Many resources are needed to download a project. Please understand that we have to compensate our server costs. Thank you in advance.
Project price only 1 $
You can buy this project and download/modify it how often you want.
org.deeplearning4j.iterativereduce.runtime.yarn.appmaster.ApplicationMasterService Maven / Gradle / Ivy
/*
*
* * Copyright 2015 Skymind,Inc.
* *
* * Licensed under the Apache License, Version 2.0 (the "License");
* * you may not use this file except in compliance with the License.
* * You may obtain a copy of the License at
* *
* * http://www.apache.org/licenses/LICENSE-2.0
* *
* * Unless required by applicable law or agreed to in writing, software
* * distributed under the License is distributed on an "AS IS" BASIS,
* * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* * See the License for the specific language governing permissions and
* * limitations under the License.
*
*/
package org.deeplearning4j.iterativereduce.runtime.yarn.appmaster;
import org.apache.avro.AvroRemoteException;
import org.apache.avro.ipc.NettyServer;
import org.apache.avro.ipc.Server;
import org.apache.avro.ipc.specific.SpecificResponder;
import org.apache.hadoop.conf.Configuration;
import org.deeplearning4j.iterativereduce.runtime.ComputableMaster;
import org.deeplearning4j.scaleout.api.ir.Updateable;
import org.deeplearning4j.iterativereduce.runtime.Utils;
import org.deeplearning4j.iterativereduce.runtime.yarn.avro.generated.*;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import java.net.InetSocketAddress;
import java.nio.ByteBuffer;
import java.util.HashMap;
import java.util.LinkedHashMap;
import java.util.Map;
import java.util.concurrent.Callable;
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.atomic.AtomicLong;
/**
* @author Michael
*
* TODO: change to ConcurrentHashMap (maybe)?
*
* TODO: Have an expiration period for all workers to check in,
* otherwise we need to bail?
*
* TODO: We need to fix the overall logic, that actually waits for
* workers to check in. For example, what happens if we have 2/3 workers
* check-in, and 1 didn't for some reason, does that mean that
* everything continues? Do we let our other workers start and waste
* their time?
*
* Additionally, what implications does that have on the RPC? Do workers
* poll for a command to tell them to start, or do we want to push that
* down to them?
*
* @param
*/
public class ApplicationMasterService implements
IterativeReduceService, Callable {
private static final Logger LOG = LoggerFactory.getLogger(ApplicationMasterService.class);
private enum WorkerState {
NONE, STARTED, RUNNING, UPDATE, WAITING, COMPLETE, ERROR
}
private enum MasterState {
WAITING, UPDATING
}
private Map workers;
private Map workersState;
private Map> workersProgress;
private Map workersUpdate;
private MasterState masterState;
private int currentUpdateId = 0;
private Map masterUpdates;
private ComputableMaster computable;
private Class updateable;
private final InetSocketAddress masterAddr;
private Server masterServer;
private CountDownLatch workersCompleted;
private AtomicLong expectedUpdates;
private Configuration conf;
private Map appConf;
// Metrics
private Map> workerMetrics;
private long mMasterTime;
private long mMasterExecutions;
private long mUpdates;
public ApplicationMasterService(InetSocketAddress masterAddr,
Map workers,
ComputableMaster computable, Class updatable,
Map appConf, Configuration conf) {
if (masterAddr == null || computable == null || updatable == null)
throw new IllegalStateException(
"masterAddress or computeUpdate cannot be null");
this.workers = workers;
this.workersCompleted = new CountDownLatch(workers.size());
this.expectedUpdates = new AtomicLong(workers.size());
this.workersState = new HashMap<>();
this.workersProgress = new HashMap<>();
this.masterState = MasterState.WAITING;
this.masterUpdates = new HashMap<>();
this.masterAddr = masterAddr;
this.computable = computable;
this.updateable = updatable;
this.appConf = appConf;
this.conf = conf;
Utils.mergeConfigs(this.appConf, this.conf);
this.computable.setup(this.conf);
// Merger workers into worker state
for (WorkerId workerId : workers.keySet()) {
workersState.put(workerId, WorkerState.NONE);
}
}
public ApplicationMasterService(InetSocketAddress masterAddr,
Map workers,
ComputableMaster computable, Class updatable,
Map appConf) {
this(masterAddr, workers, computable, updatable, appConf,
new Configuration());
}
public ApplicationMasterService(InetSocketAddress masterAddr,
Map workers,
ComputableMaster computable, Class updatable) {
this(masterAddr, workers, computable, updatable, null);
}
public Integer call() {
Thread.currentThread().setName("ApplicationMasterService Thread");
LOG.info("Starting MasterService [NettyServer] on " + masterAddr);
masterServer = new NettyServer(new SpecificResponder(
IterativeReduceService.class, this), masterAddr);
try {
workersCompleted.await();
int complete = 0;
int error = 0;
int unknown = 0;
for (WorkerState state : workersState.values()) {
switch (state) {
case COMPLETE:
complete++;
break;
case ERROR:
error++;
break;
default:
unknown++;
break;
}
}
// TODO: fix, and move here from ApplicationMaster
// computable.complete(null);
LOG.info("All workers have completed. Shutting down master service"
+ ", workersComplete=" + complete + ", workersError=" + error
+ ", workersUnknown=" + unknown);
return (error == 0 && unknown == 0) ? 0 : 1;
} catch (InterruptedException ex) {
// This will occur purposely if:
// 1. Someone calls a ExecutorService.shutdownNow()
// 2. Via our stop() method using Future.cancel()
LOG.warn("Interrupted while waiting for workers to complete", ex);
return -1;
} finally {
LOG.debug("Shutting down Netty server");
masterServer.close();
printMetrics();
}
}
public void fail() {
while (workersCompleted.getCount() > 0)
workersCompleted.countDown();
}
private void printMetrics() {
StringBuffer metrics = new StringBuffer("Master metrics:\n");
metrics.append(" MasterTime: ").append(mMasterTime).append("\n");
metrics.append(" MasterExecutions: ").append(mMasterExecutions).append("\n");
metrics.append(" UpdatesReceived: " ).append(mUpdates).append("\n");
metrics.append("\n");
// Possibly to be null if no workers have supplied metrics
if (workerMetrics == null) {
metrics.append(" Worker metrics: no workers supplied metrics!\n");
} else {
for (WorkerId wid : workerMetrics.keySet()) {
metrics.append(" Worker Metrics - ").append(Utils.getWorkerId(wid)).append(":").append("\n");
for (Map.Entry entry : workerMetrics.get(wid).entrySet()) {
metrics.append(" ").append(entry.getKey()).append(": ").append(entry.getValue()).append("\n");
}
metrics.append("\n");
}
}
LOG.info(metrics.toString());
}
@Override
public StartupConfiguration startup(WorkerId workerId)
throws AvroRemoteException {
synchronized (workersState) {
if (!workersState.containsKey(workerId)) {
throw ServiceError
.newBuilder()
.setDescription(
"Worker " + Utils.getWorkerId(workerId)
+ "unknown.").build();
}
// TODO: can a worker "start" more than once?
StartupConfiguration workerConf = workers.get(workerId);
Utils.mergeConfigs(appConf, workerConf);
LOG.debug("Got a startup call, workerId="
+ Utils.getWorkerId(workerId) + ", responded with"
+ ", batchSize=" + workerConf.getBatchSize() + ", iterations="
+ workerConf.getIterations() + ", fileSplit=[" + workerConf.getSplit().getPath()
+ ", " + workerConf.getSplit().getOffset() + "]");
workersState.put(workerId, WorkerState.STARTED);
return workerConf;
}
}
private boolean handleProgress(WorkerId workerId,
ProgressReport report) {
synchronized (workersState) {
LinkedHashMap progress = workersProgress.get(workerId);
if (progress == null)
progress = new LinkedHashMap<>();
progress.put(System.currentTimeMillis(), report);
workersProgress.put(workerId, progress);
workersState.put(workerId, WorkerState.RUNNING);
LOG.debug("Got a progress report" + ", workerId="
+ Utils.getWorkerId(workerId) + ", workerState="
+ workersState.get(workerId) + ", progressSize="
+ report.getReport().size() + ", totalReports=" + progress.size());
}
return true;
}
@Override
public boolean progress(WorkerId workerId, ProgressReport report)
throws AvroRemoteException {
return handleProgress(workerId, report);
}
@Override
public boolean update(WorkerId workerId, ByteBuffer data)
throws AvroRemoteException {
// We only want updates from workers we know are either running, or have
// started
WorkerState workerState;
synchronized (workersState) {
workerState = workersState.get(workerId);
if (workerState != null) {
if (workerState != WorkerState.RUNNING && workerState != WorkerState.STARTED) {
LOG.debug("Received an erroneous update" + ", workerId="
+ Utils.getWorkerId(workerId) + ", workerState="
+ workerState + ", length=" + data.limit());
return false;
}
}
}
LOG.info("Received update, workerId=" + Utils.getWorkerId(workerId)
+ ", workerState=" + workerState + ", length=" + data.limit());
synchronized (masterState) {
// First update, create a new update map
if (MasterState.WAITING == masterState) {
LOG.debug("Initial update for this round, initializing update map");
if (workersUpdate == null)
workersUpdate = new HashMap<>();
workersUpdate.clear();
masterState = MasterState.UPDATING;
}
}
// Duplicate update?
if (workersUpdate.containsKey(workerId)) {
LOG.warn("Received a duplicate update for, workerId="
+ Utils.getWorkerId(workerId) + ", ignoring this update");
return false;
}
// Synchronized?
T update;
try {
update = updateable.newInstance();
update.fromBytes(data);
} catch (Exception ex) {
LOG.warn("Unable to instantiate a computable object", ex);
return false;
}
synchronized (workersState) {
workersUpdate.put(workerId, update);
workersState.put(workerId, WorkerState.UPDATE);
// Our latch should have the number of currently active workers
if (workersUpdate.size() == expectedUpdates.get()) {
LOG.info("Received updates from all workers, spawning local compute thread");
// Fire off thread to compute update
// TODO: need to fix this to something more reusable
Thread updateThread = new Thread(new Runnable() {
@Override
public void run() {
long startTime, endTime;
startTime = System.currentTimeMillis();
T result = computable.compute(workersUpdate.values(),
masterUpdates.values());
endTime = System.currentTimeMillis();
LOG.info("Computed local update in " + (endTime - startTime) + "ms");
expectedUpdates.set(workersCompleted.getCount());
mMasterExecutions++;
mMasterTime += (endTime - startTime);
// check master computable to see if it wants to end early
//earlyTerminationDetected = computable.checkEarlyTerminationCondition();
synchronized (masterUpdates) {
// we pre-do this so
currentUpdateId++;
// changed around the id to inc after put
masterUpdates.put(currentUpdateId, result);
LOG.info("Adding master update for " + currentUpdateId + "");
masterState = MasterState.WAITING;
}
}
});
updateThread.setName("Compute thread");
updateThread.start();
}
}
mUpdates++;
return true;
}
@Override
public int waiting(WorkerId workerId, int lastUpdate, long waiting)
throws AvroRemoteException {
synchronized (workersState) {
workersState.put(workerId, WorkerState.WAITING);
LOG.info("Got waiting message" + ", workerId="
+ Utils.getWorkerId(workerId) + ", workerState="
+ workersState.get(workerId) + ", lastUpdate=" + lastUpdate
+ ", currentUpdateId=" + currentUpdateId
+ ", waitingFor=" + waiting);
}
if (MasterState.UPDATING == masterState && lastUpdate == currentUpdateId)
return -1;
return currentUpdateId;
}
@Override
public ByteBuffer fetch(WorkerId workerId, int updateId)
throws AvroRemoteException {
LOG.info("Received a fetch request"
+ ", workerId=" + Utils.getWorkerId(workerId)
+ ", requestedUpdateId=" + updateId);
synchronized (workersState) {
workersState.put(workerId, WorkerState.RUNNING);
}
ByteBuffer bytes = masterUpdates.get(updateId).toBytes();
bytes.rewind();
return bytes;
}
@Override
public void complete(WorkerId workerId, ProgressReport finalReport) {
handleProgress(workerId, finalReport);
workersState.put(workerId, WorkerState.COMPLETE);
LOG.info("Received complete message, workerId="
+ Utils.getWorkerId(workerId));
workersCompleted.countDown();
}
@Override
public void error(WorkerId workerId, CharSequence message) {
LOG.warn("A worker encountered an error" + ", worker="
+ Utils.getWorkerId(workerId) + ", message=" + message);
workersState.put(workerId, WorkerState.ERROR);
workersCompleted.countDown();
}
@Override
public void metricsReport(WorkerId workerId, Map metrics) {
if (workerMetrics == null)
workerMetrics = new HashMap<>();
// Bludgeon it for now, TODO: be smarter about merging them
workerMetrics.put(workerId, metrics);
}
}
