Overview

In clustered file systems, single node failure are most common, however, there are failure bursts happen from time to time. We thus propose RSTAIR code for a fast recovery for single node failure, while tolerable of burst of failures with low storage overhead.

User's Guide

R-STAIR code is implemented and integrated with Facebook's Hadoop-20.

I. Install Hadoop-20

Hadoop-20 can be downloaded from here.
Download and extract it in one of your hosts. If the above link is not working, you can download from [homepage of this project] (http://ansrlab.cse.cuhk.edu.hk/software/ear).

To simplify our description, in this instruction, we refer to the hadoop-20/ directory as hadoop directory.

Call the following command to update the environment variable.

export HADOOP_HOME=absolute address of hadoop directory
export PATH=$PATH:$HADOOP_HOME/bin

It is a good practice to include the above two lines in your ~/.bashrc file.

Add the following line to $HADOOP_HOME/conf/hadoop-env.sh

export JAVA_HOME=YOUR JAVA_HOME

II. Installation and Configuration

0. Run bash install.sh to install R-STAIR code.

1. Configure masters and slaves
   To start with, you need to make sure in your cluster every machine is public-key accessible from other machines.
   Input the IP address/hostname of master and slaves in file conf/masters and conf/slaves.

2. Create cluster topology definition script. Hadoop uses a topology definition script for rack awareness.
   The script takes IP address or host name of a machine as input and returns which rack the machine belongs to. Take a look at $HADOOP_HOME/rackAware.sh to get a feeling and change it according to your cluster topology.

To use R-STAIR code, we assume you have enough racks and enough nodes in each rack. For example, if you set n=6 and r=5, you should have at least 6 racks and at least 5 nodes per rack.

3. Configure conf/core-site.xml conf/hdfs-site.xml conf/raid.xml
   according to the instructions in the xml files.

4. Finishing: Copy the hadoop directory to every machine in the same location.

III. Configuration

To change parameters of R-STAIR code, please refer to $HADOOP_HOME/src/contrib/raid/src/java/raid-default.xml as an example.

To make R-STAIR code work properly, please also update corresponding configurations in $HADOOP_HOME/conf/hdfs-site.xml for the R-STAIR code placement.

Developer's Guide

Our modifications can be divided into three parts:

1. Coding implementation R-STAIR is implemented in C++ and integrated with HDFS-RAID through JNI. Find our implementation in src/native/src/org/apache/hadoop/util/stair.c
2. R-STAIR code placement We implement our two-dimension round-robin placement to place data/parity blocks to achieve load balance and exploit the locality property of R-STAIR codes. (src/contrib/raid/src/java/org/apache/hadoop/hdfs/server/namenode)
   • See BlockPlacementPolicyStair.java for more detail.
3. MapReduce integration We explore how to leverage the feature of R-STAIR code to improve MapReduce performance. We modify MapReduce scheduler to integrate
   • We mainly modified JobInProgress.java to leverage locality of R-STAIR code for MapReduce jobs.

Change Log

Version 1.0.0 (September, 2015): First release.