How to build a Hadoop cluster using Raspberry Pi systems

noreply@blogger.com (Anonymous) — Wed, 06 Feb 2013 12:18:00 +0000

Raspberry Pi cluster
Initial enclosure with sticky tape and rubber bands.

Background information
“Big Data” analysis is a fast and agile alternative to structured data warehouses and OLAP approaches that have dominated analytic reporting for many years.

Using Big Data approaches you can derive information from unstructured data, such as written documentation, websites and even program log files. This is the one of the key technologies used by internet search engines such as Google and Yahoo.

Hadoop is an open source Apache Big Data tool that provides both a base platform and data analysis tools. It is written in Java and the redundancy and high availability make it a feasible technology for deployment on large numbers of commodity grade computers.

Most smart phones use ARM based system on a chip processors. These are an alternate technology to Intel processors that offers both a lower cost point and power consumption. This was highlighted recently in an IEEE article: http://spectrum.ieee.org/semiconductors/processors/the-battle-between-arm-and-intel-gets-real:

“Exactly how this competition shapes up will depend not on performance or power consumption but on the ratio between the two: performance per watt. And that metric is fueling a fiery debate over the fundamental differences between Intel’s x86 chips and ARM’s processors.  ”

ARM processor technology is set to make its way into data centres. A quick search on Google shows some of the current ARM service offerings from the big boys

http://content.dell.com/us/en/enterprise/d/campaigns/project-copper

http://h17007.www1.hp.com/us/en/iss/110111.aspx

http://www.engadget.com/2011/11/02/hp-and-calxedas-moonshot-arm-servers-will-bring-all-the-boys-to/

It is also available on a smaller scale. The Raspberry Pi is one of a range of low cost ARM system on a chip solutions.

This begs the questions:

Can I build a working Hadoop Cluster using the Raspberry Pi

Does it perform as well as an Intel processor using price/performance and performance/watt?

Can I find a practical use for it at home?

I will cover the basics of building Raspberry Pi Hadoop cluster in this post.

The plan

To configure three Raspberry Pi as Hadoop data servers with the master Job Tracker and Name Node processes running on another server.

So why 3 data servers? This is less because of any special properties of prime numbers than the number of available ports on my router and my budget does not stretch much past three servers.

Fortunately we will be standing on the shoulders of others

The Raspberry Pi website contains a lot of useful information about the initial setup.

A simpler project is a WebServer farm such as this one (http://www.raspberrypi.org/phpBB3/viewtopic.php?f=26&t=24441
http://my-music.mine.nu/images/rpi_raspianwheezy_setup.pdf)

A 64 node supercomputer if you want to go hard core and have a spare box of Lego (University of Southampton) http://www.southampton.ac.uk/~sjc/raspberrypi/

Michael Noll has provided a couple of excellent tutorials on setting up Hadoop clusters using a Single node or Multiple node configuration.

Big Data for the impatient http://www.ibm.com/developerworks/data/library/techarticle/dm-1209hadoopbigdata/

The time saver here is to get this working in a single node worker image configuration first before attempting to configure the cluster. You can then copy and adjust the image to configure your cluster.

Results to date

My initial test case is the word count example program shipped with Hadoop. This provides a simple map/reduce performance test of the solution and is described in Michael Noll's Multi-Node configuration.

Untuned the runtime is approximately 20 min to count the words of the seven documents Michael suggests. This is mostly a result of JVM swapping. A little tuning brought this down to 7 minutes.

I was hoping for rough price/performance parity with an i3 Sandy bridge processor. When I complete testing I will be providing comparison data with a Samsung Series 3 NP350E5C-A01 as a single node solution. (The same test ran in around 20seconds)

Some things to try

Overclocking the Pi to 1GHz
Turning off swapping
does the -client switch provide better optimisation
Does the Oracle JDK provide better performance

Closing thoughts and open questions
Commercial grade ARM processors come with far better performance than the Raspberry Pi. A larger scale test would be required to understand if commercial ARM solutions are a feasible alternative to Intel processors for a Hadoop Big Data farm. Are the advanced functions on the Intel XEON chips worth the investment for this type of workload?

The Performance per Watt metric will most likely be important for the ongoing operational costs of data centers. The operational cost of a server is usually greater than the initial purchase price. Perhaps a balance between performance per Watt and price/performance metrics is required for large system purchases.

I have three open questions

What is the best performance I can get on the Raspberry Pi Cluster through tuning?
Did I achieve price/performance parity with an i3 processor?
What practical use can I find for my Hadoop cluster at home?

If you want to try this

I used the following equipment

Three RaspBerry Pi Model B with 512M memory running at 700Mhz

My home router

Three ethernet cables

Three iPad power supplies. A little overkill but a good power supply is required for a stable solution.

Three fast SD cards rated at 30Mb/s.

Raspbian Wheezy operating system. I used the hard float version that is not compatible with the Oracle Java runtime.

OpenJDK Java runtime

For the initial setup: 1xHDMI to DVI cable; 1x wireless keyboard & mouse

How to

Download the latest Wheezy image, write it to an SD card, wire up your RPI including a monitor, keyboard and mouse, and boot it up. The setup guides will tell you everything you need to know.
In Raspi-config

expand the root file system,
set the keyboard layout, locale, timezone and password for the “pi” user
enable ssh and set the boot behaviour to disable the desktop on boot. This sets up the RPI for headless operation.

Login as pi
We will be going commando at the console, but if you prefer, type startx to bring up the GUI then fire up a terminal window.

The rest of the instructions require root privileges
$ sudo -i Switching to root
# passwd Sets a new root password

Grab the software we will need
# apt-get install proftpd Installs an ftp server for file transfer
# apt-get install rsync A fast file transfer tool Install Java and any other software you might need
# sudo apt-get install openjdk-7-jdk I grabbed a copy of FORTRAN so that I can experiment with Beowulf clusters at a later date.
Fully upgrade the system # apt-get update && apt-get upgrade
Configure the networking by adding static IP addresses for your cluster in /etc/hosts, and rename your system by updating /etc/hostname to the first of your data servers. For example, my cluster is: 192.168.1.108 beowulf 192.168.1.90 grendel 192.168.1.91 ides 192.168.1.92 draca
At this point I followed Michael Nolls guide to setup a single node Hadoop cluster to setup and test Hadoop on both the data server and master node.
Once you have Hadoop working in single server mode on both the Data server and Master node you can clone the SD card for the remaining data servers, and configure your multi node Hadoop cluster. Don't forget to # poweroff first

Subscribe to IT Bleeding edge