c* summit 2013: hardware agnostic - cassandra on raspberry pi by andy cobley
DESCRIPTION
The raspberry Pi is a credit-card sized $25 ARM based linux box designed to teach children the basics of programming. The machine comes with a 700MHz ARM and 512Mb of memory and boots off a SD card, not much power for running the likes of a Cassandra cluster. This presentation will discuss the problems of getting Cassandra up and running on the Pi and will answer the all important question: Why on Earth would you want to do this!?TRANSCRIPT
Hardware Agnostic: Cassandra on Raspberry Pi
Andy Cobley | Lecturer, University of Dundee, Scotland
* Cassandra is hardware agnostic * So why not run it on a Raspberry Pi ? * How hard can it be ? * What can we do with it once it works?
Cassandra on Raspberry Pi
* Andy Cobley * School of Computing * University of Dundee * Twitter: @andycobley
Who Am I ?
* Single chip Linux computer * 500 Meg ram * Boots off an SD card * Ethernet port * (graphics and all you need for a general purpose computer)
Whats a Raspberry Pi ?
Pi with pound coin
* And here’s the Cassandra cluster *
And, here’s one for real
* Power Permitting !
* Cassandra is designed to be fast, fast at writing, fast at reading. * This laptop with one instance of Cassandra will do 12,000 write
operations * Raspberry Pi will do 200 !
The Bad News
* Running a external USB drive is actually worse ! * Probably be hardware feature
More bad news !
Raspberry Pi Schematic
* Oracle Java vs OpenJDK
And then there’s Java!
* Raspbian is Debian for the PI * Uses the Hard floating point accelerator * Much faster than Debian * Current Oracle JDK won’t run on it !
And Raspbian
* http://www.oracle.com/technetwork/java/embedded/downloads/javase/index.html
* Java SE Embedded version 6 * Cassandra might prefer 6 * But * https://blogs.oracle.com/henrik/entry/oracle_releases_jdk_for_linux * Preview at: * https://jdk8.java.net/fxarmpreview/
Oracle java
* Actually not much difference in performance
Hard vs Soft Float
* Cassandra uses compression for performance * Started in version 1.0
2x-‐4x reduc+on in data size 25-‐35% performance improvement on reads 5-‐10% performance improvement on writes
The Problem with compression
* Two types:
Google Snappy Compressor (Faster read/writes) DeflateCompressor (Java zip, slower , beLer compression)
* Snappy Compression not available on Pi
(requires na+ve methods, so someone might get it to work!)
Compression types
* Startup script allocates memory * Calculates based on number of processors * Pi reports Zero processors ! * Boom ! * Now fixed
And the startup script
* In Cassandra-env.sh * JVM_OPTS="$JVM_OPTS -
Djava.rmi.server.hostname=192.168.1.15” * Or else nodetool will not work between nodes
JMX Config
* C* 1.22. added UseCondCardMark as a JVM Opt * "for better lock handling especially on hotspot with multicore
processor” * In cassandra-env.sh
#if [ "$JVM_VERSION" \> "1.7" ] ; then # JVM_OPTS="$JVM_OPTS -‐XX:+UseCondCardMark" #fi
JVM OPT UseCondCardMark
* We’ve forgotten one thing * The Pi cost £25 * You can power 4 from USB hub (no need for a power supply on
each one) * So:
The Good News !
So, have a 64 node computer for £2000
University of Southhampton
* 32 node Beowolf cluster: * Joshua Kiepert, Boise University
Or this
* Adding nodes adds performance * Adding nodes adds replicas of data * BUT * Make sure your ring is balanced, * Pi’s don’t like to be unbalanced.
Adding nodes is good
* Vnodes (in 1.2) would be very nice * However at this point I haven’t got 1.2 on Pi running on a cluster
Vnodes
Performance with 3/4 nodes
Performance with 5/6 nodes
* ./stress -d 192.168.1.10,192.168.1.11,192.168.1.12 -o insert -I DeflateCompressor
* Note: nodes to use * You will get different performance if you insert to less nodes than
you have in your ring
Stress test commands
* Adding a node (in the absence of Vnodes)
Must seed form a known node Use a program to calculate new keys Bring up new node with the correct key in cassandra.yaml Use node tool to move other nodes
Adding Nodes Procedure
* Python code import sys if (len(sys.argv) > 1): num = int(sys.argv[1]) else: num = int(raw_input("How many nodes? :")) for i in range(0,num): print 'node %d: %d' % (i, (i*(2**127)/num))
Calculating keys
* Use nodetool
sudo ./nodetool -‐h 192.168.1.10 move 42535295865117307932921825928971026432
* And cleanup
./nodetool -‐h 192.168.1.10 cleanup
Moving existing nodes
* On Debian, you can free memory from the graphics chip
Cd /boot sudo cp start.elf start.elf.old sudo cp arm224_start.elf to start.elf reboot
Getting more memory
* Under Rasbian * Run with a monitor plugged for the first time * Set options for screen memory * Perhaps disable boot to GUI
Getting more Memory
* I prefer static network addresses * Edit /etc/network/interfaces iface eth0 inet sta+c address 192.168.1.41 netmask 255.255.255.0 network 192.168.1.0 broadcast 192.168.1.255 gateway 192.168.1.254
*
Network address
* Make a master SD card * Copy it ! * Make sure the master version has no data on it. * Consider ”Puppet” (though I don’t use it)
Multiple nodes
* See https://github.com/acobley/CassandraStartup * Put the file in /etc/init.d * update-rc.d cassandra defaults
Starting as a service
* So for £200 we get an 8 node C* cluster * It can be reconfigured, blown away, stress tested and generally
abused * We can simulate data racks, data centers and I hope even long
network delays. * Hopefully our upcoming MSc in Data Science will use these clusters
Pi is for teaching
* We know C* can be configured to be aware of:
Network racks Data Centers
* We know we can have replicas are stored across these racks * How can we play with this cheaply ?
C* is network aware
Proposed teaching tool
10mbs Hubb
Noise injec+on
Switch 2
Switch 1
Pi 1
Pi 2
Pi 3
Pi 1
Pi 2
Pi 3
* Cassandra wouldn’t run on a PI * It does now. * Running it on a Pi shook out some Cassandra bugs * You can run it in a secure lab
Pi is discovery
* Most important, this was pure Geeky Fun
Pi is for fun
* Data Science: * http://www.computing.dundee.ac.uk/study/postgrad/
degreedetails.asp?17
Obligatory Plug
* Raspberry Pi is cheap * C* needs some work to run on it * You can make clusters cheaply for experimentation * It’s fun !
C* is Hardware Agnostic
THANK YOU