Tuesday, 28 April 2015

Introduction to Apache HDFS and MapReduce

Introduction to Apache HDFS and MapReduce

Apache Hadoop has been originated from Google’s Whitepapers:
1.    Apache HDFS is derived from GFS  (Google File System).
2.    Apache MapReduce is derived from Google MapReduce
3.    Apache HBase is derived from Google BigTable.
Though Google has only provided the Whitepapers, without any implementation, around 90-95% of the architecture presented in these Whitepapers is applied in these three Java-based Apache projects.
HDFS and MapReduce are the two major components of Hadoop, where HDFS is from the ‘Infrastructural’ point of view and MapReduce is from the ‘Programming’ aspect. Though HDFS is at present a subproject of Apache Hadoop, it was formally developed as an infrastructure for the Apache Nutch web search engine project.
To understand the magic behind the scalability of Hadoop from one-node cluster to a thousand-nodes cluster (Yahoo! has 4,500-node cluster managing 40 petabytes of enterprise data), we need to first understand Hadoop’s file system, that is, HDFS (Hadoop Distributed File System).

What is HDFS (Hadoop Distributed File System)?

HDFS is a distributed and scalable file system designed for storing very large files with streaming data access patterns, running clusters on commodity hardware.
Very large” in this context means files that are hundreds of megabytes, gigabytes,or terabytes in size. There are Hadoop clusters running today that store petabytesof data.
Though it has many similarities with existing traditional distributed file systems, there are noticeable differences between these. Let’s look into some of the assumptions and goals/objectives behind HDFS, which also form some striking features of this incredible file system!

Assumptions and Goals/Objectives behind HDFS:

1.  Large Data Sets:

It is assumed that HDFS always needs to work with large data sets. It will be an underplay if HDFS is deployed to process several small data sets ranging in some megabytes or even a few gigabytes. So The architecture of HDFS is designed in such a way that it is best fit to store and retrieve huge amount of data. What is required is high cumulative data bandwidth and the scalability feature to spread out from a single node cluster to a hundred or a thousand-node cluster. The acid test is that HDFS should be able to manage tens of millions of files in a single occurrence.

2.  Write Once, Read Many Model:

HDFS follows the write-once, read-many approach for its files and applications. It assumes that a file in HDFS once written will not be modified, though it can be access ‘n’ number of times (though future versions of Hadoop may support this feature too)! At present, in HDFS strictly has one writer at any time. This assumption enables high throughput data access and also simplifies data coherency issues. A web crawler or a MapReduce application is best suited for HDFS.

3.  Streaming Data Access:

As HDFS works on the principle of ‘Write Once, Read Many‘, the feature of streaming data access is extremely important in HDFS. 
As HDFS is designed more for batch processing rather than interactive use by users(returning single record at a time) . The emphasis is on high throughput of data access rather than low latency of data access. HDFS focuses not so much on storing the data but how to retrieve it at the fastest possible speed, especially while analyzing Data. In HDFS, reading the complete data is more important than the time taken to fetch a single record from the data. 
A dataset is typically generated or copied from source, and then various analyses are performed on that dataset over time. Each analysis will involve a large proportion of the dataset, so the time to read the whole dataset is more important than the latency in reading the first record.


4.  Commodity Hardware:

HDFS (Hadoop Distributed File System) assumes that the cluster(s) will run on common hardware, that is, non-expensive, ordinary machines rather than high-availability systems. A great feature of Hadoop is that it can be installed in any average commodity hardware. We don’t need super computers or high-end hardware to work on Hadoop. This leads to an overall cost reduction to a great extent.

5.  Data Replication and Fault Tolerance:

HDFS works on the assumption that hardware is bound to fail at some point of time or the other. This disrupts the smooth and quick processing of large volumes of data. To overcome this obstacle, in HDFS, the files are divided into large blocks of data and each block is stored on three nodes: two on the same rack and one on a different rack for fault tolerance. A block is the amount of data stored on every data node. Though the default block size is 64MB and the replication factor is three, these are configurable per file. This redundancy enables robustness, fault detection, quick recovery, scalability, eliminating the need of RAID storage on hosts and merits of data locality.
HDFS is designed to carry on working without a noticeable interruption to the user in the face of such failure.

6.  High Throughput:

Throughput is the amount of work done in a unit time. It describes how fast the data is getting accessed from the system and it is usually used to measure performance of the system. In Hadoop HDFS, when we want to perform a task or an action, then the work is divided and shared among different systems. So, all the systems will be executing the tasks assigned to them independently and in parallel. So the work will be completed in a very short period of time. In this way, the Apache HDFS gives good throughput. By reading data in parallel, we decrease the actual time to read data tremendously.
Applications that require low-latency access to data, in the tens of milliseconds range, will not work well with HDFS.HBaseis currently a better choice for low-latency access.

7.  Moving Computation is better than Moving Data:

Hadoop HDFS works on the principle that if a computation is done by an application near the data it operates on, it is much more efficient than done far of, particularly when there are large data sets. The major advantage is reduction in the network congestion and increased overall throughput of the system. The assumption is that it is often better to locate the computation closer to where the data is located rather than moving the data to the application space. To facilitate this, Apache HDFS provides interfaces for applications to relocate themselves nearer to where the data is located.

8.  File System Namespace:

A traditional hierarchical file organization is followed by HDFS, where any user or an application can create directories and store files inside these directories. Thus, HDFS’s file system namespace hierarchy is similar to most of the other existing file systems, where one can create and delete files or relocate a file from one directory to another, or even rename a file. In general, HDFS does not support hard links or soft links, though these can be implemented if need arise.
Thus, HDFS works on these assumptions and goals in order to help the user access or process large data sets within incredibly short period of time!

9.What is HDFS block ? why is it so large ?

A disk has a block size, which is the minimum amount of data that it can read or write.Normal Filesystem blocks are typically a few kilobytes in size, whereas disk blocks are normally 512 bytes. This is generally transparent to the filesystem user who is simply reading or writing a file

HDFS also has the concept of a block, but it is a much larger unit—64 MB by default.Like in a filesystem for a single disk, files in HDFS are broken into block-sized chunks,which are stored as independent units to make a file.

HDFS blocks are large compared to disk blocks, and the reason is to minimize the cost of seeks. By making a block large enough, the time to transfer the data from the disk can be significantly longer than the time to seek to the start of the block. Thus the time to transfer a large file made of multiple blocks operates at the disk transfer rate.


10.MAP-Reduce components  (Introduction) :

MapReduce is a programming model for data processing.
Hadoop can run MapReduce programs written in various languages;Java, Ruby, Python, and C++.

To speed up the processing, we need to run parts of the program in parallel,so some processes will finish much earlier than others since the input files may vary from small to large files.so for A better approach Map-Reduce will split the input into fixed-size chunks and assign each chunk to a process.
A MapReduce job is made up of four distinct stages, executed in order:
1.client job submission,
2.map task execution,
3.shuffle and sort,
4.reduce task execution.

A MapReducejob is a unit of work that the client wants to be performed: it consists of the input data, the MapReduce program, and configuration information. Hadoop runs the job by dividing it into tasks, of which there are two types:
map tasks and reduce tasks.

There are two types of nodes that control the job execution process: a jobtracker(Master) and  number of tasktrackers(workers). The jobtracker coordinates all the jobs run on the system by scheduling tasks to run on tasktrackers.

Tasktrackers run tasks and send progress reports to the jobtracker, which keeps a record of the overall progress of each job. If a task fails, the jobtracker can reschedule it on a different tasktracker.

A master process, called the jobtrackerin HadoopMapReduce, is responsible for accepting these submissions on the name node and the task Tracker which will work on the Datanode will manage this task in to map task and Reduce Task on node where the Data located is nearer.

Hadoop divides the input to a MapReduce job into fixed-size pieces called input splits
Hadoop creates one map task for each split, which runs the userdefined map function for each record in the split.

if we are processing the splits in parallel, the processing is better load-balanced when the splits are small, since a faster machine will be able to process proportionally more splits over the course of the job than a slower machine.

if splits are too small, the overhead of managing the splits and of map task creation begins to dominate(timeconsuming) the total job execution time.
For most jobs, a good split size tends to be the size of an HDFS block, 64 MB by default although its changeable

Hadoop does its best to run the map task on a node where the input data resides in HDFS. This is called the data locality optimization because it doesn’t use valuable cluster bandwidth.

To make it simple Files which is made of Blocks are divided in to input splits Each input split will be processed by a mapper task and then each mappers output produced will  go in to shuffle and sort phase which are intermediate outputs will be processed by reduce tasks to produce the final output, and once the job is complete, the map output can be thrown away. if the node running the map task fails before the map output has been consumed by the reduce task, then Hadoop will automatically
rerun the map task on an another node to re-create the map output

Reduce tasks don’t have the advantage of data locality; the input to a single reduce task is normally the output from all mappers.

We will focus more about MAP-Reduce on the title Anatomy of Map-Reduce Program.

Also we will focus more about HDFS Architecture on the Title HDFS Architecture.



No comments:

Post a Comment