TEST_P(CpuIsolatorTest, ROOT_UserCpuUsage)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = GetParam();
Fetcher fetcher(flags);
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(
detector.get(),
containerizer.get());
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched,
DEFAULT_FRAMEWORK_INFO,
master.get()->pid,
DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
// Max out a single core in userspace. This will run for at most one
// second.
TaskInfo task = createTask(
offers.get()[0],
"while true ; do true ; done & sleep 60");
Future<TaskStatus> statusRunning;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusRunning));
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *(containers->begin());
// Wait up to 1 second for the child process to induce 1/8 of a
// second of user cpu time.
ResourceStatistics statistics;
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
statistics = usage.get();
// If we meet our usage expectations, we're done!
if (statistics.cpus_user_time_secs() >= 0.125) {
break;
}
os::sleep(Milliseconds(200));
waited += Milliseconds(200);
} while (waited < Seconds(1));
EXPECT_LE(0.125, statistics.cpus_user_time_secs());
driver.stop();
driver.join();
}
TEST_F(DiskQuotaTest, ResourceStatistics)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "posix/cpu,posix/mem,disk/du";
flags.resources = strings::format("disk(%s):10", DEFAULT_TEST_ROLE).get();
// NOTE: We can't pause the clock because we need the reaper to reap
// the 'du' subprocess.
flags.container_disk_watch_interval = Milliseconds(1);
Fetcher fetcher(flags);
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave =
StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_role(DEFAULT_TEST_ROLE);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched,
frameworkInfo,
master.get()->pid,
DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
const Offer& offer = offers.get()[0];
Resource volume = createPersistentVolume(
Megabytes(4),
DEFAULT_TEST_ROLE,
"id1",
"path1",
None(),
None(),
DEFAULT_CREDENTIAL.principal());
Resources taskResources = Resources::parse("cpus:1;mem:128").get();
taskResources += createDiskResource(
"3",
DEFAULT_TEST_ROLE,
None(),
None());
taskResources += volume;
// Create a task that uses 2MB disk.
TaskInfo task = createTask(
offer.slave_id(),
taskResources,
"dd if=/dev/zero of=file bs=1048576 count=2 && "
"dd if=/dev/zero of=path1/file bs=1048576 count=2 && "
"sleep 1000");
Future<TaskStatus> status1;
Future<TaskStatus> status2;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status1))
.WillOnce(FutureArg<1>(&status2))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.acceptOffers(
{offer.id()},
{CREATE(volume),
LAUNCH({task})});
AWAIT_READY(status1);
EXPECT_EQ(task.task_id(), status1->task_id());
EXPECT_EQ(TASK_RUNNING, status1->state());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
const ContainerID& containerId = *(containers->begin());
// Wait until disk usage can be retrieved.
Duration elapsed = Duration::zero();
while (true) {
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
ASSERT_TRUE(usage->has_disk_limit_bytes());
EXPECT_EQ(Megabytes(3), Bytes(usage->disk_limit_bytes()));
if (usage->has_disk_used_bytes()) {
EXPECT_LE(usage->disk_used_bytes(), usage->disk_limit_bytes());
}
ASSERT_EQ(2u, usage->disk_statistics().size());
bool done = true;
foreach (const DiskStatistics& statistics, usage->disk_statistics()) {
ASSERT_TRUE(statistics.has_limit_bytes());
EXPECT_EQ(
statistics.has_persistence() ? Megabytes(4) : Megabytes(3),
statistics.limit_bytes());
if (!statistics.has_used_bytes()) {
done = false;
} else {
EXPECT_GT(
statistics.has_persistence() ? Megabytes(4) : Megabytes(3),
statistics.used_bytes());
}
}
if (done) {
break;
}
ASSERT_LT(elapsed, Seconds(5));
os::sleep(Milliseconds(1));
elapsed += Milliseconds(1);
}
driver.killTask(task.task_id());
AWAIT_READY(status2);
EXPECT_EQ(task.task_id(), status2->task_id());
EXPECT_EQ(TASK_KILLED, status2->state());
driver.stop();
driver.join();
}
// This test verifies that disk quota isolator recovers properly after
// the slave restarts.
TEST_F(DiskQuotaTest, SlaveRecovery)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "posix/cpu,posix/mem,disk/du";
flags.container_disk_watch_interval = Milliseconds(1);
Fetcher fetcher(flags);
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave =
StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
// Enable checkpointing for the framework.
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true);
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
const Offer& offer = offers.get()[0];
// Create a task that uses 2MB disk.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:3").get(),
"dd if=/dev/zero of=file bs=1048576 count=2 && sleep 1000");
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offer.id(), {task});
AWAIT_READY(status);
EXPECT_EQ(task.task_id(), status->task_id());
EXPECT_EQ(TASK_RUNNING, status->state());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
const ContainerID& containerId = *(containers->begin());
// Stop the slave.
slave.get()->terminate();
Future<ReregisterExecutorMessage> reregisterExecutorMessage =
FUTURE_PROTOBUF(ReregisterExecutorMessage(), _, _);
Future<Nothing> _recover = FUTURE_DISPATCH(_, &Slave::_recover);
_containerizer = MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
containerizer.reset(_containerizer.get());
detector = master.get()->createDetector();
slave = StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
Clock::pause();
AWAIT_READY(_recover);
// Wait for slave to schedule reregister timeout.
Clock::settle();
// Ensure the executor re-registers before completing recovery.
AWAIT_READY(reregisterExecutorMessage);
// Ensure the slave considers itself recovered.
Clock::advance(flags.executor_reregistration_timeout);
// NOTE: We resume the clock because we need the reaper to reap the
// 'du' subprocess.
Clock::resume();
// Wait until disk usage can be retrieved.
Duration elapsed = Duration::zero();
while (true) {
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
ASSERT_TRUE(usage->has_disk_limit_bytes());
EXPECT_EQ(Megabytes(3), Bytes(usage->disk_limit_bytes()));
if (usage->has_disk_used_bytes()) {
EXPECT_LE(usage->disk_used_bytes(), usage->disk_limit_bytes());
// NOTE: This is to capture the regression in MESOS-2452. The data
// stored in the executor meta directory should be less than 64K.
if (usage->disk_used_bytes() > Kilobytes(64).bytes()) {
break;
}
}
ASSERT_LT(elapsed, Seconds(15));
os::sleep(Milliseconds(1));
elapsed += Milliseconds(1);
}
driver.stop();
driver.join();
}
TEST_F(MemoryPressureMesosTest, CGROUPS_ROOT_Statistics)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
// We only care about memory cgroup for this test.
flags.isolation = "cgroups/mem";
Fetcher fetcher(flags);
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave =
StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
// Run a task that triggers memory pressure event. We request 1G
// disk because we are going to write a 512 MB file repeatedly.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:256;disk:1024").get(),
"while true; do dd count=512 bs=1M if=/dev/zero of=./temp; done");
Future<TaskStatus> starting;
Future<TaskStatus> running;
Future<TaskStatus> killed;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&starting))
.WillOnce(FutureArg<1>(&running))
.WillOnce(FutureArg<1>(&killed))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offer.id(), {task});
AWAIT_READY(starting);
EXPECT_EQ(task.task_id(), starting->task_id());
EXPECT_EQ(TASK_STARTING, starting->state());
AWAIT_READY(running);
EXPECT_EQ(task.task_id(), running->task_id());
EXPECT_EQ(TASK_RUNNING, running->state());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *(containers->begin());
// Wait a while for some memory pressure events to occur.
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
if (usage->mem_low_pressure_counter() > 0) {
// We will check the correctness of the memory pressure counters
// later, because the memory-hammering task is still active
// and potentially incrementing these counters.
break;
}
os::sleep(Milliseconds(100));
waited += Milliseconds(100);
} while (waited < Seconds(5));
EXPECT_LE(waited, Seconds(5));
// Pause the clock to ensure that the reaper doesn't reap the exited
// command executor and inform the containerizer/slave.
Clock::pause();
Clock::settle();
// Stop the memory-hammering task.
driver.killTask(task.task_id());
AWAIT_READY_FOR(killed, Seconds(120));
EXPECT_EQ(task.task_id(), killed->task_id());
EXPECT_EQ(TASK_KILLED, killed->state());
// Now check the correctness of the memory pressure counters.
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
EXPECT_GE(usage->mem_low_pressure_counter(),
usage->mem_medium_pressure_counter());
EXPECT_GE(usage->mem_medium_pressure_counter(),
usage->mem_critical_pressure_counter());
Clock::resume();
driver.stop();
driver.join();
}
// This test verifies that the container will not be killed if
// disk_enforce_quota flag is false (even if the disk usage exceeds
// its quota).
TEST_F(DiskQuotaTest, NoQuotaEnforcement)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "posix/cpu,posix/mem,disk/du";
// NOTE: We can't pause the clock because we need the reaper to reap
// the 'du' subprocess.
flags.container_disk_watch_interval = Milliseconds(1);
flags.enforce_container_disk_quota = false;
Fetcher fetcher(flags);
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave =
StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
const Offer& offer = offers.get()[0];
// Create a task that uses 2MB disk.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:1").get(),
"dd if=/dev/zero of=file bs=1048576 count=2 && sleep 1000");
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offer.id(), {task});
AWAIT_READY(status);
EXPECT_EQ(task.task_id(), status->task_id());
EXPECT_EQ(TASK_RUNNING, status->state());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
const ContainerID& containerId = *(containers->begin());
// Wait until disk usage can be retrieved and the usage actually
// exceeds the limit. If the container is killed due to quota
// enforcement (which shouldn't happen), the 'usage' call will
// return a failed future, leading to a failed test.
Duration elapsed = Duration::zero();
while (true) {
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
ASSERT_TRUE(usage->has_disk_limit_bytes());
EXPECT_EQ(Megabytes(1), Bytes(usage->disk_limit_bytes()));
if (usage->has_disk_used_bytes() &&
usage->disk_used_bytes() > usage->disk_limit_bytes()) {
break;
}
ASSERT_LT(elapsed, Seconds(5));
os::sleep(Milliseconds(1));
elapsed += Milliseconds(1);
}
driver.stop();
driver.join();
}
// In this test, the framework is not checkpointed. This ensures that when we
// stop the slave, the executor is killed and we will need to recover the
// working directories without getting any checkpointed recovery state.
TEST_F(ROOT_XFS_QuotaTest, NoCheckpointRecovery)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
Fetcher fetcher(flags);
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(
detector.get(),
containerizer.get(),
flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:1").get(),
"dd if=/dev/zero of=file bs=1048576 count=1; sleep 1000");
Future<TaskStatus> runningStatus;
Future<TaskStatus> startingStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&startingStatus))
.WillOnce(FutureArg<1>(&runningStatus))
.WillOnce(Return());
driver.launchTasks(offer.id(), {task});
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), startingStatus->task_id());
EXPECT_EQ(TASK_STARTING, startingStatus->state());
AWAIT_READY(runningStatus);
EXPECT_EQ(task.task_id(), runningStatus->task_id());
EXPECT_EQ(TASK_RUNNING, runningStatus->state());
Future<ResourceUsage> usage1 =
process::dispatch(slave.get()->pid, &Slave::usage);
AWAIT_READY(usage1);
// We should have 1 executor using resources.
ASSERT_EQ(1, usage1->executors().size());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *containers->begin();
// Restart the slave.
slave.get()->terminate();
Future<SlaveReregisteredMessage> slaveReregisteredMessage =
FUTURE_PROTOBUF(SlaveReregisteredMessage(), _, _);
_containerizer = MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
containerizer.reset(_containerizer.get());
slave = StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
// Wait until slave recovery is complete.
Future<Nothing> _recover = FUTURE_DISPATCH(_, &Slave::_recover);
AWAIT_READY_FOR(_recover, Seconds(60));
// Wait until the orphan containers are cleaned up.
AWAIT_READY_FOR(containerizer.get()->wait(containerId), Seconds(60));
AWAIT_READY(slaveReregisteredMessage);
Future<ResourceUsage> usage2 =
process::dispatch(slave.get()->pid, &Slave::usage);
AWAIT_READY(usage2);
// We should have no executors left because we didn't checkpoint.
ASSERT_TRUE(usage2->executors().empty());
Try<std::list<string>> sandboxes = os::glob(path::join(
slave::paths::getSandboxRootDir(mountPoint.get()),
"*",
"frameworks",
"*",
"executors",
"*",
"runs",
"*"));
ASSERT_SOME(sandboxes);
// One sandbox and one symlink.
ASSERT_EQ(2u, sandboxes->size());
// Scan the remaining sandboxes and make sure that no projects are assigned.
foreach (const string& sandbox, sandboxes.get()) {
// Skip the "latest" symlink.
if (os::stat::islink(sandbox)) {
continue;
}
EXPECT_NONE(xfs::getProjectId(sandbox));
}
driver.stop();
driver.join();
}
// This is the same logic as ResourceStatistics, except the task should
// be allowed to exceed the disk quota, and usage statistics should report
// that the quota was exceeded.
TEST_F(ROOT_XFS_QuotaTest, ResourceStatisticsNoEnforce)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.enforce_container_disk_quota = false;
Fetcher fetcher(flags);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Try<Owned<cluster::Slave>> slave =
StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
// Create a task that uses 4MB of 3MB disk and fails if it can't
// write the full amount.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:3").get(),
"dd if=/dev/zero of=file bs=1048576 count=4 && sleep 1000");
Future<TaskStatus> startingStatus;
Future<TaskStatus> runningStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&startingStatus))
.WillOnce(FutureArg<1>(&runningStatus))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), startingStatus->task_id());
EXPECT_EQ(TASK_STARTING, startingStatus->state());
AWAIT_READY(runningStatus);
EXPECT_EQ(task.task_id(), runningStatus->task_id());
EXPECT_EQ(TASK_RUNNING, runningStatus->state());
Future<hashset<ContainerID>> containers = containerizer.get()->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *(containers->begin());
Duration diskTimeout = Seconds(5);
Timeout timeout = Timeout::in(diskTimeout);
while (true) {
Future<ResourceStatistics> usage = containerizer.get()->usage(containerId);
AWAIT_READY(usage);
ASSERT_TRUE(usage->has_disk_limit_bytes());
EXPECT_EQ(Megabytes(3), Bytes(usage->disk_limit_bytes()));
if (usage->has_disk_used_bytes()) {
if (usage->disk_used_bytes() >= Megabytes(4).bytes()) {
break;
}
}
// The stopping condition for this test is that the isolator is
// able to report that we wrote the full amount of data without
// being constrained by the task disk limit.
EXPECT_LE(usage->disk_used_bytes(), Megabytes(4).bytes());
ASSERT_FALSE(timeout.expired())
<< "Used " << Bytes(usage->disk_used_bytes())
<< " of expected " << Megabytes(4)
<< " within the " << diskTimeout << " timeout";
os::sleep(Milliseconds(100));
}
driver.stop();
driver.join();
}
// Verify that we can get accurate resource statistics from the XFS
// disk isolator.
TEST_F(ROOT_XFS_QuotaTest, ResourceStatistics)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
Fetcher fetcher(flags);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Try<Owned<cluster::Slave>> slave =
StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
// Create a task that uses 4 of 3MB disk but doesn't fail. We will verify
// that the allocated disk is filled.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:3").get(),
"dd if=/dev/zero of=file bs=1048576 count=4 || sleep 1000");
Future<TaskStatus> startingStatus;
Future<TaskStatus> runningStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&startingStatus))
.WillOnce(FutureArg<1>(&runningStatus))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), startingStatus->task_id());
EXPECT_EQ(TASK_STARTING, startingStatus->state());
AWAIT_READY(runningStatus);
EXPECT_EQ(task.task_id(), runningStatus->task_id());
EXPECT_EQ(TASK_RUNNING, runningStatus->state());
Future<hashset<ContainerID>> containers = containerizer.get()->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *(containers->begin());
Timeout timeout = Timeout::in(Seconds(5));
while (true) {
Future<ResourceStatistics> usage = containerizer.get()->usage(containerId);
AWAIT_READY(usage);
ASSERT_TRUE(usage->has_disk_limit_bytes());
EXPECT_EQ(Megabytes(3), Bytes(usage->disk_limit_bytes()));
if (usage->has_disk_used_bytes()) {
// Usage must always be <= the limit.
EXPECT_LE(usage->disk_used_bytes(), usage->disk_limit_bytes());
// Usage might not be equal to the limit, but it must hit
// and not exceed the limit.
if (usage->disk_used_bytes() >= usage->disk_limit_bytes()) {
EXPECT_EQ(
usage->disk_used_bytes(), usage->disk_limit_bytes());
EXPECT_EQ(Megabytes(3), Bytes(usage->disk_used_bytes()));
break;
}
}
ASSERT_FALSE(timeout.expired());
os::sleep(Milliseconds(100));
}
driver.stop();
driver.join();
}
// This test verifies that persistent volumes are unmounted properly
// after a checkpointed framework disappears and the slave restarts.
//
// TODO(jieyu): Even though the command task specifies a new
// filesystem root, the executor (command executor) itself does not
// change filesystem root (uses the host filesystem). We need to add a
// test to test the scenario that the executor itself changes rootfs.
TEST_F(LinuxFilesystemIsolatorMesosTest,
ROOT_RecoverOrphanedPersistentVolume)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
string registry = path::join(sandbox.get(), "registry");
AWAIT_READY(DockerArchive::create(registry, "test_image"));
slave::Flags flags = CreateSlaveFlags();
flags.resources = "cpus:2;mem:1024;disk(role1):1024";
flags.isolation = "filesystem/linux,docker/runtime";
flags.docker_registry = registry;
flags.docker_store_dir = path::join(sandbox.get(), "store");
flags.image_providers = "docker";
Fetcher fetcher(flags);
Try<MesosContainerizer*> create =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(create);
Owned<Containerizer> containerizer(create.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(
detector.get(),
containerizer.get(),
flags);
ASSERT_SOME(slave);
MockScheduler sched;
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, "role1");
frameworkInfo.set_checkpoint(true);
MesosSchedulerDriver driver(
&sched,
frameworkInfo,
master.get()->pid,
DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
string dir1 = path::join(sandbox.get(), "dir1");
ASSERT_SOME(os::mkdir(dir1));
Resource persistentVolume = createPersistentVolume(
Megabytes(64),
"role1",
"id1",
"path1",
None(),
None(),
frameworkInfo.principal());
// Create a task that does nothing for a long time.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:512").get() + persistentVolume,
"sleep 1000");
task.mutable_container()->CopyFrom(createContainerInfo(
"test_image",
{createVolumeHostPath("/tmp", dir1, Volume::RW)}));
Future<TaskStatus> statusStarting;
Future<TaskStatus> statusRunning;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusStarting))
.WillOnce(FutureArg<1>(&statusRunning))
.WillRepeatedly(DoDefault());
Future<Nothing> ack =
FUTURE_DISPATCH(_, &Slave::_statusUpdateAcknowledgement);
// Create the persistent volumes and launch task via `acceptOffers`.
driver.acceptOffers(
{offer.id()},
{CREATE(persistentVolume), LAUNCH({task})});
AWAIT_READY(statusStarting);
EXPECT_EQ(TASK_STARTING, statusStarting->state());
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
// Wait for the ACK to be checkpointed.
AWAIT_READY(ack);
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *containers->begin();
// Restart the slave.
slave.get()->terminate();
// Wipe the slave meta directory so that the slave will treat the
// above running task as an orphan.
ASSERT_SOME(os::rmdir(slave::paths::getMetaRootDir(flags.work_dir)));
Future<Nothing> _recover = FUTURE_DISPATCH(_, &Slave::_recover);
// Recreate the containerizer using the same helper as above.
containerizer.reset();
create = MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(create);
containerizer.reset(create.get());
slave = StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
// Wait until slave recovery is complete.
AWAIT_READY(_recover);
// Wait until the orphan containers are cleaned up.
AWAIT_READY(containerizer->wait(containerId));
Try<fs::MountInfoTable> table = fs::MountInfoTable::read();
ASSERT_SOME(table);
// All mount targets should be under this directory.
string directory = slave::paths::getSandboxRootDir(flags.work_dir);
// Verify that the orphaned container's persistent volume and
// the rootfs are unmounted.
foreach (const fs::MountInfoTable::Entry& entry, table->entries) {
EXPECT_FALSE(strings::contains(entry.target, directory))
<< "Target was not unmounted: " << entry.target;
}
driver.stop();
driver.join();
}
TEST_P(MemoryIsolatorTest, ROOT_MemUsage)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = GetParam();
Fetcher fetcher(flags);
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(
detector.get(),
containerizer.get());
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched,
DEFAULT_FRAMEWORK_INFO,
master.get()->pid,
DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
TaskInfo task = createTask(offers.get()[0], "sleep 120");
Future<TaskStatus> statusRunning;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusRunning));
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *(containers->begin());
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
// TODO(jieyu): Consider using a program that predictably increases
// RSS so that we can set more meaningful expectation here.
EXPECT_LT(0u, usage->mem_rss_bytes());
driver.stop();
driver.join();
}