Running Torque inside of Eucalyptus

We describe how to setup a TORQUE cluster system within a Eucalyptus cloud. The installation of Torque is not difficult, especially not since Debian provides these packages for it.

This page explains how to go some steps further and automate the installation of Torque across many machines. And it goes yet another step further in that all those machines shall be allowed to run clones of the same hard drive. And with yet another step we no longer require physical machines but allow networked virtual images: the cloud.

The here described process is applicable to any cloud image. There are several examples on how to prepare your own images. This Wiki of Debian has one such showing how to extend from within an already running instance, and ?another about how to create one from scratch. Both are educative in their own right.

Preparation

 $ source ~/.euca/eucarc

Specify a Squeeze image

 $ EMI=emi-1AF00C98

Start two instances of our Squeeze image

 $ euca-run-instances $EMI -k mykey -t c1.medium -n2
 RESERVATION    r-4488080C      myuser  myuser-default
 INSTANCE       i-57E309BE      emi-1AF00C98    0.0.0.0 0.0.0.0 pending mykey   2010-09-13T02:31:51.172Z        eki-D224100C    eri-059910F2
 INSTANCE       i-4C1F0986      emi-1AF00C98    0.0.0.0 0.0.0.0 pending mykey   2010-09-13T02:31:51.173Z        eki-D224100C    eri-059910F2

After a few seconds it will be running

 $ euca-describe-instances 
 RESERVATION    r-4488080C      myuser default
 INSTANCE       i-4C1F0986      emi-1AF00C98    192.168.0.14    192.168.0.14    running         mykey   1       c1.medium       2010-09-13T02:31:51.173Z   mycloud    eki-D224100C    eri-059910F2
 INSTANCE       i-57E309BE      emi-1AF00C98    192.168.0.15    192.168.0.15    running         mykey   0       c1.medium       2010-09-13T02:31:51.172Z   mycloud    eki-D224100C    eri-059910F2

Let's say you want to start a torque server on 192.168.0.14 and two torque worker on 192.168.0.14 and 192.168.0.15, MPI enabled

}}}

This will install all necessary torque packages in the instances. It might take a few minutes, depending on the internet connection and processor speed of the instances.

Connect to a instance as root with your key

 ssh -X -i ~/.euca/mykey.priv root@192.168.0.14

virtual: Switch to the guest user

 su - guest

Check if nodes are up

 pbsnodes

Perform some simple tests

 echo "sleep 10" | qsub
 echo "sleep 5" | qsub
 echo "hostname" | qsub
 echo "sleep 15" | qsub
 echo "hostname" | qsub
 echo "sleep 3" | qsub

Look at the queue

 qstat

Let sleep 2 worker nodes

 echo "sleep 10" | qsub -l nodes=2

Check if both nodes are in state 'job-exclusive'

 pbsnodes

During the installation phase we compiled a simple MPI Hello World program.

Start it without Torque

$ mpiexec -n 4 /tmp/hello.out
Hello MPI from the server process!
Hello MPI!
 mesg from 1 of 4 on ip-192-168-0-14
Hello MPI!
 mesg from 2 of 4 on ip-192-168-0-14
Hello MPI!
 mesg from 3 of 4 on ip-192-168-0-14

Start it with Torque (without -tm support)

cat <<EOF > mpi-test_2_1_mpirun
#PBS -N helloworld
#PBS -l nodes=2:ppn=1
cd $PBS_O_WORKDIR

/usr/bin/mpirun -np 2 --hostfile /etc/torque/hostfile -v -v -v /tmp/hello.out
EOF

qsub mpi-test_2_1_mpirun

Check the output files

cat helloworld.o*
Hello MPI from the server process!
Hello MPI!
 mesg from 1 of 2 on ip-192-168-0-15

cat helloworld.e*

Start it with Torque (with -tm support)

package is ready but not in squeeze yet

example script for setting up torque:

set -ex

#NOTES:
# in whatever mode you are running your Eucalyptus system, you will have two interfaces, one is called public(you can reach the instances from outside),
# one is called inside(you can reach the instances usually only from inside, or maybe from your front end)

# you have to specify a TORQUE SERVER and one or more TORQUE NODES, because I am accessing those instances you have to specify working IP addresses.

# example: SYSTEM mode,   both interfaces are using the same address, so there is no
# example: MANAGED modes, public and private interface are different, TORQUE has to be setup for the PRIVATE mode (intra-cloud communtcation need firewall settings)

# usage: bash start_torque.sh --verbose -s="192.168.0.13" -n="192.168.0.14,192.168.0.17,192.168.0.45" -k="~/.euca/mykey.priv"

# default
VERBOSE=0
IN_INSTANCE=0

echo `hostname` : `/sbin/ifconfig eth0 | grep "inet addr" | awk '{print $2}' | sed 's/addr\://'`

for i in $*
do
        case $i in
        -s=*|--torque-server=*)
                # remove option from string
                PUBLIC_TORQUE_SERVER_IP=`echo $i | sed 's/[-a-zA-Z0-9]*=//'`
                echo $PUBLIC_TORQUE_SERVER_IP
                ;;
        -n=*|--torque-nodes=*)
                # remove option from string
                NIPS=`echo $i | sed 's/[-a-zA-Z0-9]*=//'`
                PUBLIC_NODES_IP=`echo $NIPS | sed 's/\,/ /g'`
                echo $NIPS
                echo $PUBLIC_NODES_IP
                ;;
        -k=*|--key=*)
                # remove option from string
                KEY=`echo $i | sed 's/[-a-zA-Z0-9]*=//'`
                echo $KEY
                ;;
        --verbose)
                VERBOSE=1
                ;;
        -i|--in-instance)
                IN_INSTANCE=1
                ;;
        -m=*|--with-mpi=*)
                MPI=`echo $i | sed 's/[-a-zA-Z0-9]*=//'`
                #TODO: only 0 or 1 are feasible values
                ;;
        *)
                echo "unknown option"
                ;;
        esac
done


cat > keygen_in_instance.sh << EOF
#!/bin/bash
su guest -c 'ssh-keygen -t rsa -N "" -f /home/guest/.ssh/id_rsa'
EOF
chmod 755 keygen_in_instance.sh


# BEGIN execution on master #################################################
if [ $IN_INSTANCE -eq 0 ] ; then


    # join server and nodes
    if [[ $PUBLIC_NODES_IP == *$PUBLIC_TORQUE_SERVER_IP* ]]
    then
        ALL_INSTANCES="$PUBLIC_NODES_IP"
    else
        ALL_INSTANCES="$PUBLIC_TORQUE_SERVER_IP $PUBLIC_NODES_IP"
    fi
    echo $ALL_INSTANCES


    # copy setup-torque-script to eucalyptus instances
    for NODE_IP in `echo $ALL_INSTANCES`
    do
        echo $NODE_IP
        # make this host known to ~/.ssh/known_hosts
        eval "ssh -i $KEY -o StrictHostKeychecking=no root@$NODE_IP echo ''"
        eval "/usr/bin/scp -p -i $KEY start_torque.sh root@$NODE_IP:/root/start_torque.sh"

        # MPI example
        eval "scp -p -i $KEY compileMPI.sh helloworld.c root@$NODE_IP:/root/"

        # start script in instance
        eval "ssh -X -i $KEY root@$NODE_IP \"/root/start_torque.sh -s=\"$PUBLIC_TORQUE_SERVER_IP\" -n=\"$NIPS\" -i -m=$MPI \""
    done


    # generate keys in instances - for user guest
    for NODE_IP in `echo $ALL_INSTANCES`
    do
        echo $NODE_IP
        eval "/usr/bin/scp -p -i $KEY keygen_in_instance.sh root@$NODE_IP:/root/keygen_in_instance.sh"
        eval "ssh -X -i $KEY root@$NODE_IP \"/root/keygen_in_instance.sh\""
    done


    # distribute keys
    for NODE_IP in `echo $ALL_INSTANCES`
    do
        # distribute this key to all other nodes
        for NODE_IP2 in `echo $ALL_INSTANCES`
        do
            echo $NODE_IP2
            eval "/usr/bin/scp -p -i $KEY root@$NODE_IP:/home/guest/.ssh/id_rsa.pub /tmp/id_rsa.pub"
            eval "/usr/bin/scp -p -i $KEY /tmp/id_rsa.pub root@$NODE_IP2:/tmp/id_rsa.pub"
            eval "ssh -X -i $KEY root@$NODE_IP2 \"cat /tmp/id_rsa.pub >> /home/guest/.ssh/authorized_keys\""
            #TODO, I need an entry in known_hosts, for now the following happens from within the instances
            #eval "ssh -X -i $KEY root@$NODE_IP \"ssh -o StrictHostKeychecking=no guest@$NODE_IP2 & echo '' & wait\""
        done
        eval "ssh -X -i $KEY root@$NODE_IP /root/hosts.sh"
    done


    exit # on master don't execute commands for instances
fi
# END   execution on master #################################################


# BEGIN execution in instance ###############################################
usage()
{
cat << EOF
usage: $0 options

This script starts the torque environment.

OPTIONS:

   -h      Show this message
   -n      nodes              e.g. "192.168.0.14,192.168.0.14"
   -s      torque server ip   e.g. "192.168.0.13"
   -k      key file
   -v      Verbose
   -m      With MPI support

example: start_torque.sh --verbose -s="192.168.0.13" -n="192.168.0.14,192.168.0.17,192.168.0.45" -k="~/.euca/mykey.priv"
EOF
}

#SERVER_IP  NIPS   KEY
#if [[ -z $PUBLIC_NODES_IP ]] || [[ -z $PUBLIC_TORQUE_SERVER_IP ]]
#then
#     usage
#     exit 1
#fi

function install_package {
 PACKAGE=$1

 if [ "`dpkg-query -W -f='${Status}\n' $PACKAGE`" != "install ok installed" ] ; then
  apt-get -o Dpkg::Options::="--force-confnew" --force-yes -y install $PACKAGE
  #aptitude -y install $PACKAGE
  if [ $? -ne 0 ] ; then
  echo "aptitude install $PACKAGE failed"
  fi
 else
  echo "package $PACKAGE is already installed"
 fi
}


export DEBIAN_FRONTEND="noninteractive"
export APT_LISTCHANGES_FRONTEND="none"
API_VERSION="2008-02-01"
METADATA_URL="http://169.254.169.254/$API_VERSION/meta-data"
CURL="/usr/bin/curl"


# those variables are needed for the locales package
export LANGUAGE=en_US.UTF-8
export LANG=en_US.UTF-8
export LC_ALL=en_US.UTF-8

# for dialog frontend
export PATH=$PATH:/sbin:/usr/sbin:/usr/local/sbin
export TERM=linux


PRIVATE_TORQUE_SERVER_IP="172.16.1.2"
PRIVATE_NODES="172.16.1.2   172.16.1.3   172.16.1.4   172.16.1.5   172.16.1.6   172.16.1.7   172.16.1.8"



MODE="public"
#MODE="private"
#MODE="system"

PUBLIC_TORQUE_SERVER_HOSTNAME=ip-`echo $PUBLIC_TORQUE_SERVER_IP | sed 's/\./-/g'`
echo $PUBLIC_TORQUE_SERVER_IP $PUBLIC_TORQUE_SERVER_HOSTNAME

PRIVATE_TORQUE_SERVER_HOSTNAME=ip-`echo $PRIVATE_TORQUE_SERVER_IP | sed 's/\./-/g'`
echo $PRIVATE_TORQUE_SERVER_IP $PRIVATE_TORQUE_SERVER_HOSTNAME


#GET INSTANCE IPs, create hostnames

if [ $MODE == "public" ] ; then
 PUBLIC_INSTANCE_IP=`/sbin/ifconfig eth0 | grep "inet addr" | awk '{print $2}' | sed 's/addr\://'`
else
#PUBLIC_INSTANCE_IP=192.168.0.115
 PUBLIC_INSTANCE_IP=`curl -s $METADATA_URL/public-ipv4`
fi

#PUBLIC_INSTANCE_HOSTNAME=`curl -s $METADATA_URL/public-hostname`
PUBLIC_INSTANCE_HOSTNAME=ip-`echo $PUBLIC_INSTANCE_IP | sed 's/\./-/g'`
echo $PUBLIC_INSTANCE_IP $PUBLIC_INSTANCE_HOSTNAME

PRIVATE_INSTANCE_IP=`/sbin/ifconfig eth0 | grep "inet addr" | awk '{print $2}' | sed 's/addr\://'`
PRIVATE_INSTANCE_HOSTNAME=ip-`echo $PRIVATE_INSTANCE_IP | sed 's/\./-/g'`
echo $PRIVATE_INSTANCE_IP $PRIVATE_INSTANCE_HOSTNAME



#using PUBLIC or PRIVATE interface
if [ $MODE == "public" ] ; then
   INSTANCE_HOSTNAME=$PUBLIC_INSTANCE_HOSTNAME
   NODES=$PUBLIC_NODES_IP
   INSTANCE_IP=$PUBLIC_INSTANCE_IP
   TORQUE_SERVER_IP=$PUBLIC_TORQUE_SERVER_IP
   TORQUE_SERVER_HOSTNAME=$PUBLIC_TORQUE_SERVER_HOSTNAME
else
   if [ $MODE == "private" ] ; then
      INSTANCE_HOSTNAME=$PRIVATE_INSTANCE_HOSTNAME
      NODES=$PRIVATE_NODES
      INSTANCE_IP=$PRIVATE_INSTANCE_IP
      TORQUE_SERVER_IP=$PRIVATE_TORQUE_SERVER_IP
      TORQUE_SERVER_HOSTNAME=$PRIVATE_TORQUE_SERVER_HOSTNAME
   else
      echo "please specify private or public interface"
   fi
fi


# using Google's nameserver
echo "nameserver 8.8.8.8" >> /etc/resolv.conf


# update aptitude first
#echo "deb http://ftp.us.debian.org/debian squeeze main" > /etc/apt/sources.list
#echo "deb http://security.debian.org/ squeeze/updates main" >> /etc/apt/sources.list
#aptitude update
apt-get -o Dpkg::Options::="--force-confnew" --force-yes -y update
if [ $? -ne 0 ] ; then
echo "aptitude update failed"
fi



# get rid of some error messages because of missing locales package
install_package locales
echo "en_US.UTF-8 UTF-8" > /etc/locale.gen
locale-gen

# install portmap for NFS
install_package portmap
#TODO mount here


# install nmap
install_package nmap
nmap localhost -p 1-20000

# install lsb-release
install_package lsb-release

# Print some Information about the Operating System
DISTRIBUTOR=`lsb_release -i | awk '{print $3}'`
CODENAME=`lsb_release -c | awk '{print $2}'`
echo $DISTRIBUTOR $CODENAME


# install ntpdate
install_package ntpdate
###ntpdate pool.ntp.org
ntpdate ntp.ubuntu.com

# install OpenMPI packages
if [ $MPI -eq 1 ] ; then
   install_package "libopenmpi-dev"
   install_package "openmpi-bin"
   #compile MPI test program
   bash compileMPI.sh
fi

# make hostnames known to all the TORQUE nodes and server/scheduler


if [ $MODE == "private" ] ; then
   for NODE_IP in `echo $PRIVATE_NODES`
   do
      NODE_HOSTNAME=ip-`echo $NODE_IP | sed 's/\./-/g'`
      echo "$NODE_IP   $NODE_HOSTNAME" >> /etc/hosts
      #MPI support
      mkdir -p /etc/torque
      echo "$NODE_HOSTNAME slots=1" >> /etc/torque/hostfile
   done
fi


if [ $MODE == "public" ] ; then
   for NODE_IP in `echo $PUBLIC_NODES_IP`
   do
      NODE_HOSTNAME=ip-`echo $NODE_IP | sed 's/\./-/g'`
      if [ $INSTANCE_IP != $TORQUE_SERVER_IP ] || [ $NODE_IP != $TORQUE_SERVER_IP ]; then
         if ! egrep -q "$NODE_IP|$NODE_HOSTNAME" /etc/hosts ; then
            echo "$NODE_IP   $NODE_HOSTNAME" >> /etc/hosts
         fi
      fi
      #MPI support
      mkdir -p /etc/torque
      if ! egrep -q "$NODE_HOSTNAME" /etc/torque/hostfile ; then
         echo "$NODE_HOSTNAME slots=1" >> /etc/torque/hostfile
         echo "(su - guest -c \"ssh -t -t -o StrictHostKeychecking=no guest@$NODE_HOSTNAME echo ''\")& wait" >> /root/hosts.sh                # for key distribution
      fi
   done
   if ! egrep -q "$PUBLIC_TORQUE_SERVER_HOSTNAME" /etc/torque/hostfile ; then
      echo "(su - guest -c \"ssh -t -t -o StrictHostKeychecking=no guest@$PUBLIC_TORQUE_SERVER_HOSTNAME echo ''\")& wait" >> /root/hosts.sh   # for key distribution
   fi
fi
chmod 755 /root/hosts.sh



## on TORQUE server
if [ $INSTANCE_IP == $TORQUE_SERVER_IP ]; then
   #this one is for the scheduler, if using the public interface
   if ! egrep -q "127.0.1.1|$PUBLIC_INSTANCE_HOSTNAME" /etc/hosts ; then
      echo "127.0.1.1 $PUBLIC_INSTANCE_HOSTNAME" >> /etc/hosts
   fi

#   echo "$PRIVATE_INSTANCE_IP $PRIVATE_INSTANCE_HOSTNAME" >> /etc/hosts
else
   if ! egrep -q "$TORQUE_SERVER_IP|$TORQUE_SERVER_HOSTNAME" /etc/hosts ; then
      echo "$TORQUE_SERVER_IP $TORQUE_SERVER_HOSTNAME" >> /etc/hosts
   fi
fi


# need to set a hostname before installing torque packages
echo $INSTANCE_HOSTNAME > /etc/hostname # preserve hostname if rebooting is necessary
hostname $INSTANCE_HOSTNAME # immediately change
#getent hosts `hostname`
#PUBLIC_INSTANCE_HOSTNAME=`curl -s $METADATA_URL/public-hostname`


#echo "deb http://ftp.us.debian.org/debian sid main" > /etc/apt/sources.list
apt-get -o Dpkg::Options::="--force-confnew" --force-yes -y update
if [ $INSTANCE_IP == $TORQUE_SERVER_IP ]; then
   apt-get -o Dpkg::Options::="--force-confnew" --force-yes -y install torque-server torque-scheduler torque-client
   #aptitude -y install torque-server torque-scheduler torque-client
fi

if [[ $PUBLIC_NODES_IP ==  *$INSTANCE_IP* ]]; then
   apt-get -o Dpkg::Options::="--force-confnew" --force-yes -y install torque-mom
   #aptitude -y install torque-mom
fi


## fix /tmp directory in debian eucalyptus image
chmod 777 /tmp

## add user to all nodes
USER=auser

if id $USER > /dev/null 2>&1
then
   echo "user exist!"
else
   adduser $USER --disabled-password --gecos ""
fi



#echo $PUBLIC_TORQUE_SERVER_HOSTNAME > /etc/torque/server_name
#echo $PUBLIC_INSTANCE_HOSTNAME > /etc/hostname # preserve hostname  if rebooting is necessary
#hostname $PUBLIC_INSTANCE_HOSTNAME # immediately change


DATE=`date '+%Y%m%d'`

## for TORQUE mom
if [[ $PUBLIC_NODES_IP == *$INSTANCE_IP* ]]; then
    echo $TORQUE_SERVER_HOSTNAME > /etc/torque/server_name
    echo "\$timeout 120" > /var/spool/torque/mom_priv/config # more options possible (NFS...)
    echo "\$loglevel 5" >> /var/spool/torque/mom_priv/config # more options possible (NFS...)
    /etc/init.d/torque-mom restart
    cat /var/spool/torque/mom_logs/$DATE
fi

## for TORQUE server
if [ $INSTANCE_IP == $TORQUE_SERVER_IP ]; then
   echo $TORQUE_SERVER_HOSTNAME > /etc/torque/server_name
   rm -f /var/spool/torque/server_priv/nodes
   touch /var/spool/torque/server_priv/nodes
   for NODE_IP in `echo $NODES`
   do
      NODE_HOSTNAME=ip-`echo $NODE_IP | sed 's/\./-/g'`
      echo -ne "$NODE_HOSTNAME np=1\n" >> /var/spool/torque/server_priv/nodes
   done
   /etc/init.d/torque-server restart
   /etc/init.d/torque-scheduler restart
   qmgr -c "s s scheduling=true"
   qmgr -c "c q batch queue_type=execution"
   qmgr -c "s q batch started=true"
   qmgr -c "s q batch enabled=true"
   qmgr -c "s q batch resources_default.nodes=1"
   qmgr -c "s q batch resources_default.walltime=3600"
   # had to set this for MPI, TODO: double check
   qmgr -c "s q batch resources_min.nodes=1"
   qmgr -c "s s default_queue=batch"
   # let all nodes submit jobs, not only the server
   qmgr -c "s s allow_node_submit=true"
   #qmgr -c 'set server submit_hosts += $TORQUE_SERVER_IP'
   #qmgr -c 'set server submit_hosts += $INSTANCE_IP'

   # adding extra nodes
   #qmgr -c "create node $INSTANCE_HOSTNAME"

#debug
   cat /var/spool/torque/server_logs/$DATE
   qstat -q
   pbsnodes -a
   cat /etc/torque/server_name
fi

See also


CategoryVirtualization | CategorySystemAdministration