 1024 * 100; - const sg_size_t chunksize = (sg_size_t)1024 * 1024 * 100000; - sg_size_t remaining = ramsize; - double computed_total = 0; - - while (remaining > 0) { - sg_size_t datasize = chunksize; - if (remaining < chunksize) - datasize = remaining; - - remaining -= datasize; - 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; - } - - return computed_total; -} - -static double get_threshold_value(double bandwidth, double max_downtime) -{ - return max_downtime * bandwidth; + return static_cast(updated_size); } static int migration_tx_fun(int argc, char *argv[]) @@ -584,7 +560,7 @@ static int migration_tx_fun(int argc, char *argv[]) XBT_DEBUG("mig: tx_start"); // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;) - migration_session *ms = (migration_session *) MSG_process_get_data(MSG_process_self()); + migration_session *ms = static_cast(MSG_process_get_data(MSG_process_self())); s_vm_params_t params; static_cast(ms->vm)->parameters(¶ms); @@ -597,11 +573,9 @@ static int migration_tx_fun(int argc, char *argv[]) const double mig_speed = params.mig_speed; 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 mig_timeout = 10000000.0; - double remaining_size = (double) (ramsize + devsize); + double remaining_size = static_cast(ramsize + devsize); double threshold = 0.0; /* check parameters */ @@ -618,7 +592,7 @@ static int migration_tx_fun(int argc, char *argv[]) start_dirty_page_tracking(ms->vm); double computed_during_stage1 = 0; - if (!skip_stage1) { + if (not skip_stage1) { double clock_prev_send = MSG_get_clock(); try { @@ -646,20 +620,19 @@ static int migration_tx_fun(int argc, char *argv[]) double clock_post_send = MSG_get_clock(); mig_timeout -= (clock_post_send - clock_prev_send); if (mig_timeout < 0) { - XBT_VERB("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); + XBT_VERB("The duration of stage 1 exceeds the timeout value, skip stage 2"); skip_stage2 = 1; } /* estimate bandwidth */ double bandwidth = ramsize / (clock_post_send - clock_prev_send); - threshold = get_threshold_value(bandwidth, max_downtime); + threshold = bandwidth * max_downtime; XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold); } /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */ - if (!skip_stage2) { + if (not skip_stage2) { int stage2_round = 0; for (;;) { @@ -701,7 +674,7 @@ static int migration_tx_fun(int argc, char *argv[]) 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); + threshold = bandwidth * max_downtime; XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold); remaining_size -= sent; stage2_round += 1; @@ -726,17 +699,21 @@ static int migration_tx_fun(int argc, char *argv[]) /* Stage3: stop the VM and copy the rest of states. */ XBT_DEBUG("mig-stage3: remaining_size %f", remaining_size); - simcall_vm_suspend(ms->vm); + simgrid::vm::VirtualMachineImpl* pimpl = static_cast(ms->vm)->pimpl_vm_; + pimpl->setState(SURF_VM_STATE_RUNNING); // FIXME: this bypass of the checks in suspend() is not nice + pimpl->isMigrating = false; // FIXME: this bypass of the checks in suspend() is not nice + pimpl->suspend(SIMIX_process_self()); stop_dirty_page_tracking(ms->vm); try { XBT_DEBUG("Stage 3: Gonna send %f", remaining_size); - send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, (sg_size_t)remaining_size, ms->mbox, 3, 0, mig_speed, -1); + send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, static_cast(remaining_size), ms->mbox, 3, 0, + mig_speed, -1); } catch(xbt_ex& e) { //hostfailure (if you want to know whether this is the SRC or the DST 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); + static_cast(ms->vm)->pimpl_vm_->resume(); return 0; } @@ -747,14 +724,51 @@ static int migration_tx_fun(int argc, char *argv[]) return 0; } -static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) +/** @brief Migrate the VM to the given host. + * @ingroup msg_VMs + */ +void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm) { + /* some thoughts: + * - One approach is ... + * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will + * receive the state of the source + * VM over network. We will finally destroy the source VM. + * - This behavior is similar to the way of migration in the real world. + * Even before a migration is completed, we will see a destination VM, consuming resources. + * - We have to relocate all processes. The existing process migration code will work for this? + * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM? + * + * - Another one is ... + * We update the information of the given VM to place it to the destination physical host. + * + * The second one would be easier. + */ + + simgrid::s4u::VirtualMachine* typedVm = static_cast(vm); + simgrid::vm::VirtualMachineImpl* pimpl = typedVm->pimpl_vm_; + msg_host_t src_pm = pimpl->getPm(); + + if (src_pm->isOff()) + THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->cname(), src_pm->cname()); + if (dst_pm->isOff()) + THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->cname(), dst_pm->cname()); + if (not MSG_vm_is_running(vm)) + THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->cname()); + if (typedVm->isMigrating()) + THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->cname()); + + pimpl->isMigrating = true; + 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); + + /* We have two mailboxes. mbox is used to transfer migration data between 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. */ + ms->mbox_ctl = bprintf("__mbox_mig_ctl:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname()); + ms->mbox = bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm->cname(), src_pm->cname(), dst_pm->cname()); 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); @@ -772,84 +786,32 @@ static int do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) /* wait until the migration have finished or on error has occurred */ XBT_DEBUG("wait for reception of the final ACK (i.e. migration has been correctly performed"); msg_task_t task = nullptr; - 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); + msg_error_t ret = MSG_task_receive(&task, ms->mbox_ctl); + + pimpl->isMigrating = false; 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_ERROR("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; + THROWF(host_error, 0, "Source host '%s' failed during the migration of VM '%s'.", src_pm->cname(), vm->cname()); } else if ((ret == MSG_TRANSFER_FAILURE) || (ret == MSG_TIMEOUT)) { // MSG_TIMEOUT here means that MSG_host_is_avail() returned false. XBT_ERROR("DST crashes, throw an exception (m-control)"); - return -2; + THROWF(host_error, 0, "Destination host '%s' failed during the migration of VM '%s'.", dst_pm->cname(), + vm->cname()); } 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; -} - -/** @brief Migrate the VM to the given host. - * @ingroup msg_VMs - * - * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a MSG_task_send() before or after, - * depending on whether you want to do cold or hot migration. - */ -void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm) -{ - /* some thoughts: - * - One approach is ... - * We first create a new VM (i.e., destination VM) on the destination physical host. The destination VM will - * receive the state of the source - * VM over network. We will finally destroy the source VM. - * - This behavior is similar to the way of migration in the real world. - * Even before a migration is completed, we will see a destination VM, consuming resources. - * - We have to relocate all processes. The existing process migration code will work for this? - * - The name of the VM is a somewhat unique ID in the code. It is tricky for the destination VM? - * - * - Another one is ... - * We update the information of the given VM to place it to the destination physical host. - * - * The second one would be easier. - */ - - simgrid::surf::VirtualMachineImpl* pimpl = static_cast(vm)->pimpl_vm_; - msg_host_t old_pm = pimpl->getPm(); - - if (old_pm->isOff()) - THROWF(vm_error, 0, "Cannot start a migration from host '%s', which is offline.", sg_host_get_name(old_pm)); - - if (new_pm->isOff()) - THROWF(vm_error, 0, "Cannot start a migration to host '%s', which is offline.", sg_host_get_name(new_pm)); - - if (!MSG_vm_is_running(vm)) - THROWF(vm_error, 0, "VM(%s) is not running", sg_host_get_name(vm)); - - if (MSG_vm_is_migrating(vm)) - THROWF(vm_error, 0, "VM(%s) is already migrating", sg_host_get_name(vm)); - - pimpl->isMigrating = 1; - - int ret = do_migration(vm, old_pm, new_pm); - if (ret == -1) { - pimpl->isMigrating = 0; - THROWF(host_error, 0, "SRC host failed during migration"); - } else if (ret == -2) { - pimpl->isMigrating = 0; - THROWF(host_error, 0, "DST host failed during migration"); - } } /** @brief Immediately suspend the execution of all processes within the given VM. @@ -862,18 +824,18 @@ void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm) */ void MSG_vm_suspend(msg_vm_t vm) { - if (MSG_vm_is_migrating(vm)) - THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm)); - - simcall_vm_suspend(vm); + smx_actor_t issuer = SIMIX_process_self(); + simgrid::simix::kernelImmediate([vm,issuer]() { + static_cast(vm)->pimpl_vm_->suspend(issuer); + }); XBT_DEBUG("vm_suspend done"); if (TRACE_msg_vm_is_enabled()) { - container_t vm_container = PJ_container_get(vm->name().c_str()); + container_t vm_container = PJ_container_get(vm->cname()); type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type); val_t value = PJ_value_get_or_new("suspend", "1 0 0", type); // suspend is red - new_pajePushState(MSG_get_clock(), vm_container, type, value); + new PushStateEvent(MSG_get_clock(), vm_container, type, value); } } @@ -884,54 +846,12 @@ void MSG_vm_suspend(msg_vm_t vm) */ void MSG_vm_resume(msg_vm_t vm) { - simcall_vm_resume(vm); + static_cast(vm)->pimpl_vm_->resume(); if (TRACE_msg_vm_is_enabled()) { - container_t vm_container = PJ_container_get(vm->name().c_str()); + container_t vm_container = PJ_container_get(vm->cname()); type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type); - new_pajePopState(MSG_get_clock(), vm_container, type); - } -} - - -/** @brief Immediately save the execution of all processes within the given VM. - * @ingroup msg_VMs - * - * 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 of VM save to you. - */ -void MSG_vm_save(msg_vm_t vm) -{ - if (MSG_vm_is_migrating(vm)) - THROWF(vm_error, 0, "VM(%s) is migrating", sg_host_get_name(vm)); - - simcall_vm_save(vm); - - if (TRACE_msg_vm_is_enabled()) { - container_t vm_container = PJ_container_get(vm->name().c_str()); - type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type); - val_t value = PJ_value_get_or_new("save", "0 1 0", type); // save is green - new_pajePushState(MSG_get_clock(), vm_container, type, value); - } -} - -/** @brief Restore the execution of the VM. All processes on the VM run again. - * @ingroup msg_VMs - * - * FIXME: No restore cost occurs. If you want to simulate this too, you want to use a \ref MSG_file_read() before or - * after, depending on the exact semantic of VM restore to you. - */ -void MSG_vm_restore(msg_vm_t vm) -{ - simcall_vm_restore(vm); - - if (TRACE_msg_vm_is_enabled()) { - container_t vm_container = PJ_container_get(vm->name().c_str()); - type_t type = PJ_type_get("MSG_VM_STATE", vm_container->type); - new_pajePopState(MSG_get_clock(), vm_container, type); + new PopStateEvent(MSG_get_clock(), vm_container, type); } } @@ -940,7 +860,7 @@ void MSG_vm_restore(msg_vm_t vm) */ msg_host_t MSG_vm_get_pm(msg_vm_t vm) { - return static_cast(vm)->pimpl_vm_->getPm(); + return static_cast(vm)->pm(); } /** @brief Set a CPU bound for a given VM. @@ -971,5 +891,8 @@ msg_host_t MSG_vm_get_pm(msg_vm_t vm) */ void MSG_vm_set_bound(msg_vm_t vm, double bound) { - simcall_vm_set_bound(vm, bound); + simgrid::simix::kernelImmediate( + [vm, bound]() { static_cast(vm)->pimpl_vm_->setBound(bound); }); } + +SG_END_DECL()