hadoop job提交完成的整個過程介紹 zz

2009-11-17 11:16http://blog.chinaunix.net/u3/94300/showart_1902760.html

由於大量的使用interface，reflection，rpc proxy，所以當我們提交job給hadoop的時候，他到底是如何一步步運行的確實不太容易看明白，今天費了將近一天的功夫終於將其大概整理了出倆，爲以後繼續深入仔細閱讀源碼打下基礎。

start JobTracker and TaskTracker by bin/start-all.sh
JobTracker initialization:
read all host( master and slave )
init QueueManager to manage job queue
start JobQueueTaskScheduler to schedule job
add JobQueueJobInProcessListener to manage submited jobs
add EagerTaskInitialization, so when job was submited, JobinitThread will get first job in job queue and execute job ( throught job.initTask)
start JobTracker Server which is instance of RPC.Server, derived from org.apache.hadoop.mapred.ipc.Server, so client can submit job by creating proxy
start httpServer (third party library) so we can get status of job tracker through http
init JobTrackerInstrumentation so record measure information
start RecoverManager so we can try to recover unfinished history jobs
init CompletedJobStatusStore to persist and retrive job information
start ExpireTrackers thread to remove all expiry TaskTracker, whose heatbeat response out of time
start RetireJobs to remove finished jobs that have been around for too long ( the job status is not runing, prepare ... )
start ExpireLaunchingTasks to check task assignment failure and reassign the task
join JobTracker Server to wait for job submit request

TaskTracker initialization:
start http server to report map/reduce status by http
start TaskTrackerInstrument to record measure info
init JvmManager which can find free JVM to run task
start taskReportServer (ipc.Server) with instance as TaskTracker
create JobTracker proxy so the taskTracker can communicate with JobTracker
start mapEventsFetcher to fetch all map event
start mapLauncher and reduceLauncher ( instance of TaskLauncher) thread to monitor task queue, if new task was inserted, get first and startNewTask.
after TaskTracker initialization, start TaskTracker server which will looply communicate with JobTracker by Hearbbeat test, also get task.
check buildVersion and file system
purge map-event and reset reducer
check if the TaskTracker should be restart
get TaskAction from Heartbert message, if exist, add action to actionQueue


org.apache.hadoop.mapred.JobClient :
public static RunningJob runJob(JobConf job)

public RunningJob submitJob(JobConf job)
copy job jar to DFS
get input split
sort input split
write split info into file system
jobSubmitClient.submitJob, jobSubmitClient is client proxy of org.apache.hadoop.mapred.JobTracker, which record status of all jobs

org.apache.hadoop.mapred.JobTracker :
public JobStatus submitJob(JobID jobName)
create JobInProcess with jobID
copy job jar from DFS to local
add job to JobQueueJobInProgressListener and EagerTaskInitializationListener, EagerTaskInitializationListener will call JobInProcess.initTask to create map and reduce task based on InputSplit, and create other assistant task, such as clean-up task ect.
JobTracker call JobQueueTaskScheduler.assignTask to assign task for each request host
call JobTracker.getSetupAndCleanupTasks() to get all tasks waited for execution, call TaskInProcess.addRunningTask() to set map or reduce task
all got tasks were encapsulated in TaskActions
send TaskActions to TaskTrackers through HeartbeatReponse

TaskTracker get TaskAction from HeartbeatResponse message ( JobTracker call JobQueueTaskScheduler.assignTask to assign task for each request host)
TaskTracker.TaskInProcess.registerTask() was called so TaskAction was transformed to Tasktracker.TaskInProcess (tip)
the newly got tip was insert into task queue ( named tasksToLaunch) of TaskLauncher ( named mapLauncher and reduceLauncher), once new task was inserted:
TaskTracker.startNewTask was called
TaskTracker.localizeJob() was called
add task to job, so MapEventFetcher can get map event (further thinking)
get job jar file and create necessary working directory
unjar job jar file
TaskTracker.launchTaskForJob was called
TaskTracker.TaskInProcess.launchTask was called
TaskTracker.TaskInProces.localizeTask() was called
create local working dir
create symlink for job dir if it doesn't exist
set resolved hostname
set debug parameter if debugCommand exist
Task.createRunner (MapTask or ReduceTask) was called
TaskRunner.start()
build parameter for JVM running, JVM was selected by JvmManager which obey singleInstance model
entry of JVM is org.apache.hadoop.mapred.Child which will call, which will call Task.run()
in MapTask.run() or ReduceTask.run(), our mapper or reducer code waw call through
**************
MapRunnable runner = ReflectionUtils.newInstance(job
.getMapRunnerClass(), job);
**************
finally, we finish our map/reduce work.