Grid5000
https://www.grid5000.fr/mediawiki/index.php/Grid5000:Home
USAGE POLICY - Very important, after booking nodes (oarsub ...) run the command:outofchart
Contents
Login, job submission, deployment of image
- Select sites and clusters for experiments, using information on the Grid5000 network and the Status page
- Access is provided via access nodes access.SITE.grid5000.fr marked here as accessible from everywhere via ssh with keyboard-interactive authentication method. As soon as you are on one of the sites, you can directly ssh frontend node of any other site:
access_$ ssh frontend.SITE2
- There is no access to Internet from computing nodes (external IPs should be registered on proxy), therefore, download/update your stuff at the access nodes. Several revision control clients are available.
- Each site has a separate NFS, therefore, to run an application on several sites at once, you need to copy it scp, sftp, rsync between access or frontend nodes.
- Jobs are run from the frondend nodes, using a PBS-like system OAR. Basic commands:
- oarstat - queue status
- oarsub - job submission
- oardel - job removal
fontend_$ oarsub -I -t deploy -l [/cluster=N/]nodes=N,walltime=HH[:MM[:SS]] [-p 'PROPERTY="VALUE"']
fontend_$ oarsub BATCH_FILE -t allow_classic_ssh -l [/cluster=N/]nodes=N,walltime=HH[:MM[:SS]] [-p 'PROPERTY="VALUE"']
- The image to deploy can be created and loaded with help of a Systemimager-like system Kadeploy. Creating: described here
fontend_$ kadeploy3 -a PATH_TO_PRIVATE_IMAGE_DESC -f $OAR_FILE_NODES
Compiling and running MPI applications
- Compilation should be done on one of the reserved nodes (e.g. ssh `head -n 1 $OAR_NODEFILE`)
- Running MPI applications is described here
- mpirun/mpiexec should be run from one of the reserved nodes (e.g. ssh `head -n 1 $OAR_NODEFILE`)
Setting up new deploy image
List available images
kaenv3 -l
Then book node and launch:
oarsub -I -t deploy -l nodes=1,walltime=12 kadeploy3 -e squeeze-x64-big -f $OAR_FILE_NODES -k ssh root@`head -n 1 $OAR_NODEFILE`
default password: grid5000
edit /etc/apt/sources.list
apt-get update apt-get upgrade
apt-get install libtool autoconf automake mc colorgcc ctags libboost-serialization-dev libboost-graph-dev libatlas-base-dev gfortran vim gdb valgrind screen subversion iperf bc gsl-bin libgsl0-dev
Possibly also install (for using extrae):
libxml2-dev binutils-dev libunwind7-dev
Compiled for sources by us:
gsl-1.14 (download: ftp://ftp.gnu.org/gnu/gsl/)Now with squeeze it is in repository.
./configure && make && make install
./configure --enable-shared --enable-sharedlibs=gcc --with-pm=mpd make && make install
Mpich2 installed to:
Installing MPE2 include files to /usr/local/include Installing MPE2 libraries to /usr/local/lib Installing MPE2 utility programs to /usr/local/bin Installing MPE2 configuration files to /usr/local/etc Installing MPE2 system utility programs to /usr/local/sbin Installing MPE2 man to /usr/local/share/man Installing MPE2 html to /usr/local/share/doc/ Installed MPE2 in /usr/local
- hwloc (and lstopo) (download: http://www.open-mpi.org/software/hwloc/v1.2/)
compile from sources. To get xml support install libxml2-dev and pkg-config
apt-get install libxml2-dev pkg-config tar -xzvf hwloc-1.1.1.tar.gz cd hwloc-1.1.1 ./configure && make && make install
Change root password.
rm sources from root dir.
Edit the "message of the day"
vi /etc/motd.tail
echo 90 > /proc/sys/vm/overcommit_ratio echo 2 > /proc/sys/vm/overcommit_memory date >> release
Cleanup
apt-get clean rm /etc/udev/rules.d/*-persistent-net.rules
Make image
ssh root@node tgz-g5k > $HOME/grid5000/imagename.tgz
make appropriate .env file.
kaenv3 -p lenny-x64-nfs -u deploy > lenny-x64-custom-2.3.env
GotoBLAS2
When compiling gotoblas on a node without direct internet access get this error:wget http://www.netlib.org/lapack/lapack-3.1.1.tgz
--2011-05-19 03:11:03-- http://www.netlib.org/lapack/lapack-3.1.1.tgz
Resolving www.netlib.org... 160.36.58.108
Connecting to www.netlib.org|160.36.58.108|:80... failed: Connection timed out.
Retrying.
--2011-05-19 03:14:13-- (try: 2) http://www.netlib.org/lapack/lapack-3.1.1.tgz
Connecting to www.netlib.org|160.36.58.108|:80... failed: Connection timed out.
Retrying.
...
Fix by downloading http://www.netlib.org/lapack/lapack-3.1.1.tgz to the GotoBLAS2 source directory and editing this line in the Makefile
184c184 < -wget http://www.netlib.org/lapack/lapack-3.1.1.tgz --- > # -wget http://www.netlib.org/lapack/lapack-3.1.1.tgz
GotoBLAS needs to be compiled individualy for each unique machine - ie each cluster. Add the following to .bashrc
export CLUSTER=`hostname |sed 's/\([a-z]*\).*/\1/'` LD_LIBRARY_PATH=$HOME/lib/$CLUSTER:$HOME/lib:/usr/local/lib:$LD_LIBRARY_PATH export LIBRARY_PATH=$HOME/lib/$CLUSTER:$HOME/lib:/usr/local/lib:$LIBRARY_PATH
Run the following script once on each cluster:
#! /bin/bash
echo "Compiling gotoblas for cluster: $CLUSTER"
cd $HOME/src
if [ ! -d "$CLUSTER" ]; then
mkdir $CLUSTER
fi
cd $CLUSTER
tar -xzf ../Goto*.tar.gz
cd Goto*
make &> m.log
if [ ! -d "$HOME/lib/$CLUSTER" ]; then
mkdir $HOME/lib/$CLUSTER
fi
cp libgoto2.so $HOME/lib/$CLUSTER
echo results
ls -d $HOME/src/$CLUSTER
ls $HOME/src/$CLUSTER
ls -d $HOME/lib/$CLUSTER
ls $HOME/lib/$CLUSTER
note: for newer processors this may fail. If it is a NEHALEM processor try:
make clean make TARGET=NEHALEM
Paging and the OOM-Killer
When doing exhaustion of available memory experiments, problems can occur with over-commit. See HCL cluster#Paging_and_the_OOM-Killer for more detail.
Example of experiment setup across several sites
Sources of all files mentioned below is available at: Grid5000:sources.
Pick one head node as the main head node (I use grenoble, but any will do). Setup sources
cd dave/fupermod-1.1.0 make clean ./configure --with-cblas=goto --prefix=/usr/local/
Reserve 2 nodes from all clusters on a 3 cluster site:
oarsub -r "2011-07-25 11:01:01" -t deploy -l cluster=3/nodes=2,walltime=11:59:00
Automate with:
for a in 2 3 4; do for i in `cat sites.$a`; do echo $a $i; ssh $i oarsub -r "2011-07-25 11:01:01" -t deploy -l cluster=$a/nodes=2,walltime=11:59:00; done; done
Then on each site, where xxx is site name:
kadeploy3 -a $HOME/grid5000/lenny-dave.env -f $OAR_NODE_FILE --output-ok-nodes deployed.xxx
Gather deployed files to a head node:
for i in `cat ~/sites `; do echo $i; scp $i:deployed* . ; done cat deployed.* > deployed.all
Copy cluster specific libs to each deployed node /usr/local/lib dir with script
copy_local_libs.sh deployed.all
Copy source files to root dir of each deployed node. Then make install each (node ssh -f does this in parallel)
for i in `cat ~/deployed.all`; do echo $i; rsync -aP ~/dave/fupermod-1.1.0 root@$i: ; done for i in `cat ~/deployed.all`; do echo $i; ssh -f root@$i "cd fupermod-1.1.0 ; make all install" ; done
ssh to the first node
ssh `head -n1 deployed.all` n=$(cat deployed.all |wc -l) mpdboot --totalnum=$n --file=$HOME/deployed.all mpdtrace
cd dave/data/ mpirun -n $n /usr/local/bin/partitioner -l /usr/local/lib/libmxm_col.so -a0 -D10000 -o N=100
Cleanup after:
for i in `cat ~/sites `; do echo $i; ssh $i rm deployed.* ; done
Check network speed
apt-get install iperf
Choose which network interface to use
mpirun --mca btl self,openib ...
or
mpirun --mca btl self,tcp ...