I'm looking for some specific information regarding the chain of events when running a MapReduce job on a Hadoop cluster.
Let's assume that my Reduce tasks are on the verge of completion. After my last reducer has written its output to the output file, how many replicas of the output file are there?
What exactly happens after the last reducer has finished writing to the output file. When does the NameNode request the respective Data Nodes to replicate the output file? And how is the Name Node informed that the output file is ready? Who conveys that information to the NameNode?
Thank you!
The Reduce tasks write output to HDFS. They do this by first communicating with the name node to request a block. The name node then tells the reducer which data nodes to write to, and then the reducer actually sends the data directly to the first data node, which then sends it to the second data node, which sends it to the third node. Typically the name node will keep things local, so the first data node is probably the same machine that is running the reduce task.
Once the reducer has finished writing outputs, and the data nodes have confirmed this, the reducer itself will tell the job tracker that it has finished via periodic heartbeat communication.
To understand the basics of HDFS replication, have a read over replica placement in the HDFS architecture document. In a nutshell, the NameNode will try to use the same rack to minimize latency.
Related
1. When assigning a task to Task Tracker for processing, the Job Tracker first tries to locate a Task Tracker with a free slots on the same server that has the data node containing the data (to ensure the data locality)
2. If it does not find this Task Tracker, it looks for a Task Tracker on another node in the same rack before it goes across the racks to locate a Task Tracker.
Thumb rule: Processing logic only will reach to data for processing.
Assuming that the Task tracker started on across the racks where corresponding processing data not available, So in this scenario, how processing logic (program) reaches to data, instead of data reaches to processing logic (program)?
When the data are not available locally, they need to be transferred through the network. Data locality is not a rule (remote nodes cannot run the program), but a goal (always prefer the local nodes, containing the data, to run the process related with this data chunk), since transferring the data (many GBs) is more costly than traferring the code (a few KBs).
I tried to understand map reduce anatomy from various books/blogs.. But I am not getting a clear idea.
What happens when I submit a job to the cluster using this command:
..Loaded the files into hdfs already
bin/hadoop jar /usr/joe/wordcount.jar org.myorg.WordCount /usr/joe/wordcount/input /usr/joe/wordcount/output
Can anyone explain the sequence of opreations that happens right from the client and inside the cluster?
The processes goes like this :
1- The client configures and sets up the job via Job and submits it to the JobTracker.
2- Once the job has been submitted the JobTracker assigns a job ID to this job.
3- Then the output specification of the job is verified. For example, if the output directory has not been specified or it already exists, the job is not submitted and an error is thrown to the MapReduce program.
4- Once this is done, InputSplits for the job are created(based on the InputFormat you are using). If the splits cannot be computed, because the input paths don’t exist, for example, then the job is not submitted and an error is thrown to the MapReduce program.
5- Based on the number of InputSplits, map tasks are created and each InputSplits gets processed by one map task.
6- Then the resources which are required to run the job are copied across the cluster like the the job JAR file, the configuration file etc. The job JAR is copied with a high replication factor (which defaults to 10) so that there are lots of copies across the cluster for the tasktrackers to access when they run tasks for the job.
7- Then based on the location of the data blocks, that are going to get processed, JobTracker directs TaskTrackers to run map tasks on that very same DataNode where that particular data block is present. If there are no free CPU slots on that DataNode, the data is moved to a nearby DataNode with free slots and the processes is continued without having to wait.
8- Once the map phase starts individual records(key-value pairs) from each InputSplit start getting processed by the Mapper one by one completing the entire InputSplit.
9- Once the map phase gets over, the output undergoes shuffle, sort and combine. After this the reduce phase starts giving you the final output.
Below is the pictorial representation of the entire process :
Also, I would suggest you to go through this link.
HTH
How does a node processing running the mapper knows that it has to send some key-value output to node A (running the reducer) & some to node B (running another reducer)?
Is there somewhere a reducer node list is maintained by the the JobTracker?
If yes, how does it chooses a node to run the reducer?
A Mapper doesn't really know where to send the data, it focuses on 2 things:
Writes the data to disk. Initially the map output is buffered in memory, and once it hits a certain threshold it gets flushed to disk. But right before going to disk, the data is partitioned by taking a hash of the output key which corresponds to which Reducer it will be sent to.
Once a map task is done it will notify the parent task tracker to say it's done, which will then notify the job tracker itself. So the job tracker has the complete mapping between map outputs and task trackers.
From there, when a Reducer starts, it will keep asking the job tracker for the map outputs corresponding to his partition until it has retrieved them all. Whenever a map output is available, the reduce task will start copying it, and gradually merge as it copies.
If this is still unclear, I will advise looking at the reference book on Hadoop which has a whole chapter describing this part, here is a schema extracted from it that could help you visualize what happens in the shuffle step:
The mappers do not send the data to the reducers, rather the reducers pull the data from the task trackers where successful map tasks ran.
The Job Tracker, when allocating a reducer task to a task tracker, knows where the successful map tasks ran, and can compile a list of task tracker and map attempt task results to pull.
Does it have any measurable effect on resources whether I submit a bunch of hadoop jobs from different client servers or all from the same one? I would think not since all the work is done in the cluster. Is this correct?
The only thing which is resource intensive on the client submitting to the Hadoop cluster is the calculation of the input splits. When the input data is huge or when too many jobs are submitted from the same client then because of the input split calculations, the job submission might become a bit slow.
I am not able to recall the Hadoop release or the parameter, but a configurable parameter was included to move the calculation of the input splits from the client submitting a job to the Hadoop cluster.
It really shouldn't matter where you submit your jobs from. The client itself doesn't do much, it uses RPC protocol to contact the services, and then just sits idle until the job is finished.
Also, the most important is what kind of scheduler you use to allocate resource, which is probably going to make the most significant difference and decide which resources to allocate to which job. More on job scheduling here.
I don't think you can move the input split calculation into Job Tracker in 'Classic' version. In YARN, you can move it using
"yarn.app.mapreduce.am.compute-splits-in-cluster"
I am guessing, Hadoop people didn't want to overload Job tracker with input split creation. Similar to the design decision of not assigning too much work for Namenode in HDFS.
In YARN, every job gets its own Application Master, so no worries about overloading a SPOF/bottleneck master like job tracker.
In reference to the original question, the client job would have to reach out to the namenode to get the block locations (I have see parts of code on block storage class calling data node for some meta data...not sure whether these happen during input split creation or in task tracker node) . This can become an issue if you are handling a lot of jobs on the same client node.
If you are using YARN, there would be a slight performance increase if all these communications happen inside the cluster.
Need to check how Oozie handles this issue.
Hopefully, this helps!
Arun
I have a Fully-Distributed Hadoop cluster with 4 nodes.When I submit my job to Jobtracker which decide 12 map tasks will be cool for my job,something strange happens.The 12 map tasks always running on a single node instead of running on the entire cluster.Before I ask the question ,I have already done the things below:
Try different Job
Run start-balance.sh to rebalance the cluster
But it does not work,so I hope someone can tell me why and how to fix it.
If all the blocks of input data files are in that node, the scheduler with prioritize the same node
Apparently the source data files is in one data node now. It could't be the balancer's fault. From what I can see, your hdfs must only have one replication or you are not in a Fully-Distributed Hadoop cluster.
Check how your input is being split. You may only have one input split, meaning that only one Node will be used to process the data. You can test this by adding more input files to your stem and placing them on different nodes, then checking which nodes are doing the work.
If that doesn't work, check to make sure that your cluster is configured correctly. Specifically, check that your name node has paths to your other nodes set in its slaves file, and that each slave node has your name node set in its masters file.