Fault injection tool for hadoop cluster from yahoo anarchyape
- Java JDK >= 1.7
- pdsh
- stress-ng for CPU and Memory hog
[Java with maven]
mvn package
mv target/anarchyape.jar ape.jar
[Java manual]
cd src/main/java
# download log4j-1.4.12.jar and apache commons-cli-1.2.jar
java -cp . ape/Main.java
rm ape.jar
# Use either (preferred):
jar cfm ape.jar META-INF/MANIFEST.MF ape META-INF/services org
# either :
javac -cp .:log4j-1.4.12.jar:commons-cli-1.2.jar ape/*.java
[Perl]
perl Makefile.PL
cpan -i JSON
make
make test
make install
[Perl] ./ape.pl [remote_ip_list_file]
[Java]
java -jar ape.jar [commands]
log file: anarchyape.log
(old way)
java -cp .:log4j-1.4.12.jar ape/Main
(Local run)
java -jar ape.jar -L -S 100 5
injects slow network with delay 100 milliseconds for 5 seconds.
(Remote run) Install pdsh:
yum install pdsh
apt-get install pdsh
cp ape /usr/local/bin/ape
chmod go-rwx /usr/local/bin/ape
java -jar ape.jar -R node1,node2,node3 -S 100 5 eth0
# creates a script to run on the remote hosts:
pdsh -Rssh -w node1,node2,node3 '/usr/local/bin/ape -L -S 100 5 eth0'
It seems not working as the remote hosts do not have /usr/local/bin/ape file.
Remote nodes can be specified in XML format: cluster-ip-list.xml
Currently, to create a scenario, the user constructs a shell script specifying the types of errors to be injected or failures to be simulated, one after another. A sample line in a scenario file could be as follows:
java -jar ape.jar -remote cluster-ip-list.xml -F lambda -k lambda
where the -F is a “Fork Bomb” injection, the -k is a “Kill
One Node” command, and the lambda specifies the failure rates.
Users can define lambda parameters by computing Mean Time Between Failures (MTBF) of a system. MTBF is defined to be the average (or expected) lifetime of a system and is one of the key decision-making criteria for data center infrastructure systems [1]. Equipment in data centers is going to fail, and MTBF helps with predicting which systems are the likeliest to fail at any given moment. Based on previous failure statistics, users can develop an estimate of MTBF for various equipment failures; however, determining MTBFs for many software failures is challenging.
[1] W. Torell and V. Avelar. Performing effective MTBF comparisons for data center infrastructure. http://www.apcmedia.com/salestools/ASTE-5ZYQF2_R1_EN.pdf.
Here are some common failures in Hadoop environments:
# Data node is killed
# Application Master (AM) is killed
# Application Master is suspended
# Node Manager (NM) is killed
# Node Manager is suspended
# Data node is suspended
# Tasktracker is suspended
# Node panics and restarts
# Node hangs and does not restart
# Random thread within data node is killed
# Random thread within data node is suspended
# Random thread within tasktracker is killed
# Random thread within tasktracker is suspended
# Network becomes slow
tc qdisc add dev eth0 root netem delay 100.0ms && sleep 30.0 && tc qdisc del dev eth0 root netem
# Network is dropping significant numbers of packets
# Network disconnect (simulate cable pull)
iptables -A INPUT -p tcp --dport 8080 -j DROP # block port
iptables -D INPUT -p tcp --dport 8080 -j DROP # unblock port
# One disk gets VERY slow
# CPU hog consumes x% of CPU cycles, for example by running stress-ng command remotely with :
stress-ng -c 1 --verify -t 1m -v
# Mem hog consumes x% of memory, for example by running stress-ng command remotely with :
stress-ng --vm 4 --vm-bytes 90% --vm-method all --verify -t 1m -v
# Corrupt blocks of a given file on rw-mounted disk
Command line options:
usage: ape [options] ... <failure command>
options:
-u,--udp-flood <hostname> <port> <duration> Flood the target
hostname with a DoS attack. For proper effect, use the -R flag and
designate more than one host.
-C,--corrupt-block <meta/ord> <size> <offset> Corrupt a random HDFS
block file with a size in bytes as the 2nd arg and offset in bytes as the
3rd argument
-d,--network-disconnect <time> <nic> Disconnect the network
for a certain period of time specified in the argument on a given network
interface, and then resumes
-p,--network-drop <percentage> <duration> <nic> Drops a specified
percentage of all inbound network packets for a duration specified in
seconds.
-c,--corrupt-file <file> <size> <offset> Corrupt the file given
the address as the first argument, size as the 2nd arg, and offset as the
3rd argument
-e,--continue-node <NodeType> Continues a tasktracker
or a datanode at the given hostname that has already been suspended
-k,--kill-node <nodetype> Kills a datanode,
tasktracker, jobtracker, or namenode.
-s,--suspend-node <NodeType> Suspends a tasktracker
or a datanode at the given hostname
-r,--remount Remounts all
filesystems as read-only
-S,--network-slow <delay> <duration> <nic> Delay all network
packet delivery by a specified amount of time (in milliseconds) for a
period specified in seconds on a given network interface
-F,--forkbomb Hangs a host by
executing a fork bomb
-L,--local Run commands locally
-P,--panic Forces a kernel panic
and does not restart the system.
-R,--remote <HostnameList> Run commands remotely
-V,--version Displays the version
number
-h,--help Displays this help menu
-t,--touch Touches a file called
/tmp/foo.tst
-v,--verbose Turn on verbose mode
command:
-fb <lambda> -k <lambda> fork bomb
-kp <lambda> -k <lambda> kernel panic
-r <lambda> -k <lambda> remount root as read only
-kn <lambda> -k <lambda> kill a node process
-dos <lambda> -k <lambda> denial of service by launching 4 bombarding threads
-cb <lambda> -k <lambda> corrupt a random HDFS block
-cf <lambda> -k <lambda> corrupt a file at the given address
-nic <lambda> -k <lambda> interface
-p <lambda> -k <lambda> packet drop
#java -jar ape.jar -h
#java -jar ape.jar -L -p,--network-drop <percentage> <duration> <nic>
java -jar ape.jar -L -p 80 30 eth0
#java -jar ape.jar -R -S <delay in millisec> <duration in sec> <nic>
#java -jar ape.jar -L -S 100 30 eth0
#java -jar ape.jar -L -d,--network-disconnect <time in seconds> <nic>
#java -jar ape.jar -L -d 30 eth0
#java -jar ape.jar -L --remount
#java -jar ape.jar -L --panic
#java -jar ape.jar -L --forkbomb
#java -jar ape.jar -L -c,--corrupt-file <file> <size> <offset>
#java -jar ape.jar -L --corrupt-file <file> <size> <offset>
#java -jar ape.jar -L -C,--corrupt-block <meta/ord> <size> <offset>
#java -jar ape.jar -R precise386 -S 100 30 eth0
#java -jar ape.jar -R cluser-ip-list.xml -S 100 5 eth0
#java -jar ape.jar -remote cluster-ip-list.xml -fb lambda -k lambda
#java -jar ape.jar -remote cluster-ip-list.xml -F lambda
#java -cp .:log4j.jar ape/Main -h
#java -cp .:log4j.jar ape/Main -R slaves
#java -cp .:log4j.jar ape/Main -V
#java -jar ape.jar -h
#java -jar ape.jar -R -S <delay in millisec> <duration in sec> <nic>
java -jar ape.jar -R precise386 -S 100 30 eth0
#java -jar ape.jar -R cluser-ip-list.xml -S 100 5 eth0
#java -jar ape.jar -remote cluster-ip-list.xml -fb lambda -k lambda
#java -jar ape.jar -remote cluster-ip-list.xml -F lambda
#java -jar ape.jar -L -v
#java -jar ape.jar -R -F
#java -jar ape.jar -R -S 1 1 eth0
#java -cp .:log4j.jar ape/Main -h
#java -cp .:log4j.jar ape/Main -R slaves
#java -cp .:log4j.jar ape/Main -remote slaves -v
#java -cp .:log4j.jar ape/Main -V