@Override
public void addWorker(String worker) {
heartbeat.put(worker, System.currentTimeMillis());
if (!workers.contains(worker)) {
log.info("Adding worker " + worker);
workers.add(worker);
log.info("Number of workers is now " + workers.size());
}
}
@Override
public boolean addJobToCurrent(Job j) throws Exception {
IAtomicReference<Job> r = h.getAtomicReference("job-" + j.getWorkerId());
if (r.get() != null || !r.isNull()) {
boolean sent = false;
while (!sent) {
// always update
for (String s : workers()) {
if (jobFor(s) == null) {
log.info(
"Redirecting worker "
+ j.getWorkerId()
+ " to "
+ s
+ " due to work already being allocated");
r = h.getAtomicReference("job-" + s);
j.setWorkerId(s);
sent = true;
}
}
}
}
r.set(j);
// iterate over jobs without the work/data
j.setWork(null);
jobs.add(j);
return true;
}
/**
* Adds an update to the current mini batch
*
* @param id the id of the worker who did the update
* @param update the update to add
*/
@Override
public void addUpdate(String id, E update) {
try {
updateSaver().save(id, update);
} catch (Exception e) {
throw new RuntimeException(e);
}
updates.add(id);
}
@Override
public void removeWorker(String worker) {
workers.remove(worker);
if (jobFor(worker) != null) {
try {
clearJob(worker);
} catch (Exception e) {
log.warn("Unable to clear job for worker with id" + worker);
}
}
}
@Override
public void clearJob(String id) throws Exception {
if (id == null) {
log.warn("No job to clear; was null, returning");
return;
}
IAtomicReference<Job> jRef = h.getAtomicReference("job-" + id);
if (jRef.isNull()) return;
jRef.clear();
log.info("Destroyed job ref " + id);
Job remove = null;
for (Job j : jobs) {
if (j.getWorkerId().equals(id)) {
remove = j;
break;
}
}
jobs.remove(remove);
}
@Override
public int numWorkers() {
int num = workers.size();
return num;
}
@Override
public void availableForWork(String id) {
if (!workers.contains(id)) workers.add(id);
}
@Override
public void addTopic(String topic) throws Exception {
topics.add(topic);
}
public BaseHazelCastStateTracker(String connectionString, String type, int stateTrackerPort)
throws Exception {
log.info(
"Setting up hazelcast with type "
+ type
+ " connection string "
+ connectionString
+ " and port "
+ stateTrackerPort);
if (type.equals("master") && !PortTaken.portTaken(stateTrackerPort)) {
// sets up a proper connection string for reference wrt external actors needing a reference
if (connectionString.equals("master")) {
String host = InetAddress.getLocalHost().getHostName();
this.connectionString = host + ":" + stateTrackerPort;
}
this.hazelCastPort = stateTrackerPort;
config = hazelcast();
h = Hazelcast.newHazelcastInstance(config);
h.getCluster()
.addMembershipListener(
new MembershipListener() {
@Override
public void memberAdded(MembershipEvent membershipEvent) {
log.info("Member added " + membershipEvent.toString());
}
@Override
public void memberRemoved(MembershipEvent membershipEvent) {
log.info("Member removed " + membershipEvent.toString());
}
@Override
public void memberAttributeChanged(MemberAttributeEvent memberAttributeEvent) {
log.info("Member changed " + memberAttributeEvent.toString());
}
});
} else if (type.equals("master") && PortTaken.portTaken(stateTrackerPort))
throw new IllegalStateException(
"Specified type was master and the port specified was taken, please specify a different port");
else {
setConnectionString(connectionString);
log.info("Connecting to hazelcast on " + connectionString);
ClientConfig client = new ClientConfig();
client.getNetworkConfig().addAddress(connectionString);
h = HazelcastClient.newHazelcastClient(client);
}
this.type = type;
jobs = h.getList(JOBS);
workers = h.getList(WORKERS);
// we can make the assumption workers isn't empty because
// the master node by default comes with a applyTransformToDestination of workers
if (!this.type.equals("master")) {
while (workers.isEmpty()) {
log.warn("Waiting for data sync...");
Thread.sleep(1000);
}
log.info("Workers is " + workers.size());
}
begunTraining = h.getAtomicReference(BEGUN);
miniBatchSize = h.getAtomicReference(INPUT_SPLIT);
workerEnabled = h.getMap(WORKER_ENABLED);
replicate = h.getList(REPLICATE_WEIGHTS);
topics = h.getList(TOPICS);
updates = h.getList(UPDATES);
heartbeat = h.getMap(HEART_BEAT);
master = h.getAtomicReference(RESULT);
isPretrain = h.getAtomicReference(IS_PRETRAIN);
numTimesPretrain = h.getAtomicReference(NUM_TIMES_RUN_PRETRAIN);
numTimesPretrainRan = h.getAtomicReference(NUM_TIMES_PRETRAIN_RAN);
done = h.getAtomicReference(DONE);
validationEpochs = h.getAtomicReference(VALIDATION_EPOCHS);
improvementThreshold = h.getAtomicReference(IMPROVEMENT_THRESHOLD);
bestLoss = h.getAtomicReference(BEST_LOSS);
earlyStop = h.getAtomicReference(EARLY_STOP);
patience = h.getAtomicReference(PATIENCE);
patienceIncrease = h.getAtomicReference(PATIENCE_INCREASE);
numBatches = h.getAtomicReference(NUM_BATCHES_SO_FAR_RAN);
// applyTransformToDestination defaults only when master, otherwise, overrides previous values
if (type.equals("master")) {
begunTraining.set(false);
saver = createUpdateSaver();
numTimesPretrainRan.set(0);
numTimesPretrain.set(1);
isPretrain.set(true);
done.set(false);
resource = new StateTrackerDropWizardResource(this);
bestLoss.set(Double.POSITIVE_INFINITY);
earlyStop.set(true);
patience.set(40.0);
patienceIncrease.set(2.0);
improvementThreshold.set(0.995);
validationEpochs.set((int) Math.min(10, patience() / 2));
numBatches.set(0);
}
}
/**
* Tracks worker ids that need state replication
*
* @param workerId the worker id to replicate
* @return the list of worker ids that need state replication
*/
@Override
public boolean needsReplicate(String workerId) {
return replicate.contains(workerId);
}
/**
* Adds a worker to the list to be replicate d
*
* @param workerId the worker id to add
*/
@Override
public void addReplicate(String workerId) {
if (!replicate.contains(workerId)) replicate.add(workerId);
}
/**
* Updates the status of the worker to not needing replication
*
* @param workerId the worker id to update
*/
@Override
public void doneReplicating(String workerId) {
replicate.remove(workerId);
}