perftools-setup

Pre-requisities:

1. HADOOP YARN Setup should be completed and HADOOP_HOME should be set in the environment variable.
2. Make sure the nodes are set for password-less SSH both ways(master->slaves & slaves->master).
3. Since we use the environment variables a lot in our scripts, make sure to comment out the portion following this statement in your ~/.bashrc , If not running interactively, don't do anything
4. Kindly refer to the setups & scripts provided in https://github.com/kmadhugit/hadoop-cluster-utils before proceeding further as the utility scripts provided in the repository are needed here.

Installations:

• Monitoring utility nmon & operf are required to collect performance data. To install follows the steps,

  git clone https://github.com/josiahsams/perftools-setup
  
  cd perftools-setup
  
  ./install.sh
  
  . ~/.bashrc

Note: The install.sh script will perform the following,

• nmon & operfin all the nodes (master + slaves)
• Configure to run pid_monitor in master node.
• include the scripts part of this repo to the PATH

# CAUTION: oprofile should be rebuild and installed every time when the system is moved to a different JDK level. 
# The old oprofile must be uninstalled before installing the new version. 

# To uninstall the old oprofile and reinstall with the new JDK level
  AN "$WORKDIR/perftools-setup/oprofile/oprofile_uninstall.sh -r $WORKDIR"

# To just uninstall the old oprofile installed do
  AN "$WORKDIR/perftools-setup/oprofile/oprofile_uninstall.sh $WORKDIR"

nmon recording.

• Two scripts are provided for nmon recording and they can be invoked as follows,

# To Start nmon recording in all the nodes of the cluster.

startnmon.sh <dir_name>

# To Stop nmon recording in all the nodes of the cluster.
# This script is also capable of collecting all the nmon recordings 
# and place it in the directory provided by user and keep it in an archive format.

stopnmon.sh <dir_name>

oprofile

Before starting Spark Application profiling add the following configuration parameter to either spark-submit or spark-sql or spark-shell,

Set OPERFLIB to ${WORKDIR}/oprofile/oprofile_install/lib

--conf spark.executor.extraJavaOptions="-agentpath:${OPERFLIB}/oprofile/libjvmti_oprofile.so" 
--conf spark.executor.extraLibraryPath=${OPERFLIB}                                
--driver-library-path ${OPERFLIB}                                                
--driver-java-options "-agentpath:${OPERFLIB}/oprofile/libjvmti_oprofile.so"

• To enable oprofile for profiling spark applications:

oprofile_start.sh

• To disable oprofile for spark applications

oprofile_stop.sh

pid_monitor

To invoke pid_monitor in your cluster, run the following command,

pmon <command/script with arguments>

eg:- pmon run_single.sh q73 6 18 23g tpcds1t

Note:

• This command pmon will start necessary data collection agent in all nodes(master+slaves)
• On completion it will collect all the performance data from the node and make it available in master.
• All the perf data can be viewed in web under http://masternode/rundir