TYPED_TEST(CpuIsolatorTest, SystemCpuUsage)
{
Flags flags;
Try<Isolator*> isolator = TypeParam::create(flags);
CHECK_SOME(isolator);
// A PosixLauncher is sufficient even when testing a cgroups isolator.
Try<Launcher*> launcher = PosixLauncher::create(flags);
ExecutorInfo executorInfo;
executorInfo.mutable_resources()->CopyFrom(
Resources::parse("cpus:1.0").get());
ContainerID containerId;
containerId.set_value("system_cpu_usage");
AWAIT_READY(isolator.get()->prepare(containerId, executorInfo));
Try<string> dir = os::mkdtemp();
ASSERT_SOME(dir);
const string& file = path::join(dir.get(), "mesos_isolator_test_ready");
// Generating random numbers is done by the kernel and will max out a single
// core and run almost exclusively in the kernel, i.e., system time.
string command = "cat /dev/urandom > /dev/null & "
"touch " + file + "; " // Signals the command is running.
"sleep 60";
int pipes[2];
ASSERT_NE(-1, ::pipe(pipes));
lambda::function<int()> inChild = lambda::bind(&execute, command, pipes);
Try<pid_t> pid = launcher.get()->fork(containerId, inChild);
ASSERT_SOME(pid);
// Reap the forked child.
Future<Option<int> > status = process::reap(pid.get());
// Continue in the parent.
::close(pipes[0]);
// Isolate the forked child.
AWAIT_READY(isolator.get()->isolate(containerId, pid.get()));
// Now signal the child to continue.
int buf;
ASSERT_LT(0, ::write(pipes[1], &buf, sizeof(buf)));
::close(pipes[1]);
// Wait for the command to start.
while (!os::exists(file));
// Wait up to 1 second for the child process to induce 1/8 of a second of
// system cpu time.
ResourceStatistics statistics;
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage = isolator.get()->usage(containerId);
AWAIT_READY(usage);
statistics = usage.get();
// If we meet our usage expectations, we're done!
if (statistics.cpus_system_time_secs() >= 0.125) {
break;
}
os::sleep(Milliseconds(200));
waited += Milliseconds(200);
} while (waited < Seconds(1));
EXPECT_LE(0.125, statistics.cpus_system_time_secs());
// Ensure all processes are killed.
AWAIT_READY(launcher.get()->destroy(containerId));
// Make sure the child was reaped.
AWAIT_READY(status);
// Let the isolator clean up.
AWAIT_READY(isolator.get()->cleanup(containerId));
delete isolator.get();
delete launcher.get();
CHECK_SOME(os::rmdir(dir.get()));
}
// This test has been temporarily disabled due to MESOS-1257.
TEST_F(ExternalContainerizerTest, DISABLED_Launch)
{
Try<PID<Master> > master = this->StartMaster();
ASSERT_SOME(master);
Flags testFlags;
slave::Flags flags = this->CreateSlaveFlags();
flags.isolation = "external";
flags.containerizer_path =
testFlags.build_dir + "/src/examples/python/test-containerizer";
MockExternalContainerizer containerizer(flags);
Try<PID<Slave> > slave = this->StartSlave(&containerizer, flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get(), DEFAULT_CREDENTIAL);
Future<FrameworkID> frameworkId;
EXPECT_CALL(sched, registered(&driver, _, _))
.WillOnce(FutureArg<1>(&frameworkId));
Future<vector<Offer> > offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(frameworkId);
AWAIT_READY(offers);
EXPECT_NE(0u, offers.get().size());
TaskInfo task;
task.set_name("isolator_test");
task.mutable_task_id()->set_value("1");
task.mutable_slave_id()->CopyFrom(offers.get()[0].slave_id());
task.mutable_resources()->CopyFrom(offers.get()[0].resources());
Resources resources(offers.get()[0].resources());
Option<Bytes> mem = resources.mem();
ASSERT_SOME(mem);
Option<double> cpus = resources.cpus();
ASSERT_SOME(cpus);
const std::string& file = path::join(flags.work_dir, "ready");
// This task induces user/system load in a child process by
// running top in a child process for ten seconds.
task.mutable_command()->set_value(
#ifdef __APPLE__
// Use logging mode with 30,000 samples with no interval.
"top -l 30000 -s 0 2>&1 > /dev/null & "
#else
// Batch mode, with 30,000 samples with no interval.
"top -b -d 0 -n 30000 2>&1 > /dev/null & "
#endif
"touch " + file + "; " // Signals that the top command is running.
"sleep 60");
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status))
.WillRepeatedly(Return()); // Ignore rest for now.
Future<ContainerID> containerId;
EXPECT_CALL(containerizer, launch(_, _, _, _, _, _, _, _))
.WillOnce(DoAll(FutureArg<0>(&containerId),
Invoke(&containerizer,
&MockExternalContainerizer::_launch)));
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(containerId);
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status.get().state());
// Wait for the task to begin inducing cpu time.
while (!os::exists(file));
ExecutorID executorId;
executorId.set_value(task.task_id().value());
// We'll wait up to 10 seconds for the child process to induce
// 1/8 of a second of user and system cpu time in total.
// TODO(bmahler): Also induce rss memory consumption, by re-using
// the balloon framework.
ResourceStatistics statistics;
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage = containerizer.usage(containerId.get());
AWAIT_READY(usage);
statistics = usage.get();
// If we meet our usage expectations, we're done!
// NOTE: We are currently getting dummy-data from the test-
// containerizer python script matching these expectations.
// TODO(tillt): Consider working with real data.
if (statistics.cpus_user_time_secs() >= 0.120 &&
statistics.cpus_system_time_secs() >= 0.05 &&
statistics.mem_rss_bytes() >= 1024u) {
break;
}
os::sleep(Milliseconds(100));
waited += Milliseconds(100);
} while (waited < Seconds(10));
EXPECT_GE(statistics.cpus_user_time_secs(), 0.120);
EXPECT_GE(statistics.cpus_system_time_secs(), 0.05);
EXPECT_EQ(statistics.cpus_limit(), cpus.get());
EXPECT_GE(statistics.mem_rss_bytes(), 1024u);
EXPECT_EQ(statistics.mem_limit_bytes(), mem.get().bytes());
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
driver.killTask(task.task_id());
AWAIT_READY(status);
EXPECT_EQ(TASK_KILLED, status.get().state());
driver.stop();
driver.join();
this->Shutdown();
}
TYPED_TEST(CpuIsolatorTest, SystemCpuUsage)
{
slave::Flags flags;
Try<Isolator*> isolator = TypeParam::create(flags);
CHECK_SOME(isolator);
// A PosixLauncher is sufficient even when testing a cgroups isolator.
Try<Launcher*> launcher = PosixLauncher::create(flags);
ExecutorInfo executorInfo;
executorInfo.mutable_resources()->CopyFrom(
Resources::parse("cpus:1.0").get());
ContainerID containerId;
containerId.set_value(UUID::random().toString());
// Use a relative temporary directory so it gets cleaned up
// automatically with the test.
Try<string> dir = os::mkdtemp(path::join(os::getcwd(), "XXXXXX"));
ASSERT_SOME(dir);
ContainerConfig containerConfig;
containerConfig.mutable_executor_info()->CopyFrom(executorInfo);
containerConfig.set_directory(dir.get());
AWAIT_READY(isolator.get()->prepare(
containerId,
containerConfig));
const string& file = path::join(dir.get(), "mesos_isolator_test_ready");
// Generating random numbers is done by the kernel and will max out a single
// core and run almost exclusively in the kernel, i.e., system time.
string command = "cat /dev/urandom > /dev/null & "
"touch " + file + "; " // Signals the command is running.
"sleep 60";
int pipes[2];
ASSERT_NE(-1, ::pipe(pipes));
vector<string> argv(3);
argv[0] = "sh";
argv[1] = "-c";
argv[2] = command;
Try<pid_t> pid = launcher.get()->fork(
containerId,
"sh",
argv,
Subprocess::FD(STDIN_FILENO),
Subprocess::FD(STDOUT_FILENO),
Subprocess::FD(STDERR_FILENO),
None(),
None(),
lambda::bind(&childSetup, pipes),
None());
ASSERT_SOME(pid);
// Reap the forked child.
Future<Option<int> > status = process::reap(pid.get());
// Continue in the parent.
ASSERT_SOME(os::close(pipes[0]));
// Isolate the forked child.
AWAIT_READY(isolator.get()->isolate(containerId, pid.get()));
// Now signal the child to continue.
char dummy;
ASSERT_LT(0, ::write(pipes[1], &dummy, sizeof(dummy)));
ASSERT_SOME(os::close(pipes[1]));
// Wait for the command to start.
while (!os::exists(file));
// Wait up to 1 second for the child process to induce 1/8 of a second of
// system cpu time.
ResourceStatistics statistics;
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage = isolator.get()->usage(containerId);
AWAIT_READY(usage);
statistics = usage.get();
// If we meet our usage expectations, we're done!
if (statistics.cpus_system_time_secs() >= 0.125) {
break;
}
os::sleep(Milliseconds(200));
waited += Milliseconds(200);
} while (waited < Seconds(1));
EXPECT_LE(0.125, statistics.cpus_system_time_secs());
// Ensure all processes are killed.
AWAIT_READY(launcher.get()->destroy(containerId));
// Make sure the child was reaped.
AWAIT_READY(status);
// Let the isolator clean up.
AWAIT_READY(isolator.get()->cleanup(containerId));
delete isolator.get();
delete launcher.get();
}
TEST_P(CpuIsolatorTest, ROOT_SystemCpuUsage)
{
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());
// Generating random numbers is done by the kernel and will max out
// a single core and run almost exclusively in the kernel, i.e.,
// system time.
TaskInfo task = createTask(
offers.get()[0],
"cat /dev/urandom > /dev/null & 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_system_time_secs() >= 0.125) {
break;
}
os::sleep(Milliseconds(200));
waited += Milliseconds(200);
} while (waited < Seconds(1));
EXPECT_LE(0.125, statistics.cpus_system_time_secs());
driver.stop();
driver.join();
}
TYPED_TEST(IsolatorTest, Usage)
{
Try<PID<Master> > master = this->StartMaster();
ASSERT_SOME(master);
TypeParam isolator;
slave::Flags flags = this->CreateSlaveFlags();
Try<PID<Slave> > slave = this->StartSlave(&isolator, flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get(), DEFAULT_CREDENTIAL);
Future<FrameworkID> frameworkId;
EXPECT_CALL(sched, registered(&driver, _, _))
.WillOnce(FutureArg<1>(&frameworkId));
Future<vector<Offer> > offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(frameworkId);
AWAIT_READY(offers);
EXPECT_NE(0u, offers.get().size());
TaskInfo task;
task.set_name("isolator_test");
task.mutable_task_id()->set_value("1");
task.mutable_slave_id()->MergeFrom(offers.get()[0].slave_id());
task.mutable_resources()->MergeFrom(offers.get()[0].resources());
Resources resources(offers.get()[0].resources());
Option<Bytes> mem = resources.mem();
ASSERT_SOME(mem);
Option<double> cpus = resources.cpus();
ASSERT_SOME(cpus);
const std::string& file = path::join(flags.work_dir, "ready");
// This task induces user/system load in a child process by
// running top in a child process for ten seconds.
task.mutable_command()->set_value(
#ifdef __APPLE__
// Use logging mode with 30,000 samples with no interval.
"top -l 30000 -s 0 2>&1 > /dev/null & "
#else
// Batch mode, with 30,000 samples with no interval.
"top -b -d 0 -n 30000 2>&1 > /dev/null & "
#endif
"touch " + file + "; " // Signals that the top command is running.
"sleep 60");
vector<TaskInfo> tasks;
tasks.push_back(task);
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
driver.launchTasks(offers.get()[0].id(), tasks);
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status.get().state());
// Wait for the task to begin inducing cpu time.
while (!os::exists(file));
ExecutorID executorId;
executorId.set_value(task.task_id().value());
// We'll wait up to 10 seconds for the child process to induce
// 1/8 of a second of user and system cpu time in total.
// TODO(bmahler): Also induce rss memory consumption, by re-using
// the balloon framework.
ResourceStatistics statistics;
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage =
process::dispatch(
(Isolator*) &isolator, // TODO(benh): Fix after reaper changes.
&Isolator::usage,
frameworkId.get(),
executorId);
AWAIT_READY(usage);
statistics = usage.get();
// If we meet our usage expectations, we're done!
if (statistics.cpus_user_time_secs() >= 0.125 &&
statistics.cpus_system_time_secs() >= 0.125 &&
statistics.mem_rss_bytes() >= 1024u) {
break;
}
os::sleep(Milliseconds(100));
waited += Milliseconds(100);
} while (waited < Seconds(10));
EXPECT_GE(statistics.cpus_user_time_secs(), 0.125);
EXPECT_GE(statistics.cpus_system_time_secs(), 0.125);
EXPECT_EQ(statistics.cpus_limit(), cpus.get());
EXPECT_GE(statistics.mem_rss_bytes(), 1024u);
EXPECT_EQ(statistics.mem_limit_bytes(), mem.get().bytes());
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
driver.killTask(task.task_id());
AWAIT_READY(status);
EXPECT_EQ(TASK_KILLED, status.get().state());
driver.stop();
driver.join();
this->Shutdown(); // Must shutdown before 'isolator' gets deallocated.
}
// TODO(bmahler): Add additional tests:
// 1. Check that the data has been published to statistics.
// 2. Check that metering is occurring on subsequent resource data.
TEST(MonitorTest, WatchUnwatch)
{
FrameworkID frameworkId;
frameworkId.set_value("framework");
ExecutorID executorId;
executorId.set_value("executor");
ExecutorInfo executorInfo;
executorInfo.mutable_executor_id()->CopyFrom(executorId);
executorInfo.mutable_framework_id()->CopyFrom(frameworkId);
executorInfo.set_name("name");
executorInfo.set_source("source");
ResourceStatistics initialStatistics;
initialStatistics.set_cpus_user_time_secs(0);
initialStatistics.set_cpus_system_time_secs(0);
initialStatistics.set_cpus_limit(2.5);
initialStatistics.set_mem_rss_bytes(0);
initialStatistics.set_mem_limit_bytes(2048);
initialStatistics.set_timestamp(Clock::now().secs());
ResourceStatistics statistics;
statistics.set_cpus_nr_periods(100);
statistics.set_cpus_nr_throttled(2);
statistics.set_cpus_user_time_secs(4);
statistics.set_cpus_system_time_secs(1);
statistics.set_cpus_throttled_time_secs(0.5);
statistics.set_cpus_limit(2.5);
statistics.set_mem_rss_bytes(1024);
statistics.set_mem_limit_bytes(2048);
statistics.set_timestamp(
initialStatistics.timestamp() +
slave::RESOURCE_MONITORING_INTERVAL.secs());
TestingIsolator isolator;
process::spawn(isolator);
Future<Nothing> usage1, usage2;
EXPECT_CALL(isolator, usage(frameworkId, executorId))
.WillOnce(DoAll(FutureSatisfy(&usage1),
Return(initialStatistics)))
.WillOnce(DoAll(FutureSatisfy(&usage2),
Return(statistics)));
slave::ResourceMonitor monitor(&isolator);
// We pause the clock first in order to make sure that we can
// advance time below to force the 'delay' in
// ResourceMonitorProcess::watch to execute.
process::Clock::pause();
monitor.watch(
frameworkId,
executorId,
executorInfo,
slave::RESOURCE_MONITORING_INTERVAL);
// Now wait for ResouorceMonitorProcess::watch to finish so we can
// advance time to cause collection to begin.
process::Clock::settle();
process::Clock::advance(slave::RESOURCE_MONITORING_INTERVAL);
process::Clock::settle();
AWAIT_READY(usage1);
// Wait until the isolator has finished returning the statistics.
process::Clock::settle();
// The second collection will populate the cpus_usage.
process::Clock::advance(slave::RESOURCE_MONITORING_INTERVAL);
process::Clock::settle();
AWAIT_READY(usage2);
// Wait until the isolator has finished returning the statistics.
process::Clock::settle();
process::UPID upid("monitor", process::ip(), process::port());
Future<Response> response = process::http::get(upid, "usage.json");
AWAIT_EXPECT_RESPONSE_STATUS_EQ(OK().status, response);
AWAIT_EXPECT_RESPONSE_HEADER_EQ(
"application/json",
"Content-Type",
response);
// TODO(bmahler): Verify metering directly through statistics.
AWAIT_EXPECT_RESPONSE_BODY_EQ(
strings::format(
"[{"
"\"executor_id\":\"executor\","
"\"executor_name\":\"name\","
"\"framework_id\":\"framework\","
"\"resource_usage\":{"
"\"cpu_time\":%g,"
"\"cpu_usage\":%g,"
"\"memory_rss\":%lu"
"},"
"\"source\":\"source\""
"}]",
statistics.cpus_system_time_secs() + statistics.cpus_user_time_secs(),
(statistics.cpus_system_time_secs() +
statistics.cpus_user_time_secs()) /
slave::RESOURCE_MONITORING_INTERVAL.secs(),
statistics.mem_rss_bytes()).get(),
response);
response = process::http::get(upid, "statistics.json");
AWAIT_EXPECT_RESPONSE_STATUS_EQ(OK().status, response);
AWAIT_EXPECT_RESPONSE_HEADER_EQ(
"application/json",
"Content-Type",
response);
// TODO(bmahler): Verify metering directly through statistics.
AWAIT_EXPECT_RESPONSE_BODY_EQ(
strings::format(
"[{"
"\"executor_id\":\"executor\","
"\"executor_name\":\"name\","
"\"framework_id\":\"framework\","
"\"source\":\"source\","
"\"statistics\":{"
"\"cpus_limit\":%g,"
"\"cpus_nr_periods\":%d,"
"\"cpus_nr_throttled\":%d,"
"\"cpus_system_time_secs\":%g,"
"\"cpus_throttled_time_secs\":%g,"
"\"cpus_user_time_secs\":%g,"
"\"mem_limit_bytes\":%lu,"
"\"mem_rss_bytes\":%lu"
"}"
"}]",
statistics.cpus_limit(),
statistics.cpus_nr_periods(),
statistics.cpus_nr_throttled(),
statistics.cpus_system_time_secs(),
statistics.cpus_throttled_time_secs(),
statistics.cpus_user_time_secs(),
statistics.mem_limit_bytes(),
statistics.mem_rss_bytes()).get(),
response);
// Ensure the monitor stops polling the isolator.
monitor.unwatch(frameworkId, executorId);
// Wait until ResourceMonitorProcess::unwatch has completed.
process::Clock::settle();
// This time, Isolator::usage should not get called.
EXPECT_CALL(isolator, usage(frameworkId, executorId))
.Times(0);
process::Clock::advance(slave::RESOURCE_MONITORING_INTERVAL);
process::Clock::settle();
response = process::http::get(upid, "usage.json");
AWAIT_EXPECT_RESPONSE_STATUS_EQ(OK().status, response);
AWAIT_EXPECT_RESPONSE_HEADER_EQ(
"application/json",
"Content-Type",
response);
AWAIT_EXPECT_RESPONSE_BODY_EQ("[]", response);
}