/**
* We lost the task tracker! All task-tracker structures have already been updated. Just process
* the contained tasks and any jobs that might be affected.
*/
void lostTaskTracker(String trackerName) {
LOG.info("Lost tracker '" + trackerName + "'");
TreeSet lostTasks = (TreeSet) trackerToTaskMap.get(trackerName);
trackerToTaskMap.remove(trackerName);
if (lostTasks != null) {
for (Iterator it = lostTasks.iterator(); it.hasNext(); ) {
String taskId = (String) it.next();
TaskInProgress tip = (TaskInProgress) taskidToTIPMap.get(taskId);
// Tell the job to fail the relevant task
JobInProgress job = tip.getJob();
job.failedTask(tip, taskId, trackerName);
}
}
}
public static void startTracker(Configuration conf) throws IOException {
if (tracker != null) throw new IOException("JobTracker already running.");
while (true) {
try {
tracker = new JobTracker(conf);
break;
} catch (IOException e) {
LOG.log(Level.WARNING, "Starting tracker", e);
}
try {
Thread.sleep(1000);
} catch (InterruptedException e) {
}
}
tracker.offerService();
}
/** Start the JobTracker process, listen on the indicated port */
JobTracker(Configuration conf) throws IOException {
//
// Grab some static constants
//
maxCurrentTasks = conf.getInt("mapred.tasktracker.tasks.maximum", 2);
RETIRE_JOB_INTERVAL = conf.getLong("mapred.jobtracker.retirejob.interval", 24 * 60 * 60 * 1000);
RETIRE_JOB_CHECK_INTERVAL = conf.getLong("mapred.jobtracker.retirejob.check", 60 * 1000);
TASK_ALLOC_EPSILON = conf.getFloat("mapred.jobtracker.taskalloc.loadbalance.epsilon", 0.2f);
PAD_FRACTION = conf.getFloat("mapred.jobtracker.taskalloc.capacitypad", 0.1f);
MIN_SLOTS_FOR_PADDING = 3 * maxCurrentTasks;
// This is a directory of temporary submission files. We delete it
// on startup, and can delete any files that we're done with
this.conf = conf;
JobConf jobConf = new JobConf(conf);
this.systemDir = jobConf.getSystemDir();
this.fs = FileSystem.get(conf);
FileUtil.fullyDelete(fs, systemDir);
fs.mkdirs(systemDir);
// Same with 'localDir' except it's always on the local disk.
jobConf.deleteLocalFiles(SUBDIR);
// Set ports, start RPC servers, etc.
InetSocketAddress addr = getAddress(conf);
this.localMachine = addr.getHostName();
this.port = addr.getPort();
this.interTrackerServer = RPC.getServer(this, addr.getPort(), 10, false, conf);
this.interTrackerServer.start();
Properties p = System.getProperties();
for (Iterator it = p.keySet().iterator(); it.hasNext(); ) {
String key = (String) it.next();
String val = (String) p.getProperty(key);
LOG.info("Property '" + key + "' is " + val);
}
this.infoPort = conf.getInt("mapred.job.tracker.info.port", 50030);
this.infoServer = new JobTrackerInfoServer(this, infoPort);
this.infoServer.start();
this.startTime = System.currentTimeMillis();
new Thread(this.expireTrackers).start();
new Thread(this.retireJobs).start();
new Thread(this.initJobs).start();
}
/**
* Accept and process a new TaskTracker profile. We might have known about the TaskTracker
* previously, or it might be brand-new. All task-tracker structures have already been updated.
* Just process the contained tasks and any jobs that might be affected.
*/
void updateTaskStatuses(TaskTrackerStatus status) {
for (Iterator it = status.taskReports(); it.hasNext(); ) {
TaskStatus report = (TaskStatus) it.next();
TaskInProgress tip = (TaskInProgress) taskidToTIPMap.get(report.getTaskId());
if (tip == null) {
LOG.info(
"Serious problem. While updating status, cannot find taskid " + report.getTaskId());
} else {
JobInProgress job = tip.getJob();
job.updateTaskStatus(tip, report);
if (report.getRunState() == TaskStatus.SUCCEEDED) {
job.completedTask(tip, report.getTaskId());
} else if (report.getRunState() == TaskStatus.FAILED) {
// Tell the job to fail the relevant task
job.failedTask(tip, report.getTaskId(), status.getTrackerName());
}
}
}
}
///////////////////////////////////////////////////////
// Maintain lookup tables; called by JobInProgress
// and TaskInProgress
///////////////////////////////////////////////////////
void createTaskEntry(String taskid, String taskTracker, TaskInProgress tip) {
LOG.info(
"Adding task '"
+ taskid
+ "' to tip "
+ tip.getTIPId()
+ ", for tracker '"
+ taskTracker
+ "'");
// taskid --> tracker
taskidToTrackerMap.put(taskid, taskTracker);
// tracker --> taskid
TreeSet taskset = (TreeSet) trackerToTaskMap.get(taskTracker);
if (taskset == null) {
taskset = new TreeSet();
trackerToTaskMap.put(taskTracker, taskset);
}
taskset.add(taskid);
// taskid --> TIP
taskidToTIPMap.put(taskid, tip);
}
/**
* A tracker wants to know if there's a Task to run. Returns a task we'd like the TaskTracker to
* execute right now.
*
* <p>Eventually this function should compute load on the various TaskTrackers, and incorporate
* knowledge of DFS file placement. But for right now, it just grabs a single item out of the
* pending task list and hands it back.
*/
public synchronized Task pollForNewTask(String taskTracker) {
//
// Compute average map and reduce task numbers across pool
//
int avgMaps = 0;
int avgReduces = 0;
int numTaskTrackers;
TaskTrackerStatus tts;
synchronized (taskTrackers) {
numTaskTrackers = taskTrackers.size();
tts = (TaskTrackerStatus) taskTrackers.get(taskTracker);
}
if (numTaskTrackers > 0) {
avgMaps = totalMaps / numTaskTrackers;
avgReduces = totalReduces / numTaskTrackers;
}
int totalCapacity = numTaskTrackers * maxCurrentTasks;
//
// Get map + reduce counts for the current tracker.
//
if (tts == null) {
LOG.warning("Unknown task tracker polling; ignoring: " + taskTracker);
return null;
}
int numMaps = tts.countMapTasks();
int numReduces = tts.countReduceTasks();
//
// In the below steps, we allocate first a map task (if appropriate),
// and then a reduce task if appropriate. We go through all jobs
// in order of job arrival; jobs only get serviced if their
// predecessors are serviced, too.
//
//
// We hand a task to the current taskTracker if the given machine
// has a workload that's equal to or less than the averageMaps
// +/- TASK_ALLOC_EPSILON. (That epsilon is in place in case
// there is an odd machine that is failing for some reason but
// has not yet been removed from the pool, making capacity seem
// larger than it really is.)
//
synchronized (jobsByArrival) {
if ((numMaps < maxCurrentTasks) && (numMaps <= (avgMaps + TASK_ALLOC_EPSILON))) {
int totalNeededMaps = 0;
for (Iterator it = jobsByArrival.iterator(); it.hasNext(); ) {
JobInProgress job = (JobInProgress) it.next();
if (job.getStatus().getRunState() != JobStatus.RUNNING) {
continue;
}
Task t = job.obtainNewMapTask(taskTracker, tts);
if (t != null) {
return t;
}
//
// Beyond the highest-priority task, reserve a little
// room for failures and speculative executions; don't
// schedule tasks to the hilt.
//
totalNeededMaps += job.desiredMaps();
double padding = 0;
if (totalCapacity > MIN_SLOTS_FOR_PADDING) {
padding = Math.min(maxCurrentTasks, totalNeededMaps * PAD_FRACTION);
}
if (totalNeededMaps + padding >= totalCapacity) {
break;
}
}
}
//
// Same thing, but for reduce tasks
//
if ((numReduces < maxCurrentTasks) && (numReduces <= (avgReduces + TASK_ALLOC_EPSILON))) {
int totalNeededReduces = 0;
for (Iterator it = jobsByArrival.iterator(); it.hasNext(); ) {
JobInProgress job = (JobInProgress) it.next();
if (job.getStatus().getRunState() != JobStatus.RUNNING) {
continue;
}
Task t = job.obtainNewReduceTask(taskTracker, tts);
if (t != null) {
return t;
}
//
// Beyond the highest-priority task, reserve a little
// room for failures and speculative executions; don't
// schedule tasks to the hilt.
//
totalNeededReduces += job.desiredReduces();
double padding = 0;
if (totalCapacity > MIN_SLOTS_FOR_PADDING) {
padding = Math.min(maxCurrentTasks, totalNeededReduces * PAD_FRACTION);
}
if (totalNeededReduces + padding >= totalCapacity) {
break;
}
}
}
}
return null;
}