@Override
public void launchTask(
ByteBuffer message,
TFullTaskId taskId,
TUserGroupInfo user,
TResourceVector estimatedResources)
throws TException {
LOG.info("Submitting task " + taskId.getTaskId() + "at " + System.currentTimeMillis());
// We want to add accounting for task start here, even though the task is actually
// queued. Note that this won't be propagated to the node monitor until another task
// finishes.
synchronized (resourceUsage) {
TResources.addTo(resourceUsage, estimatedResources);
}
// Note we ignore user here
executor.submit(new TaskRunnable(taskId.requestId, taskId, message, estimatedResources));
}
@Override
public void run() {
if (startTime == -1) {
startTime = System.currentTimeMillis();
}
long taskStart = System.currentTimeMillis();
NodeMonitorService.Client client = null;
try {
client = TClients.createBlockingNmClient(NM_HOST, NM_PORT);
} catch (IOException e) {
LOG.fatal("Error creating NM client", e);
}
int tasks = numTasks.addAndGet(1);
double taskRate = ((double) tasks) * 1000 / (System.currentTimeMillis() - startTime);
LOG.debug("Aggregate task rate: " + taskRate);
Random r = new Random();
long benchmarkStart = System.currentTimeMillis();
runBenchmark(benchmarkId, benchmarkIterations, r);
LOG.debug("Benchmark runtime: " + (System.currentTimeMillis() - benchmarkStart));
// Update bookkeeping for task finish
synchronized (resourceUsage) {
TResources.subtractFrom(resourceUsage, taskResources);
}
HashMap<TUserGroupInfo, TResourceVector> out = new HashMap<TUserGroupInfo, TResourceVector>();
// Inform NM of resource usage
out.put(user, resourceUsage);
try {
client.tasksFinished(Lists.newArrayList(taskId));
} catch (TException e) {
e.printStackTrace();
}
client.getInputProtocol().getTransport().close();
client.getOutputProtocol().getTransport().close();
LOG.debug("Task running for " + (System.currentTimeMillis() - taskStart) + " ms");
}
/**
* A prototype Sparrow backend.
*
* <p>This backend is capable of performing a number of benchmark tasks, each representing distinct
* resource consumption profiles. It initiates a thrift server with a bounded size thread pool (of
* at most {@code WORKER_THREADS} threads). To makes sure that we never queue tasks, we additionally
* spawn a new thread each time a task is launched. In the future, we will have launchTask()
* directly execute the task and rely on queuing in the underlying thread pool to queue if task
* launches exceed capacity.
*/
public class ProtoBackend implements BackendService.Iface {
/** Benchmark which, on each iteration, runs 1 million random floating point multiplications. */
public static int BENCHMARK_TYPE_FP_CPU = 1;
/**
* Benchmark which allocates a heap buffer of 200 million bytes, then on each iteration accesses 1
* million contiguous bytes of the buffer, starting at a random offset.
*/
public static int BENCHMARK_TYPE_RANDOM_MEMACCESS = 2;
// NOTE: we do not use an enum for the above because it is not possible to serialize
// an enum with our current simple serialization technique.
/**
* Tracks the total number of tasks launched since execution began. Updated on each task launch.
* This is helpful for diagnosing unwanted queuing in various parts of the system (i.e. if we
* notice the backend is launching fewer tasks than we expect based on the frontend task launch
* rate).
*/
public static AtomicInteger numTasks = new AtomicInteger(0);
public static long startTime = -1;
private static final int DEFAULT_LISTEN_PORT = 20101;
/**
* This indicates how many threads can concurrently be answering function calls from the node
* monitor. Each node monitor client gets a dedicated thread, so this should be no less than the
* expected number of clients. Each task is launched in a new thread..
*/
private static final int THRIFT_WORKER_THREADS = 4;
private static final int TASK_WORKER_THREADS = 4;
private static final String APP_ID = "testApp";
/** We assume we are speaking to local Node Manager. */
private static final String NM_HOST = "localhost";
private static int NM_PORT;
private static Client client;
private static final Logger LOG = Logger.getLogger(ProtoBackend.class);
private static final ExecutorService executor = Executors.newFixedThreadPool(TASK_WORKER_THREADS);
/**
* Thread spawned for each task. It runs for a given amount of time (and adds its resources to the
* total resources for that time) then stops. It updates the NodeMonitor when it launches and
* again when it finishes.
*/
private class TaskRunnable implements Runnable {
private int benchmarkId;
private int benchmarkIterations;
private TResourceVector taskResources;
private TFullTaskId taskId;
public TaskRunnable(
String requestId, TFullTaskId taskId, ByteBuffer message, TResourceVector taskResources) {
this.benchmarkId = message.getInt();
this.benchmarkIterations = message.getInt();
this.taskResources = taskResources;
this.taskId = taskId;
}
@Override
public void run() {
if (startTime == -1) {
startTime = System.currentTimeMillis();
}
long taskStart = System.currentTimeMillis();
NodeMonitorService.Client client = null;
try {
client = TClients.createBlockingNmClient(NM_HOST, NM_PORT);
} catch (IOException e) {
LOG.fatal("Error creating NM client", e);
}
int tasks = numTasks.addAndGet(1);
double taskRate = ((double) tasks) * 1000 / (System.currentTimeMillis() - startTime);
LOG.debug("Aggregate task rate: " + taskRate);
Random r = new Random();
long benchmarkStart = System.currentTimeMillis();
runBenchmark(benchmarkId, benchmarkIterations, r);
LOG.debug("Benchmark runtime: " + (System.currentTimeMillis() - benchmarkStart));
// Update bookkeeping for task finish
synchronized (resourceUsage) {
TResources.subtractFrom(resourceUsage, taskResources);
}
HashMap<TUserGroupInfo, TResourceVector> out = new HashMap<TUserGroupInfo, TResourceVector>();
// Inform NM of resource usage
out.put(user, resourceUsage);
try {
client.tasksFinished(Lists.newArrayList(taskId));
} catch (TException e) {
e.printStackTrace();
}
client.getInputProtocol().getTransport().close();
client.getOutputProtocol().getTransport().close();
LOG.debug("Task running for " + (System.currentTimeMillis() - taskStart) + " ms");
}
}
/**
* Run the benchmark identified by {@code benchmarkId} for {@code iterations} iterations using
* random generator {@code r}. Return true if benchmark is recognized and false otherwise.
*/
public static boolean runBenchmark(int benchmarkId, int iterations, Random r) {
if (benchmarkId == BENCHMARK_TYPE_RANDOM_MEMACCESS) {
LOG.debug("Running random access benchmark for " + iterations + " iterations.");
runRandomMemAcessBenchmark(iterations, r);
} else if (benchmarkId == BENCHMARK_TYPE_FP_CPU) {
LOG.debug("Running CPU benchmark for " + iterations + " iterations.");
runFloatingPointBenchmark(iterations, r);
} else {
LOG.error("Received unrecognized benchmark type");
return false;
}
return true;
}
/**
* Benchmark that runs random floating point multiplications for the specified amount of
* "iterations", where each iteration is one millisecond.
*/
public static void runFloatingPointBenchmark(int iterations, Random r) {
int runtimeMillis = iterations;
long startTime = System.nanoTime();
int opsPerIteration = 1000;
/* We keep a running result here and print it out so that the JVM doesn't
* optimize all this computation away. */
float result = r.nextFloat();
while (System.nanoTime() - startTime < runtimeMillis * 1000 * 1000) {
for (int j = 0; j < opsPerIteration; j++) {
// On each iteration, perform a floating point multiplication
float x = r.nextFloat();
float y = r.nextFloat();
result += (x * y);
}
}
LOG.debug("Benchmark result " + result);
}
/**
* Benchmark which allocates a heap buffer of 200 million bytes, then on each iteration accesses 1
* million contiguous bytes of the buffer, starting at a random offset.
*/
public static void runRandomMemAcessBenchmark(int iterations, Random r) {
// 2 hundred million byte buffer
int buffSize = 1000 * 1000 * 200;
byte[] buff = new byte[buffSize];
// scan 1 million bytes at a time
int runLength = 1000 * 1000;
// We keep a running result here and print it out so that the JVM doesn't
// optimize all this computation away.
byte result = 1;
for (int i = 0; i < iterations; i++) {
// On each iteration, start at a random index, and scan runLength contiguous
// bytes, potentially wrapping if we hit the end of the buffer.
int start = r.nextInt(buff.length);
for (int j = 0; j < runLength; j++) {
result = (byte) (result ^ buff[(start + j) % (buff.length - 1)]);
}
}
LOG.debug("Benchmark result " + result);
}
private TUserGroupInfo user; // We force all tasks to be run by same user
private TResourceVector resourceUsage = TResources.createResourceVector(0, 0);
public ProtoBackend() {
LOG.debug("Created");
this.user = new TUserGroupInfo();
user.setUser("*");
user.setGroup("*");
}
@Override
public void launchTask(
ByteBuffer message,
TFullTaskId taskId,
TUserGroupInfo user,
TResourceVector estimatedResources)
throws TException {
LOG.info("Submitting task " + taskId.getTaskId() + "at " + System.currentTimeMillis());
// We want to add accounting for task start here, even though the task is actually
// queued. Note that this won't be propagated to the node monitor until another task
// finishes.
synchronized (resourceUsage) {
TResources.addTo(resourceUsage, estimatedResources);
}
// Note we ignore user here
executor.submit(new TaskRunnable(taskId.requestId, taskId, message, estimatedResources));
}
public static void main(String[] args) throws IOException, TException {
OptionParser parser = new OptionParser();
parser.accepts("c", "configuration file").withRequiredArg().ofType(String.class);
parser.accepts("help", "print help statement");
OptionSet options = parser.parse(args);
if (options.has("help")) {
parser.printHelpOn(System.out);
System.exit(-1);
}
// Logger configuration: log to the console
BasicConfigurator.configure();
LOG.setLevel(Level.DEBUG);
LOG.debug("debug logging on");
Configuration conf = new PropertiesConfiguration();
if (options.has("c")) {
String configFile = (String) options.valueOf("c");
try {
conf = new PropertiesConfiguration(configFile);
} catch (ConfigurationException e) {
}
}
// Start backend server
BackendService.Processor<BackendService.Iface> processor =
new BackendService.Processor<BackendService.Iface>(new ProtoBackend());
int listenPort = conf.getInt("listen_port", DEFAULT_LISTEN_PORT);
NM_PORT = conf.getInt("node_monitor_port", NodeMonitorThrift.DEFAULT_NM_THRIFT_PORT);
TServers.launchThreadedThriftServer(listenPort, THRIFT_WORKER_THREADS, processor);
// Register server
client = TClients.createBlockingNmClient(NM_HOST, NM_PORT);
try {
client.registerBackend(APP_ID, "localhost:" + listenPort);
LOG.debug("Client successfullly registered");
} catch (TTransportException e) {
LOG.debug("Error while registering backend: " + e.getMessage());
}
}
}
