#include <xbt/ex.hpp>
+#include "simgrid/plugins/live_migration.h"
#include "src/instr/instr_private.hpp"
-#include "src/msg/msg_private.hpp"
#include "src/plugins/vm/VirtualMachineImpl.hpp"
#include "src/plugins/vm/VmHostExt.hpp"
extern "C" {
-struct s_dirty_page {
- double prev_clock = 0.0;
- double prev_remaining = 0.0;
- msg_task_t task = nullptr;
-};
-typedef s_dirty_page* dirty_page_t;
-
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg, "Cloud-oriented parts of the MSG API");
const char* MSG_vm_get_name(msg_vm_t vm)
return vm->getCname();
}
+/** @brief Get the physical host of a given VM.
+ * @ingroup msg_VMs
+ */
+msg_host_t MSG_vm_get_pm(msg_vm_t vm)
+{
+ return vm->getPm();
+}
+
/** \ingroup m_vm_management
* \brief Set the parameters of a given host
*
}
/* **** Check state of a VM **** */
+void MSG_vm_set_bound(msg_vm_t vm, double bound)
+{
+ vm->setBound(bound);
+}
static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
{
return vm->pimpl_vm_ != nullptr && vm->getState() == state;
return vm;
}
-/** @brief Destroy a VM. Destroy the VM object from the simulation.
+/** @brief Start a vm (i.e., boot the guest operating system)
* @ingroup msg_VMs
+ *
+ * If the VM cannot be started (because of memory over-provisioning), an exception is generated.
*/
-void MSG_vm_destroy(msg_vm_t vm)
+void MSG_vm_start(msg_vm_t vm)
{
- if (vm->isMigrating())
- 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);
-
- /* Then, destroy the VM object */
- simgrid::simix::kernelImmediate([vm]() { vm->destroy(); });
-
+ vm->start();
if (TRACE_msg_vm_is_enabled()) {
- container_t container = simgrid::instr::Container::byName(vm->getName());
- container->removeFromParent();
- delete container;
+ simgrid::instr::StateType* state = simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE");
+ state->addEntityValue("start", "0 0 1"); // start is blue
+ state->pushEvent("start");
}
}
-/** @brief Start a vm (i.e., boot the guest operating system)
+/** @brief Immediately suspend the execution of all processes within the given VM.
* @ingroup msg_VMs
*
- * If the VM cannot be started (because of memory over-provisioning), an exception is generated.
+ * 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_start(msg_vm_t vm)
+void MSG_vm_suspend(msg_vm_t vm)
{
- vm->start();
+ vm->suspend();
if (TRACE_msg_vm_is_enabled()) {
simgrid::instr::StateType* state = simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE");
- state->addEntityValue("start", "0 0 1"); // start is blue
- state->pushEvent("start");
+ state->addEntityValue("suspend", "1 0 0"); // suspend is red
+ state->pushEvent("suspend");
}
}
+/** @brief Resume the execution of the VM. All processes on the VM run again.
+ * @ingroup msg_VMs
+ *
+ * No resume cost occurs.
+ */
+void MSG_vm_resume(msg_vm_t vm)
+{
+ vm->resume();
+ if (TRACE_msg_vm_is_enabled())
+ simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE")->popEvent();
+}
+
/** @brief Immediately kills all processes within the given VM.
* @ingroup msg_VMs
*
*/
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); });
+ vm->shutdown();
+}
+
+/** @brief Destroy a VM. Destroy the VM object from the simulation.
+ * @ingroup msg_VMs
+ */
+void MSG_vm_destroy(msg_vm_t vm)
+{
+ if (vm->isMigrating())
+ THROWF(vm_error, 0, "Cannot destroy VM '%s', which is migrating.", vm->getCname());
+
+ /* First, terminate all processes on the VM if necessary */
+ vm->shutdown();
- // Make sure that processes in the VM are killed in this scheduling round before processing (eg with the VM destroy)
- MSG_process_sleep(0.);
+ /* Then, destroy the VM object */
+ vm->destroy();
+
+ if (TRACE_msg_vm_is_enabled()) {
+ container_t container = simgrid::instr::Container::byName(vm->getName());
+ container->removeFromParent();
+ delete container;
+ }
}
static std::string get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
return 0;
}
-static void start_dirty_page_tracking(msg_vm_t vm)
-{
- 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);
- dp->prev_clock = MSG_get_clock();
- dp->prev_remaining = remaining;
- XBT_DEBUG("%s@%s remaining %f", elm.first.c_str(), vm->getCname(), remaining);
- }
-}
-
-static void stop_dirty_page_tracking(msg_vm_t vm)
-{
- 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)
-{
- 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.c_str(), vm->getCname(), 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)
-{
- double total = 0;
-
- 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 clock = MSG_get_clock();
-
- total += get_computed(key, vm, dp, remaining, clock);
-
- dp->prev_remaining = remaining;
- dp->prev_clock = clock;
- }
-
- total += vm->pimpl_vm_->dp_updated_by_deleted_tasks;
-
- XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)", stage_for_fancy_debug,
- stage2_round_for_fancy_debug, total, vm->pimpl_vm_->dp_updated_by_deleted_tasks);
-
- vm->pimpl_vm_->dp_updated_by_deleted_tasks = 0;
-
- return total;
-}
-
-// TODO Is this code redundant with the information provided by
-// msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
-/** @brief take care of the dirty page tracking, in case we're adding a task to a migrating VM */
-void MSG_host_add_task(msg_host_t host, msg_task_t task)
-{
- simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
- if (vm == nullptr)
- return;
-
- double remaining = MSG_task_get_initial_flops_amount(task);
- std::string key = simgrid::xbt::string_printf("%s-%p", task->name, task);
-
- dirty_page_t dp = new s_dirty_page;
- dp->task = task;
- if (vm->pimpl_vm_->dp_enabled) {
- dp->prev_clock = MSG_get_clock();
- dp->prev_remaining = remaining;
- }
- vm->pimpl_vm_->dp_objs.insert({key, dp});
- XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key.c_str(), host->getCname(), remaining,
- vm->pimpl_vm_->dp_enabled);
-}
-
-void MSG_host_del_task(msg_host_t host, msg_task_t task)
-{
- simgrid::s4u::VirtualMachine* vm = dynamic_cast<simgrid::s4u::VirtualMachine*>(host);
- if (vm == nullptr)
- return;
-
- std::string key = simgrid::xbt::string_printf("%s-%p", task->name, task);
- dirty_page_t dp = nullptr;
- auto dp_obj = vm->pimpl_vm_->dp_objs.find(key);
- if (dp_obj != vm->pimpl_vm_->dp_objs.end())
- dp = dp_obj->second;
- xbt_assert(dp && dp->task == 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 clock = MSG_get_clock();
- double updated = get_computed(key, vm, dp, remaining, clock); // was host instead of vm
-
- vm->pimpl_vm_->dp_updated_by_deleted_tasks += updated;
- }
-
- vm->pimpl_vm_->dp_objs.erase(key);
- delete dp;
-
- XBT_DEBUG("del %s on %s", key.c_str(), host->getCname());
-}
-
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,
const std::string& mbox, int stage, int stage2_round, double mig_speed,
double timeout)
/* Stage1: send all memory pages to the destination. */
XBT_DEBUG("mig-stage1: remaining_size %zu", remaining_size);
- start_dirty_page_tracking(ms->vm);
+ sg_vm_start_dirty_page_tracking(ms->vm);
double computed_during_stage1 = 0;
double clock_prev_send = MSG_get_clock();
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);
+ computed_during_stage1 = sg_vm_lookup_computed_flops(ms->vm);
if (sent < ramsize) {
XBT_VERB("mig-stage1: timeout, force moving to stage 3");
} 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);
+ sg_vm_stop_dirty_page_tracking(ms->vm);
return 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(ms->vm, 2, stage2_round);
+ double computed = sg_vm_lookup_computed_flops(ms->vm);
updated_size = get_updated_size(computed, dp_rate, dp_cap);
}
// 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);
+ sg_vm_stop_dirty_page_tracking(ms->vm);
return 0;
}
double clock_post_send = MSG_get_clock();
updated_size, (clock_post_send - clock_prev_send));
remaining_size -= sent;
- double computed = lookup_computed_flop_counts(ms->vm, 2, stage2_round);
+ double computed = sg_vm_lookup_computed_flops(ms->vm);
updated_size = get_updated_size(computed, dp_rate, dp_cap);
remaining_size += updated_size;
break;
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);
+ sg_vm_stop_dirty_page_tracking(ms->vm);
try {
XBT_DEBUG("Stage 3: Gonna send %zu bytes", remaining_size);
xbt_assert(get_mig_task_name(vm, src_pm, dst_pm, 4) == task->name);
MSG_task_destroy(task);
}
-
-/** @brief Immediately suspend 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.
- *
- * No suspension cost occurs.
- */
-void MSG_vm_suspend(msg_vm_t vm)
-{
- 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
- state->pushEvent("suspend");
- }
-}
-
-/** @brief Resume the execution of the VM. All processes on the VM run again.
- * @ingroup msg_VMs
- *
- * No resume cost occurs.
- */
-void MSG_vm_resume(msg_vm_t vm)
-{
- vm->resume();
- if (TRACE_msg_vm_is_enabled())
- simgrid::instr::Container::byName(vm->getName())->getState("MSG_VM_STATE")->popEvent();
-}
-
-/** @brief Get the physical host of a given VM.
- * @ingroup msg_VMs
- */
-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); });
-}
}