mstress
Folders and files
| Name | Name | Last commit date | ||
|---|---|---|---|---|
parent directory.. | ||||
# # $Id$ # # Author: Thilee Subramaniam # # Copyright 2012 Quantcast Corp. # # This file is part of Kosmos File System (KFS). # # Licensed under the Apache License, Version 2.0 # (the "License"); you may not use this file except in compliance with # the License. You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or # implied. See the License for the specific language governing # permissions and limitations under the License. # MSTRESS : A framework for metaserver/namenode benchmarking ========================================================== Contents: [1] Framework description [2] Files in this direcotry [3] Running benchmark [4] Setting up DFS metaserver/namenode [1] Framework ============= The mstress master would invoke mstress.py in slave mode on the client hosts through SSH. Each mstress slave would invoke the necessary number of load-generating clients, which would stress the meta server. +-----------------------------+ | +-------------------+ | | | mstress.py +-----+----------------------+ | | (--mode master) +-----+------------------+ | | +-------------------+ | | | | (master host) | | | +-----------------------------+ | | | | +--------------------------------------+ | | | | | | +-----------+ | +--------------+ +--------------+ | | | | |<---------+-|mstress_client|<--| mstress.py |<-+--+ | | | | +--------------+ |(--mode slave)| | | | DFS meta | | +--------------+ | | | server | | (client host 1) | | | | +--------------------------------------+ | | | | | | | | | +--------------------------------------+ | | | | +--------------+ +--------------+ | | | |<-----------|mstress_client|<--| mstress.py |<-+------+ +-----------+ | +--------------+ |(--mode slave)| | | +--------------+ | | (client host 2) | +--------------------------------------+ The clients will do file or directory tree creation, stat, or directory walk as specified by the benchmark plan. [2] Files ========= - mstress_initialize.sh Helper script to be used before compiling the source and deploying the mstress bundle. Do ./mstress_initialize.sh --h to see options. - Makefile Used to build the QFS stress client (C++) and HDFS stress client (Java). Ensure that $JAVA_HOME is set correctly. - mstress_client.cc Produces the mstress_client binary that actually drives the QFS metaserver. Build using the provided Makefile ('make ccclient') See 'Benchmarking Procedure' below for details. - MStress_Client.java Produces the java MStress_Client for HDFS namenode. Build using the provided Makefile ('make javaclient') See 'Benchmarking Procedure' below for details. - mstress_prepare_master_clients.sh Helper script used to copy the mstress directory to a list of hosts. To be used after running make. - mstress_plan.py Used to generate a plan file for benchmarking. Args: client hosts list, number of clients per client host, file tree depth, nodes per level etc. The generated plan file is also copied to the /tmp firectory of the participating client hosts. Do ./mstress_plan.py --help to see all options. - mstress.py Used to run the metaserver test with the help of the plan file. Args: dfs server host & port, planfile etc. This script invokes mstress.py on the remote host through SSH. For this reason, the mstress path should be the same on the participating hosts. Do ./mstress.py --help to see all options. - mstress_run.py Essentially a wrapper around mstress_plan.py and mstress.py Args: client hosts list and DFS server:port information. Do mstress_run.py --help to see usage. - mstress_sample_run.sh Used to run sample benchmarks on given QFS and HDFS servers by launching clients on localhost. Essentially a wrapper around mstress_initialize.sh, make, mstress_prepare_master_clients.sh, and mstress.run.py. - mstress_cleanup.py Used to clean up the plan files and log files created on participating hosts. Do ./mstress_cleanup.py --help to see usage. [3] Benchmarking Procedure ========================== In reality, benchmark would use separate physical machines each for compiling, running the DFS server, running mstress master, and load generating clients. The procedure below assumes different machines, but one can also run all on the same box, "localhost". (1) Setup the QFS metaserver and HDFS namenode with the help of section [4] "Setting up DFS metaserver/namenode" below. (2) You should have SSH key authentication set up on the hosts involved so that the scripts can do password/passphrase-less login. (3) On the build machine, ensure that you have the Cloudera HDFS client jars. This is typically at /usr/lib/hadoop/client/*.jars. If you don't have them, install them by, 1. Add the following to /etc/yum.repos.d/thirdparty.repo (sudo needed) ----------------------------------- [cloudera-cdh4] name=Cloudera's Distribution for Hadoop, Version 4 baseurl=http://archive.cloudera.com/cdh4/redhat/6/x86_64/cdh/4/ gpgkey = http://archive.cloudera.com/cdh4/redhat/6/x86_64/cdh/RPM-GPG-KEY-cloudera gpgcheck = 1 ----------------------------------- 2. sudo yum install hadoop-client (4) On the build host, execute 'mstress_initialize.sh' to set up jar paths. ./mstress_initialize.sh /usr/lib/hadoop/client/ (5) On the build host, compile and install QFS using the steps described in https://github.com/quantcast/qfs/wiki/Developer-Documentation. Then change directory to benchmarks/mstress, and just issuing 'make' should build the Java/C++ clients. To manually build C++ client: - Assuming the QFS code is in ~/code/qfs, compile and install QFS using the steps described in https://github.com/quantcast/qfs/wiki/Developer-Documentation - cd ~/code/qfs/benchmarks/mstress - QFS_BUILD_INCLUDE=~/code/qfs/build/include \ QFS_BUILD_STATLIB=~/code/qfs/build/lib/static \ BOOST_LIBRARY_DIR=/opt/local/lib \ make If you encounter any build problem, ensure that your QFS_BUILD_INCLUDE etc. refer to valid paths. To manually build MStress_Client.java - Compile MStress_client.java with hadoop-client jars in the class path. theCP=$(echo mstress_hdfs_client_jars/*.jar | sed 's/ /:/g') javac -cp $theCP MStress_Client.java (6) Determine the master and load generating client hosts that you want to use to connect to the DFS server. This could just be "localhost" if you want to run the benchmark locally. (7) From the build host, use "mstress_prepare_master_clients.sh" to copy your mstress directory to the participating hosts. Note: Do './mstress_prepare_master_clients.sh localhost' localhost-only run. The mstress directory paths should be the same on master and client hosts. (8) On the master host change directory to ~/mstress Create a plan file using mstress_plan.py. Do ./mstress_plan.py --help to see example usage. Eg: ./mstress_plan.py -c localhost,127.0.0.1 -n 3 -t file -l 2 -i 10 -n 139 This will create a plan that creates 2 levels of 10 inodes each by 3 processes on 2 hosts. Since each client creates 110 inodes (10 directories with 10 files each) and since there are 6 clients (3 x 2), this plan is to create 660 inodes on the DFS server. The planfile will pick N files to stat per client such that (N x client-host-count x clients-per-host) is just enough to meet 139. The plan file gets copied to the /tmp directory where you run it. It will also get copied to the participating client hosts in the '-c' option. (9) Checklist: check the presence of, - the plan file on master host and client hosts (step 8 does this for you) - the mstress_client binaries (QFS and HDFS clients) on master and all client hosts (step 7). (10) Run the benchmark from the master with mstress.py. Do ./mstress.py --help to see options. Eg: ./mstress.py -f qfs -s <metahost> -p <metaport> -a </tmp/something.plan> ./mstress.py -f hdfs -s <namehost> -p <nameport> -a </tmp/something.plan> (11) The benchmark name, progress, and time taken will be printed out. [4] DFS Server Setup ==================== [4.1] QFS Metaserver Setup ------------------------- You can setup the QFS metaserver using the steps described in https://github.com/quantcast/qfs/wiki If you want to set up a simple metaserver for local testing, please use the script ~/code/qfs/examples/sampleservers/sample_setup.py. [4.2] HDFS Namenode Setup ------------------------- This will setup the HDFS namenode to listen on port 40000. The webUI will run on default port 50070. The installation used here is based on Cloudera's CDH4 release. (1) Ensure java is installed, and $JAVA_HOME is set. (2) Add the following to /etc/yum.repos.d/thirdparty.repo (sudo needed) ----------------------------------- [cloudera-cdh4] name=Cloudera's Distribution for Hadoop, Version 4 baseurl=http://archive.cloudera.com/cdh4/redhat/6/x86_64/cdh/4/ gpgkey = http://archive.cloudera.com/cdh4/redhat/6/x86_64/cdh/RPM-GPG-KEY-cloudera gpgcheck = 1 ----------------------------------- (3) Install hadoop-hdfs-namenode and update the configs. sudo yum install hadoop-hdfs-namenode sudo mv /etc/hadoop/conf /etc/hadoop/conf.orig sudo cp -r /etc/hadoop/conf.empty /etc/hadoop/conf (4) Update /etc/hadoop/conf/core-site.xml (enter your server name instead of 10.20.30.255) ---------------------------------- <configuration> <property> <name>fs.default.name</name> <value>hdfs://10.20.30.255:40000</value> </property> </configuration> ---------------------------------- (5) Edit /etc/hadoop/conf/hdfs-site.xml, fix or ensure that there is a "file://" prefix to avoid warnings. ---------------------------------- <configuration> <property> <name>dfs.name.dir</name> <value>file:///var/lib/hadoop-hdfs/cache/hdfs/dfs/name</value> </property> </configuration> ---------------------------------- (6) Format the namenode: sudo service hadoop-hdfs-namenode init (7) Start namenode. sudo service hadoop-hdfs-namenode start (8) Now namenode should be running. Confirm this by running, ps aux | grep java sudo netstat -pan | grep 40000 (9) To administer the files and directories, /usr/lib/hadoop/bin/hadoop fs -ls / (10) The user with write access on this namenode is "hdfs". Therefore, give write permission to "/" folder (for mstress benchmark to use) by logging in as "hdfs" user. sudo bash su hdfs JAVA_HOME=<java-home> /usr/lib/hadoop/bin/hadoop fs -chmod 777 / exit (11) Now the namenode is ready for running benchmarks.