> [VM00:WRK02:(4) 10.280841] [msg_test/INFO] WRK02 executed task(Task02)
> [VM01:WRK03:(5) 10.361121] [msg_test/INFO] WRK03 received task(Task03) from mailbox(MBOX:WRK03)
> [node-0.acme.org:master:(1) 10.361121] [msg_test/INFO] # Migrate all VMs to PM(node-1.acme.org)
-> [node-1.acme.org:__pr_mig_tx:VM00(node-1.acme.org-node-1.acme.org):(7) 10.361121] [msg_vm/WARNING] use the default max_downtime value 30ms
+> [node-1.acme.org:__pr_mig_tx:VM00(node-1.acme.org-node-1.acme.org):(7) 10.361121] [vm_live_migration/WARNING] use the default max_downtime value 30ms
> [VM01:WRK03:(5) 10.371121] [msg_test/INFO] WRK03 executed task(Task03)
-> [node-10.acme.org:__pr_mig_tx:VM01(node-10.acme.org-node-1.acme.org):(9) 19.682922] [msg_vm/WARNING] use the default max_downtime value 30ms
+> [node-10.acme.org:__pr_mig_tx:VM01(node-10.acme.org-node-1.acme.org):(9) 19.682922] [vm_live_migration/WARNING] use the default max_downtime value 30ms
> [node-0.acme.org:master:(1) 28.561942] [msg_test/INFO] # Migrate all VMs to PM(node-10.acme.org)
-> [node-1.acme.org:__pr_mig_tx:VM00(node-1.acme.org-node-10.acme.org):(11) 28.561942] [msg_vm/WARNING] use the default max_downtime value 30ms
-> [node-1.acme.org:__pr_mig_tx:VM01(node-1.acme.org-node-10.acme.org):(13) 37.440963] [msg_vm/WARNING] use the default max_downtime value 30ms
+> [node-1.acme.org:__pr_mig_tx:VM00(node-1.acme.org-node-10.acme.org):(11) 28.561942] [vm_live_migration/WARNING] use the default max_downtime value 30ms
+> [node-1.acme.org:__pr_mig_tx:VM01(node-1.acme.org-node-10.acme.org):(13) 37.440963] [vm_live_migration/WARNING] use the default max_downtime value 30ms
> [node-0.acme.org:master:(1) 46.319984] [msg_test/INFO] # Shutdown the half of worker processes gracefully. The remaining half will be forcibly killed.
> [VM00:WRK00:(2) 46.327790] [msg_test/INFO] WRK00 received task(finalize) from mailbox(MBOX:WRK00)
> [VM01:WRK01:(3) 46.335596] [msg_test/INFO] WRK01 received task(finalize) from mailbox(MBOX:WRK01)
> [132.765801] (5:__pr_mig_tx:VM0(Fafard-Tremblay)@Fafard) use the default max_downtime value 30ms
> [146.111793] (1:master_@Fafard) VM0 migrated: Fafard->Tremblay in 13.346 s
> [146.111793] (1:master_@Fafard) Test: Migrate two VMs at once from PM0 to PM1
-> [146.111793] (9:__pr_mig_tx:VM0(Fafard-Tremblay)@Fafard) use the default max_downtime value 30ms
+> [146.111793] (8:__pr_mig_tx:VM0(Fafard-Tremblay)@Fafard) use the default max_downtime value 30ms
> [146.111793] (11:__pr_mig_tx:VM1(Fafard-Tremblay)@Fafard) use the default max_downtime value 30ms
-> [411.566271] (8:mig_wrk@Fafard) VM1 migrated: Fafard->Tremblay in 265.454 s
+> [411.566271] (9:mig_wrk@Fafard) VM1 migrated: Fafard->Tremblay in 265.454 s
> [411.566271] (6:mig_wrk@Fafard) VM0 migrated: Fafard->Tremblay in 265.454 s
> [10146.111793] (1:master_@Fafard) Test: Migrate two VMs at once to different PMs
-> [10146.111793] (15:__pr_mig_tx:VM0(Fafard-Tremblay)@Fafard) use the default max_downtime value 30ms
+> [10146.111793] (14:__pr_mig_tx:VM0(Fafard-Tremblay)@Fafard) use the default max_downtime value 30ms
> [10146.111793] (17:__pr_mig_tx:VM1(Fafard-Bourassa)@Fafard) use the default max_downtime value 30ms
-> [10362.620589] (14:mig_wrk@Fafard) VM1 migrated: Fafard->Bourassa in 216.509 s
+> [10362.620589] (15:mig_wrk@Fafard) VM1 migrated: Fafard->Bourassa in 216.509 s
> [10411.547334] (12:mig_wrk@Fafard) VM0 migrated: Fafard->Tremblay in 265.436 s
> [20146.111793] (0:maestro@) Bye (simulation time 20146.1)
* under the terms of the license (GNU LGPL) which comes with this package. */
#include "simgrid/s4u.hpp"
+#include "simgrid/plugins/live_migration.h"
#include "simgrid/s4u/VirtualMachine.hpp"
+
XBT_LOG_NEW_DEFAULT_CATEGORY(s4u_test, "Messages specific for this s4u example");
static void computation_fun()
{
s4u_Host* pm0 = simgrid::s4u::Host::by_name("Fafard");
s4u_Host* pm1 = simgrid::s4u::Host::by_name("Tremblay");
+ s4u_Host* pm2 = simgrid::s4u::Host::by_name("Bourassa");
XBT_INFO("## Test 1 (started): check computation on normal PMs");
vm1->destroy();
XBT_INFO("## Test 5 (ended)");
+ XBT_INFO("## Test 6 (started): Check migration impact (not yet implemented neither on the CPU resource nor on the"
+ " network one");
+ XBT_INFO("### Relocate VM0 between PM0 and PM1");
+ vm0 = new simgrid::s4u::VirtualMachine("VM0", pm0, 1);
+ s_vm_params_t params;
+ memset(¶ms, 0, sizeof(params));
+ vm0->setParameters(¶ms);
+ vm0->setRamsize(1L * 1024 * 1024 * 1024); // 1GiB
+
+ vm0->start();
+ launch_communication_worker(vm0, pm2);
+ simgrid::s4u::this_actor::sleep_for(0.01);
+ sg_vm_migrate(vm0, pm1);
+ simgrid::s4u::this_actor::sleep_for(0.01);
+ sg_vm_migrate(vm0, pm0);
+ simgrid::s4u::this_actor::sleep_for(5);
+ vm0->destroy();
+ XBT_INFO("## Test 6 (ended)");
}
int main(int argc, char* argv[])
{
simgrid::s4u::Engine e(&argc, argv);
+ sg_vm_live_migration_plugin_init();
e.loadPlatform(argv[1]); /* - Load the platform description */
simgrid::s4u::Actor::createActor("master_", simgrid::s4u::Host::by_name("Fafard"), master_main);
> [VM1:comm_rx:(34) 44.291085] [s4u_test/INFO] VM0:comm_tx to VM1:comm_rx => 0.291085 sec
> [VM1:comm_rx:(32) 44.291085] [s4u_test/INFO] VM0:comm_tx to VM1:comm_rx => 0.291085 sec
> [Fafard:master_:(1) 49.000000] [s4u_test/INFO] ## Test 5 (ended)
-> [49.000000] [s4u_test/INFO] Simulation time 49
+> [Fafard:master_:(1) 49.000000] [s4u_test/INFO] ## Test 6 (started): Check migration impact (not yet implemented neither on the CPU resource nor on the network one
+> [Fafard:master_:(1) 49.000000] [s4u_test/INFO] ### Relocate VM0 between PM0 and PM1
+> [Fafard:__pr_mig_tx:VM0(Fafard-Tremblay):(38) 49.010000] [vm_live_migration/WARNING] use the default max_downtime value 30ms
+> [Bourassa:comm_rx:(36) 49.204993] [s4u_test/INFO] VM0:comm_tx to Bourassa:comm_rx => 0.204993 sec
+> [Tremblay:__pr_mig_tx:VM0(Tremblay-Fafard):(40) 191.674258] [vm_live_migration/WARNING] use the default max_downtime value 30ms
+> [Fafard:master_:(1) 339.199251] [s4u_test/INFO] ## Test 6 (ended)
+> [339.199251] [s4u_test/INFO] Simulation time 339.199
/* Shutdown the guest operating system. */
XBT_PUBLIC(void) MSG_vm_shutdown(msg_vm_t vm);
-XBT_PUBLIC(void) MSG_vm_migrate(msg_vm_t vm, msg_host_t destination);
-
/* Suspend the execution of the VM, but keep its state on memory. */
XBT_PUBLIC(void) MSG_vm_suspend(msg_vm_t vm);
XBT_PUBLIC(void) MSG_vm_resume(msg_vm_t vm);
XBT_PUBLIC(void) sg_vm_start_dirty_page_tracking(sg_vm_t vm);
XBT_PUBLIC(void) sg_vm_stop_dirty_page_tracking(sg_vm_t vm);
XBT_PUBLIC(double) sg_vm_lookup_computed_flops(sg_vm_t vm);
+XBT_PUBLIC(void) sg_vm_migrate(sg_vm_t vm, sg_host_t dst_pm);
#define MSG_vm_live_migration_plugin_init() sg_vm_live_migration_plugin_init()
+#define MSG_vm_migrate(vm, dst_pm) sg_vm_migrate(vm, dst_pm)
SG_END_DECL()
#include "jmsg_vm.h"
#include "jmsg_host.h"
#include "jxbt_utilities.hpp"
+#include "simgrid/plugins/live_migration.h"
#include "src/plugins/vm/VirtualMachineImpl.hpp"
#include "xbt/ex.hpp"
}
}
-static std::string get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
-{
- return std::string("__pr_mig_tx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
-}
-
-static std::string get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
-{
- return std::string("__pr_mig_rx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
-}
-
-static std::string get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
-{
- return std::string("__task_mig_stage") + std::to_string(stage) + ":" + vm->getCname() + "(" + src_pm->getCname() +
- "-" + dst_pm->getCname() + ")";
-}
-
-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. */
- std::string mbox_ctl;
- /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
- std::string mbox;
-};
-
-static int migration_rx_fun(int argc, char *argv[])
-{
- XBT_DEBUG("mig: rx_start");
-
- // The structure has been created in the do_migration function and should only be freed in the same place ;)
- migration_session* ms = static_cast<migration_session*>(MSG_process_get_data(MSG_process_self()));
-
- bool received_finalize = false;
-
- std::string finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
- while (not received_finalize) {
- msg_task_t task = nullptr;
- int ret = MSG_task_recv(&task, ms->mbox.c_str());
-
- if (ret != MSG_OK) {
- // An error occurred, clean the code and return
- // The owner did not change, hence the task should be only destroyed on the other side
- return 0;
- }
-
- if (finalize_task_name == task->name)
- received_finalize = 1;
-
- MSG_task_destroy(task);
- }
-
- // 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 beginning and free it at the end of each function)
- simgrid::s4u::VirtualMachine* vm = ms->vm;
- 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, dst_pm]() {
- /* Update the vm location */
- /* precopy migration makes the VM temporally paused */
- xbt_assert(vm->getState() == SURF_VM_STATE_SUSPENDED);
-
- /* Update the vm location and resume it */
- vm->pimpl_vm_->setPm(dst_pm);
- 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)", ms->vm->getCname(), ms->src_pm->getCname(), ms->dst_pm->getCname());
-
- if (TRACE_msg_vm_is_enabled()) {
- static long long int counter = 0;
- std::string key = std::to_string(counter);
- counter++;
-
- // start link
- container_t msg = simgrid::instr::Container::byName(vm->getName());
- simgrid::instr::Container::getRoot()->getLink("MSG_VM_LINK")->startEvent(msg, "M", key);
-
- // destroy existing container of this vm
- container_t existing_container = simgrid::instr::Container::byName(vm->getName());
- existing_container->removeFromParent();
- delete existing_container;
-
- // create new container on the new_host location
- new simgrid::instr::Container(vm->getCname(), "MSG_VM", simgrid::instr::Container::byName(ms->dst_pm->getName()));
-
- // end link
- msg = simgrid::instr::Container::byName(vm->getName());
- simgrid::instr::Container::getRoot()->getLink("MSG_VM_LINK")->endEvent(msg, "M", key);
- }
-
- // Inform the SRC that the migration has been correctly performed
- std::string task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
- msg_task_t task = MSG_task_create(task_name.c_str(), 0, 0, nullptr);
- msg_error_t ret = MSG_task_send(task, ms->mbox_ctl.c_str());
- 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_DEBUG("mig: rx_done");
- return 0;
-}
-
-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)
-{
- sg_size_t sent = 0;
- std::string task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
- msg_task_t task = MSG_task_create(task_name.c_str(), 0, static_cast<double>(size), nullptr);
-
- double clock_sta = MSG_get_clock();
-
- msg_error_t ret;
- if (mig_speed > 0)
- ret = MSG_task_send_with_timeout_bounded(task, mbox.c_str(), timeout, mig_speed);
- else
- ret = MSG_task_send(task, mbox.c_str());
-
- if (ret == MSG_OK) {
- sent = size;
- } else if (ret == MSG_TIMEOUT) {
- sg_size_t remaining = static_cast<sg_size_t>(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)", vm->getCname(), stage);
- MSG_task_destroy(task);
- THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", vm->getCname(), stage);
- }else if(ret == MSG_TRANSFER_FAILURE){
- XBT_DEBUG("DST host failed during migration of %s (stage %d)", vm->getCname(), stage);
- MSG_task_destroy(task);
- THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", vm->getCname(), 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);
-
- return sent;
-}
-
-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);
- if (updated_size > dp_cap) {
- updated_size = dp_cap;
- }
-
- return static_cast<sg_size_t>(updated_size);
-}
-
-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 = static_cast<migration_session*>(MSG_process_get_data(MSG_process_self()));
-
- double host_speed = ms->vm->getPm()->getSpeed();
- s_vm_params_t params;
- ms->vm->getParameters(¶ms);
- 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;
-
- size_t remaining_size = ramsize;
- size_t threshold = 0.0;
-
- /* check parameters */
- if (ramsize == 0)
- XBT_WARN("migrate a VM, but ramsize is zero");
-
- if (max_downtime <= 0) {
- XBT_WARN("use the default max_downtime value 30ms");
- max_downtime = 0.03;
- }
-
- /* Stage1: send all memory pages to the destination. */
- XBT_DEBUG("mig-stage1: remaining_size %zu", remaining_size);
- sg_vm_start_dirty_page_tracking(ms->vm);
-
- double computed_during_stage1 = 0;
- 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 = sg_vm_lookup_computed_flops(ms->vm);
-
- 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)
- sg_vm_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 = 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) {
-
- int stage2_round = 0;
- for (;;) {
-
- 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 = sg_vm_lookup_computed_flops(ms->vm);
- updated_size = get_updated_size(computed, dp_rate, dp_cap);
- }
-
- 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);
-
- /* 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 %zu (%s threshold %zu)", stage2_round, remaining_size,
- (remaining_size < threshold) ? "<" : ">", threshold);
- if (remaining_size < threshold)
- break;
-
- sg_size_t sent = 0;
- double clock_prev_send = MSG_get_clock();
- try {
- XBT_DEBUG("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 (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)
- sg_vm_stop_dirty_page_tracking(ms->vm);
- return 0;
- }
- double clock_post_send = MSG_get_clock();
-
- if (sent == updated_size) {
- /* timeout did not happen */
- double bandwidth = updated_size / (clock_post_send - clock_prev_send);
- threshold = bandwidth * max_downtime;
- 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);
- 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_VERB("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 = sg_vm_lookup_computed_flops(ms->vm);
- updated_size = get_updated_size(computed, dp_rate, dp_cap);
- remaining_size += updated_size;
- break;
- } else
- XBT_CRITICAL("bug");
- }
- }
-
- /* Stage3: stop the VM and copy the rest of states. */
- 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
- pimpl->suspend(SIMIX_process_self());
- sg_vm_stop_dirty_page_tracking(ms->vm);
-
- try {
- 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->resume();
- return 0;
- }
-
- // 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_DEBUG("mig: tx_done");
-
- return 0;
-}
-
-/** @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.
- */
-
- 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());
- if (dst_pm->isOff())
- THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->getCname(), dst_pm->getCname());
- if (not MSG_vm_is_running(vm))
- THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->getCname());
- if (vm->isMigrating())
- THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->getCname());
-
- vm->pimpl_vm_->isMigrating = true;
-
- migration_session ms;
- ms.vm = vm;
- ms.src_pm = src_pm;
- ms.dst_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 =
- simgrid::xbt::string_printf("__mbox_mig_ctl:%s(%s-%s)", vm->getCname(), src_pm->getCname(), dst_pm->getCname());
- ms.mbox = simgrid::xbt::string_printf("__mbox_mig_src_dst:%s(%s-%s)", vm->getCname(), src_pm->getCname(),
- dst_pm->getCname());
-
- std::string pr_rx_name = get_mig_process_rx_name(vm, src_pm, dst_pm);
- std::string pr_tx_name = get_mig_process_tx_name(vm, src_pm, dst_pm);
-
- MSG_process_create(pr_rx_name.c_str(), migration_rx_fun, &ms, dst_pm);
-
- MSG_process_create(pr_tx_name.c_str(), migration_tx_fun, &ms, src_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_task_receive(&task, ms.mbox_ctl.c_str());
-
- vm->pimpl_vm_->isMigrating = false;
-
- 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
- THROWF(host_error, 0, "Source host '%s' failed during the migration of VM '%s'.", src_pm->getCname(),
- vm->getCname());
- } 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)");
- THROWF(host_error, 0, "Destination host '%s' failed during the migration of VM '%s'.", dst_pm->getCname(),
- vm->getCname());
- }
-
- xbt_assert(get_mig_task_name(vm, src_pm, dst_pm, 4) == task->name);
- MSG_task_destroy(task);
-}
}
--- /dev/null
+/* Copyright (c) 2013-2017. The SimGrid Team. All rights reserved. */
+
+/* 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. */
+
+#include <simgrid/plugins/live_migration.h>
+#include <simgrid/s4u.hpp>
+#include <simgrid/s4u/VirtualMachine.hpp>
+#include <src/instr/instr_private.hpp>
+#include <src/plugins/vm/VirtualMachineImpl.hpp>
+#include <src/plugins/vm/VmLiveMigration.hpp>
+#include <xbt/ex.hpp>
+
+XBT_LOG_NEW_DEFAULT_CATEGORY(vm_live_migration, "S4U virtual machines live migration");
+
+namespace simgrid {
+namespace vm {
+
+void MigrationRx::operator()()
+{
+ XBT_DEBUG("mig: rx_start");
+ bool received_finalize = false;
+
+ std::string finalize_task_name =
+ std::string("__mig_stage3:") + vm_->getCname() + "(" + src_pm_->getCname() + "-" + dst_pm_->getCname() + ")";
+
+ while (not received_finalize) {
+ std::string* payload = static_cast<std::string*>(mbox->get());
+
+ if (finalize_task_name == *payload)
+ received_finalize = true;
+
+ delete payload;
+ }
+
+ // 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 beginning and free it at the end of each function)
+ simgrid::s4u::VirtualMachine* vm = vm_;
+ simgrid::s4u::Host* dst_pm = 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, dst_pm]() {
+ /* Update the vm location */
+ /* precopy migration makes the VM temporally paused */
+ xbt_assert(vm->getState() == SURF_VM_STATE_SUSPENDED);
+
+ /* Update the vm location and resume it */
+ vm->pimpl_vm_->setPm(dst_pm);
+ 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)", vm_->getCname(), src_pm_->getCname(), dst_pm_->getCname());
+
+ if (TRACE_msg_vm_is_enabled()) {
+ static long long int counter = 0;
+ std::string key = std::to_string(counter);
+ counter++;
+
+ // start link
+ container_t msg = simgrid::instr::Container::byName(vm->getName());
+ simgrid::instr::Container::getRoot()->getLink("MSG_VM_LINK")->startEvent(msg, "M", key);
+
+ // destroy existing container of this vm
+ container_t existing_container = simgrid::instr::Container::byName(vm->getName());
+ existing_container->removeFromParent();
+ delete existing_container;
+
+ // create new container on the new_host location
+ new simgrid::instr::Container(vm->getCname(), "MSG_VM", simgrid::instr::Container::byName(dst_pm_->getName()));
+
+ // end link
+ msg = simgrid::instr::Container::byName(vm->getName());
+ simgrid::instr::Container::getRoot()->getLink("MSG_VM_LINK")->endEvent(msg, "M", key);
+ }
+ // Inform the SRC that the migration has been correctly performed
+ std::string* payload = new std::string("__mig_stage4:");
+ *payload = *payload + vm_->getCname() + "(" + src_pm_->getCname() + "-" + dst_pm_->getCname() + ")";
+
+ mbox_ctl->put(payload, 0);
+
+ XBT_DEBUG("mig: rx_done");
+}
+
+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);
+ if (updated_size > dp_cap) {
+ updated_size = dp_cap;
+ }
+
+ return static_cast<sg_size_t>(updated_size);
+}
+
+sg_size_t MigrationTx::sendMigrationData(sg_size_t size, int stage, int stage2_round, double mig_speed, double timeout)
+{
+ sg_size_t sent = size;
+ std::string* msg = new std::string("__mig_stage");
+ *msg = *msg + std::to_string(stage) + ":" + vm_->getCname() + "(" + src_pm_->getCname() + "-" + dst_pm_->getCname() +
+ ")";
+
+ double clock_sta = s4u::Engine::getClock();
+
+ s4u::Activity* comm = nullptr;
+ try {
+ if (mig_speed > 0)
+ comm = mbox->put_init(msg, size)->setRate(mig_speed)->wait(timeout);
+ else
+ comm = mbox->put_async(msg, size)->wait();
+ } catch (xbt_ex& e) {
+ if (comm) {
+ sg_size_t remaining = static_cast<sg_size_t>(comm->getRemains());
+ XBT_VERB("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu", timeout, remaining, size);
+ sent -= remaining;
+ }
+ }
+
+ double clock_end = s4u::Engine::getClock();
+ 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);
+
+ return sent;
+}
+
+void MigrationTx::operator()()
+{
+ XBT_DEBUG("mig: tx_start");
+
+ double host_speed = vm_->getPm()->getSpeed();
+ s_vm_params_t params;
+ vm_->getParameters(¶ms);
+ const sg_size_t ramsize = 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;
+
+ size_t remaining_size = ramsize;
+ size_t threshold = 0.0;
+
+ /* check parameters */
+ if (ramsize == 0)
+ XBT_WARN("migrate a VM, but ramsize is zero");
+
+ if (max_downtime <= 0) {
+ XBT_WARN("use the default max_downtime value 30ms");
+ max_downtime = 0.03;
+ }
+
+ /* Stage1: send all memory pages to the destination. */
+ XBT_DEBUG("mig-stage1: remaining_size %zu", remaining_size);
+ sg_vm_start_dirty_page_tracking(vm_);
+
+ double computed_during_stage1 = 0;
+ double clock_prev_send = s4u::Engine::getClock();
+
+ 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 = sendMigrationData(ramsize, 1, 0, mig_speed, -1);
+ remaining_size -= sent;
+ computed_during_stage1 = sg_vm_lookup_computed_flops(vm_);
+
+ 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)
+ sg_vm_stop_dirty_page_tracking(vm_);
+ return;
+ }
+
+ double clock_post_send = s4u::Engine::getClock();
+ 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) {
+
+ int stage2_round = 0;
+ for (;;) {
+ 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 = sg_vm_lookup_computed_flops(vm_);
+ updated_size = get_updated_size(computed, dp_rate, dp_cap);
+ }
+
+ 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);
+
+ /* 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 %zu (%s threshold %zu)", stage2_round, remaining_size,
+ (remaining_size < threshold) ? "<" : ">", threshold);
+ if (remaining_size < threshold)
+ break;
+
+ sg_size_t sent = 0;
+ double clock_prev_send = s4u::Engine::getClock();
+ try {
+ XBT_DEBUG("Stage 2, gonna send %llu", updated_size);
+ sent = sendMigrationData(updated_size, 2, stage2_round, mig_speed, mig_timeout);
+ } 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)
+ sg_vm_stop_dirty_page_tracking(vm_);
+ return;
+ }
+ double clock_post_send = s4u::Engine::getClock();
+
+ if (sent == updated_size) {
+ /* timeout did not happen */
+ double bandwidth = updated_size / (clock_post_send - clock_prev_send);
+ threshold = bandwidth * max_downtime;
+ 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);
+ 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_VERB("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 = sg_vm_lookup_computed_flops(vm_);
+ updated_size = get_updated_size(computed, dp_rate, dp_cap);
+ remaining_size += updated_size;
+ break;
+ } else
+ XBT_CRITICAL("bug");
+ }
+ }
+
+ /* Stage3: stop the VM and copy the rest of states. */
+ XBT_DEBUG("mig-stage3: remaining_size %zu", remaining_size);
+ vm_->suspend();
+ sg_vm_stop_dirty_page_tracking(vm_);
+
+ try {
+ XBT_DEBUG("Stage 3: Gonna send %zu bytes", remaining_size);
+ sendMigrationData(remaining_size, 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)
+ vm_->resume();
+ return;
+ }
+
+ // 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_DEBUG("mig: tx_done");
+}
+}
+}
+
+SG_BEGIN_DECL()
+void sg_vm_migrate(simgrid::s4u::VirtualMachine* vm, simgrid::s4u::Host* dst_pm)
+{
+ simgrid::s4u::Host* 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());
+ if (dst_pm->isOff())
+ THROWF(vm_error, 0, "Cannot migrate VM '%s' to host '%s', which is offline.", vm->getCname(), dst_pm->getCname());
+ if (vm->getState() != SURF_VM_STATE_RUNNING)
+ THROWF(vm_error, 0, "Cannot migrate VM '%s' that is not running yet.", vm->getCname());
+ if (vm->isMigrating())
+ THROWF(vm_error, 0, "Cannot migrate VM '%s' that is already migrating.", vm->getCname());
+
+ std::string rx_name =
+ std::string("__pr_mig_rx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
+ std::string tx_name =
+ std::string("__pr_mig_tx:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")";
+
+ simgrid::s4u::ActorPtr rx =
+ simgrid::s4u::Actor::createActor(rx_name.c_str(), dst_pm, simgrid::vm::MigrationRx(vm, dst_pm));
+ simgrid::s4u::ActorPtr tx =
+ simgrid::s4u::Actor::createActor(tx_name.c_str(), src_pm, simgrid::vm::MigrationTx(vm, 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");
+ simgrid::s4u::MailboxPtr mbox_ctl = simgrid::s4u::Mailbox::byName(
+ std::string("__mbox_mig_ctl:") + vm->getCname() + "(" + src_pm->getCname() + "-" + dst_pm->getCname() + ")");
+ delete static_cast<std::string*>(mbox_ctl->get());
+
+ tx->join();
+ rx->join();
+
+ vm->pimpl_vm_->isMigrating = false;
+}
+}
--- /dev/null
+/* Copyright (c) 2004-2016. The SimGrid Team. All rights reserved. */
+
+/* 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. */
+
+#include "simgrid/s4u.hpp"
+
+#ifndef VM_LIVE_MIGRATION_HPP_
+#define VM_LIVE_MIGRATION_HPP_
+
+namespace simgrid {
+namespace vm {
+
+class MigrationRx {
+ /* The miration_rx process uses mbox_ctl to let the caller of do_migration() know the completion of the migration. */
+ s4u::MailboxPtr mbox_ctl;
+ /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
+ s4u::MailboxPtr mbox;
+ s4u::VirtualMachine* vm_;
+ s4u::Host* src_pm_ = nullptr;
+ s4u::Host* dst_pm_ = nullptr;
+
+public:
+ explicit MigrationRx(s4u::VirtualMachine* vm, s4u::Host* dst_pm) : vm_(vm), dst_pm_(dst_pm)
+ {
+ src_pm_ = vm_->getPm();
+
+ mbox_ctl = s4u::Mailbox::byName(std::string("__mbox_mig_ctl:") + vm_->getCname() + "(" + src_pm_->getCname() + "-" +
+ dst_pm_->getCname() + ")");
+ mbox = s4u::Mailbox::byName(std::string("__mbox_mig_src_dst:") + vm_->getCname() + "(" + src_pm_->getCname() + "-" +
+ dst_pm_->getCname() + ")");
+ }
+ void operator()();
+};
+
+class MigrationTx {
+ /* The migration_rx and migration_tx processes use mbox to transfer migration data. */
+ s4u::MailboxPtr mbox;
+ s4u::VirtualMachine* vm_;
+ s4u::Host* src_pm_ = nullptr;
+ s4u::Host* dst_pm_ = nullptr;
+
+public:
+ explicit MigrationTx(s4u::VirtualMachine* vm, s4u::Host* dst_pm) : vm_(vm), dst_pm_(dst_pm)
+ {
+ src_pm_ = vm_->getPm();
+ mbox = s4u::Mailbox::byName(std::string("__mbox_mig_src_dst:") + vm_->getCname() + "(" + src_pm_->getCname() + "-" +
+ dst_pm_->getCname() + ")");
+ }
+ void operator()();
+ sg_size_t sendMigrationData(sg_size_t size, int stage, int stage2_round, double mig_speed, double timeout);
+};
+}
+}
+#endif
/* 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. */
+#include "simgrid/s4u/Actor.hpp"
#include "src/instr/instr_private.hpp"
#include "src/plugins/vm/VirtualMachineImpl.hpp"
#include "src/plugins/vm/VmHostExt.hpp"
#include "src/simix/smx_host_private.hpp"
#include "src/surf/cpu_cas01.hpp"
+#include <src/plugins/vm/VmLiveMigration.hpp>
XBT_LOG_NEW_DEFAULT_CATEGORY(s4u_vm, "S4U virtual machines");
> [Fafard:master_:(1) 49.000000] [msg_test/INFO] ## Test 5 (ended)
> [Fafard:master_:(1) 49.000000] [msg_test/INFO] ## Test 6 (started): Check migration impact (not yet implemented neither on the CPU resource nor on the network one
> [Fafard:master_:(1) 49.000000] [msg_test/INFO] ### Relocate VM0 between PM0 and PM1
-> [Fafard:__pr_mig_tx:VM0(Fafard-Tremblay):(38) 49.010000] [msg_vm/WARNING] use the default max_downtime value 30ms
+> [Fafard:__pr_mig_tx:VM0(Fafard-Tremblay):(38) 49.010000] [vm_live_migration/WARNING] use the default max_downtime value 30ms
> [Bourassa:comm_rx:(36) 49.204993] [msg_test/INFO] VM0:comm_tx to Bourassa:comm_rx => 0.204993 sec
-> [Tremblay:__pr_mig_tx:VM0(Tremblay-Fafard):(40) 191.674258] [msg_vm/WARNING] use the default max_downtime value 30ms
+> [Tremblay:__pr_mig_tx:VM0(Tremblay-Fafard):(40) 191.674258] [vm_live_migration/WARNING] use the default max_downtime value 30ms
> [Fafard:master_:(1) 339.199251] [msg_test/INFO] ## Test 6 (ended)
> [339.199251] [msg_test/INFO] Bye (simulation time 339.199)
src/plugins/vm/VirtualMachineImpl.cpp
src/plugins/vm/VmHostExt.hpp
src/plugins/vm/VmHostExt.cpp
+ src/plugins/vm/VmLiveMigration.cpp
+ src/plugins/vm/VmLiveMigration.hpp
)
set(SIMIX_GENERATED_SRC src/simix/popping_generated.cpp )
