X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/ecd5f7562caf1d443bf22788fa5f4fac408776ec..b9d2f7277242a82702b73e4ccdf63aec98855b5e:/src/msg/msg_vm.cpp diff --git a/src/msg/msg_vm.cpp b/src/msg/msg_vm.cpp index 00d1d02f14..34ab03d4a7 100644 --- a/src/msg/msg_vm.cpp +++ b/src/msg/msg_vm.cpp @@ -31,7 +31,6 @@ typedef s_dirty_page* dirty_page_t; XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API"); -/* **** ******** GENERAL ********* **** */ const char* MSG_vm_get_name(msg_vm_t vm) { return vm->getCname(); @@ -59,10 +58,19 @@ void MSG_vm_get_params(msg_vm_t vm, vm_params_t params) vm->getParameters(params); } +void MSG_vm_set_ramsize(msg_vm_t vm, size_t size) +{ + vm->setRamsize(size); +} +size_t MSG_vm_get_ramsize(msg_vm_t vm) +{ + return vm->getRamsize(); +} + /* **** Check state of a VM **** */ static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state) { - return vm->pimpl_vm_ != nullptr && vm->pimpl_vm_->getState() == state; + return vm->pimpl_vm_ != nullptr && vm->getState() == state; } /** @brief Returns whether the given VM has just created, not running. @@ -113,18 +121,12 @@ msg_vm_t MSG_vm_create(msg_host_t pm, const char* name, int coreAmount, int rams /* For the moment, intensity_rate is the percentage against the migration bandwidth */ - msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, coreAmount); + msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, coreAmount, static_cast(ramsize) * 1024 * 1024); s_vm_params_t params; - params.ncpus = 0; - params.ramsize = static_cast(ramsize) * 1024 * 1024; - params.overcommit = 0; - params.devsize = 0; - params.skip_stage1 = 0; - params.skip_stage2 = 0; params.max_downtime = 0.03; params.mig_speed = static_cast(mig_netspeed) * 1024 * 1024; // mig_speed params.dp_intensity = static_cast(dp_intensity) / 100; - params.dp_cap = params.ramsize * 0.9; // assume working set memory is 90% of ramsize + params.dp_cap = vm->getRamsize() * 0.9; // assume working set memory is 90% of ramsize XBT_DEBUG("migspeed : %f intensity mem : %d", params.mig_speed, dp_intensity); vm->setParameters(¶ms); @@ -142,7 +144,11 @@ 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, 1); + msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, 1); + s_vm_params_t params; + memset(¶ms, 0, sizeof(params)); + vm->setParameters(¶ms); + return vm; } /** @brief Create a new VM object with the default parameters, but with a specified amount of cores * @ingroup msg_VMs* @@ -154,7 +160,11 @@ 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); + msg_vm_t vm = new simgrid::s4u::VirtualMachine(name, pm, coreAmount); + s_vm_params_t params; + memset(¶ms, 0, sizeof(params)); + vm->setParameters(¶ms); + return vm; } /** @brief Destroy a VM. Destroy the VM object from the simulation. @@ -166,11 +176,10 @@ void MSG_vm_destroy(msg_vm_t vm) THROWF(vm_error, 0, "Cannot destroy VM '%s', which is migrating.", vm->getCname()); /* First, terminate all processes on the VM if necessary */ - if (MSG_vm_is_running(vm)) - MSG_vm_shutdown(vm); + vm->shutdown(); /* Then, destroy the VM object */ - simgrid::simix::kernelImmediate([vm]() { vm->destroy(); }); + vm->destroy(); if (TRACE_msg_vm_is_enabled()) { container_t container = simgrid::instr::Container::byName(vm->getName()); @@ -203,11 +212,7 @@ void MSG_vm_start(msg_vm_t vm) */ void MSG_vm_shutdown(msg_vm_t vm) { - smx_actor_t issuer = SIMIX_process_self(); - simgrid::simix::kernelImmediate([vm, issuer]() { vm->pimpl_vm_->shutdown(issuer); }); - - // Make sure that processes in the VM are killed in this scheduling round before processing (eg with the VM destroy) - MSG_process_sleep(0.); + vm->shutdown(); } static std::string get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm) @@ -277,11 +282,11 @@ static int migration_rx_fun(int argc, char *argv[]) simgrid::simix::kernelImmediate([vm, dst_pm]() { /* Update the vm location */ /* precopy migration makes the VM temporally paused */ - xbt_assert(vm->pimpl_vm_->getState() == SURF_VM_STATE_SUSPENDED); + xbt_assert(vm->getState() == SURF_VM_STATE_SUSPENDED); /* Update the vm location and resume it */ vm->pimpl_vm_->setPm(dst_pm); - vm->pimpl_vm_->resume(); + vm->resume(); }); @@ -331,13 +336,13 @@ static int migration_rx_fun(int argc, char *argv[]) static void start_dirty_page_tracking(msg_vm_t vm) { - vm->pimpl_vm_->dp_enabled = 1; + vm->pimpl_vm_->dp_enabled = true; if (vm->pimpl_vm_->dp_objs.empty()) return; for (auto const& elm : vm->pimpl_vm_->dp_objs) { dirty_page_t dp = elm.second; - double remaining = MSG_task_get_remaining_work_ratio(dp->task); + double remaining = MSG_task_get_flops_amount(dp->task); dp->prev_clock = MSG_get_clock(); dp->prev_remaining = remaining; XBT_DEBUG("%s@%s remaining %f", elm.first.c_str(), vm->getCname(), remaining); @@ -346,7 +351,7 @@ static void start_dirty_page_tracking(msg_vm_t vm) static void stop_dirty_page_tracking(msg_vm_t vm) { - vm->pimpl_vm_->dp_enabled = 0; + vm->pimpl_vm_->dp_enabled = false; } static double get_computed(const std::string& key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock) @@ -367,7 +372,7 @@ static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug for (auto const& elm : vm->pimpl_vm_->dp_objs) { const std::string& key = elm.first; dirty_page_t dp = elm.second; - double remaining = MSG_task_get_remaining_work_ratio(dp->task); + double remaining = MSG_task_get_flops_amount(dp->task); double clock = MSG_get_clock(); @@ -396,7 +401,7 @@ void MSG_host_add_task(msg_host_t host, msg_task_t task) if (vm == nullptr) return; - double remaining = MSG_task_get_initial_flops_amount(task); + double remaining = MSG_task_get_flops_amount(task); std::string key = simgrid::xbt::string_printf("%s-%p", task->name, task); dirty_page_t dp = new s_dirty_page; @@ -426,7 +431,7 @@ void MSG_host_del_task(msg_host_t host, msg_task_t task) /* If we are in the middle of dirty page tracking, we record how much computation has been done until now, and keep * the information for the lookup_() function that will called soon. */ if (vm->pimpl_vm_->dp_enabled) { - double remaining = MSG_task_get_remaining_work_ratio(task); + double remaining = MSG_task_get_flops_amount(task); double clock = MSG_get_clock(); double updated = get_computed(key, vm, dp, remaining, clock); // was host instead of vm @@ -506,22 +511,20 @@ static int migration_tx_fun(int argc, char *argv[]) // Note that the ms structure has been allocated in do_migration and hence should be freed in the same function ;) migration_session* ms = static_cast(MSG_process_get_data(MSG_process_self())); - double host_speed = ms->vm->pimpl_vm_->getPm()->getSpeed(); + double host_speed = ms->vm->getPm()->getSpeed(); s_vm_params_t params; ms->vm->getParameters(¶ms); - const sg_size_t ramsize = params.ramsize; - const sg_size_t devsize = params.devsize; - const int skip_stage1 = params.skip_stage1; - int skip_stage2 = params.skip_stage2; + const sg_size_t ramsize = ms->vm->getRamsize(); const double dp_rate = host_speed ? (params.mig_speed * params.dp_intensity) / host_speed : 1; const double dp_cap = params.dp_cap; const double mig_speed = params.mig_speed; double max_downtime = params.max_downtime; double mig_timeout = 10000000.0; + bool skip_stage2 = false; - double remaining_size = static_cast(ramsize + devsize); - double threshold = 0.0; + size_t remaining_size = ramsize; + size_t threshold = 0.0; /* check parameters */ if (ramsize == 0) @@ -533,48 +536,44 @@ static int migration_tx_fun(int argc, char *argv[]) } /* Stage1: send all memory pages to the destination. */ - XBT_DEBUG("mig-stage1: remaining_size %f", remaining_size); + XBT_DEBUG("mig-stage1: remaining_size %zu", remaining_size); start_dirty_page_tracking(ms->vm); double computed_during_stage1 = 0; - if (not skip_stage1) { - double clock_prev_send = MSG_get_clock(); - - try { - /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this - * transfer exceeds the timeout value. */ - XBT_VERB("Stage 1: Gonna send %llu bytes", 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_VERB("mig-stage1: timeout, force moving to stage 3"); - skip_stage2 = 1; - } else if (sent > ramsize) - XBT_CRITICAL("bug"); + double clock_prev_send = MSG_get_clock(); - } - 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) - 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_VERB("The duration of stage 1 exceeds the timeout value, skip stage 2"); - skip_stage2 = 1; - } + try { + /* At stage 1, we do not need timeout. We have to send all the memory pages even though the duration of this + * transfer exceeds the timeout value. */ + XBT_VERB("Stage 1: Gonna send %llu bytes", 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_VERB("mig-stage1: timeout, force moving to stage 3"); + skip_stage2 = true; + } else if (sent > ramsize) + XBT_CRITICAL("bug"); + + } 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) + stop_dirty_page_tracking(ms->vm); + return 0; + } - /* estimate bandwidth */ - double bandwidth = ramsize / (clock_post_send - clock_prev_send); - threshold = bandwidth * max_downtime; - XBT_DEBUG("actual bandwidth %f (MB/s), threshold %f", bandwidth / 1024 / 1024, threshold); + 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, skip stage 2"); + skip_stage2 = true; } + /* estimate bandwidth */ + double bandwidth = ramsize / (clock_post_send - clock_prev_send); + threshold = bandwidth * max_downtime; + XBT_DEBUG("actual bandwidth %f (MB/s), threshold %zu", bandwidth / 1024 / 1024, threshold); /* Stage2: send update pages iteratively until the size of remaining states becomes smaller than threshold value. */ if (not skip_stage2) { @@ -596,7 +595,7 @@ static int migration_tx_fun(int argc, char *argv[]) /* 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, + XBT_DEBUG("mig-stage2.%d: remaining_size %zu (%s threshold %zu)", stage2_round, remaining_size, (remaining_size < threshold) ? "<" : ">", threshold); if (remaining_size < threshold) break; @@ -620,7 +619,7 @@ static int migration_tx_fun(int argc, char *argv[]) /* timeout did not happen */ double bandwidth = updated_size / (clock_post_send - clock_prev_send); threshold = bandwidth * max_downtime; - XBT_DEBUG("actual bandwidth %f, threshold %f", bandwidth / 1024 / 1024, threshold); + XBT_DEBUG("actual bandwidth %f, threshold %zu", bandwidth / 1024 / 1024, threshold); remaining_size -= sent; stage2_round += 1; mig_timeout -= (clock_post_send - clock_prev_send); @@ -643,7 +642,7 @@ 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); + XBT_DEBUG("mig-stage3: remaining_size %zu", remaining_size); simgrid::vm::VirtualMachineImpl* pimpl = 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 @@ -651,14 +650,13 @@ static int migration_tx_fun(int argc, char *argv[]) stop_dirty_page_tracking(ms->vm); try { - XBT_DEBUG("Stage 3: Gonna send %f bytes", remaining_size); - send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, static_cast(remaining_size), ms->mbox, 3, 0, - mig_speed, -1); + XBT_DEBUG("Stage 3: Gonna send %zu bytes", remaining_size); + send_migration_data(ms->vm, ms->src_pm, ms->dst_pm, 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) - ms->vm->pimpl_vm_->resume(); + ms->vm->resume(); return 0; } @@ -690,7 +688,7 @@ void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm) * The second one would be easier. */ - msg_host_t src_pm = vm->pimpl_vm_->getPm(); + msg_host_t src_pm = vm->getPm(); if (src_pm->isOff()) THROWF(vm_error, 0, "Cannot migrate VM '%s' from host '%s', which is offline.", vm->getCname(), src_pm->getCname()); @@ -758,11 +756,7 @@ void MSG_vm_migrate(msg_vm_t vm, msg_host_t dst_pm) */ void MSG_vm_suspend(msg_vm_t vm) { - smx_actor_t issuer = SIMIX_process_self(); - simgrid::simix::kernelImmediate([vm, issuer]() { vm->pimpl_vm_->suspend(issuer); }); - - XBT_DEBUG("vm_suspend done"); - + vm->suspend(); if (TRACE_msg_vm_is_enabled()) { simgrid::instr::StateType* state = simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE"); state->addEntityValue("suspend", "1 0 0"); // suspend is red @@ -777,8 +771,7 @@ void MSG_vm_suspend(msg_vm_t vm) */ void MSG_vm_resume(msg_vm_t vm) { - vm->pimpl_vm_->resume(); - + vm->resume(); if (TRACE_msg_vm_is_enabled()) simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE")->popEvent(); } @@ -791,34 +784,8 @@ msg_host_t MSG_vm_get_pm(msg_vm_t vm) return vm->getPm(); } -/** @brief Set a CPU bound for a given VM. - * @ingroup msg_VMs - * - * 1. Note that in some cases MSG_task_set_bound() may not intuitively work for VMs. - * - * For example, - * 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, - * Task1\@PM0 gets 50%. - * Task2\@VM0 gets 25%. - * This is NOT 75% for Task1\@PM0 and 25% for Task2\@VM0, respectively. - * - * This is because a VM has the dummy CPU action in the PM layer. Putting a task on the VM does not affect the bound of - * the dummy CPU action. The bound of the dummy CPU action is unlimited. - * - * 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 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. - * - * The current solution is to use MSG_vm_set_bound(), which allows us to directly set the bound of the dummy CPU action. - * - * 2. Note that bound == 0 means no bound (i.e., unlimited). But, if a host has multiple CPU cores, the CPU share of a - * computation task (or a VM) never exceeds the capacity of a CPU core. - */ void MSG_vm_set_bound(msg_vm_t vm, double bound) { - simgrid::simix::kernelImmediate([vm, bound]() { vm->pimpl_vm_->setBound(bound); }); + vm->setBound(bound); } }