gmldry / repairboost-code

This is the implementation of RepairBoost described in our paper "Boosting Full-Node Repair in Erasure-Coded Storage" appeared in USENIX ATC'21.

RepairBoost

This is the implementation of RepairBoost described in our paper "Boosting Full-Node Repair in Erasure-Coded Storage" appeared in USENIX ATC'21. RepairBoost is a scheduling framework that can assist existing linear erasure codes and repair algorithms to boost the full-node repair performance. Please contact [email protected] if you have any questions.

1. Install

We have tested RepairBoost on Ubuntu16.04 LTS.

1.1 Common

• g++ & make & libtool & autoconf & git

$ sudo apt-get install g++ cmake libtool autoconf git

• gf-complete

$ git clone https://github.com/ceph/gf-complete.git
$ cd gf-complete
$ ./autogen.sh && ./configure && make && sudo make install

• redis-3.2.8

Download redis-3.2.8 and install it.

$ sudo wget http://download.redis.io/releases/redis-3.2.8.tar.gz
$ tar -zxvf redis-3.2.8.tar.gz
$ cd redis-3.2.8
$ make && sudo make install

Install redis as a background daemon. You can just use the default settings.

$ cd utils
$ sudo ./install_server.sh

Configure redis to be remotely accessible.

$ sudo /etc/init.d/redis_6379 stop

Edit /etc/redis/6379.conf. Find the line with bind 127.0.0.0 and modify it to bind 0.0.0.0, then start redis.

$ sudo /etc/init.d/redis_6379 start

• hiredis-1.0.0

Download hiredis-1.0.0 and install it.

$ sudo wget https://github.com/redis/hiredis/archive/refs/tags/v1.0.0.tar.gz 
$ tar -zxvf v1.0.0.tar.gz
$ cd hiredis-1.0.0
$ make && sudo make install

1.2 Compile RepairBoost

After finishing the preparations above, download and compile the source code.

$ cd repairboost-code
$ make

2. Standalone Test

2.1 Prerequisites

Configuration File

We configure the settings of RepairBoost via the configuration file config.xml in XML format.

Property	Description
erasure.code.type	Three types are implemented: RS, LRC, and BUTTERFLY (only n=6, k=4).
erasure.code.k	The number of data chunks.
erasure.code.n	Total number of chunks in a stripe.
lrc.code.l	The number of groups, only valid in LRC. Default is 0.
encode.matrix.file	The path of the encoding matrix file. Absolute path is required.
packet.size	The size of a packet in units of bytes.
packet.count	The number of packets in a chunk. A chunk is partitioned into many smaller packets.
repair.method	Three single-chunk repair methods are implemented: cr (i.e., conventional repair), ppr and path(i.e., ecpipe).
file.system.type	Two types are supported: standalone and HDFS3.
meta.stripe.dir	The directory where the stripe metadata is stored. Absolute path is required.
block.directory	The directory where the coded chunks are stored. Absolute path is required.
coordinator.address	IP address of the coordinator.
helpers.address	IP addresses of all nodes in the system.
local.ip.address	IP address of the node itself.

Encoding Matrix File

For RS codes and LRC codes, we read the encoding coefficients from the encoding matrix file. The file contains an (n-k)×k matrix that specifies the coefficients for generating n-k parity chunks from k data chunks.

The file repairboost-code/conf/rsEncMat_6_9 shows a encoding matrix for RS(9, 6) (n=9, k=6):

Our example uses the file repairboost-code/conf/rsEncMat_3_4 to construct a simple coding matrix for RS(4, 3)(n=4, k=3):

Create Erasure-Coded Chunks

Before testing our standalone system, we need to create stripes of erasure-coded chunks and corresponding metadata. We have provided programs for generating chunks of different codes under repairboost-code/test/.

$ cd repairboost-code/test
$ make
$ dd if=/dev/urandom iflag=fullblock of=input.txt bs=64M count=3
$ ./createdata_rs ../conf/rsEncMat_3_4 input.txt 3 4

For more detail, you can run the command ./createdata_rs (or ./createdata_lrc, etc.) to get the usage of the programs.

Take RS(4, 3) (n=4, k=3) as an example. In repairboost-code/test, we create 3 files of uncoded chunks and 1 file of coded chunks. We can distribute the 4 chunks across 4 nodes, each of which stores one chunk under the path specified by block.directory. The coordinator stores the stripe metadata under the path specified by meta.stripe.dir.

The naming of data files and metadata should meet the requirements. If there are 2 stripes, the data of the stripes are named as follows:

stripe_0_file_k1   stripe_0_file_k2   stripe_0_file_k3   stripe_0_file_m1   
stripe_1_file_k1   stripe_1_file_k2   stripe_1_file_k3   stripe_1_file_m1   

The coordinator stores 4 files for each stripe: rs:stripe_0_file_k1_1001, rs:stripe_0_file_k2_1001, rs:stripe_0_file_k3_1001 and rs:stripe_0_file_m1_1002, all of which have the following content:

stripe_0_file_k1_1001:stripe_0_file_k2_1001:stripe_0_file_k3_1001:stripe_0_file_m1_1002

The file name rs:file_k1_1001 means that the chunk uses Reed-Solomon (RS) code and the chunk name is file_k1. The tail 1001 (resp. 1002) means the chunk is an uncoded chunk (resp. coded chunk).

2.2 Run

Start RepairBoost

The start script is in repairboost-code/scripts/.

$ python scripts/start.py

Tip: If an error occurs when loading the shared library libgf_complete.so.1, you may need to set up the environment as follows.

$ sudo cp /usr/local/lib/libgf_complete.so.1 /usr/lib/ && sudo ldconfig

Full-node Repair Test

Use ssh to connect to any node (not the coordinator) to send a full-node repair request.

$ ./ECClient

Stop RepairBoost

The stop script is in repairboost-code/scripts/.

$ python scripts/stop.py

3. Hadoop-3 Integration

3.1 Prerequisites

• isa-l

$ git clone https://github.com/01org/isa-l.git
$ cd isal
$ ./autogen.sh && ./configure && make && sudo make install

• java8

$ sudo apt-get purge openjdk*
$ sudo apt-get install software-properties-common
$ sudo add-apt-repository  ppa:ts.sch.gr/ppa 
$ sudo apt-get update 
$ sudo apt-get install oracle-java8-installer  
$ sudo apt install oracle-java8-set-default

Then configure the environment variables for java.

export JAVA_HOME=/usr/lib/jvm/java-8-oracle
export PATH=$JAVA_HOME/bin:$PATH

Test

$ java -version
java version "1.8.0_212"

• maven

Download apache-maven-3.5.4 on available mirror.

$ tar -zxvf apache-maven-3.5.4-bin.tar.gz
$ sudo mv apache-maven-3.5.4 /opt/
$ sudo ln -s /opt/apache-maven-3.5.4 /opt/apache-maven

Then configure the environment variables for maven.

export MAVEN_HOME=/opt/apache-maven
export PATH=$MAVEN_HOME/bin:$PATH

Test

$ mvn help:system
BUILD SUCESS

• protobuf-2.5.0

Download protobuf-2.5.0. (required)

$ tar -zxvf protobuf-2.5.0.tar.gz
$ cd protobuf-2.5.0
$ sudo mkdir /usr/local/protoc-2.5.0/
$ ./configure --prefix=/usr/local/protoc-2.5.0/
$ make && sudo make install

Then configure the environment variables for maven.

export PROTOC_HOME=/usr/local/protoc-2.5.0
export PATH=$PROTOC_HOME/bin:$PATH

Test

$ protoc --version
libprotoc 2.5.0

• hadoop-3.1.4-src

Download hadoop-3.1.4-src on available mirror.

$ tar -zxvf hadoop-3.1.4-src.tar.gz

Then configure the environment variable.

export HADOOP_SRC_DIR=/path/to/hadoop-3.1.4-src
export HADOOP_HOME=$HADOOP_SRC_DIR/hadoop-dist/target/hadoop-3.1.4
export PATH=$HADOOP_HOME/bin:$HADOOP_HOME/sbin:$PATH

Tip: Native libs used to rebuild hadoop.

$ sudo apt-get -y install build-essential autoconf automake libtool cmake zlib1g-dev pkg-config libssl-dev

Edit repairboostcode/hadoop-3.1.4-integrate/install.sh with the proper directory to where you installed hadoop-3.1.4-src. Then execute the script to rebuild hadoop.

$ ./install.sh

3.2 Hadoop Configuration

The following shows an example to configure Hadoop-3.1.4 with 6 nodes.

IP Address	Role in Hadoop	Role in standalone
192.168.0.201	NameNode	Coordinator
192.168.0.202	DataNode	Helper
192.168.0.203	DataNode	Helper
192.168.0.204	DataNode	Helper
192.168.0.205	DataNode	Helper
192.168.0.206	DataNode	Helper
You should modify the following configure files in /path/to/hadoop-3.1.4-src/hadoop-dist/target/hadoop-3.1.1/etc/hadoop.

• core-site.xml

<property><name>fs.defaultFS</name><value>hdfs://192.168.0.201:9000</value></property>
<property><name>hadoop.tmp.dir</name><value>/path/to/hadoop-3.1.4</value></property>

• hadoop-env.sh

export JAVA_HOME="/usr/lib/jvm/java-8-oracle" 

• hdfs-site.xml

<property><name>dfs.client.use.datanode.hostname</name><value>true</value></property>
<property><name>dfs.replication</name><value>1</value></property>
<property><name>dfs.blocksize</name><value>67108864</value></property>
<property><name>repairboost.coordinator</name><value>192.168.0.201</value></property>
<property><name>dfs.datanode.ec.reconstruction.stripedread.buffer.size</name><value>1048576</value></property>
<property><name>dfs.datanode.ec.repairboost</name><value>true</value></property>
<property><name>repairboost.packetsize</name><value>1048576</value></property>
<property><name>repairboost.packetcnt</name><value>64</value></property>

• user_ec_policies.xml

We provide sample configuration for RS-3-1-1024k erasure code policy.

• workers

This file contains multiple lines, each of which is a DataNode IP address.

192.168.0.202
192.168.0.203
192.168.0.204
192.168.0.205
192.168.0.206

3.3 Run

Start Hadoop

• Format the Hadoop cluster.

$ hdfs namenode -format
$ start-dfs.sh
$ hdfs dfsadmin -report 

If the report result indicates that there are 5 DataNodes, then the Hadoop cluster starts correctly.

• Set erasure coding policy.

$ hdfs ec -addPolicies -policyFile /path/to/user_ec_policies.xml
$ hdfs ec -enablePolicy -policy RS-3-1-1024k
$ hdfs dfs -mkdir /ec_test
$ hdfs ec -setPolicy -path /ec_test -policy RS-3-1-1024k

• Write data into HDFS.

$ dd if=/dev/urandom iflag=fullblock of=file.txt bs=64M count=3
$ hdfs dfs -put file.txt /ec_test/testfile

Check the writed data.

$ hdfs fsck / -files -blocks -locations

Start RepairBoost

• Get the directory where the erasure-coded data is stored in Hadoop.

$ ssh datanode
$ find -name "finalized"

You can ssh to any DataNode to execute the command.

• Modify the value of file.system.type, block.directory, and helpers.address in the configuration file config.xml.

• Start.

$ cd repairboost-code
$ python scripts/start.py

Full-node Repair Test

• Stop a datanode in Hadoop.

$ ssh datanode
$ hdfs --daemon stop datanode

You can ssh to any DataNode to execute the command.

• Check the repaired data.

$ hdfs fsck / -files -blocks -locations

Stop

$ stop-dfs.sh
$ cd repairboost-code && python scripts/stop.py