/**
* JobTracker.submitJob() kicks off a new job.
*
* <p>Create a 'JobInProgress' object, which contains both JobProfile and JobStatus. Those two
* sub-objects are sometimes shipped outside of the JobTracker. But JobInProgress adds info that's
* useful for the JobTracker alone.
*
* <p>We add the JIP to the jobInitQueue, which is processed asynchronously to handle
* split-computation and build up the right TaskTracker/Block mapping.
*/
public synchronized JobStatus submitJob(String jobFile) throws IOException {
totalSubmissions++;
JobInProgress job = new JobInProgress(jobFile, this, this.conf);
synchronized (jobs) {
synchronized (jobsByArrival) {
synchronized (jobInitQueue) {
jobs.put(job.getProfile().getJobId(), job);
jobsByArrival.add(job);
jobInitQueue.add(job);
jobInitQueue.notifyAll();
}
}
}
return job.getStatus();
}
/**
* 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);
}
}
}
/**
* The run method lives for the life of the JobTracker, and removes Jobs that are not still
* running, but which finished a long time ago.
*/
public void run() {
while (shouldRun) {
try {
Thread.sleep(RETIRE_JOB_CHECK_INTERVAL);
} catch (InterruptedException ie) {
}
synchronized (jobs) {
synchronized (jobInitQueue) {
synchronized (jobsByArrival) {
for (Iterator it = jobs.keySet().iterator(); it.hasNext(); ) {
String jobid = (String) it.next();
JobInProgress job = (JobInProgress) jobs.get(jobid);
if (job.getStatus().getRunState() != JobStatus.RUNNING
&& job.getStatus().getRunState() != JobStatus.PREP
&& (job.getFinishTime() + RETIRE_JOB_INTERVAL < System.currentTimeMillis())) {
it.remove();
jobInitQueue.remove(job);
jobsByArrival.remove(job);
}
}
}
}
}
}
}
public synchronized JobStatus getJobStatus(String jobid) {
JobInProgress job = (JobInProgress) jobs.get(jobid);
if (job != null) {
return job.getStatus();
} else {
return null;
}
}
public synchronized JobProfile getJobProfile(String jobid) {
JobInProgress job = (JobInProgress) jobs.get(jobid);
if (job != null) {
return job.getProfile();
} else {
return null;
}
}
public Vector completedJobs() {
Vector v = new Vector();
for (Iterator it = jobs.values().iterator(); it.hasNext(); ) {
JobInProgress jip = (JobInProgress) it.next();
JobStatus status = jip.getStatus();
if (status.getRunState() == JobStatus.SUCCEEDED) {
v.add(jip);
}
}
return v;
}
public Vector runningJobs() {
Vector v = new Vector();
for (Iterator it = jobs.values().iterator(); it.hasNext(); ) {
JobInProgress jip = (JobInProgress) it.next();
JobStatus status = jip.getStatus();
if (status.getRunState() == JobStatus.RUNNING) {
v.add(jip);
}
}
return v;
}
public synchronized TaskReport[] getReduceTaskReports(String jobid) {
JobInProgress job = (JobInProgress) jobs.get(jobid);
if (job == null) {
return new TaskReport[0];
} else {
Vector reports = new Vector();
Vector completeReduceTasks = job.reportTasksInProgress(false, true);
for (Iterator it = completeReduceTasks.iterator(); it.hasNext(); ) {
TaskInProgress tip = (TaskInProgress) it.next();
reports.add(tip.generateSingleReport());
}
Vector incompleteReduceTasks = job.reportTasksInProgress(false, false);
for (Iterator it = incompleteReduceTasks.iterator(); it.hasNext(); ) {
TaskInProgress tip = (TaskInProgress) it.next();
reports.add(tip.generateSingleReport());
}
return (TaskReport[]) reports.toArray(new TaskReport[reports.size()]);
}
}
/**
* 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());
}
}
}
}
public void run() {
while (shouldRun) {
JobInProgress job = null;
synchronized (jobInitQueue) {
if (jobInitQueue.size() > 0) {
job = (JobInProgress) jobInitQueue.elementAt(0);
jobInitQueue.remove(job);
} else {
try {
jobInitQueue.wait(JOBINIT_SLEEP_INTERVAL);
} catch (InterruptedException iex) {
}
}
}
try {
if (job != null) {
job.initTasks();
}
} catch (Exception e) {
LOG.log(Level.WARNING, "job init failed", e);
job.kill();
}
}
}
public synchronized void killJob(String jobid) {
JobInProgress job = (JobInProgress) jobs.get(jobid);
job.kill();
}
/**
* 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;
}